DUALVIEW：利用广泛证据和输出特定证据进行基于大语言模型的多输出文本回归

# DualView: Leveraging Broad and Output-Specific Evidence for LLM-based Multi-Output Text Regression
Source: [https://arxiv.org/html/2610.19206](https://arxiv.org/html/2610.19206)
Junyi Jessy LiMatthew LeaseAffiliation:The University of Texas at AustinAffiliation:\{sam\.su, jessy, ml\}@utexas\.edu

###### Abstract

While work on LLM\-based regression has largely focused on scalar prediction, multi\-output text regression remains less explored\. We presentDualView, a parameter\-efficient approach to leverage LLMs for multi\-output text regression\.DualViewextracts broad and output\-specific representations from a frozen LLM, then uses a dual\-branch architecture to jointly predict numerical scores for all targets\. Across five regression tasks and eight LLMs \(up to 122B\),DualViewoutperforms seven supervised regression baselines while requiring 82–1842×\\timesfewer fitted parameters than the three strongest baselines\. These improvements extend across individual output dimensions and persist even with limited training data\.DualViewatop a smaller open\-weight LLM also outperforms unsupervised prediction with Gemini 3\.1 Pro\. Ablations show that broad and output\-specific evidence provide complementary predictive benefits, while lightweight shared predictors achieve strong performance with fewer trainable parameters\.

## 1Introduction

Text regression aims to predict continuous numerical targets from textual inputs\. While recent work has largely focused on predicting scalar targets\([Nafar et al\., 2025](https://arxiv.org/html/2610.19206#bib.bib27);[Lukasik et al\., 2025](https://arxiv.org/html/2610.19206#bib.bib14);[Su and Lease, 2026](https://arxiv.org/html/2610.19206#bib.bib12);[Chen et al\., 2026b](https://arxiv.org/html/2610.19206#bib.bib13);[Akhauri et al\., 2026](https://arxiv.org/html/2610.19206#bib.bib11);[Zhang et al\., 2026a](https://arxiv.org/html/2610.19206#bib.bib10);[Zuo et al\., 2026](https://arxiv.org/html/2610.19206#bib.bib9);[Chung et al\., 2026](https://arxiv.org/html/2610.19206#bib.bib8)\), predicting multiple numerical targets for the same input—*multi\-output text regression*\(MOTR\)—remains less explored\.

Building better MOTR systems can benefit fine\-grained evaluations of generated text, where standard protocols assess free\-form outputs along multiple dimensions such as fluency, factuality, coherence, and relevance\([Liu et al\., 2023b](https://arxiv.org/html/2610.19206#bib.bib18);[Hashemi et al\., 2024](https://arxiv.org/html/2610.19206#bib.bib7);[Kim et al\., 2025](https://arxiv.org/html/2610.19206#bib.bib6);[Chen et al\., 2026a](https://arxiv.org/html/2610.19206#bib.bib46)\)\. More specialized tasks exhibit the same multi\-output structure: machine translation can be evaluated for accuracy, fluency, and style\([Popov et al\., 2025](https://arxiv.org/html/2610.19206#bib.bib53);[Shayegh et al\., 2025](https://arxiv.org/html/2610.19206#bib.bib54)\), generated code for correctness and efficiency\([Qiu et al\., 2025](https://arxiv.org/html/2610.19206#bib.bib47);[Ye et al\., 2026](https://arxiv.org/html/2610.19206#bib.bib55)\), and research ideas for novelty, relevance, and feasibility\([Si et al\., 2025](https://arxiv.org/html/2610.19206#bib.bib48);[Baek et al\., 2025](https://arxiv.org/html/2610.19206#bib.bib52)\)\.Figure[1](https://arxiv.org/html/2610.19206#S1.F1)illustrates the latter two examples\.

Figure 1:Multi\-output text regressionpredicts multiple numerical targets for the same textual artifact\. Examples include code generation and research ideation, where the same program or paper is evaluated along dimensions with potentially different units and scales\.However, developing general\-purpose MOTR methods presents several challenges\. First, existing text regression methods often fine\-tune all LLM parameters, making them costly to scale to larger models\([Lukasik et al\., 2025](https://arxiv.org/html/2610.19206#bib.bib14);[Chen et al\., 2026b](https://arxiv.org/html/2610.19206#bib.bib13)\)\. Second, training one scalar predictor for each output scales adaptation costs with the number of targets, motivating more compact multi\-output predictors\([Do et al\., 2024a](https://arxiv.org/html/2610.19206#bib.bib25)\)\. Third, existing supervised MOTR methods often employ specialized architectures that may be less suitable across heterogeneous domains\. For example, GAT\-AES\([Li and Ng, 2025](https://arxiv.org/html/2610.19206#bib.bib5)\)uses handcrafted features and a fully connected graph that exploits strong correlations among essay\-quality criteria\. However, this architecture is less suitable for code quality estimation, where essay\-specific features do not apply, and targets such as code correctness and peak memory may not be strongly correlated\.

In this work, we proposeDualView, a supervised MOTR approach that shifts specialization from domain\-specific prediction architectures to evidence extraction\. Specifically,DualViewextracts complementary broad and output\-specific representations from an LLM\. A unified dual\-branch model leverages these representations to predict all output dimensions, sharing predictors across targets instead of training separate models \(seeFigure[2](https://arxiv.org/html/2610.19206#S1.F2)\)\. Finally, by freezing the LLM,DualViewtrains only lightweight modules and scales better to larger LLMs\.

We evaluateDualViewon five quality estimation tasks spanning summarization\([Min et al\., 2025](https://arxiv.org/html/2610.19206#bib.bib56)\), machine translation\([Popov et al\., 2025](https://arxiv.org/html/2610.19206#bib.bib53)\), code generation\([Ouyang et al\., 2026](https://arxiv.org/html/2610.19206#bib.bib59)\), dialogue\([Kirk et al\., 2024](https://arxiv.org/html/2610.19206#bib.bib58)\), and research ideation\([Lu et al\., 2026](https://arxiv.org/html/2610.19206#bib.bib57)\)\.DualViewoutperforms seven supervised state\-of\-the\-art regression baselines across the five datasets and eight LLM backbones \(8–122B\)\. Relative to the second\-best method, GAT\-AES\([Li and Ng, 2025](https://arxiv.org/html/2610.19206#bib.bib5)\),DualViewimproves aggregate Pearson correlation by 18\.4% and NRMSE by 2\.8% \(Table[1](https://arxiv.org/html/2610.19206#S5.T1)\)\. Output\-level results further show thatDualViewoutperforms the strongest supervised baselines on nearly all targets \(Figure[3](https://arxiv.org/html/2610.19206#S4.F3)\)\.

By freezing the LLM and sharing predictors across targets,DualViewrequires only 5\.1K–57\.4K fitted parameters, compared with 82–1842×\\timesmore for the strongest supervised baselines \(Table[2](https://arxiv.org/html/2610.19206#S5.T2)\)\. Training\-data efficiency experiments show thatDualViewcontinues to outperform the strongest baselines even with limited supervision \(Figure[4](https://arxiv.org/html/2610.19206#S5.F4)\)\.DualViewoutperforms Gemini 3\.1 Pro under zero\- and few\-shot LLMaaJ \(Table[3](https://arxiv.org/html/2610.19206#S5.T3)\)\.

![Refer to caption](https://arxiv.org/html/2610.19206v1/dualview_architecture.png)Figure 2:DualViewarchitecture\.DualViewqueries a frozen LLM with a broad prompt and focused prompts to extract complementary representations\. The “Broad View” jointly predicts all output dimensions from one representation using a reduced\-rank predictor\. On the other hand, the “Focused View” predicts each output dimension individually using a shared linear predictor with target\-specific representations\. The two branches predict standardized scores, which are summed and rescaled to obtain the final predictions\.
## 2Related Work

### 2\.1LLM\-Based Text Regression

Existing supervised, LLM\-based text regression methods broadly fall into three families\. i\)*Autoregressive decoding*generates numerical values as text\([Vacareanu et al\., 2024](https://arxiv.org/html/2610.19206#bib.bib30);[Song and Bahri, 2025](https://arxiv.org/html/2610.19206#bib.bib15);[Chen et al\., 2026b](https://arxiv.org/html/2610.19206#bib.bib13)\)\. ii\)*Regression\-aware inference*derives point estimates from the model’s output distribution under a regression objective\([Lukasik et al\., 2024](https://arxiv.org/html/2610.19206#bib.bib26);[Lukasik et al\., 2025](https://arxiv.org/html/2610.19206#bib.bib14)\)\. iii\)*Predictive heads*map LLM hidden representations directly to continuous values\([Tang et al\., 2025](https://arxiv.org/html/2610.19206#bib.bib29);[Su and Lease, 2026](https://arxiv.org/html/2610.19206#bib.bib12)\)\.

Our work differs in three key respects\. 1\) While most supervised methods focus on scalar prediction, we study the joint prediction of multiple numerical targets\. 2\) Existing multi\-output methods often introduce specialized prediction architectures such as target\-specific experts and graph\-based interactions\([Wang and Liu, 2025](https://arxiv.org/html/2610.19206#bib.bib16);[Li and Ng, 2025](https://arxiv.org/html/2610.19206#bib.bib5)\)\. In contrast,DualViewshifts specialization to evidence extraction, using simple predictors over complementary representations from a frozen LLM\. 3\) Recent supervised methods only evaluate LLMs with up to 32B parameters\([Su and Lease, 2026](https://arxiv.org/html/2610.19206#bib.bib12);[Zhang et al\., 2026a](https://arxiv.org/html/2610.19206#bib.bib10)\), whereas we experiment with LLMs up to 122B parameters\.

### 2\.2Fine\-Grained Text Evaluation

Fine\-grained text evaluation \(also known as multi\-dimensional, multi\-criteria, or multi\-trait evaluation\) assesses the same text input along multiple numerical axes\([Fabbri et al\., 2021](https://arxiv.org/html/2610.19206#bib.bib24);[Liu et al\., 2023b](https://arxiv.org/html/2610.19206#bib.bib18)\)\. By assessing distinct aspects of quality, fine\-grained evaluation helps identify specific weaknesses, reveal performance trade\-offs, and guide targeted revisions\([Hashemi et al\., 2024](https://arxiv.org/html/2610.19206#bib.bib7);[Kim et al\., 2025](https://arxiv.org/html/2610.19206#bib.bib6);[Kola et al\., 2026](https://arxiv.org/html/2610.19206#bib.bib19)\)\. However, obtaining human annotations or execution\-based metrics \(e\.g\., code correctness and peak memory\) across multiple dimensions can be expensive, motivating automated evaluation methods that learn from limited supervision\. We show thatDualViewconsistently outperforms supervised baselines even with as few as 64 training examples \(§\\lx@sectionsign[E\.3](https://arxiv.org/html/2610.19206#A5.SS3)\)\.

### 2\.3Multi\-Task Learning

Multi\-task learning \(MTL\) trains a single model to perform multiple related tasks simultaneously\([Ding et al\., 2023](https://arxiv.org/html/2610.19206#bib.bib44);[Hu et al\., 2025](https://arxiv.org/html/2610.19206#bib.bib20);[Hu et al\., 2026](https://arxiv.org/html/2610.19206#bib.bib21)\)\. Multi\-output text regression is closely related but imposes a more structured prediction problem\. Whereas MTL may combine tasks with different inputs, outputs, or objectives, MOTR predicts thesameset of continuous targets from thesametext input\. This shifts the focus from knowledge transfer across heterogeneous tasks to balancing shared and target\-specific information within the same prediction problem\. Regression\-capable MTL methods such as UniSO\-N\([Tan et al\., 2025](https://arxiv.org/html/2610.19206#bib.bib22)\)and MT\-ICL\([Zhu et al\., 2026a](https://arxiv.org/html/2610.19206#bib.bib23)\)provide general mechanisms for sharing across tasks, but do not specifically exploit this same\-input, fixed\-target structure\. In contrast,DualViewexploits this structure directly through a dual\-branch architecture that leverages complementary evidence extracted from frozen LLMs\.§\\lx@sectionsign[B](https://arxiv.org/html/2610.19206#A2)details additional comparisons\.

## 3DualView

### 3\.1Problem Formulation

Let𝒮=\{\(xi,𝐲i\)\}i=1N\\mathcal\{S\}=\\\{\(x\_\{i\},\\mathbf\{y\}\_\{i\}\)\\\}\_\{i=1\}^\{N\}denote a training set of textual inputsxi∈𝒳x\_\{i\}\\in\\mathcal\{X\}paired with target vectors𝐲i=\(yi​1,…,yi​D\)∈ℝD\\mathbf\{y\}\_\{i\}=\(y\_\{i1\},\\ldots,y\_\{iD\}\)\\in\\mathbb\{R\}^\{D\}\. Let𝒬=\(d1,…,dD\)\\mathcal\{Q\}=\(d\_\{1\},\\ldots,d\_\{D\}\)denote an ordered set of target descriptions, wheredjd\_\{j\}specifies the meaning and numerical scale of targetjj\. Given𝒬\\mathcal\{Q\}, multi\-output text regression aims to learn a predictorfREG:𝒳→ℝDf\_\{\\mathrm\{REG\}\}:\\mathcal\{X\}\\to\\mathbb\{R\}^\{D\}that jointly predicts allDDtargets for each input\.

Following[Tang et al\. \(2025\)](https://arxiv.org/html/2610.19206#bib.bib29)and[Su and Lease \(2026\)](https://arxiv.org/html/2610.19206#bib.bib12), we standardize each target using its training\-set meanμj\\mu\_\{j\}and standard deviationσj\\sigma\_\{j\}:

zi​j=yi​j−μjσj,y^i​j=μj\+σj​z^i​j\.z\_\{ij\}=\\frac\{y\_\{ij\}\-\\mu\_\{j\}\}\{\\sigma\_\{j\}\},\\qquad\\widehat\{y\}\_\{ij\}=\\mu\_\{j\}\+\\sigma\_\{j\}\\widehat\{z\}\_\{ij\}\.All predictors are trained on standardized targets\. At inference, predictions are rescaled byσj\\sigma\_\{j\}and shifted byμj\\mu\_\{j\}to restore their original scales\.

### 3\.2Evidence Extraction

Rather than tailoring architectures to each domain,DualViewapplies simple predictors to complementary evidence extracted from a frozen LLM\. We first extract a broad representation to capture evidence relevant to all targets\. Since different targets may depend on different aspects of the input, we also extract output\-specific representations to identify cues particularly relevant to each target\.

Lets⁡\(𝒬\)s\(\\mathcal\{Q\}\)denote a shared task instruction specifying all prediction targets, and letq⁡\(dj\)q\(d\_\{j\}\)denote a short instruction requesting only an assessment of targetjj\. For each inputxix\_\{i\}, we construct a broad prompt andDDfocused prompts:

Pib\\displaystyle P\_\{i\}^\{b\}=\[s⁡\(𝒬\);xi\],\\displaystyle=\[s\(\\mathcal\{Q\}\);x\_\{i\}\],Pi​jf\\displaystyle P\_\{ij\}^\{f\}=\[s⁡\(𝒬\);xi;q⁡\(dj\)\]\.\\displaystyle=\[s\(\\mathcal\{Q\}\);x\_\{i\};q\(d\_\{j\}\)\]\.
For the broad view, we mean\-pool model hidden states across all tokens of the broad prompt:

𝐉i=1\|Pib\|​∑t=1\|Pib\|𝐡t​\(Pib\)∈ℝH\.\\mathbf\{J\}\_\{i\}=\\frac\{1\}\{\|P\_\{i\}^\{b\}\|\}\\sum\_\{t=1\}^\{\|P\_\{i\}^\{b\}\|\}\\mathbf\{h\}\_\{t\}\(P\_\{i\}^\{b\}\)\\in\\mathbb\{R\}^\{H\}\.
For the focused view, we extract the hidden state at the last token of each focused prompt, immediately before generation:

𝐂i​j=𝐡\|Pi​jf\|​\(Pi​jf\)∈ℝH\.\\mathbf\{C\}\_\{ij\}=\\mathbf\{h\}\_\{\|P\_\{ij\}^\{f\}\|\}\(P\_\{ij\}^\{f\}\)\\in\\mathbb\{R\}^\{H\}\.Here, we use the final hidden state rather than mean pooling because most tokens precede the target\-specific instruction and are shared across targets\. Averaging their representations thus could dilute the effect of the focused instruction\.

##### Shared\-prefix reuse\.

The broad and focused prompts share the task instructions, target descriptions, and input\. We therefore cache the shared prefix’s model states \(e\.g\., attention KV pairs\) and reuse them across prompts\. Rather than independently evaluatingD\+1D\+1complete prompts, we process the shared prefix once and evaluate only the remaining suffix of each prompt\. See§\\lx@sectionsign[C](https://arxiv.org/html/2610.19206#A3)for details on extraction and normalization\.

### 3\.3Broad Vector Predictor

The broad branch predicts allDDtargets jointly from𝐉i∈ℝH\\mathbf\{J\}\_\{i\}\\in\\mathbb\{R\}^\{H\}using a linear predictor:

𝐳^ib=𝐖b⊤​𝐉i\.\\widehat\{\\mathbf\{z\}\}^\{\\,\\mathrm\{b\}\}\_\{i\}=\\mathbf\{W\}\_\{\\mathrm\{b\}\}^\{\\top\}\\mathbf\{J\}\_\{i\}\.An unconstrained𝐖b∈ℝH×D\\mathbf\{W\}\_\{\\mathrm\{b\}\}\\in\\mathbb\{R\}^\{H\\times D\}effectively fits a separate linear predictor for each target\. To obtain a compact multivariate predictor, we adopt reduced\-rank regression\([Izenman, 1975](https://arxiv.org/html/2610.19206#bib.bib3)\), factorizing𝐖b=𝐀𝐁\\mathbf\{W\}\_\{\\mathrm\{b\}\}=\\mathbf\{A\}\\mathbf\{B\}, where𝐀∈ℝH×k\\mathbf\{A\}\\in\\mathbb\{R\}^\{H\\times k\}and𝐁∈ℝk×D\\mathbf\{B\}\\in\\mathbb\{R\}^\{k\\times D\}, withk<Dk<D\. Each targetjjuses a coefficient vector𝐀𝐛j\\mathbf\{A\}\\mathbf\{b\}\_\{j\}, where𝐛j\\mathbf\{b\}\_\{j\}is thejj\-th column of𝐁\\mathbf\{B\}\. Thus, all targets are predicted from the samekklinear projections of𝐉i\\mathbf\{J\}\_\{i\}, but each target assigns different weights to these projections\. This factorization provides a compact multivariate predictor by sharingkklinear projections across outputs\.

### 3\.4Shared Focused Predictor

The focused branch uses a shared linear predictor to estimate targetjjfrom𝐂i​j∈ℝH\\mathbf\{C\}\_\{ij\}\\in\\mathbb\{R\}^\{H\}:

z^i​jf=𝐯⊤​𝐂i​j\.\\widehat\{z\}^\{\\,\\mathrm\{f\}\}\_\{ij\}=\\mathbf\{v\}^\{\\top\}\\mathbf\{C\}\_\{ij\}\.Because𝐂i​j\\mathbf\{C\}\_\{ij\}captures evidence specific to targetjj, a shared coefficient vector𝐯∈ℝH\\mathbf\{v\}\\in\\mathbb\{R\}^\{H\}can produce distinct predictions for different output targets\. Sharing one predictor also allows supervision from all targets to inform a common scoring function, while requiring fewer parameters than fitting separate predictors for each target\.

Notably, the two branches serve complementary purposes: the broad branch learns common predictive directions across different targets, while the focused branch pools supervision across targets to score their specialized evidence\.

### 3\.5Training

We jointly train the broad and focused predictors to exploit their complementary evidence\. The final prediction for targetjjcombines both branches:

z^i​j=𝐉i⊤​𝐀𝐛j\+𝐯⊤​𝐂i​j\+aj,\\widehat\{z\}\_\{ij\}=\\mathbf\{J\}\_\{i\}^\{\\top\}\\mathbf\{A\}\\mathbf\{b\}\_\{j\}\+\\mathbf\{v\}^\{\\top\}\\mathbf\{C\}\_\{ij\}\+a\_\{j\},whereaja\_\{j\}is a target\-specific intercept\. We minimize regularized squared error on the combined predictions, subject torank⁡\(𝐖b\)≤k\\operatorname\{rank\}\(\\mathbf\{W\}\_\{\\mathrm\{b\}\}\)\\leq k:

min𝐖b,𝐯,𝐚\\displaystyle\\min\_\{\\mathbf\{W\}\_\{\\mathrm\{b\}\},\\,\\mathbf\{v\},\\,\\mathbf\{a\}\}1N​D​∑i=1N∑j=1D\(zi​j−z^i​j\)2\\displaystyle\\frac\{1\}\{ND\}\\sum\_\{i=1\}^\{N\}\\sum\_\{j=1\}^\{D\}\(z\_\{ij\}\-\\widehat\{z\}\_\{ij\}\)^\{2\}\+λb​‖𝐖b‖F2\+λf​‖𝐯‖22,\\displaystyle\+\\lambda\_\{\\mathrm\{b\}\}\\\|\\mathbf\{W\}\_\{\\mathrm\{b\}\}\\\|\_\{F\}^\{2\}\+\\lambda\_\{\\mathrm\{f\}\}\\\|\\mathbf\{v\}\\\|\_\{2\}^\{2\},where𝐖b=𝐀𝐁\\mathbf\{W\}\_\{\\mathrm\{b\}\}=\\mathbf\{A\}\\mathbf\{B\},𝐚=\(a1,…,aD\)\\mathbf\{a\}=\(a\_\{1\},\\ldots,a\_\{D\}\), andλb,λf\\lambda\_\{\\mathrm\{b\}\},\\lambda\_\{\\mathrm\{f\}\}control regularization\. We optimize this objective through alternating minimization\. Starting from a reduced\-rank ridge solution with𝐯=𝟎\\mathbf\{v\}=\\mathbf\{0\}, we alternate between fitting the shared predictor to the broad predictor’s residuals and refitting the broad predictor to the focused predictor’s residuals\. We repeat until convergence and select the rank and regularization parameters using the validation split\.

## 4Experimental Setup

##### Models\.

We assessDualViewover eight LLM backbones: Gemma 4 E4B and 31B\([Gemma Team, 2026](https://arxiv.org/html/2610.19206#bib.bib60)\), Qwen3\.5 9B and 122B\-A10B\([Qwen Team, 2026](https://arxiv.org/html/2610.19206#bib.bib61)\), Qwen3\.6 27B, Llama 3\.3 70B Instruct\([Grattafiori et al\., 2024](https://arxiv.org/html/2610.19206#bib.bib65)\), Qwen2\.5 72B Instruct\([Qwen Team, 2025](https://arxiv.org/html/2610.19206#bib.bib62)\), and GPT OSS 120B\([OpenAI, 2025](https://arxiv.org/html/2610.19206#bib.bib63)\)\. We freeze all LLM backbones and fit only lightweight regression predictors\. We further compare against Gemini 3\.1 Pro in§\\lx@sectionsign[5\.4](https://arxiv.org/html/2610.19206#S5.SS4)\.

##### Datasets\.

We evaluateDualViewon five fine\-grained quality estimation tasks: MSumBench for summarization\([Min et al\., 2025](https://arxiv.org/html/2610.19206#bib.bib56)\), RATE for translation\([Popov et al\., 2025](https://arxiv.org/html/2610.19206#bib.bib53)\), CreativeEval for research ideation\([Lu et al\., 2026](https://arxiv.org/html/2610.19206#bib.bib57)\), DSCodeBench for code generation\([Ouyang et al\., 2026](https://arxiv.org/html/2610.19206#bib.bib59)\), and PRISM for dialogue\([Kirk et al\., 2024](https://arxiv.org/html/2610.19206#bib.bib58)\)\.§\\lx@sectionsign[D\.2](https://arxiv.org/html/2610.19206#A4.SS2)reports dataset split sizes and target ranges \(Table[8](https://arxiv.org/html/2610.19206#A4.T8)\)\.

##### Baselines\.

We evaluate against seven supervised baselines: RELISH\([Su and Lease, 2026](https://arxiv.org/html/2610.19206#bib.bib12)\), GenRe2\([Chen et al\., 2026b](https://arxiv.org/html/2610.19206#bib.bib13)\), T\-MES\([Wang and Liu, 2025](https://arxiv.org/html/2610.19206#bib.bib16)\), GAT\-AES\([Li and Ng, 2025](https://arxiv.org/html/2610.19206#bib.bib5)\), SAJA\([Kola et al\., 2026](https://arxiv.org/html/2610.19206#bib.bib19)\), UniSO\-N\([Tan et al\., 2025](https://arxiv.org/html/2610.19206#bib.bib22)\), and MT\-ICL\([Zhu et al\., 2026a](https://arxiv.org/html/2610.19206#bib.bib23)\)\. We use frozen LLMs for all supervised methods and adapt their prediction modules to the multi\-output setting where necessary \(§[D\.3](https://arxiv.org/html/2610.19206#A4.SS3)\)\. We also report zero\- and four\-shot LLM\-as\-a\-judge prompting baselines\.

##### Metrics\.

We report Pearson correlation and normalized root mean squared error \(NRMSE\)\. Following prior work\([Lin et al\., 2023](https://arxiv.org/html/2610.19206#bib.bib51);[Allison et al\., 2023](https://arxiv.org/html/2610.19206#bib.bib49);[Holzmüller et al\., 2024](https://arxiv.org/html/2610.19206#bib.bib50)\), we normalize RMSE by the training\-set standard deviation to fairly compare across datasets and targets with different scales\. We additionally report Spearman correlation in the appendix to assess rank\-order agreement\.§\\lx@sectionsign[D\.4](https://arxiv.org/html/2610.19206#A4.SS4)provides further details\.

Figure 3:Predictive Performance across datasets and dimensions\.Each spoke reports Pearson correlation for one dataset aspect, macro\-averaged over 8 LLM backbones and 3 training seeds\. We compareDualViewagainst the three strongest supervised baselines\. Larger radial values are better, where scales vary across datasets for visibility\.

## 5Key Results

### 5\.1Predictive Performance

Table 1:Predictive performanceaggregated across 8 LLMs and 5 datasets spanning 21 output targets\. Values are macro\-averaged over datasets and LLMs, with standard deviations across three seeds \(±\\pm\)\. Lower \(↓\\downarrow\) NRMSE and higher \(↑\\uparrow\) Pearson correlation are better\.##### Aggregate performance\.

Table[1](https://arxiv.org/html/2610.19206#S5.T1)shows thatDualViewachieves the strongest aggregate performance across five datasets and eight LLM backbones\. Compared with GAT\-AES, the strongest supervised baseline,DualViewimproves Pearson correlation by 18\.4% and NRMSE by 2\.8%\. The substantially higher NRMSE of zero\- and few\-shot prompting reflects large out\-of\-range predictions\.

##### Performance across targets\.

Since our MOTR setting emphasizes predicting multiple targets per dataset, a robust method should ideally perform well onalltargets\. WhileTable[1](https://arxiv.org/html/2610.19206#S5.T1)’s results usefully summarize predictive performance on average, simple arithmetic averaging across target dimensions risks strong performance on some dimensions masking weak performance on others\. Thus,Figure[3](https://arxiv.org/html/2610.19206#S4.F3)comparesDualViewwith the three strongest supervised baselines across all 21 output targets and reports Pearson correlations macro\-averaged over eight LLMs\.DualViewachieves the strongest performance on 20 of 21 targets, with MSumBench faithfulness effectively tied at 0\.675\.

##### Performance across LLM–target pairs\.

Figure[3](https://arxiv.org/html/2610.19206#S4.F3)still averages performance across LLMs, which can hide inconsistent results for individual LLM–target pairs\.Table[14](https://arxiv.org/html/2610.19206#A5.T14)in§\\lx@sectionsign[E\.1\.1](https://arxiv.org/html/2610.19206#A5.SS1.SSS1.Px2)reports two additional fine\-grained measures against the three strongest supervised baselines\. Thestrict win rateis the percentage of LLM–target pairs on which a method outperforms all competitors\. Theworst\-output gapis the largest performance difference from the best method across targets within each dataset–LLM setting, averaged across settings\.DualViewachieves the highest win rates of 65\.9% for Pearson and 71\.4% for NRMSE, compared with 8\.9% and 7\.7% for GAT\-AES\. Our mean worst\-output gaps are also smaller: 0\.039 versus 0\.149 for Pearson and 0\.016 versus 0\.055 for NRMSE\. These results show thatDualViewwins more consistently with smaller target\-level gaps\.

DualViewalso outperforms all supervised baselines under aggregation schemes that place greater weight on poorly predicted targets \(Table[15](https://arxiv.org/html/2610.19206#A5.T15)\)\. Detailed results for every dataset–LLM combination are reported inTables[11](https://arxiv.org/html/2610.19206#A5.T11),[12](https://arxiv.org/html/2610.19206#A5.T12), and[13](https://arxiv.org/html/2610.19206#A5.T13)\(§\\lx@sectionsign[E\.1](https://arxiv.org/html/2610.19206#A5.SS1)\), with complete breakdowns in§\\lx@sectionsign[I](https://arxiv.org/html/2610.19206#A9)\.

### 5\.2Trainable Footprint

Table 2:Trainable parameter countsforDualViewand the three strongest supervised baselines\. Ranges span 8 LLMs, 5 datasets, and 3 training seeds\. Baseline ratios are computed relative toDualViewfor each LLM, then macro\-averaged\.Figure 4:Predictive Performance as a function of training\-set size\.Curves show Pearson correlation averaged over 8 LLMs within each seed, then averaged across three seeds, with whiskers denoting one standard deviation\. We compareDualViewwith the three strongest aggregate supervised baselines across five datasets\. The horizontal axis is logarithmic, and Full denotes each dataset’s complete training split\.§\\lx@sectionsign[E\.3](https://arxiv.org/html/2610.19206#A5.SS3)provides additional details\.Table 3:Comparisons to Gemini 3\.1 Pro\.We compareDualViewatop Gemma 4 31B \(best open LLM\) to Gemini 3\.1 Pro under zero\-shot and few\-shot LLMaaJ across five datasets\. Metrics are macro\-averaged across all output dimensions within each dataset, with standard deviations across three seeds \(±\\pm\)\.Table[2](https://arxiv.org/html/2610.19206#S5.T2)shows thatDualViewrequires only 5\.1K–57\.4K fitted parameters, compared with 1\.84M–20\.98M for the three strongest supervised baselines \(82–1842×\\timesmore\)\. Reduced\-rank and shared predictors avoid fitting separate high\-dimensional models for each target\. Notably, the trainable footprintscales with representation width rather than backbone parameter count, facilitating adaptation to larger LLMs\.Table[17](https://arxiv.org/html/2610.19206#A5.T17)in§\\lx@sectionsign[E\.2](https://arxiv.org/html/2610.19206#A5.SS2)provides detailed counts for all baselines across datasets and LLMs\.

### 5\.3Training\-Data Efficiency

We also investigate training\-data efficiency by examining how predictive performance varies as the number of training examples increases\. We compareDualViewagainst the three strongest supervised baselines and report Pearson correlation macro\-averaged across all 8 LLM backbones and 3 seeds for each dataset\. We refit methods on nested training subsets \(e\.g\., larger subsets must contain all examples from smaller ones\), starting at 64 instances and doubling the budget up to each dataset’s full training split\. We hold the validation and test splits fixed across budgets\.

Figure[4](https://arxiv.org/html/2610.19206#S5.F4)shows thatDualViewconsistently outperforms the three strongest baselines at every evaluated training budget\.DualView’s lead on MSumBench narrows but remains stable or widens on the other four datasets\.Figures[7](https://arxiv.org/html/2610.19206#A5.F7),[8](https://arxiv.org/html/2610.19206#A5.F8), and[9](https://arxiv.org/html/2610.19206#A5.F9)\(§\\lx@sectionsign[E\.3](https://arxiv.org/html/2610.19206#A5.SS3)\) compareDualViewagainst all seven supervised baselines for the three metrics\.

### 5\.4Comparison with Gemini 3\.1 Pro

Table[3](https://arxiv.org/html/2610.19206#S5.T3)comparesDualViewusing Gemma 4 31B with Gemini 3\.1 Pro under zero\-shot and few\-shot prompting\. Relative to the stronger few\-shot condition,DualViewimproves Pearson correlation and NRMSE by 30\.7% and 19\.8%, respectively\. This result shows that supervised prediction with a smaller open LLM can outperform unsupervised prompting using a frontier proprietary LLM\.

## 6Ablations

§\\lx@sectionsign[6\.1](https://arxiv.org/html/2610.19206#S6.SS1)examines the complementarity of broad and focused evidence\.§\\lx@sectionsign[6\.2](https://arxiv.org/html/2610.19206#S6.SS2)ablates other training strategies\.§\\lx@sectionsign[6\.3](https://arxiv.org/html/2610.19206#S6.SS3)investigates the importance of target\-specific evidence alignment\. Finally,§\\lx@sectionsign[6\.4](https://arxiv.org/html/2610.19206#S6.SS4)assesses parameter sharing across outputs\. Unless otherwise stated, ablation results are macro\-averaged across five datasets and eight LLM backbones with standard deviations across three training seeds \(±\\pm\)\.

Table 4:Evidence complementarity\.We compare broad and focused evidence individually, using both constrained and unconstrained predictors, against their combination inDualView\.### 6\.1Evidence Complementarity

We first investigate whether broad and focused representations provide complementary predictive information\. We compare the reduced\-rank broad and shared focused predictors individually against their combination, using the same regularized squared\-error objective\. To assess whether single\-view performance is limited by predictor capacity, we also evaluate full\-rank broad regression and independent focused predictors for each target\. All variants are refitted using their respective features\.

Table[4](https://arxiv.org/html/2610.19206#S6.T4)shows that combining both evidence outperforms either alone, even with more capacity\. Compared with the strongest single\-view configuration \(full\-rank broad ridge\),DualViewimproves Pearson correlation by 11\.3% and reduces NRMSE by 2\.1%\. These results support the complementarity of the two evidence views, suggesting that target\-focused representations capture predictive information not fully captured by the broad view\.

Table 5:Training strategy ablations\.We compare our alternating optimization to two simpler strategies\. Separate fitting andα\\alphafusion train both predictors independently and combine their predictions using a validation\-tuned weight\. Fixed\-basis fitting holds the broad basis𝐀\\mathbf\{A\}fixed while jointly optimizing broad projection weights𝐁\\mathbf\{B\}and focused ridge weights𝐯\\mathbf\{v\}\.
### 6\.2Training Strategy Ablations

We also examine whether our alternating optimization improves over two simpler training strategies\.Separate fitting \+ fusiontrains the broad and focused predictors independently, then combines their predictions using a validation\-tuned weightα\\alpha\.Fixed\-basis joint fittingfirst trains the broad predictor alone, then fixes its learned basis𝐀\\mathbf\{A\}while jointly fitting the target\-specific weights𝐁\\mathbf\{B\}and focused predictor𝐯\\mathbf\{v\}\. All strategies use the same extracted representations and rank constraints\.

Table[5](https://arxiv.org/html/2610.19206#S6.T5)shows that alternating joint fitting achieves the strongest performance, improving Pearson correlation by 3\.1% and 4\.3% over separate fitting and fixed\-basis joint fitting, respectively, while reducing NRMSE by 0\.8% and 0\.9%\. Notably, fixed\-basis joint fitting does not improve over separate fitting, suggesting that joint optimization alone is insufficient\. Instead, the results suggest that adapting the broad predictive directions to focused evidence helps the two branches make more complementary contributions\.

Table 6:Target\-specific evidence alignment\.We compare matched and mismatched focused evidence, with and without broad evidence\. Mismatching assigns each target another target’s focused representation through a fixed cyclic permutation\.DualView\(\*\) uses both evidence views with matched semantics\.
### 6\.3Target\-Specific Evidence Alignment

The focused view assumes that target\-specific prompting extracts evidence particularly useful for the corresponding output\. To test this, we introduce a “mismatched” control that assigns each target to another target’s focused representation through cyclic permutation\. We refit focused predictors using their permuted features with and without the broad predictor, applying the permutation consistently during training and evaluation\.

Table[6](https://arxiv.org/html/2610.19206#S6.T6)shows that mismatched focused representations degrade predictive performance in both the focused\-only and the dual\-branch configurations\. The performance loss is more pronounced in the combined setting: Pearson drops by 10\.0% versus 5\.3%, and NRMSE increases by 1\.8% versus 1\.2%, respectively\.

Table 7:Parameter\-sharing ablations\.We compare reduced\- and full\-rank broad predictors with shared and independent focused predictors\.DualView\(\*\) uses reduced\-rank broad and shared focused predictors\.
### 6\.4Sharing Across Outputs

Finally, we assess the accuracy–parameter trade\-off introduced by the two forms of sharing inDualView\. We compare reduced\-rank and full\-rank broad predictors, crossed with shared and target\-specific focused predictors\. All configurations use the same extracted evidence and are fitted jointly using theL2L\_\{2\}\-regularized objective in§\\lx@sectionsign[3\.5](https://arxiv.org/html/2610.19206#S3.SS5)\.Table[7](https://arxiv.org/html/2610.19206#S6.T7)shows thatDualViewachieves comparable predictive performance to less constrained alternatives while fitting the fewest parameters\.

Parameter sharing reduces the trainable footprint in two ways\. First, the reduced\-rank broad predictor usesk⁡\(H\+D\)k\(H\+D\)rather thanH​DHDcoefficients for an unrestricted predictor, withk⁡\(H\+D\)<H​Dk\(H\+D\)<HDin all evaluated configurations\. Second, sharing the focused predictor reduces its coefficients fromH​DHDtoHH\. IncludingDDtarget\-specific intercepts,DualViewrequires only\(k\+1\)​\(H\+D\)\(k\+1\)\(H\+D\)parameters, compared with2​H​D\+D2HD\+Dfor full\-rank broad and independent focused predictors\.

## 7Discussion and Future Work

##### When does supervised adaptation pay off?

Comparisons with Gemini raise a practical cost question beyond predictive performance\. Direct prompting requires no task\-specific training, whereasDualViewincurs an upfront fitting cost but can use a smaller open\-weight LLM for repeated inference\. In principle, if the supervised system has lower marginal inference cost, this upfront cost can be amortized over sufficient requests\. ForDualView, however, the trade\-off also depends on the number of outputs: processingNNinputs requiresNNbroad andN​DNDfocused representations, although shared\-prefix reuse avoids redundant backbone computation\. The preferred configuration, potentially including hybrid strategies that balance latency and predictive performance, therefore depends on the practical deployment context\.

##### Extensions to distribution elicitation\.

Our multi\-output formulation may extend from predicting multiple semantic targets to predicting multiple points of a conditional distribution\.[Vedula et al\. \(2025\)](https://arxiv.org/html/2610.19206#bib.bib64)explored LLM\-based quantile regression for text\-based price prediction, and[Zhu et al\. \(2026b\)](https://arxiv.org/html/2610.19206#bib.bib2)extend it to general text\-to\-distribution settings using dense quantile grids\. Specifically, for quantilesτ1,…,τN\{\\tau\_\{1\},\\ldots,\\tau\_\{N\}\}, a text input maps to a vector of quantile estimates\[qτ1​\(x\),…,qτN​\(x\)\]\[q\_\{\\tau\_\{1\}\}\(x\),\\ldots,q\_\{\\tau\_\{N\}\}\(x\)\]\. To leverageDualView, we can treat each quantile level as a separate target and extract broad and quantile\-specific representations from the LLM\. Downstream ridge predictors can be applied as\-is since they are domain\-agnostic\. Training could use pinball loss with a monotonicity constraint to prevent quantile crossing\. We leave text\-to\-distribution extensions to future work\.

##### Integrating multi\-objective optimization\.

Multi\-output regression can also be viewed through a multi\-objective lens:DualViewuses the sameL2L\_\{2\}\-regularized objectives across targets, but each target contributes separate loss terms that can induce conflicting updates to shared parameters\. Multi\-objective optimization \(MOO\) methods offer a natural way to address such conflicts by reweighting or modifying output\-wise gradients during shared\-parameter training\([Sener and Koltun, 2018](https://arxiv.org/html/2610.19206#bib.bib31);[Efroni et al\., 2025](https://arxiv.org/html/2610.19206#bib.bib37);[Li et al\., 2025](https://arxiv.org/html/2610.19206#bib.bib39)\)\. MOO methods are thus complementary toDualView, modifying how supervision is balanced across outputs without affecting how to extract features from the LLM backbone\. Integrating MOO strategies is left for future work\.

##### Transfer to unseen targets\.

The shared focused predictor may generalize to unseen target descriptions without retraining\. However, the broad predictor uses coefficients learned for a fixed output set, so extending the fullDualViewmodel to new targets would require additional adaptation\. This would be another interesting direction to pursue\.

## 8Conclusion

We proposedDualView, a lightweight approach to multi\-output text regression that leverages broad and output\-specific representations from a frozen LLM\. A rank\-constrained multivariate predictor captures shared structure across outputs, while a shared scalar predictor uses evidence specialized to each target\. Across five tasks and eight LLM backbones,DualViewoutperforms seven supervised baselines while fitting substantially fewer parameters\. These advantages extend across nearly all output dimensions and persist even with limited supervision\. Ablations show that the two evidence views are complementary and that parameter sharing preserves performance with fewer parameters\.

## Limitations

##### Sensitivity to prompt design\.

BecauseDualViewextracts representations from LLMs, the extracted vectors may be sensitive to changes in prompt formatting, which can affect predictions\. We use fixed prompt templates for each dataset and do not evaluate alternative formulations\.

##### Dependence on LLM representations\.

DualViewrequires access to model internals to extract representations and thus may not work with black\-box models\. Because the LLM backbone is frozen,DualView’s predictive performance also depends on the quality of its pretrained representations and may degrade on unfamiliar domains\.

##### Nonconvex optimization\.

The rank\-constrained joint objective specified in§\\lx@sectionsign[3\.5](https://arxiv.org/html/2610.19206#S3.SS5)is nonconvex\. Although each alternating update optimally fits one predictor while holding the other fixed, our optimization strategy does not guarantee a globally optimal joint solution\.

##### Computational cost\.

WhileDualViewfits fewer parameters \(§\\lx@sectionsign[5\.2](https://arxiv.org/html/2610.19206#S5.SS2)\) versus supervised baselines, actual computational costs depend on the practical deployment environment\. Specifically,DualViewrequires one broad representation andDDfocused representations per input\. Although shared\-prefix reuse avoids redundant computation, theDDfocused suffixes still require separate processing\. Thus, inference computation grows approximately linearly with the number of targets, and fewer fitted parameters do not necessarily translate into lower inference latency or memory use\.

## References

- Akhauriet al\.\(2026\)Y\. Akhauri, X\. Song, A\. Wongpanich, B\. Lewandowski, and M\. S\. AbdelfattahRegression language models for code\.InProceedings of the 43rd International Conference on Machine Learning,External Links:[Link](https://icml.cc/virtual/2026/poster/60978)Cited by:[§1](https://arxiv.org/html/2610.19206#S1.p1.1)\.
- Allisonet al\.\(2023\)R\. Allison, A\. Stephenson, S\. F, and E\. O\. Pyzer\-KnappLeveraging locality and robustness to achieve massively scalable gaussian process regression\.InAdvances in Neural Information Processing Systems,A\. Oh, T\. Naumann, A\. Globerson, K\. Saenko, M\. Hardt, and S\. Levine \(Eds\.\),Vol\.36,pp\. 18906–18931\.External Links:[Document](https://dx.doi.org/10.52202/075280-0829),[Link](https://proceedings.neurips.cc/paper_files/paper/2023/file/3c2b60a3f269c404e9329ee119f2d34a-Paper-Conference.pdf)Cited by:[§D\.4](https://arxiv.org/html/2610.19206#A4.SS4.p1.1),[§4](https://arxiv.org/html/2610.19206#S4.SS0.SSS0.Px4.p1.1)\.
- Aymanet al\.\(2026\)A\. Ayman, A\. Mukhopadhyay, and A\. LaszkaEnsemble prediction of task affinity for efficient multi\-task learning\.InInternational Conference on Learning Representations,C\. Vondrick, B\. Hariharan, C\. Raffel, L\. Pinto, D\. Yang, and A\. Faust \(Eds\.\),Vol\.2026,pp\. 111821–111843\.External Links:[Link](https://proceedings.iclr.cc/paper_files/paper/2026/file/b595b039d8f3e04649e45879f487e707-Paper-Conference.pdf)Cited by:[§G\.1](https://arxiv.org/html/2610.19206#A7.SS1.p1.1)\.
- Baeket al\.\(2025\)J\. Baek, S\. K\. Jauhar, S\. Cucerzan, and S\. J\. HwangResearchAgent: iterative research idea generation over scientific literature with large language models\.InProceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies \(Volume 1: Long Papers\),L\. Chiruzzo, A\. Ritter, and L\. Wang \(Eds\.\),Albuquerque, New Mexico,pp\. 6709–6738\.External Links:[Document](https://dx.doi.org/10.18653/v1/2025.naacl-long.342),ISBN 979\-8\-89176\-189\-6,[Link](https://aclanthology.org/2025.naacl-long.342/)Cited by:[§1](https://arxiv.org/html/2610.19206#S1.p2.1)\.
- Ban and Ji \(2024\)H\. Ban and K\. JiFair resource allocation in multi\-task learning\.InProceedings of the 41st International Conference on Machine Learning,R\. Salakhutdinov, Z\. Kolter, K\. Heller, A\. Weller, N\. Oliver, J\. Scarlett, and F\. Berkenkamp \(Eds\.\),Proceedings of Machine Learning Research, Vol\.235,pp\. 2715–2731\.External Links:[Link](https://proceedings.mlr.press/v235/ban24a.html)Cited by:[§G\.1](https://arxiv.org/html/2610.19206#A7.SS1.p1.1)\.
- Chenet al\.\(2026a\)J\. Chen, Y\. Lu, X\. Wang, H\. Zeng, J\. Huang, J\. Gesi, Y\. Xu, B\. Yao, and D\. WangMulti\-agent\-as\-judge: aligning LLM\-agent\-based automated evaluation with multi\-dimensional human evaluation\.InProceedings of the 64th Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),M\. Liakata, V\. P\. Moreira, J\. Zhang, and D\. Jurgens \(Eds\.\),San Diego, California, United States,pp\. 17389–17416\.External Links:[Document](https://dx.doi.org/10.18653/v1/2026.acl-long.790),ISBN 979\-8\-89176\-390\-6,[Link](https://aclanthology.org/2026.acl-long.790/)Cited by:[§1](https://arxiv.org/html/2610.19206#S1.p2.1)\.
- Chenet al\.\(2026b\)M\. Chen, S\. Tang, R\. Tan, Z\. Li, J\. Chen, K\. Xue, and C\. QianBeyond token\-level supervision: unlocking the potential of decoding\-based regression via reinforcement learning\.InProceedings of the 43rd International Conference on Machine Learning,External Links:[Link](https://icml.cc/virtual/2026/poster/60986)Cited by:[§D\.3](https://arxiv.org/html/2610.19206#A4.SS3.SSS0.Px2),[§1](https://arxiv.org/html/2610.19206#S1.p1.1),[§1](https://arxiv.org/html/2610.19206#S1.p3.1),[§2\.1](https://arxiv.org/html/2610.19206#S2.SS1.p1.1),[§4](https://arxiv.org/html/2610.19206#S4.SS0.SSS0.Px3.p1.1)\.
- Chianget al\.\(2025\)C\. Chiang, H\. Lee, and M\. LukasikTRACT: regression\-aware fine\-tuning meets chain\-of\-thought reasoning for LLM\-as\-a\-judge\.InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),W\. Che, J\. Nabende, E\. Shutova, and M\. T\. Pilehvar \(Eds\.\),Vienna, Austria,pp\. 2934–2952\.External Links:[Document](https://dx.doi.org/10.18653/v1/2025.acl-long.147),ISBN 979\-8\-89176\-251\-0,[Link](https://aclanthology.org/2025.acl-long.147/)Cited by:[§B\.1](https://arxiv.org/html/2610.19206#A2.SS1.p2.1)\.
- Chunget al\.\(2026\)J\. Chung, S\. KIM, Y\. Jo, J\. Park, D\. Min, and Y\. YuTeaching metric distance to discrete autoregressive language models\.InInternational Conference on Learning Representations,C\. Vondrick, B\. Hariharan, C\. Raffel, L\. Pinto, D\. Yang, and A\. Faust \(Eds\.\),Vol\.2026,pp\. 11633–11655\.External Links:[Link](https://proceedings.iclr.cc/paper_files/paper/2026/file/13b45b44e26c353c64cba9529bf4724f-Paper-Conference.pdf)Cited by:[§B\.1](https://arxiv.org/html/2610.19206#A2.SS1.p2.1),[§1](https://arxiv.org/html/2610.19206#S1.p1.1)\.
- Dinget al\.\(2023\)Y\. Ding, M\. Bexte, and A\. HorbachScore it all together: a multi\-task learning study on automatic scoring of argumentative essays\.InFindings of the Association for Computational Linguistics: ACL 2023,A\. Rogers, J\. Boyd\-Graber, and N\. Okazaki \(Eds\.\),Toronto, Canada,pp\. 13052–13063\.External Links:[Document](https://dx.doi.org/10.18653/v1/2023.findings-acl.825),[Link](https://aclanthology.org/2023.findings-acl.825/)Cited by:[§B\.3](https://arxiv.org/html/2610.19206#A2.SS3.p1.1),[§2\.3](https://arxiv.org/html/2610.19206#S2.SS3.p1.1)\.
- Doet al\.\(2024a\)H\. Do, Y\. Kim, and G\. LeeAutoregressive score generation for multi\-trait essay scoring\.InFindings of the Association for Computational Linguistics: EACL 2024,Y\. Graham and M\. Purver \(Eds\.\),St\. Julian’s, Malta,pp\. 1659–1666\.External Links:[Document](https://dx.doi.org/10.18653/v1/2024.findings-eacl.115),[Link](https://aclanthology.org/2024.findings-eacl.115/)Cited by:[§B\.2](https://arxiv.org/html/2610.19206#A2.SS2.p2.1),[§1](https://arxiv.org/html/2610.19206#S1.p3.1)\.
- Doet al\.\(2024b\)H\. Do, S\. Ryu, and G\. LeeAutoregressive multi\-trait essay scoring via reinforcement learning with scoring\-aware multiple rewards\.InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing,Y\. Al\-Onaizan, M\. Bansal, and Y\. Chen \(Eds\.\),Miami, Florida, USA,pp\. 16427–16438\.External Links:[Document](https://dx.doi.org/10.18653/v1/2024.emnlp-main.917),[Link](https://aclanthology.org/2024.emnlp-main.917/)Cited by:[§G\.1](https://arxiv.org/html/2610.19206#A7.SS1.p2.1)\.
- Efroniet al\.\(2025\)Y\. Efroni, B\. Kretzu, D\. R\. Jiang, J\. Bhandari, Z\. Zhu, and K\. UllrichAligned multi objective optimization\.InProceedings of the 42nd International Conference on Machine Learning,A\. Singh, M\. Fazel, D\. Hsu, S\. Lacoste\-Julien, F\. Berkenkamp, T\. Maharaj, K\. Wagstaff, and J\. Zhu \(Eds\.\),Proceedings of Machine Learning Research, Vol\.267,pp\. 14989–15017\.External Links:[Link](https://proceedings.mlr.press/v267/efroni25a.html)Cited by:[§G\.1](https://arxiv.org/html/2610.19206#A7.SS1.p1.1),[§7](https://arxiv.org/html/2610.19206#S7.SS0.SSS0.Px3.p1.1)\.
- Fabbriet al\.\(2021\)A\. R\. Fabbri, W\. Kryściński, B\. McCann, C\. Xiong, R\. Socher, and D\. RadevSummEval: re\-evaluating summarization evaluation\.Transactions of the Association for Computational Linguistics9,pp\. 391–409\.External Links:[Document](https://dx.doi.org/10.1162/tacl%5Fa%5F00373),[Link](https://aclanthology.org/2021.tacl-1.24/)Cited by:[§B\.2](https://arxiv.org/html/2610.19206#A2.SS2.p1.1),[§2\.2](https://arxiv.org/html/2610.19206#S2.SS2.p1.1)\.
- Gemma Team \(2026\)Gemma TeamGemma 4 Technical Report\.Note:arXiv preprint arXiv:2607\.02770External Links:[Link](https://arxiv.org/abs/2607.02770)Cited by:[§D\.1](https://arxiv.org/html/2610.19206#A4.SS1.p1.1),[§4](https://arxiv.org/html/2610.19206#S4.SS0.SSS0.Px1.p1.1)\.
- Grattafioriet al\.\(2024\)A\. Grattafiori, A\. Dubey, A\. Jauhri, A\. Pandey, A\. Kadian, A\. Al\-Dahle, A\. Letman, A\. Mathur, A\. Schelten, A\. Vaughan,et al\.The Llama 3 herd of models\.arXiv preprint arXiv:2407\.21783\.External Links:[Link](https://arxiv.org/abs/2407.21783)Cited by:[§D\.1](https://arxiv.org/html/2610.19206#A4.SS1.p1.1),[§4](https://arxiv.org/html/2610.19206#S4.SS0.SSS0.Px1.p1.1)\.
- Hashemiet al\.\(2024\)H\. Hashemi, J\. Eisner, C\. Rosset, B\. Van Durme, and C\. KedzieLLM\-rubric: a multidimensional, calibrated approach to automated evaluation of natural language texts\.InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),L\. Ku, A\. Martins, and V\. Srikumar \(Eds\.\),Bangkok, Thailand,pp\. 13806–13834\.External Links:[Document](https://dx.doi.org/10.18653/v1/2024.acl-long.745),[Link](https://aclanthology.org/2024.acl-long.745/)Cited by:[§1](https://arxiv.org/html/2610.19206#S1.p2.1),[§2\.2](https://arxiv.org/html/2610.19206#S2.SS2.p1.1)\.
- Heet al\.\(2024\)Y\. He, S\. Zhou, G\. Zhang, H\. Yun, Y\. Xu, B\. Zeng, T\. Chilimbi, and H\. ZhaoRobust multi\-task learning with excess risks\.InProceedings of the 41st International Conference on Machine Learning,R\. Salakhutdinov, Z\. Kolter, K\. Heller, A\. Weller, N\. Oliver, J\. Scarlett, and F\. Berkenkamp \(Eds\.\),Proceedings of Machine Learning Research, Vol\.235,pp\. 18094–18114\.External Links:[Link](https://proceedings.mlr.press/v235/he24n.html)Cited by:[§G\.1](https://arxiv.org/html/2610.19206#A7.SS1.p1.1)\.
- Holzmülleret al\.\(2024\)D\. Holzmüller, L\. Grinsztajn, and I\. SteinwartBetter by default: strong pre\-tuned mlps and boosted trees on tabular data\.InAdvances in Neural Information Processing Systems,A\. Globerson, L\. Mackey, D\. Belgrave, A\. Fan, U\. Paquet, J\. Tomczak, and C\. Zhang \(Eds\.\),Vol\.37,pp\. 26577–26658\.External Links:[Document](https://dx.doi.org/10.52202/079017-0837),[Link](https://proceedings.neurips.cc/paper_files/paper/2024/file/2ee1c87245956e3eaa71aaba5f5753eb-Paper-Conference.pdf)Cited by:[§D\.4](https://arxiv.org/html/2610.19206#A4.SS4.p1.1),[§4](https://arxiv.org/html/2610.19206#S4.SS0.SSS0.Px4.p1.1)\.
- Huet al\.\(2026\)D\. Hu, L\. Wei, H\. Xiao, S\. Hu, and Y\. ZhangMulti\-task representation alignment on language understanding: a mutual information perspective\.InProceedings of the 64th Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),M\. Liakata, V\. P\. Moreira, J\. Zhang, and D\. Jurgens \(Eds\.\),San Diego, California, United States,pp\. 45722–45740\.External Links:[Document](https://dx.doi.org/10.18653/v1/2026.acl-long.2120),ISBN 979\-8\-89176\-390\-6,[Link](https://aclanthology.org/2026.acl-long.2120/)Cited by:[§B\.3](https://arxiv.org/html/2610.19206#A2.SS3.p1.1),[§2\.3](https://arxiv.org/html/2610.19206#S2.SS3.p1.1)\.
- Huet al\.\(2025\)D\. Hu, L\. Wei, W\. Zhou, and S\. HuImpartial multi\-task representation learning via variance\-invariant probabilistic decoding\.InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),W\. Che, J\. Nabende, E\. Shutova, and M\. T\. Pilehvar \(Eds\.\),Vienna, Austria,pp\. 19883–19897\.External Links:[Document](https://dx.doi.org/10.18653/v1/2025.acl-long.975),ISBN 979\-8\-89176\-251\-0,[Link](https://aclanthology.org/2025.acl-long.975/)Cited by:[§B\.3](https://arxiv.org/html/2610.19206#A2.SS3.p1.1),[§2\.3](https://arxiv.org/html/2610.19206#S2.SS3.p1.1)\.
- Izenman \(1975\)A\. J\. IzenmanReduced\-rank regression for the multivariate linear model\.Journal of Multivariate Analysis5\(2\),pp\. 248–264\.External Links:[Document](https://dx.doi.org/10.1016/0047-259x%2875%2990042-1),ISSN 0047\-259X,[Link](http://dx.doi.org/10.1016/0047-259X(75)90042-1)Cited by:[§3\.3](https://arxiv.org/html/2610.19206#S3.SS3.p1.2)\.
- Jeong and Yoon \(2025\)W\. Jeong and K\. YoonSelective task group updates for multi\-task optimization\.InInternational Conference on Learning Representations,Y\. Yue, A\. Garg, N\. Peng, F\. Sha, and R\. Yu \(Eds\.\),Vol\.2025,pp\. 95046–95079\.External Links:[Link](https://proceedings.iclr.cc/paper_files/paper/2025/file/ed3cd2520148b577039adfade82a5566-Paper-Conference.pdf)Cited by:[§G\.1](https://arxiv.org/html/2610.19206#A7.SS1.p1.1)\.
- Kimet al\.\(2025\)S\. Kim, J\. Suk, J\. Y\. Cho, S\. Longpre, C\. Kim, D\. Yoon, G\. Son, Y\. Cho, S\. Shafayat, J\. Baek, S\. H\. Park, H\. Hwang, J\. Jo, H\. Cho, H\. Shin, S\. Lee, H\. Oh, N\. Lee, N\. Ho, S\. J\. Joo, M\. Ko, Y\. Lee, H\. Chae, J\. Shin, J\. Jang, S\. Ye, B\. Y\. Lin, S\. Welleck, G\. Neubig, M\. Lee, K\. Lee, and M\. SeoThe BiGGen bench: a principled benchmark for fine\-grained evaluation of language models with language models\.InProceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies \(Volume 1: Long Papers\),L\. Chiruzzo, A\. Ritter, and L\. Wang \(Eds\.\),Albuquerque, New Mexico,pp\. 5877–5919\.External Links:[Document](https://dx.doi.org/10.18653/v1/2025.naacl-long.303),ISBN 979\-8\-89176\-189\-6,[Link](https://aclanthology.org/2025.naacl-long.303/)Cited by:[§1](https://arxiv.org/html/2610.19206#S1.p2.1),[§2\.2](https://arxiv.org/html/2610.19206#S2.SS2.p1.1)\.
- Kirket al\.\(2024\)H\. R\. Kirk, A\. Whitefield, P\. Röttger, A\. Bean, K\. Margatina, J\. Ciro, R\. Mosquera, M\. Bartolo, A\. Williams, H\. He, B\. Vidgen, and S\. HaleThe prism alignment dataset: what participatory, representative and individualised human feedback reveals about the subjective and multicultural alignment of large language models\.InAdvances in Neural Information Processing Systems,A\. Globerson, L\. Mackey, D\. Belgrave, A\. Fan, U\. Paquet, J\. Tomczak, and C\. Zhang \(Eds\.\),Vol\.37,pp\. 105236–105344\.External Links:[Document](https://dx.doi.org/10.52202/079017-3342),[Link](https://proceedings.neurips.cc/paper_files/paper/2024/file/be2e1b68b44f2419e19f6c35a1b8cf35-Paper-Datasets_and_Benchmarks_Track.pdf)Cited by:[§D\.2](https://arxiv.org/html/2610.19206#A4.SS2.SSS0.Px6),[§1](https://arxiv.org/html/2610.19206#S1.p5.1),[§4](https://arxiv.org/html/2610.19206#S4.SS0.SSS0.Px2.p1.1)\.
- Kolaet al\.\(2026\)S\. Kola, P\. K\. Sharma, S\. Dey, B\. Mishra, and M\. DatarSAJA: a simple approach to judge alignment for LLM\-as\-a\-judge\.InProceedings of the 64th Annual Meeting of the Association for Computational Linguistics \(Volume 6: Industry Track\),Y\. Li, G\. Rehm, and M\. Tu \(Eds\.\),San Diego, California, USA,pp\. 646–664\.External Links:[Document](https://dx.doi.org/10.18653/v1/2026.acl-industry.45),ISBN 979\-8\-89176\-394\-4,[Link](https://aclanthology.org/2026.acl-industry.45/)Cited by:[§B\.2](https://arxiv.org/html/2610.19206#A2.SS2.p1.1),[§D\.3](https://arxiv.org/html/2610.19206#A4.SS3.SSS0.Px5),[§2\.2](https://arxiv.org/html/2610.19206#S2.SS2.p1.1),[§4](https://arxiv.org/html/2610.19206#S4.SS0.SSS0.Px3.p1.1)\.
- Liet al\.\(2025\)C\. Li, H\. Zhang, Y\. Xu, H\. Xue, X\. Ao, and Q\. HeGradient\-adaptive policy optimization: towards multi\-objective alignment of large language models\.InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),W\. Che, J\. Nabende, E\. Shutova, and M\. T\. Pilehvar \(Eds\.\),Vienna, Austria,pp\. 11214–11232\.External Links:[Document](https://dx.doi.org/10.18653/v1/2025.acl-long.549),ISBN 979\-8\-89176\-251\-0,[Link](https://aclanthology.org/2025.acl-long.549/)Cited by:[§7](https://arxiv.org/html/2610.19206#S7.SS0.SSS0.Px3.p1.1)\.
- Li and Ng \(2025\)S\. Li and V\. NgGraph\-based multi\-trait essay scoring\.InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing,C\. Christodoulopoulos, T\. Chakraborty, C\. Rose, and V\. Peng \(Eds\.\),Suzhou, China,pp\. 33325–33351\.External Links:[Document](https://dx.doi.org/10.18653/v1/2025.emnlp-main.1691),ISBN 979\-8\-89176\-332\-6,[Link](https://aclanthology.org/2025.emnlp-main.1691/)Cited by:[§B\.2](https://arxiv.org/html/2610.19206#A2.SS2.p2.1),[§D\.3](https://arxiv.org/html/2610.19206#A4.SS3.SSS0.Px4),[§1](https://arxiv.org/html/2610.19206#S1.p3.1),[§1](https://arxiv.org/html/2610.19206#S1.p5.1),[§2\.1](https://arxiv.org/html/2610.19206#S2.SS1.p2.1),[§4](https://arxiv.org/html/2610.19206#S4.SS0.SSS0.Px3.p1.1)\.
- Linet al\.\(2023\)V\. Lin, L\. Morency, D\. Dimitriadis, and S\. SharmaCounterfactual augmentation for multimodal learning under presentation bias\.InFindings of the Association for Computational Linguistics: EMNLP 2023,H\. Bouamor, J\. Pino, and K\. Bali \(Eds\.\),Singapore,pp\. 592–606\.External Links:[Document](https://dx.doi.org/10.18653/v1/2023.findings-emnlp.43),[Link](https://aclanthology.org/2023.findings-emnlp.43/)Cited by:[§D\.4](https://arxiv.org/html/2610.19206#A4.SS4.p1.1),[§4](https://arxiv.org/html/2610.19206#S4.SS0.SSS0.Px4.p1.1)\.
- Liuet al\.\(2023a\)B\. Liu, Y\. Feng, P\. Stone, and Q\. LiuFAMO: fast adaptive multitask optimization\.InAdvances in Neural Information Processing Systems,A\. Oh, T\. Naumann, A\. Globerson, K\. Saenko, M\. Hardt, and S\. Levine \(Eds\.\),Vol\.36,pp\. 57226–57243\.External Links:[Document](https://dx.doi.org/10.52202/075280-2500),[Link](https://proceedings.neurips.cc/paper_files/paper/2023/file/b2fe1ee8d936ac08dd26f2ff58986c8f-Paper-Conference.pdf)Cited by:[§G\.1](https://arxiv.org/html/2610.19206#A7.SS1.p1.1)\.
- Liuet al\.\(2021\)B\. Liu, X\. Liu, X\. Jin, P\. Stone, and Q\. LiuConflict\-averse gradient descent for multi\-task learning\.InAdvances in Neural Information Processing Systems,M\. Ranzato, A\. Beygelzimer, Y\. Dauphin, P\.S\. Liang, and J\. W\. Vaughan \(Eds\.\),Vol\.34,pp\. 18878–18890\.External Links:[Link](https://proceedings.neurips.cc/paper_files/paper/2021/file/9d27fdf2477ffbff837d73ef7ae23db9-Paper.pdf)Cited by:[§G\.1](https://arxiv.org/html/2610.19206#A7.SS1.p1.1)\.
- Liuet al\.\(2023b\)Y\. Liu, D\. Iter, Y\. Xu, S\. Wang, R\. Xu, and C\. ZhuG\-eval: NLG evaluation using gpt\-4 with better human alignment\.InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing,H\. Bouamor, J\. Pino, and K\. Bali \(Eds\.\),Singapore,pp\. 2511–2522\.External Links:[Document](https://dx.doi.org/10.18653/v1/2023.emnlp-main.153),[Link](https://aclanthology.org/2023.emnlp-main.153/)Cited by:[§B\.2](https://arxiv.org/html/2610.19206#A2.SS2.p1.1),[§1](https://arxiv.org/html/2610.19206#S1.p2.1),[§2\.2](https://arxiv.org/html/2610.19206#S2.SS2.p1.1)\.
- Luet al\.\(2026\)L\. Lu, M\. Liu, P\. C\. Lu, Y\. Tian, S\. Sun, and N\. PengRethinking creativity evaluation: a critical analysis of existing creativity evaluations\.InProceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics \(Volume 1: Long Papers\),V\. Demberg, K\. Inui, and L\. Marquez \(Eds\.\),Rabat, Morocco,pp\. 6329–6352\.External Links:[Document](https://dx.doi.org/10.18653/v1/2026.eacl-long.297),ISBN 979\-8\-89176\-380\-7,[Link](https://aclanthology.org/2026.eacl-long.297/)Cited by:[§D\.2](https://arxiv.org/html/2610.19206#A4.SS2.SSS0.Px4),[§1](https://arxiv.org/html/2610.19206#S1.p5.1),[§4](https://arxiv.org/html/2610.19206#S4.SS0.SSS0.Px2.p1.1)\.
- Lukasiket al\.\(2025\)M\. Lukasik, Z\. Meng, H\. Narasimhan, Y\. Chang, A\. K\. Menon, F\. Yu, and S\. KumarBetter autoregressive regression with llms via regression\-aware fine\-tuning\.InInternational Conference on Learning Representations,Y\. Yue, A\. Garg, N\. Peng, F\. Sha, and R\. Yu \(Eds\.\),Vol\.2025,pp\. 69871–69901\.External Links:[Link](https://proceedings.iclr.cc/paper_files/paper/2025/file/ae3bb0adcdce28afb5f7e86a79b6cffa-Paper-Conference.pdf)Cited by:[§B\.1](https://arxiv.org/html/2610.19206#A2.SS1.p2.1),[§1](https://arxiv.org/html/2610.19206#S1.p1.1),[§1](https://arxiv.org/html/2610.19206#S1.p3.1),[§2\.1](https://arxiv.org/html/2610.19206#S2.SS1.p1.1)\.
- Lukasiket al\.\(2024\)M\. Lukasik, H\. Narasimhan, A\. K\. Menon, F\. Yu, and S\. KumarRegression aware inference with LLMs\.InFindings of the Association for Computational Linguistics: EMNLP 2024,Y\. Al\-Onaizan, M\. Bansal, and Y\. Chen \(Eds\.\),Miami, Florida, USA,pp\. 13667–13678\.External Links:[Document](https://dx.doi.org/10.18653/v1/2024.findings-emnlp.799),[Link](https://aclanthology.org/2024.findings-emnlp.799/)Cited by:[§2\.1](https://arxiv.org/html/2610.19206#S2.SS1.p1.1)\.
- Minet al\.\(2025\)H\. Min, Y\. Lee, M\. Ban, J\. Deng, N\. H\. Kim, T\. Yun, H\. Su, J\. Cai, and H\. SongTowards multi\-dimensional evaluation of LLM summarization across domains and languages\.InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),W\. Che, J\. Nabende, E\. Shutova, and M\. T\. Pilehvar \(Eds\.\),Vienna, Austria,pp\. 14417–14450\.External Links:[Document](https://dx.doi.org/10.18653/v1/2025.acl-long.702),ISBN 979\-8\-89176\-251\-0,[Link](https://aclanthology.org/2025.acl-long.702/)Cited by:[§D\.2](https://arxiv.org/html/2610.19206#A4.SS2.SSS0.Px2),[§1](https://arxiv.org/html/2610.19206#S1.p5.1),[§4](https://arxiv.org/html/2610.19206#S4.SS0.SSS0.Px2.p1.1)\.
- Nafaret al\.\(2025\)A\. Nafar, K\. B\. Venable, and P\. KordjamshidiLearning vs retrieval: the role of in\-context examples in regression with large language models\.InProceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies \(Volume 1: Long Papers\),L\. Chiruzzo, A\. Ritter, and L\. Wang \(Eds\.\),Albuquerque, New Mexico,pp\. 8206–8229\.External Links:[Document](https://dx.doi.org/10.18653/v1/2025.naacl-long.417),ISBN 979\-8\-89176\-189\-6,[Link](https://aclanthology.org/2025.naacl-long.417/)Cited by:[§1](https://arxiv.org/html/2610.19206#S1.p1.1)\.
- Navonet al\.\(2022\)A\. Navon, A\. Shamsian, I\. Achituve, H\. Maron, K\. Kawaguchi, G\. Chechik, and E\. FetayaMulti\-task learning as a bargaining game\.InProceedings of the 39th International Conference on Machine Learning,K\. Chaudhuri, S\. Jegelka, L\. Song, C\. Szepesvari, G\. Niu, and S\. Sabato \(Eds\.\),Proceedings of Machine Learning Research, Vol\.162,pp\. 16428–16446\.External Links:[Link](https://proceedings.mlr.press/v162/navon22a.html)Cited by:[§G\.1](https://arxiv.org/html/2610.19206#A7.SS1.p1.1)\.
- OpenAI \(2025\)OpenAIgpt\-oss\-120b & gpt\-oss\-20b Model Card\.Note:arXiv preprint arXiv:2508\.10925External Links:[Link](https://arxiv.org/abs/2508.10925)Cited by:[§D\.1](https://arxiv.org/html/2610.19206#A4.SS1.p1.1),[§4](https://arxiv.org/html/2610.19206#S4.SS0.SSS0.Px1.p1.1)\.
- Ouyanget al\.\(2026\)S\. Ouyang, D\. HUANG, J\. Guo, Z\. Sun, Q\. Zhu, and J\. M\. ZhangDSCodeBench: a realistic benchmark for data science code generation\.Proceedings of the AAAI Conference on Artificial Intelligence40\(38\),pp\. 32628–32636\.External Links:[Document](https://dx.doi.org/10.1609/aaai.v40i38.40540),ISSN 2159\-5399,[Link](http://dx.doi.org/10.1609/aaai.v40i38.40540)Cited by:[§D\.2](https://arxiv.org/html/2610.19206#A4.SS2.SSS0.Px5),[§1](https://arxiv.org/html/2610.19206#S1.p5.1),[§4](https://arxiv.org/html/2610.19206#S4.SS0.SSS0.Px2.p1.1)\.
- Piskorzet al\.\(2026\)J\. Piskorz, K\. Kobalczyk, and M\. van der SchaarEliciting numerical predictive distributions of llms without auto\-regression\.InInternational Conference on Learning Representations,C\. Vondrick, B\. Hariharan, C\. Raffel, L\. Pinto, D\. Yang, and A\. Faust \(Eds\.\),Vol\.2026,pp\. 135557–135582\.External Links:[Link](https://proceedings.iclr.cc/paper_files/paper/2026/file/db79b6c50430a52fbab9d63efb8433ca-Paper-Conference.pdf)Cited by:[§B\.1](https://arxiv.org/html/2610.19206#A2.SS1.p3.1)\.
- Popovet al\.\(2025\)D\. Popov, V\. Negodin, E\. Enikeeva, I\. Matrosova, N\. Karpachev, and M\. RyabininRefined assessment for translation evaluation: rethinking machine translation evaluation in the era of human\-level systems\.InFindings of the Association for Computational Linguistics: EMNLP 2025,C\. Christodoulopoulos, T\. Chakraborty, C\. Rose, and V\. Peng \(Eds\.\),Suzhou, China,pp\. 22079–22095\.External Links:[Document](https://dx.doi.org/10.18653/v1/2025.findings-emnlp.1203),ISBN 979\-8\-89176\-335\-7,[Link](https://aclanthology.org/2025.findings-emnlp.1203/)Cited by:[§D\.2](https://arxiv.org/html/2610.19206#A4.SS2.SSS0.Px3),[§1](https://arxiv.org/html/2610.19206#S1.p2.1),[§1](https://arxiv.org/html/2610.19206#S1.p5.1),[§4](https://arxiv.org/html/2610.19206#S4.SS0.SSS0.Px2.p1.1)\.
- Qiuet al\.\(2025\)R\. Qiu, W\. Zeng, J\. Ezick, C\. Lott, and H\. TongHow efficient is llm\-generated code? a rigorous & high\-standard benchmark\.InInternational Conference on Learning Representations,Y\. Yue, A\. Garg, N\. Peng, F\. Sha, and R\. Yu \(Eds\.\),Vol\.2025,pp\. 2233–2261\.External Links:[Link](https://proceedings.iclr.cc/paper_files/paper/2025/file/06694da057cb15fef11542270a592627-Paper-Conference.pdf)Cited by:[§1](https://arxiv.org/html/2610.19206#S1.p2.1)\.
- Qwen Team \(2025\)Qwen TeamQwen2\.5 Technical Report\.Note:arXiv preprint arXiv:2412\.15115External Links:[Link](https://arxiv.org/abs/2412.15115)Cited by:[§D\.1](https://arxiv.org/html/2610.19206#A4.SS1.p1.1),[§4](https://arxiv.org/html/2610.19206#S4.SS0.SSS0.Px1.p1.1)\.
- Qwen Team \(2026\)Qwen TeamQwen3\. 5\-omni technical report\.arXiv preprint arXiv:2604\.15804\.Cited by:[§D\.1](https://arxiv.org/html/2610.19206#A4.SS1.p1.1),[§4](https://arxiv.org/html/2610.19206#S4.SS0.SSS0.Px1.p1.1)\.
- Saad\-Falconet al\.\(2025\)J\. Saad\-Falcon, R\. P\. Vivek, W\. Berrios, N\. S\. Naik, M\. Franklin, B\. Vidgen, A\. Singh, D\. Kiela, and S\. MehriLMUNIT: fine\-grained evaluation with natural language unit tests\.InFindings of the Association for Computational Linguistics: EMNLP 2025,C\. Christodoulopoulos, T\. Chakraborty, C\. Rose, and V\. Peng \(Eds\.\),Suzhou, China,pp\. 3303–3324\.External Links:[Document](https://dx.doi.org/10.18653/v1/2025.findings-emnlp.176),ISBN 979\-8\-89176\-335\-7,[Link](https://aclanthology.org/2025.findings-emnlp.176/)Cited by:[§G\.1](https://arxiv.org/html/2610.19206#A7.SS1.p2.1)\.
- Sener and Koltun \(2018\)O\. Sener and V\. KoltunMulti\-task learning as multi\-objective optimization\.InAdvances in Neural Information Processing Systems,S\. Bengio, H\. Wallach, H\. Larochelle, K\. Grauman, N\. Cesa\-Bianchi, and R\. Garnett \(Eds\.\),Vol\.31\.External Links:[Link](https://proceedings.neurips.cc/paper_files/paper/2018/file/432aca3a1e345e339f35a30c8f65edce-Paper.pdf)Cited by:[§G\.1](https://arxiv.org/html/2610.19206#A7.SS1.p1.1),[§7](https://arxiv.org/html/2610.19206#S7.SS0.SSS0.Px3.p1.1)\.
- Shayeghet al\.\(2025\)B\. Shayegh, J\. Peter, D\. Vilar, T\. Domhan, J\. Juraska, M\. Freitag, and L\. MouFeeding two birds or favoring one? adequacy–fluency tradeoffs in evaluation and meta\-evaluation of machine translation\.InProceedings of the Tenth Conference on Machine Translation,B\. Haddow, T\. Kocmi, P\. Koehn, and C\. Monz \(Eds\.\),Suzhou, China,pp\. 269–285\.External Links:[Document](https://dx.doi.org/10.18653/v1/2025.wmt-1.16),ISBN 979\-8\-89176\-341\-8,[Link](https://aclanthology.org/2025.wmt-1.16/)Cited by:[§1](https://arxiv.org/html/2610.19206#S1.p2.1)\.
- Siet al\.\(2025\)C\. Si, D\. Yang, and T\. HashimotoCan llms generate novel research ideas? a large\-scale human study with 100\+ nlp researchers\.InInternational Conference on Learning Representations,Y\. Yue, A\. Garg, N\. Peng, F\. Sha, and R\. Yu \(Eds\.\),Vol\.2025,pp\. 94003–94092\.External Links:[Link](https://proceedings.iclr.cc/paper_files/paper/2025/file/ea94957d81b1c1caf87ef5319fa6b467-Paper-Conference.pdf)Cited by:[§1](https://arxiv.org/html/2610.19206#S1.p2.1)\.
- Song and Bahri \(2025\)X\. Song and D\. BahriDecoding\-based regression\.Transactions on Machine Learning Research\.External Links:ISSN 2835\-8856,[Link](https://openreview.net/forum?id=avUQ8jguxg)Cited by:[§2\.1](https://arxiv.org/html/2610.19206#S2.SS1.p1.1)\.
- Su and Lease \(2026\)Y\. Su and M\. LeaseRELISH: LLM REgression with a latent iterative state head\.InThird Conference on Language Modeling,External Links:[Link](https://openreview.net/forum?id=IebWFoQwdz)Cited by:[§C\.1](https://arxiv.org/html/2610.19206#A3.SS1.p1.2),[§D\.3](https://arxiv.org/html/2610.19206#A4.SS3.SSS0.Px1),[§1](https://arxiv.org/html/2610.19206#S1.p1.1),[§2\.1](https://arxiv.org/html/2610.19206#S2.SS1.p1.1),[§2\.1](https://arxiv.org/html/2610.19206#S2.SS1.p2.1),[§3\.1](https://arxiv.org/html/2610.19206#S3.SS1.p2.1),[§4](https://arxiv.org/html/2610.19206#S4.SS0.SSS0.Px3.p1.1)\.
- Tanet al\.\(2025\)R\. Tan, M\. Chen, K\. Xue, Y\. Wang, Y\. Wang, F\. Sheng, and C\. QianTowards universal offline black\-box optimization via learning language model embeddings\.InProceedings of the 42nd International Conference on Machine Learning,A\. Singh, M\. Fazel, D\. Hsu, S\. Lacoste\-Julien, F\. Berkenkamp, T\. Maharaj, K\. Wagstaff, and J\. Zhu \(Eds\.\),Proceedings of Machine Learning Research, Vol\.267,pp\. 58499–58544\.External Links:[Link](https://proceedings.mlr.press/v267/tan25b.html)Cited by:[§B\.3](https://arxiv.org/html/2610.19206#A2.SS3.p3.1),[§D\.3](https://arxiv.org/html/2610.19206#A4.SS3.SSS0.Px6),[§2\.3](https://arxiv.org/html/2610.19206#S2.SS3.p1.1),[§4](https://arxiv.org/html/2610.19206#S4.SS0.SSS0.Px3.p1.1)\.
- Tanget al\.\(2025\)E\. Tang, B\. Yang, and X\. SongUnderstanding LLM embeddings for regression\.Transactions on Machine Learning Research\.External Links:ISSN 2835\-8856,[Link](https://openreview.net/forum?id=Wt6Iz5XNIO)Cited by:[§B\.1](https://arxiv.org/html/2610.19206#A2.SS1.p3.1),[§C\.1](https://arxiv.org/html/2610.19206#A3.SS1.p1.2),[§2\.1](https://arxiv.org/html/2610.19206#S2.SS1.p1.1),[§3\.1](https://arxiv.org/html/2610.19206#S3.SS1.p2.1)\.
- Vacareanuet al\.\(2024\)R\. Vacareanu, V\. Negru, V\. Suciu, and M\. SurdeanuFrom words to numbers: your large language model is secretly a capable regressor when given in\-context examples\.InFirst Conference on Language Modeling,External Links:[Link](https://openreview.net/forum?id=LzpaUxcNFK)Cited by:[§2\.1](https://arxiv.org/html/2610.19206#S2.SS1.p1.1)\.
- Vedulaet al\.\(2025\)N\. Vedula, D\. Dhyani, L\. Jalali, B\. N\. Oreshkin, M\. Bayati, and S\. MalmasiQuantile regression with large language models for price prediction\.InFindings of the Association for Computational Linguistics: ACL 2025,W\. Che, J\. Nabende, E\. Shutova, and M\. T\. Pilehvar \(Eds\.\),Vienna, Austria,pp\. 12396–12415\.External Links:[Document](https://dx.doi.org/10.18653/v1/2025.findings-acl.641),ISBN 979\-8\-89176\-256\-5,[Link](https://aclanthology.org/2025.findings-acl.641/)Cited by:[§7](https://arxiv.org/html/2610.19206#S7.SS0.SSS0.Px2.p1.1)\.
- Wanget al\.\(2024\)H\. Wang, W\. Xiong, T\. Xie, H\. Zhao, and T\. ZhangInterpretable preferences via multi\-objective reward modeling and mixture\-of\-experts\.InFindings of the Association for Computational Linguistics: EMNLP 2024,Y\. Al\-Onaizan, M\. Bansal, and Y\. Chen \(Eds\.\),Miami, Florida, USA,pp\. 10582–10592\.External Links:[Document](https://dx.doi.org/10.18653/v1/2024.findings-emnlp.620),[Link](https://aclanthology.org/2024.findings-emnlp.620/)Cited by:[§G\.1](https://arxiv.org/html/2610.19206#A7.SS1.p2.1)\.
- Wang and Liu \(2025\)J\. Wang and J\. LiuT\-MES: trait\-aware mix\-of\-experts representation learning for multi\-trait essay scoring\.InProceedings of the 31st International Conference on Computational Linguistics,O\. Rambow, L\. Wanner, M\. Apidianaki, H\. Al\-Khalifa, B\. D\. Eugenio, and S\. Schockaert \(Eds\.\),Abu Dhabi, UAE,pp\. 1224–1236\.External Links:[Link](https://aclanthology.org/2025.coling-main.81/)Cited by:[§B\.2](https://arxiv.org/html/2610.19206#A2.SS2.p2.1),[§D\.3](https://arxiv.org/html/2610.19206#A4.SS3.SSS0.Px3),[§2\.1](https://arxiv.org/html/2610.19206#S2.SS1.p2.1),[§4](https://arxiv.org/html/2610.19206#S4.SS0.SSS0.Px3.p1.1)\.
- Yeet al\.\(2026\)T\. Ye, T\. Ma, X\. Zhang, H\. Yu, J\. Yin, and W\. WangA problem\-oriented perspective and anchor verification for code optimization\.InInternational Conference on Learning Representations,C\. Vondrick, B\. Hariharan, C\. Raffel, L\. Pinto, D\. Yang, and A\. Faust \(Eds\.\),Vol\.2026,pp\. 47925–47957\.External Links:[Link](https://proceedings.iclr.cc/paper_files/paper/2026/file/4f5aa418348643f2041a55bba9afa7dc-Paper-Conference.pdf)Cited by:[§1](https://arxiv.org/html/2610.19206#S1.p2.1)\.
- Yuet al\.\(2020\)T\. Yu, S\. Kumar, A\. Gupta, S\. Levine, K\. Hausman, and C\. FinnGradient surgery for multi\-task learning\.InAdvances in Neural Information Processing Systems,H\. Larochelle, M\. Ranzato, R\. Hadsell, M\.F\. Balcan, and H\. Lin \(Eds\.\),Vol\.33,pp\. 5824–5836\.External Links:[Link](https://proceedings.neurips.cc/paper_files/paper/2020/file/3fe78a8acf5fda99de95303940a2420c-Paper.pdf)Cited by:[§G\.1](https://arxiv.org/html/2610.19206#A7.SS1.p1.1)\.
- Yuchiet al\.\(2026\)F\. Yuchi, L\. Du, and J\. EisnerLLMs know more about numbers than they can say\.InProceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics \(Volume 2: Short Papers\),V\. Demberg, K\. Inui, and L\. Marquez \(Eds\.\),Rabat, Morocco,pp\. 659–673\.External Links:[Document](https://dx.doi.org/10.18653/v1/2026.eacl-short.47),ISBN 979\-8\-89176\-381\-4,[Link](https://aclanthology.org/2026.eacl-short.47/)Cited by:[§B\.1](https://arxiv.org/html/2610.19206#A2.SS1.p3.1)\.
- Zhanget al\.\(2026a\)Y\. Zhang, T\. Chen, M\. Zhou, O\. Leong, Y\. N\. Wu, and M\. LukasikREAL: regression\-aware reinforcement learning for llm\-as\-a\-judge\.InProceedings of the 43rd International Conference on Machine Learning,External Links:[Link](https://icml.cc/virtual/2026/poster/62715)Cited by:[§B\.1](https://arxiv.org/html/2610.19206#A2.SS1.p2.1),[§1](https://arxiv.org/html/2610.19206#S1.p1.1),[§2\.1](https://arxiv.org/html/2610.19206#S2.SS1.p2.1)\.
- Zhanget al\.\(2026b\)Z\. Zhang, H\. Liu, X\. Li, Z\. Dai, J\. Zeng, F\. Wang, M\. Lin, R\. Chandradevan, L\. Wu, Z\. Li, C\. Luo, Z\. Wu, X\. Tang, Q\. He, and S\. WangBradley\-terry and multi\-objective reward modeling are complementary\.InInternational Conference on Learning Representations,C\. Vondrick, B\. Hariharan, C\. Raffel, L\. Pinto, D\. Yang, and A\. Faust \(Eds\.\),Vol\.2026,pp\. 22736–22776\.External Links:[Link](https://proceedings.iclr.cc/paper_files/paper/2026/file/2703e3329fef63ed7690f1b8145f25aa-Paper-Conference.pdf)Cited by:[§G\.1](https://arxiv.org/html/2610.19206#A7.SS1.p2.1)\.
- Zhuet al\.\(2026a\)Q\. Zhu, E\. K\. Oermann, and K\. ChoMulti\-task bayesian in\-context learning\.InProceedings of the 43rd International Conference on Machine Learning,External Links:[Link](https://icml.cc/virtual/2026/poster/63501)Cited by:[§B\.3](https://arxiv.org/html/2610.19206#A2.SS3.p3.1),[§D\.3](https://arxiv.org/html/2610.19206#A4.SS3.SSS0.Px7),[§2\.3](https://arxiv.org/html/2610.19206#S2.SS3.p1.1.3),[§4](https://arxiv.org/html/2610.19206#S4.SS0.SSS0.Px3.p1.1)\.
- Zhuet al\.\(2026b\)Y\. Zhu, Y\. Zhuang, N\. Vedula, D\. Dhyani, S\. Xu, M\. Bayati, B\. Wang, and S\. MalmasiText\-to\-distribution prediction with quantile tokens and neighbor context\.InProceedings of the 64th Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),M\. Liakata, V\. P\. Moreira, J\. Zhang, and D\. Jurgens \(Eds\.\),San Diego, California, United States,pp\. 16632–16648\.External Links:[Document](https://dx.doi.org/10.18653/v1/2026.acl-long.758),ISBN 979\-8\-89176\-390\-6,[Link](https://aclanthology.org/2026.acl-long.758/)Cited by:[§7](https://arxiv.org/html/2610.19206#S7.SS0.SSS0.Px2.p1.1)\.
- Zuoet al\.\(2026\)Z\. Zuo, L\. Yue, W\. Zheng, C\. Wang, and X\. LiuEnhancing numerical prediction in llms via smooth mmd alignment\.InProceedings of the 43rd International Conference on Machine Learning,External Links:[Link](https://icml.cc/virtual/2026/poster/64143)Cited by:[§B\.1](https://arxiv.org/html/2610.19206#A2.SS1.p2.1),[§1](https://arxiv.org/html/2610.19206#S1.p1.1)\.

## Appendix AAppendix Organization

We organize the appendix as follows:

1. 1\.Extended Related Work\(§\\lx@sectionsign[B](https://arxiv.org/html/2610.19206#A2)\) further discusses LLM\-based text regression methods, fine\-grained text evaluation, and multi\-task learning\.
2. 2\.Additional Method Details\(§\\lx@sectionsign[C](https://arxiv.org/html/2610.19206#A3)\) details evidence extraction, feature standardization, broad and focused predictors, and alternating training optimization\.
3. 3\.Additional Experimental Setup\(§\\lx@sectionsign[D](https://arxiv.org/html/2610.19206#A4)\) describes the LLMs, datasets, baselines, evaluation metrics, and hyperparameters\.
4. 4\.Additional Key Results\(§\\lx@sectionsign[E](https://arxiv.org/html/2610.19206#A5)\) provides more fine\-grained predictive results, parameter counts, training\-data efficiency results, and comparisons with Gemini 3\.1 Pro\.
5. 5\.Additional Ablations \(§\\lx@sectionsignF\) examines strong linear readout baselines and the accuracy–parameter\-sharing trade\-off inDualView\.
6. 6\.Expanded Discussion\(§\\lx@sectionsign[G](https://arxiv.org/html/2610.19206#A7)\) discusses multi\-objective optimization as a possible extension ofDualView\.
7. 7\.Dataset Prompt Templates\(§\\lx@sectionsign[H](https://arxiv.org/html/2610.19206#A8)\) specifies the broad prompts and target\-specific suffixes used for all five datasets\.
8. 8\.Output\-Level Performance Tables\(§\\lx@sectionsign[I](https://arxiv.org/html/2610.19206#A9)\) report NRMSE, Pearson, and Spearman results for each output, dataset, and LLM across all supervised methods\.

## Appendix BExtended Related Work

### B\.1LLM\-based Text Regression

The three families described in§\\lx@sectionsign[2\.1](https://arxiv.org/html/2610.19206#S2.SS1)mainly categorize LLM\-based text regression methods by how they obtain numerical predictions\. Another important consideration is how they fine\-tune the underlying LLM\. Autoregressive decoding and regression\-aware inference typically update the LLM through supervised fine\-tuning or reinforcement learning that updates all model parameters, whereas predictive heads typically freeze the base LLM and only train the auxiliary modules\.

SinceDualViewfreezes the LLM, we adopt the same setting for all supervised baselines\. We therefore exclude methods whose primary contributions involve fine\-tuning the LLM, since freezing would eliminate the central mechanism of their original formulations\. Specifically, we acknowledge but exclude RAFT\([Lukasik et al\., 2025](https://arxiv.org/html/2610.19206#bib.bib14)\), TRACT\([Chiang et al\., 2025](https://arxiv.org/html/2610.19206#bib.bib28)\), REAL\([Zhang et al\., 2026a](https://arxiv.org/html/2610.19206#bib.bib10)\), SmoothMMD\([Zuo et al\., 2026](https://arxiv.org/html/2610.19206#bib.bib9)\), and Dist2Loss\([Chung et al\., 2026](https://arxiv.org/html/2610.19206#bib.bib8)\)\.

DualViewuses lightweight regression modules trained atop frozen LLMs\. This approach is motivated by prior work demonstrating the effectiveness of LLM hidden representations for downstream regression\([Tang et al\., 2025](https://arxiv.org/html/2610.19206#bib.bib29)\), recovering numerical predictive statistics directly from internal states\([Piskorz et al\., 2026](https://arxiv.org/html/2610.19206#bib.bib1)\), and showing that numerical information encoded internally may not be reliably expressed through generated answers\([Yuchi et al\., 2026](https://arxiv.org/html/2610.19206#bib.bib17)\)\. Building on these findings,DualViewextends existing scalar text regression work using simple predictors based on extracted broad and target\-specific representations\.

### B\.2Fine\-Grained Text Evaluation

As a core application ofDualView, works like SummEval\([Fabbri et al\., 2021](https://arxiv.org/html/2610.19206#bib.bib24)\)and G\-Eval\([Liu et al\., 2023b](https://arxiv.org/html/2610.19206#bib.bib18)\)have long motivated assessing machine\-generated text across multiple quality dimensions\. One supervised baseline we compare to is SAJA\([Kola et al\., 2026](https://arxiv.org/html/2610.19206#bib.bib19)\), which learns to align judge\-derived rubric features with human annotations using traditional machine learning models such as random forest\. In contrast,DualViewdirectly uses prompt\-conditioned hidden representations from a frozen LLM rather than generated rubric scores, extracting both broad and target\-specific evidence for downstream regression\.

Most supervised methods for fine\-grained text evaluation focus on automated essay scoring \(AES\), where each essay receives scores for traits such as grammar, vocabulary, coherence, and fluency\. Representative approaches include autoregressive trait\-score generation\([Do et al\., 2024a](https://arxiv.org/html/2610.19206#bib.bib25)\), T\-MES’s trait\-aware experts\([Wang and Liu, 2025](https://arxiv.org/html/2610.19206#bib.bib16)\), and GAT\-AES’s graph modeling\([Li and Ng, 2025](https://arxiv.org/html/2610.19206#bib.bib5)\)\. We adapt T\-MES and GAT\-AES to the frozen\-backbone setting and evaluate them across heterogeneous domains\. The key difference forDualViewis that we use simple predictors \(ridge\) that can work well across heterogeneous domains\. Notably,DualViewshifts the burden of specialization to extracting complementary input representations \(features\) from frozen LLMs\.

### B\.3Sharing Across Tasks and Outputs

Multi\-task learning \(MTL\) typically entails learning one model to perform heterogeneous tasks\. Examples include representation sharing across language\-understanding tasks\([Hu et al\., 2025](https://arxiv.org/html/2610.19206#bib.bib20);[Hu et al\., 2026](https://arxiv.org/html/2610.19206#bib.bib21)\)and joint argument\-span detection, type prediction, and quality scoring\([Ding et al\., 2023](https://arxiv.org/html/2610.19206#bib.bib44)\)\.

Multi\-output text regression can be viewed as a structured instance of MTL in which every input has the same fixed set of continuous targets\. This implies that information sharing between “tasks" may be more intuitive, since output formats are homogeneous \(e\.g\., the same numerical loss applies to all output dimensions\)\. Nevertheless, it remains unclear how to share supervision across outputs while still allowing each prediction to retain information relevant to its specific target\.

We compare with two regression\-capable MTL approaches\. MT\-ICL learns a transformer\-based Bayesian predictor from sequences of tasks\([Zhu et al\., 2026a](https://arxiv.org/html/2610.19206#bib.bib23)\), whereas UniSO\-N learns numerical surrogates across heterogeneous optimization tasks using language\-model representations\([Tan et al\., 2025](https://arxiv.org/html/2610.19206#bib.bib22)\)\.DualViewdiffers by exploiting MOTR’s shared\-input structure through two complementary forms of sharing\. Namely, a reduced\-rank broad predictor learns common predictive patterns across outputs, while a shared predictor predicts individual targets from focused representations\.

## Appendix CAdditional Method Details

### C\.1Feature Extraction

Per§\\lx@sectionsign[3\.2](https://arxiv.org/html/2610.19206#S3.SS2), for each inputxix\_\{i\}, we construct a broad promptPib=\[s⁡\(𝒬\);xi\]P\_\{i\}^\{b\}=\[s\(\\mathcal\{Q\}\);x\_\{i\}\]andDDfocused promptsPi​jf=\[s⁡\(𝒬\);xi;q⁡\(dj\)\]P\_\{ij\}^\{f\}=\[s\(\\mathcal\{Q\}\);x\_\{i\};q\(d\_\{j\}\)\]\. We then extract representations:

𝐉i=1\|Pib\|​∑t=1\|Pib\|𝐡t​\(Pib\),𝐂i​j=𝐡\|Pi​jf\|​\(Pi​jf\)\.\\mathbf\{J\}\_\{i\}=\\frac\{1\}\{\|P\_\{i\}^\{b\}\|\}\\sum\_\{t=1\}^\{\|P\_\{i\}^\{b\}\|\}\\mathbf\{h\}\_\{t\}\(P\_\{i\}^\{b\}\),\\qquad\\mathbf\{C\}\_\{ij\}=\\mathbf\{h\}\_\{\|P\_\{ij\}^\{f\}\|\}\(P\_\{ij\}^\{f\}\)\.Both representations have shapeℝH\\mathbb\{R\}^\{H\}and use hidden states from the penultimate LLM layer \(directly preceding the language modeling head\)\. The broad view effectively mean\-pools all prompt tokens, while the focused view retrieves the last\-token hidden state after appending the target\-specific instruction\. Notably, during extraction, we do not autoregressively generate numerical answers\. Our design of using pooled LLM representations for downstream regression follows prior work\([Tang et al\., 2025](https://arxiv.org/html/2610.19206#bib.bib29);[Su and Lease, 2026](https://arxiv.org/html/2610.19206#bib.bib12)\)\.

##### Shared\-prefix reuse\.

By design, the focused prompt is just the broad prompt plus instructions to identify target\-specific evidence\. This means that we can cache the broad prompt’s model activations \(e\.g\., attention key–value pairs, recurrent states\) to avoid redundant computation\. Each focused representation then requires processing only the target\-specific continuations\.

### C\.2Evidence and Label Standardization

We standardize representations using training\-split statistics\. For the broad evidence, define

𝐦J\\displaystyle\\mathbf\{m\}\_\{J\}=1N​∑i𝐉i,\\displaystyle=\\frac\{1\}\{N\}\\sum\_\{i\}\\mathbf\{J\}\_\{i\},ℓJ\\displaystyle\\ell\_\{J\}=max⁡\{1N​∑i‖𝐉i−𝐦J‖22,ϵJ\},\\displaystyle=\\max\\\!\\left\\\{\\sqrt\{\\frac\{1\}\{N\}\\sum\_\{i\}\\\|\\mathbf\{J\}\_\{i\}\-\\mathbf\{m\}\_\{J\}\\\|\_\{2\}^\{2\}\},\\epsilon\_\{J\}\\right\\\},𝐉~i\\displaystyle\\widetilde\{\\mathbf\{J\}\}\_\{i\}=\(𝐉i−𝐦J\)/ℓJ\.\\displaystyle=\(\\mathbf\{J\}\_\{i\}\-\\mathbf\{m\}\_\{J\}\)/\\ell\_\{J\}\.For focused evidence, one mean vector and one scalar scale are shared across targets:

𝐦C\\displaystyle\\mathbf\{m\}\_\{C\}=1N​D​∑i,j𝐂i​j,\\displaystyle=\\frac\{1\}\{ND\}\\sum\_\{i,j\}\\mathbf\{C\}\_\{ij\},ℓC\\displaystyle\\ell\_\{C\}=max⁡\{1N​D​H​∑i,j‖𝐂i​j−𝐦C‖22,ϵC\},\\displaystyle=\\max\\\!\\left\\\{\\sqrt\{\\frac\{1\}\{NDH\}\\sum\_\{i,j\}\\\|\\mathbf\{C\}\_\{ij\}\-\\mathbf\{m\}\_\{C\}\\\|\_\{2\}^\{2\}\},\\epsilon\_\{C\}\\right\\\},𝐂~i​j\\displaystyle\\widetilde\{\\mathbf\{C\}\}\_\{ij\}=\(𝐂i​j−𝐦C\)/ℓC\.\\displaystyle=\(\\mathbf\{C\}\_\{ij\}\-\\mathbf\{m\}\_\{C\}\)/\\ell\_\{C\}\.The positive constantsϵJ\\epsilon\_\{J\}andϵC\\epsilon\_\{C\}prevent division by zero\. For each targetjj, we also compute the training\-label meanμj\\mu\_\{j\}and population standard deviationσj\\sigma\_\{j\}, then setzi​j=\(yi​j−μj\)/σjz\_\{ij\}=\(y\_\{ij\}\-\\mu\_\{j\}\)/\\sigma\_\{j\}\. All statistics are fixed at evaluation time\. In the main\-text equations, we omit the tildes on the preprocessed representations for readability\.

### C\.3Rank\-Constrained Broad Predictor

Let𝐗∈ℝN×H\\mathbf\{X\}\\in\\mathbb\{R\}^\{N\\times H\}contain the preprocessed broad representations𝐉~i⊤\\widetilde\{\\mathbf\{J\}\}\_\{i\}^\{\\top\}as rows\. In each broad fitting step, let𝐑∈ℝN×D\\mathbf\{R\}\\in\\mathbb\{R\}^\{N\\times D\}denote the residual targets after subtracting the current focused predictions\. We solve

minrank⁡\(𝐖b\)≤k𝐚∈ℝD⁡‖𝐑−𝐗𝐖b−𝟏​𝐚⊤‖F2\+γb​‖𝐖b‖F2,\\min\_\{\\begin\{subarray\}\{c\}\\operatorname\{rank\}\(\\mathbf\{W\}\_\{\\mathrm\{b\}\}\)\\leq k\\\\ \\mathbf\{a\}\\in\\mathbb\{R\}^\{D\}\\end\{subarray\}\}\\\|\\mathbf\{R\}\-\\mathbf\{X\}\\mathbf\{W\}\_\{\\mathrm\{b\}\}\-\\mathbf\{1\}\\mathbf\{a\}^\{\\top\}\\\|\_\{F\}^\{2\}\+\\gamma\_\{\\mathrm\{b\}\}\\\|\\mathbf\{W\}\_\{\\mathrm\{b\}\}\\\|\_\{F\}^\{2\},whereγb=N​D​λb\\gamma\_\{\\mathrm\{b\}\}=ND\\lambda\_\{\\mathrm\{b\}\}and𝐚\\mathbf\{a\}is an unregularized, target\-specific intercept vector\. Since the columns of𝐗\\mathbf\{X\}are centered, the optimal intercept equals the column means of𝐑\\mathbf\{R\}:

𝐚=1N​𝐑⊤​𝟏,𝐑c=𝐑−𝟏​𝐚⊤\.\\mathbf\{a\}=\\frac\{1\}\{N\}\\mathbf\{R\}^\{\\top\}\\mathbf\{1\},\\qquad\\mathbf\{R\}\_\{c\}=\\mathbf\{R\}\-\\mathbf\{1\}\\mathbf\{a\}^\{\\top\}\.We then solve reduced\-rank ridge on𝐑c\\mathbf\{R\}\_\{c\}:

𝐂γ\\displaystyle\\mathbf\{C\}\_\{\\gamma\}=\(𝐗⊤​𝐗\+γb​𝐈\)−1​𝐗⊤​𝐑c,\\displaystyle=\(\\mathbf\{X\}^\{\\top\}\\mathbf\{X\}\+\\gamma\_\{\\mathrm\{b\}\}\\mathbf\{I\}\)^\{\-1\}\\mathbf\{X\}^\{\\top\}\\mathbf\{R\}\_\{c\},𝐊\\displaystyle\\mathbf\{K\}=𝐑c⊤​𝐗𝐂γ\.\\displaystyle=\\mathbf\{R\}\_\{c\}^\{\\top\}\\mathbf\{X\}\\mathbf\{C\}\_\{\\gamma\}\.If𝐐k∈ℝD×k\\mathbf\{Q\}\_\{k\}\\in\\mathbb\{R\}^\{D\\times k\}contains the leading orthonormal eigenvectors of𝐊\\mathbf\{K\}, a reduced\-rank ridge solution is

𝐀=𝐂γ​𝐐k,𝐁=𝐐k⊤,𝐖b=𝐀𝐁\.\\mathbf\{A\}=\\mathbf\{C\}\_\{\\gamma\}\\mathbf\{Q\}\_\{k\},\\qquad\\mathbf\{B\}=\\mathbf\{Q\}\_\{k\}^\{\\top\},\\qquad\\mathbf\{W\}\_\{\\mathrm\{b\}\}=\\mathbf\{A\}\\mathbf\{B\}\.Here𝐀∈ℝH×k\\mathbf\{A\}\\in\\mathbb\{R\}^\{H\\times k\}learnskkcommon predictive directions and𝐁∈ℝk×D\\mathbf\{B\}\\in\\mathbb\{R\}^\{k\\times D\}assigns weights to them for each target\. The broad contribution to the standardized prediction is

𝐳^ib=𝐁⊤​𝐀⊤​𝐉~i∈ℝD\.\\widehat\{\\mathbf\{z\}\}^\{\\,\\mathrm\{b\}\}\_\{i\}=\\mathbf\{B\}^\{\\top\}\\mathbf\{A\}^\{\\top\}\\widetilde\{\\mathbf\{J\}\}\_\{i\}\\in\\mathbb\{R\}^\{D\}\.
##### Rank and storage\.

The reduced\-rank factorization storesk⁡\(H\+D\)k\(H\+D\)coefficients, versusH​DHDfor a full\-rank broad predictor\. This reduces model complexity whenk⁡\(H\+D\)<H​Dk\(H\+D\)<HD, as we empirically confirm in our evaluated configurations\.

### C\.4Shared Focused Predictor

The focused branch uses one linear coefficient vector𝐯∈ℝH\\mathbf\{v\}\\in\\mathbb\{R\}^\{H\}for all targets:

z^i​jf=𝐯⊤​𝐂~i​j\.\\widehat\{z\}\_\{ij\}^\{\\,\\mathrm\{f\}\}=\\mathbf\{v\}^\{\\top\}\\widetilde\{\\mathbf\{C\}\}\_\{ij\}\.Because the focused representations are conditioned on target instructions, the same predictor can assign different scores to different targets\. The shared coefficients are estimated from allN​DNDexample–target pairs\.

With the broad predictor and intercepts fixed, the residuals are:

ei​j=zi​j−𝐉~i⊤​𝐀𝐛j−aj\.e\_\{ij\}=z\_\{ij\}\-\\widetilde\{\\mathbf\{J\}\}\_\{i\}^\{\\top\}\\mathbf\{A\}\\mathbf\{b\}\_\{j\}\-a\_\{j\}\.The focused fitting step solves

min⁡∑i,j𝐯⁡\(ei​j−𝐯⊤​𝐂~i​j\)2\+γf​‖𝐯‖22,\\min\_\{\\mathbf\{v\}\}\\ \\sum\_\{i,j\}\(e\_\{ij\}\-\\mathbf\{v\}^\{\\top\}\\widetilde\{\\mathbf\{C\}\}\_\{ij\}\)^\{2\}\+\\gamma\_\{\\mathrm\{f\}\}\\\|\\mathbf\{v\}\\\|\_\{2\}^\{2\},whereγf=N​D​λf\\gamma\_\{\\mathrm\{f\}\}=ND\\lambda\_\{\\mathrm\{f\}\}\. Its ridge solution is

𝐯=\(∑i,j𝐂~i​j​𝐂~i​j⊤\+γf​𝐈\)−1​∑i,j𝐂~i​j​ei​j\.\\mathbf\{v\}=\\left\(\\sum\_\{i,j\}\\widetilde\{\\mathbf\{C\}\}\_\{ij\}\\widetilde\{\\mathbf\{C\}\}\_\{ij\}^\{\\top\}\+\\gamma\_\{\\mathrm\{f\}\}\\mathbf\{I\}\\right\)^\{\-1\}\\sum\_\{i,j\}\\widetilde\{\\mathbf\{C\}\}\_\{ij\}e\_\{ij\}\.Thus, the focused branch hasHHfitted coefficients, rather thanH​DHDfor separate target\-specific linear predictors\.

### C\.5Training and Inference

##### Joint objective\.

The standardized prediction combines the two branches additively:

z^i​j=𝐉~i⊤​𝐀𝐛j\+𝐯⊤​𝐂~i​j\+aj\.\\widehat\{z\}\_\{ij\}=\\widetilde\{\\mathbf\{J\}\}\_\{i\}^\{\\top\}\\mathbf\{A\}\\mathbf\{b\}\_\{j\}\+\\mathbf\{v\}^\{\\top\}\\widetilde\{\\mathbf\{C\}\}\_\{ij\}\+a\_\{j\}\.As specified in§\\lx@sectionsign[3\.5](https://arxiv.org/html/2610.19206#S3.SS5), we minimize

min𝐖b,𝐯,𝐚rank⁡\(𝐖b\)≤k\\displaystyle\\min\_\{\\begin\{subarray\}\{c\}\\mathbf\{W\}\_\{\\mathrm\{b\}\},\\,\\mathbf\{v\},\\,\\mathbf\{a\}\\\\ \\operatorname\{rank\}\(\\mathbf\{W\}\_\{\\mathrm\{b\}\}\)\\leq k\\end\{subarray\}\}1N​D​∑i,j\(zi​j−z^i​j\)2\\displaystyle\\frac\{1\}\{ND\}\\sum\_\{i,j\}\(z\_\{ij\}\-\\widehat\{z\}\_\{ij\}\)^\{2\}\+λb​‖𝐖b‖F2\+λf​‖𝐯‖22,\\displaystyle\+\\lambda\_\{\\mathrm\{b\}\}\\\|\\mathbf\{W\}\_\{\\mathrm\{b\}\}\\\|\_\{F\}^\{2\}\+\\lambda\_\{\\mathrm\{f\}\}\\\|\\mathbf\{v\}\\\|\_\{2\}^\{2\},with𝐖b=𝐀𝐁\\mathbf\{W\}\_\{\\mathrm\{b\}\}=\\mathbf\{A\}\\mathbf\{B\}\.

##### Alternating residual fitting\.

We initialize𝐯=𝟎\\mathbf\{v\}=\\mathbf\{0\}and fit the reduced\-rank broad predictor to the standardized targets\. We then alternate two ridge updates: \(i\) fit𝐯\\mathbf\{v\}to the residuals of the current broad predictor and intercepts, using the solution in§\\lx@sectionsign[C\.4](https://arxiv.org/html/2610.19206#A3.SS4); and \(ii\) subtract the current focused predictions from the standardized targets,

Ri​j=zi​j−𝐯⊤​𝐂~i​j,R\_\{ij\}=z\_\{ij\}\-\\mathbf\{v\}^\{\\top\}\\widetilde\{\\mathbf\{C\}\}\_\{ij\},then refit𝐀\\mathbf\{A\},𝐁\\mathbf\{B\}, and𝐚\\mathbf\{a\}using the rank\-constrained solution in§\\lx@sectionsign[C\.3](https://arxiv.org/html/2610.19206#A3.SS3)\. Both steps minimize the same combined objective with the other branch fixed\. We repeat until convergence and selectkk,λb\\lambda\_\{\\mathrm\{b\}\}, andλf\\lambda\_\{\\mathrm\{f\}\}on the validation split\.

##### Inference\.

During inference, we restore the predicted standardized scores back to their original target units:

y^i​j=μj\+σj​z^i​j\.\\widehat\{y\}\_\{ij\}=\\mu\_\{j\}\+\\sigma\_\{j\}\\widehat\{z\}\_\{ij\}\.

## Appendix DAdditional Experimental Setup

### D\.1Models

We evaluateDualViewand baselines on eight frozen LLM backbones: Gemma 4 E4B Instruct111[https://huggingface\.co/google/gemma\-4\-E4B\-it](https://huggingface.co/google/gemma-4-E4B-it)and Gemma 4 31B Instruct222[https://huggingface\.co/google/gemma\-4\-31B\-it](https://huggingface.co/google/gemma-4-31B-it)\([Gemma Team, 2026](https://arxiv.org/html/2610.19206#bib.bib60)\), Qwen3\.5 9B333[https://huggingface\.co/Qwen/Qwen3\.5\-9B](https://huggingface.co/Qwen/Qwen3.5-9B)and Qwen3\.5 122B\-A10B444[https://huggingface\.co/Qwen/Qwen3\.5\-122B\-A10B](https://huggingface.co/Qwen/Qwen3.5-122B-A10B)\([Qwen Team, 2026](https://arxiv.org/html/2610.19206#bib.bib61)\), Qwen3\.6 27B555[https://huggingface\.co/Qwen/Qwen3\.6\-27B](https://huggingface.co/Qwen/Qwen3.6-27B), Llama 3\.3 70B Instruct666[https://huggingface\.co/meta\-llama/Llama\-3\.3\-70B\-Instruct](https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct)\([Grattafiori et al\., 2024](https://arxiv.org/html/2610.19206#bib.bib65)\), Qwen2\.5 72B Instruct777[https://huggingface\.co/Qwen/Qwen2\.5\-72B\-Instruct](https://huggingface.co/Qwen/Qwen2.5-72B-Instruct)\([Qwen Team, 2025](https://arxiv.org/html/2610.19206#bib.bib62)\), and GPT OSS 120B888[https://huggingface\.co/openai/gpt\-oss\-120b](https://huggingface.co/openai/gpt-oss-120b)\([OpenAI, 2025](https://arxiv.org/html/2610.19206#bib.bib63)\)\.

### D\.2Datasets

Table[8](https://arxiv.org/html/2610.19206#A4.T8)reports the partitions for each dataset\.Tables[22](https://arxiv.org/html/2610.19206#A8.T22)–[26](https://arxiv.org/html/2610.19206#A8.T26)in§\\lx@sectionsign[H](https://arxiv.org/html/2610.19206#A8)give the broad and focused prompt templates for these datasets\.

##### Partitioning

Several benchmarks contain multiple model outputs associated with the same underlying source\. Randomly partitioning individual examples could therefore place related instances in both training and evaluation sets\. To avoid this form of overlap, we construct the training, validation, and test partitions at the source\-group level\. Specifically, we partition examples by programming problem \(DSCodeBench\), source document \(MSumBench\), source segment \(RATE\), conversation \(PRISM\), or paper \(CreativeEval\), keeping all examples from each group in the same split\.

We repeat experiments using three random seeds \(42, 1234, and 2026\)\. For each seed, we independently construct train, validation, and test partitions at the source\-group level, following the grouping strategy described above\. All supervised methods use identical partitions within each seed to ensure matched comparisons\. Counts are identical across seeds\.

SinceDualViewuses closed\-form regression updates with a fixed initialization, its fitting procedure is deterministic given the extracted representations and data partitions\. Thus, variation across seeds reflects differences in data partitioning rather than stochastic optimization of the proposed predictor forDualView\. For the other supervised baselines, however, differences might reflect both partitioning and stochastic optimization from gradient descent\.

Table 8:Dataset split sizes and outputs\.Multiple candidate outputs associated with a source group are assigned to the same partition to avoid leakage\. For DSCodeBench, Memory denotes peak resident memory, and Integral denotes its time integral\.
##### MSumBench\([Min et al\., 2025](https://arxiv.org/html/2610.19206#bib.bib56)\)\.

MSumBench999[https://github\.com/DISL\-Lab/MSumBench](https://github.com/DISL-Lab/MSumBench)contains machine\-generated summaries in English and Chinese across six domains: CNN/DailyMail news, GovReport reports, PubMed medical literature, MultiWOZ booking dialogue, MeetingBank meetings, and MediaSum interviews\.

The dataset provides three human\-annotated targets:completeness, the fraction of annotated source key facts covered by the summary;conciseness, the fraction of summary sentences containing at least one source key fact; andfaithfulness, the fraction of summary sentences without factual errors\. All fractions are in 0–1\.

##### RATE\([Popov et al\., 2025](https://arxiv.org/html/2610.19206#bib.bib53)\)\.

The WMT24 English\-\-Russian task101010[https://huggingface\.co/datasets/yandex/wmt24\-en\-ru\-rate](https://huggingface.co/datasets/yandex/wmt24-en-ru-rate)predicts expert assessments of a candidate translation along human\-annotated dimensions of accuracy, fluency, and style, each on a 0–100 scale\. Accuracy concerns preserving source meaning, fluency concerns the quality of the Russian translation, and style concerns preserving the source’s manner of expression\.

##### CreativeEval\([Lu et al\., 2026](https://arxiv.org/html/2610.19206#bib.bib57)\)\.

CreativeEval111111[https://github\.com/lichun\-19/creative\_eval](https://github.com/lichun-19/creative_eval)contains human\-written introductions from ICLR 2024 submissions\. The authors form a set of 513 accepted and 513 rejected papers and match them across 16 soundness/presentation bins\. The target dimensions are real average human reviewer scores spanning contribution, soundness, and presentation scores on a 1–4 scale, together with an overall score on a 1–10 scale\. Notably, averaging reviewer scores yields continuous targets even when individual judgments use a discrete Likert scale\.

##### DSCodeBench\([Ouyang et al\., 2026](https://arxiv.org/html/2610.19206#bib.bib59)\)\.

DSCodeBench121212[https://github\.com/ShuyinOuyang/DSCodeBench](https://github.com/ShuyinOuyang/DSCodeBench)asks a model to implement a Python solution to a realistic data\-science programming problem derived from GitHub\. We collect execution\-based targets for this study using a sandbox during dataset construction\. Our motivation is to evaluate whether we can predict machine\-generated code behavior without execution at inference\.

Targets include the fraction of unit tests passed \(from a fixed 200\-test suite\), wall\-clock time \(seconds\), peak sampled process\-tree resident memory \(bytes\), and the time integral of resident memory \(byte\-seconds\)\. We sample the summed resident set size \(RSS\) at a nominal interval of 5 ms, taking the maximum sampled RSS as peak memory and using trapezoidal integration to compute the memory integral\.

Correctness lies in\[0,1\]\[0,1\], while the resource measurements are nonnegative\. We impose a 300\-second execution timeout\. We drop programs with execution exceptions and programs that execute beyond the 300\-second timeout to avoid artificially inflating or deflating the measured wall\-clock\. We apply all preprocessing rules before partitioning\.

##### PRISM\([Kirk et al\., 2024](https://arxiv.org/html/2610.19206#bib.bib58)\)\.

PRISM131313[https://github\.com/HannahKirk/prism\-alignment](https://github.com/HannahKirk/prism-alignment)evaluates an AI assistant’s first\-turn response to unguided, values\-guided, and controversy\-guided conversations\. Responses are evaluated along seven dimensions on a 1–100 scale:

1. 1\.Values: the extent to which the response reflected the participant’s values or cultural perspective
2. 2\.Fluency: the extent to which the response was well written and coherent
3. 3\.Factuality: the extent to which the response was factual and informative
4. 4\.Safety: the extent to which the response was safe and did not risk harm to the participant or others
5. 5\.Diversity: the extent to which the response summarized multiple viewpoints
6. 6\.Creativity: the extent to which the response was creative and inspiring
7. 7\.Helpfulness: the extent to which the response was helpful and relevant\.

For regression, each retained conversation contributes its selected top\-rated first\-turn response and the corresponding seven performance\-attribute ratings\. The dataset only releases the seven\-dimensional annotations for the response to the opening message, and not for the full conversation or every candidate response\. We retain only examples with an available selected response and valid ratings for all seven attributes, excluding examples with missing ratings\.

We partition examples by conversation, ensuring that each conversation appears in only one split\. Participant\-level disjointness is not enforced, so the evaluation measures generalization to unseen conversations, potentially involving the same participants represented in the training data\.

### D\.3Baseline Adaptations

Because the supervised baselines differ in their original training procedures \(supervised fine\-tuning of all LLM parameters\) and prediction settings \(mostly scalar\), we adapt them to operate with frozen LLM backbones and multi\-output text regression targets\.

##### RELISH\([Su and Lease, 2026](https://arxiv.org/html/2610.19206#bib.bib12)\)\.

RELISH iteratively refines a learned latent state through cross\-attention over frozen LLM token representations and maps the final state to ascalarnumerical prediction\. To support joint prediction of multiple target dimensions, we replace the scalar linear readout with aDD\-dimensional linear readout\. All other components and loss functions are not changed\.

##### GenRe2\([Chen et al\., 2026b](https://arxiv.org/html/2610.19206#bib.bib13)\)\.

GenRe2 trains autoregressive numerical decoders forscalarregression using supervised initialization followed by sequence\-level reinforcement learning\. While most experiments fine\-tune the LLM backbone, we note that its code\-regression experiment already freezes the LLM encoder, making GenRe2 amenable to our frozen LLM setting\. To support multi\-output settings, we train a compact decoder on pooled broad\-prompt features and condition it on output identity\. One shared decoder predicts all outputs, with a rank\-16 residual adapter for each output after each decoder block to permit specialization\. We use normalized base\-2 numerical tokens and median aggregation of sampled predictions\.

##### T\-MES\([Wang and Liu, 2025](https://arxiv.org/html/2610.19206#bib.bib16)\)\.

T\-MES is designed for multi\-trait automated essay scoring with a shared pretrained text encoder, token\-level mixture\-of\-experts routing, and trait\-specific scalar readouts\. Its objective combines regression with trait\-consistency, representation\-correlation, and expert\-diversity terms\. We only substitute frozen broad\-prompt LLM token states for its essay encoder representations since its original multi\-output structure carries over directly\. We also adapt the output scalar readouts to our datasets’ target ranges using scaled sigmoid predictions for bounded outputs and softplus predictions for nonnegative unbounded outputs\.

##### GAT\-AES\([Li and Ng, 2025](https://arxiv.org/html/2610.19206#bib.bib5)\)\.

GAT\-AES is also designed for multi\-trait essay scoring by passing messages among essay\-embedding nodes, handcrafted\-feature nodes, and learned trait nodes\. Each training example supplies one embedding node obtained by projecting the last\-token state of the broad prompt, with edges from this node to the output nodes and between output nodes\. We omit the essay\-specific handcrafted\-feature nodes\. We freeze the LLM while updating the other components, including the projection module, graph layers, and scalar readouts for each output dimension\. As in T\-MES, output links accommodate our bounded and nonnegative unbounded targets\.

##### SAJA\([Kola et al\., 2026](https://arxiv.org/html/2610.19206#bib.bib19)\)\.

SAJA aligns LLM judgments with human labels by prompting for a structured vector of rubric scores and fitting a lightweight calibration model\. Thus, SAJA already operates without updating the LLM or accessing its hidden states\. We preserve this generated\-score interface and use its random\-forest regression approach\. Specifically, each forest shares the same extracted inputs but has separate parameters\.

##### UniSO\-N\([Tan et al\., 2025](https://arxiv.org/html/2610.19206#bib.bib22)\)\.

UniSO\-N functions as a numerical surrogate for offline black\-box optimization across heterogeneous generative tasks\. Its numerical\-regression variant maps language\-model embeddings to a scalar using a shared MLP, where the embedder is fine\-tuned with contrastive and Lipschitz objectives\. After freezing the LLM, the embedder is no longer updated\. Like RELISH, we replace the final projection layer of the scalar MLP with aDD\-dimensional linear projection to avoid fitting separate models for each output target\.

##### MT\-ICL\([Zhu et al\., 2026a](https://arxiv.org/html/2610.19206#bib.bib23)\)\.

Multi\-Task Bayesian In\-Context Learning trains a numerical Transformer to predict from labeled examples of a target task and auxiliary tasks\. The original work studies synthetic inference problems and spatiotemporal regression\. To apply MT\-ICL to text, we replace numerical covariates with frozen output\-specific LLM representations and treat each output dimension as a task\. One shared Transformer serves every output and returns its predictive mean in the original target units\.

### D\.4Metrics and Aggregation

We report normalized root mean squared error \(NRMSE\), Pearsonrr, and Spearmanρ\\rho\. Because the criteria use different units and score scales, raw RMSEs are not directly comparable\. We therefore normalize RMSE by the training standard deviation, following prior work\([Allison et al\., 2023](https://arxiv.org/html/2610.19206#bib.bib49);[Holzmüller et al\., 2024](https://arxiv.org/html/2610.19206#bib.bib50);[Lin et al\., 2023](https://arxiv.org/html/2610.19206#bib.bib51)\):

NRMSEj=1n​∑i\(y^i​j−yi​j\)2σjtrain\.\\mathrm\{NRMSE\}\_\{j\}=\\frac\{\\sqrt\{\\frac\{1\}\{n\}\\sum\_\{i\}\(\\hat\{y\}\_\{ij\}\-y\_\{ij\}\)^\{2\}\}\}\{\\sigma\_\{j\}^\{\\mathrm\{train\}\}\}\.
NRMSE is unitless, where lower error is better\. For aggregation, we first compute metrics within each dataset–LLM–seed setting, averaging the relevant output dimensions and settings within that seed, then report the mean and sample standard deviations across three seeds\. Outputs receive equal weight within a dataset, and aggregate results weight datasets and LLMs equally\. That is, datasets with more dimensions or sample sizes do not receive more weight\.

### D\.5Hyperparameters and Model Selection

##### DualView

We fit the reduced\-rank broad predictor and shared linear focused predictor through alternating residual minimization, as described in§\\lx@sectionsign[C\.5](https://arxiv.org/html/2610.19206#A3.SS5)\. Both ridge subproblems are solved analytically, with the broad predictor’s rank and both regularization penalties selected on the validation split\.We use a validation utility that averages NRMSE and one minus Pearson correlation\.Table[10](https://arxiv.org/html/2610.19206#A4.T10)summarizes the hyperparameter settings\.

Table 9:Baseline Hyperparameters\.Braced values denote searched alternatives\. Whenever possible, we try to follow standard selection criteria except for methods that explicitly minimize some text\-based generative loss\.
##### Baseline training and selection\.

Table[9](https://arxiv.org/html/2610.19206#A4.T9)summarizes the recorded configurations of the seven supervised baselines\. The methods use different objectives and model\-selection criteria, reflecting their respective prediction mechanisms\. MT\-ICL uses 32 training support examples per task, with up to 64 query examples per episode\.

Table 10:DualViewHyperparameters\.The broad rankkkconstrains the output coefficient matrix to rank at mostk<Dk<D\. The ridge penaltiesγb\\gamma\_\{\\mathrm\{b\}\}andγf\\gamma\_\{\\mathrm\{f\}\}correspond to the unaveraged squared\-error subproblems in§\\lx@sectionsign[C\.3](https://arxiv.org/html/2610.19206#A3.SS3)and§\\lx@sectionsign[C\.4](https://arxiv.org/html/2610.19206#A3.SS4), respectively\.

### D\.6Computational Budget

We conduct experiments on NVIDIA Grace Hopper GH200 nodes, with a 72\-core Arm\-based Grace CPU and approximately 237GB of memory\. All supervised experiments are completed within one thousand GPU hours\.

## Appendix EAdditional Key Results

The following analyses supplement the main results in the same order as the main text: predictive performance and output\-level analyses, trainable footprint, training\-data efficiency, and comparisons with Gemini 3\.1 Pro\.

### E\.1Predictive Performance

Tables[11](https://arxiv.org/html/2610.19206#A5.T11),[12](https://arxiv.org/html/2610.19206#A5.T12), and[13](https://arxiv.org/html/2610.19206#A5.T13)expandTable[1](https://arxiv.org/html/2610.19206#S5.T1)with Pearson correlation, Spearman correlation, and NRMSE, respectively, for all seven supervised baselines across every dataset–LLM combination\. Each dataset and LLM receives equal weight in the overall average\.

Figure 5:Output\-level Spearman correlation\.Each spoke reports Spearman correlation for one dataset output, macro\-averaged over 8 LLM backbones and 3 training seeds\. We compareDualViewagainst the three strongest supervised baselines\. Larger radial values are better, where scales vary across datasets for visibility\.Figure 6:Output\-level normalized RMSE \(NRMSE\)\.Each spoke reports NRMSE for one dataset output, macro\-averaged over 8 LLM backbones and 3 training seeds\. We compareDualViewagainst the three strongest supervised baselines\. Smaller radial values are better, where scales vary across datasets for visibility\.Table 11:Pearson correlation across five datasets\(MSumBench, RATE, CreativeEval, DSCodeBench, and PRISM\) and eight LLMs \(Gemma 4 E4B, Qwen3\.5 9B, Qwen3\.6 27B, Gemma 4 31B, Llama 3\.3 70B Instruct, Qwen2\.5 72B Instruct, GPT OSS 120B, and Qwen3\.5 122B A10B\)\. Values are means with standard deviations across three seeds \(±\\pm\)\. Bold marks the best value in each row\.DualViewwins are marked in light green whereas baseline wins are highlighted in yellow\. Average LLM gives the within\-dataset average, and Overall gives the average of all settings\.Table 12:Spearman correlation across five datasets\(MSumBench, RATE, CreativeEval, DSCodeBench, and PRISM\) and eight LLMs \(Gemma 4 E4B, Qwen3\.5 9B, Qwen3\.6 27B, Gemma 4 31B, Llama 3\.3 70B Instruct, Qwen2\.5 72B Instruct, GPT OSS 120B, and Qwen3\.5 122B A10B\)\. Values are means with standard deviations across three seeds \(±\\pm\)\. Bold marks the best value in each row\.DualViewwins are marked in light green whereas baseline wins are highlighted in yellow\. Average LLM gives the within\-dataset average, and Overall gives the average of all settings\.Table 13:Normalized RMSE \(NRMSE\) across five datasets\(MSumBench, RATE, CreativeEval, DSCodeBench, and PRISM\) and eight LLMs \(Gemma 4 E4B, Qwen3\.5 9B, Qwen3\.6 27B, Gemma 4 31B, Llama 3\.3 70B Instruct, Qwen2\.5 72B Instruct, GPT OSS 120B, and Qwen3\.5 122B A10B\)\. Values are means with standard deviations across three seeds \(±\\pm\)\. Bold marks the best value in each row\.DualViewwins are marked in light green whereas baseline wins are highlighted in yellow\. Average LLM gives the within\-dataset average, and Overall gives the average of all settings\.#### E\.1\.1Fine\-grained Performance Analyses

To complement Table[1](https://arxiv.org/html/2610.19206#S5.T1), we examine howDualViewperforms across different output targets and whether gains persist under aggregation schemes that penalize weak targets\.

##### Output\-level analyses\.

Similar toFigure[3](https://arxiv.org/html/2610.19206#S4.F3),Figures[5](https://arxiv.org/html/2610.19206#A5.F5)and[6](https://arxiv.org/html/2610.19206#A5.F6)report Spearman correlation and NRMSE results forDualViewacross five datasets spanning 21 output targets\.DualViewstill wins most targets, leading Spearman correlation on 19 of 21 outputs and NRMSE on 20 of 21 compared to the three strongest supervised baselines\. The two Spearman deficits are DSCodeBench correctness and runtime, where GenRe2 outperformsDualViewby 0\.008 and 0\.018, respectively\. For NRMSE, the sole exception is MSumBench faithfulness, whereDualViewis effectively tied with GAT\-AES \(< 0\.01 difference\)\.

##### Strict wins and worst\-output gaps\.

These measures assess consistency across individual LLM–target comparisons since averaging across LLMs can still mask weak performance\. Astrict winoccurs when a method outperforms every other method on a given output, dataset, LLM, metric, and seed \(excluding ties\)\. We compute the winning percentage within each dataset–LLM setting and macro\-average these percentages across settings\.

To measure loss severity, we define each method’soutput\-level gaprelative to the best\-performing method on the same output\. This gap is the nonnegative difference in Pearson or Spearman correlation, or the excess NRMSE\. For each dataset–LLM setting and seed, theworst\-output gapis the largest such gap across targets\. We then macro\-average the worst\-output gaps across dataset–LLM settings and report mean±\\pmsample standard deviation across three seeds\.

Table[14](https://arxiv.org/html/2610.19206#A5.T14)shows thatDualViewhas higher strict win rates and smaller mean worst\-output gaps than the three strongest supervised baselines\. For Pearson and NRMSE, its win rates are 65\.9% and 71\.4%, respectively, versus 8\.9% and 7\.7% for GAT\-AES\. Its corresponding mean worst\-output gaps are 0\.039 and 0\.016, versus 0\.149 and 0\.055 for GAT\-AES\. These measures indicate thatDualViewwins more frequently and incurs smaller worst\-target deficits when it does not rank first\. Win rates do not add up to 100% because we compared to all seven supervised baselines but only show the three strongest\.

Table 14:Output\-level wins and gaps\.Top: percentage of LLM–target comparisons on which a method strictly outperforms all competitors \(↑\\uparrow\)\. Win rates do not sum to 100% because we only show values for the three strongest supervised baselines, while win rates are computed by comparingDualViewto all seven supervised baselines\. Bottom: largest target\-level gap relative to the best method \(↓\\downarrow\)\. Results are macro\-averaged across 5 datasets and 8 LLMs within each seed, with standard deviations \(±\\pm\) over three seeds\.
##### Power\-mean aggregation\.

For a given dataset, LLM, seed, and evaluation metric, letaja\_\{j\}denote the score on outputjj\. We aggregate across itsDDoutputs using the power mean

Mq​\(𝐚\)=\(1D​∑j=1Dajq\)1/q\.M\_\{q\}\(\\mathbf\{a\}\)=\\left\(\\frac\{1\}\{D\}\\sum\_\{j=1\}^\{D\}a\_\{j\}^\{q\}\\right\)^\{1/q\}\.The main results use the arithmetic mean \(q=1q=1\)\. To emphasize weak outputs, we instead use the harmonic mean \(q=−1q=\-1\) for positive correlation scores and the root mean square \(q=2q=2\) for NRMSE\. When any correlation is zero, its negative\-order power mean is defined as zero\. We observe no negative correlations empirically\. Unlike the arithmetic mean, the harmonic mean penalizes small correlations more strongly\. Conversely, RMS assigns greater weight to large prediction errors\.

Power means are computed within each dataset–LLM–seed setting before macro\-averaging datasets and LLMs within each seed\. We report means and sample standard deviations across three seeds\.Table[15](https://arxiv.org/html/2610.19206#A5.T15)shows thatDualViewoutperforms other supervised baselines even when weaker outputs are penalized more heavily\.

Table 15:Predictive performance under weak\-output penalties\.We report harmonic means \(q=−1q=\-1,↑\\uparrow\) for Pearson and Spearman and root mean square for NRMSE \(q=2q=2,↓\\downarrow\)\. Average Dataset weights the five datasets equally\. Entries report mean±\\pmsample standard deviation across three seeds\.
##### Sensitivity to stronger penalties on weak correlations\.

We further examine power means usingq=−2q=\-2, which place greater emphasis on outputs with low positive correlations\. SinceM−2≤M−1≤M1M\_\{\-2\}\\leq M\_\{\-1\}\\leq M\_\{1\}, this provides a more stringent check than harmonic aggregation\.Table[16](https://arxiv.org/html/2610.19206#A5.T16)shows thatDualViewremains the strongest method even when weak Pearson and Spearman performances are penalized more heavily\.

Table 16:Sensitivity to a stronger weak\-output correlation penalty\.We report power means using orderq=−2q=\-2to penalize small correlations more strongly than the harmonic means \(q=−1q=\-1\) inTable[15](https://arxiv.org/html/2610.19206#A5.T15)\. Values are macro\-averaged across eight LLMs with sample standard deviation across three seeds \(±\\pm\)\.

### E\.2Trainable Footprint

Table[17](https://arxiv.org/html/2610.19206#A5.T17)supplementsTable[2](https://arxiv.org/html/2610.19206#S5.T2)with fitted parameter counts for the supervised baselines across all eight backbones\. Ranges reflect variation across datasets and seeds\. SAJA is omitted because it uses random forests rather than neural regression modules\.

ForDualView, the reduced\-rank broad predictor hask⁡\(H\+D\)k\(H\+D\)coefficients, the shared focused predictor hasHH, and the target\-specific intercepts haveDD, yielding\(k\+1\)​\(H\+D\)\(k\+1\)\(H\+D\)fitted parameters total\. Thus, the trainable footprint depends on the model hidden sizeHH, number of targetsDD, and selected rankkk, but not directly on the size of the underlying LLM backbone\.

Table 17:Absolute fitted head coefficients for all neural comparators\.Baseline counts are in millions \(M\), whereasDualViewcounts are in thousands \(K\)\. Ranges span different configurations across datasets and seeds, if any\. SAJA fits random forests and therefore has no neural coefficient count\.
### E\.3Training\-Data Efficiency

Figures[7](https://arxiv.org/html/2610.19206#A5.F7),[8](https://arxiv.org/html/2610.19206#A5.F8), and[9](https://arxiv.org/html/2610.19206#A5.F9)extendFigure[4](https://arxiv.org/html/2610.19206#S5.F4)to all seven supervised baselines, reporting Pearson correlation, Spearman correlation, and NRMSE as functions of training\-set size\. We vary the training budget but do not change the validation and test splits\. Each dot represents macro\-averages across eight LLMs within each seed, with whiskers showing standard deviation across three seeds\.

Figure 7:Training\-data efficiency: Pearson\.Each curve averages outputs across eight LLMs within each seed, then shows the mean and sample standard deviation across three seeds\. Full denotes the complete training split\.Figure 8:Training\-data efficiency: Spearman\.Each curve averages outputs across eight LLMs within each seed, then shows the mean and sample standard deviation across three seeds\. Full denotes the complete training split\.Figure 9:Training\-data efficiency: NRMSE\.Each curve averages outputs across eight LLMs within each seed, then shows the mean and sample standard deviation across three seeds\. Full denotes the complete training split\.
### E\.4Comparison with Gemini 3\.1 Pro

Table[18](https://arxiv.org/html/2610.19206#A5.T18)extendsTable[3](https://arxiv.org/html/2610.19206#S5.T3)by reporting Spearman correlation in addition to NRMSE and Pearson correlation\. We compare zero\- and four\-shot Gemini 3\.1 Pro withDualViewusing Gemma 4 31B on each dataset\. We selected this model forDualViewbased on validation results, since comparing Gemini against results aggregated across multipleDualViewbackbones would not be an apples\-to\-apples comparison\.

Table 18:Comparisons to Gemini 3\.1 Pro\.We compareDualViewto Gemini 3\.1 Pro under zero\-shot and few\-shot \(4 demonstrations\) settings across five datasets\.DualViewresults use Gemma 4 31B\. Metrics are macro\-averaged across dimensions within each dataset with standard deviations across three seeds \(±\\pm\)\.

## Appendix FAdditional Ablations

Table 19:Linear readout comparisons\.Single\-view controls use either broad or focused representations\. Joint two\-view configurations vary the broad rank constraint and focused parameter sharing\. Values are macro\-averaged across five datasets and eight LLMs, with standard deviations across three seeds\. We include GAT\-AES as the strongest adapted supervised baseline\.### F\.1Strong Linear Baselines

To assess how much ofDualView’s advantage comes from the extracted representations and regularized linear prediction, we compare against single\-view ridge baselines and less constrained two\-view predictors\.Table[19](https://arxiv.org/html/2610.19206#A6.T19)consolidates the corresponding results from the evidence\-complementarity and parameter\-sharing ablations\.

##### Strength of single\-view linear predictors\.

Full\-rank ridge regression on broad representations achieves a Pearson correlation of 0\.388 and NRMSE of 0\.892, outperforming all seven adapted supervised baselines in aggregate\. Focused\-only ridge predictors are also competitive: the shared version achieves a Pearson correlation of 0\.377, while independent predictors achieve 0\.379 and an NRMSE of 0\.893\. These results establish that regularized linear readouts of frozen LLM representations are strong baselines by themselves\.

##### Value of combining evidence\.

Nevertheless,DualViewoutperforms full\-rank broad ridge, with a 11\.3% higher Pearson and a 2\.1% lower NRMSE\. Although a simple ridge predictor on broad representations already outperforms the strongest supervised baseline \(GAT\-AES\), combining broad and target\-specific representations yields further gains\. This finding supports the complementary predictive value of the two evidence views\.

##### Role of parameter sharing\.

Less constrained two\-view predictors perform similarly toDualView\. Consequently, the parameter\-sharing constraints should primarily be interpreted as an efficiency mechanism: they substantially reduce the trainable footprint while preserving comparable observed predictive performance\. Nevertheless, these comparisons are not measurements of total computational savings\.

### F\.2Statistical Reliability of Ablation Improvements

We assess the statistical reliability of two key ablation findings: a\) the benefit of combining broad and output\-specific evidence compared to full rank broad\-only ridge \(Table[4](https://arxiv.org/html/2610.19206#S6.T4)\), and b\) the benefit of alternating joint optimization over separate fitting withα\\alphafusion \(Table[5](https://arxiv.org/html/2610.19206#S6.T5)\)\. Specifically, we conduct a paired, source\-group\-level bootstrap analysis on test predictions\. We resample source groups with replacement within each dataset, using identical resamples for each pair of compared methods to preserve within\-group dependencies\. For each of 5,000 bootstrap replicates, we recompute Pearson correlation and NRMSE and aggregate across all outputs, LLM backbones, datasets, and seeds following the protocol in §[D\.4](https://arxiv.org/html/2610.19206#A4.SS4)\.

We report 95% confidence intervals \(CIs\) for paired performance differences, computed asDualViewminus the baseline for Pearson correlation and the reverse for NRMSE, so that positive values indicate improvements in both metrics\. Table[20](https://arxiv.org/html/2610.19206#A6.T20)shows thatDualView’s gains are likely statistically significant, since all intervals exclude zero\.

Table 20:95% Confidence Intervalsfor improvements ofDualViewcompared to ablation alternatives\. Positive values indicate that differences are likely statistically significant\.Compared with full\-rank broad\-only ridge,DualViewimproves aggregate Pearson correlation from 0\.3876 to 0\.4322 \(\+0\.0447\) and reduces NRMSE from 0\.8919 to 0\.8732 \(\+0\.0186\)\. The corresponding 95% CIs are\[0\.0328,0\.0497\]\[0\.0328,0\.0497\]and\[0\.0134,0\.0229\]\[0\.0134,0\.0229\], respectively, supporting the complementary predictive value of broad and output\-specific representations\.

Compared with separate fitting followed byα\\alphafusion,DualViewfurther improves Pearson correlation from 0\.419 to 0\.432 \(\+0\.013\) and reduces NRMSE from 0\.880 to 0\.873 \(a reduction of 0\.007\)\. The corresponding 95% CIs are\[0\.0053,0\.0177\]\[0\.0053,0\.0177\]and\[0\.0039,0\.0109\]\[0\.0039,0\.0109\], respectively, further supporting that alternating joint optimization outperforms separate fitting and fusion\.

## Appendix GExpanded Discussion

### G\.1Multi\-Objective Optimization

Multi\-objective formulations typically apply to balance different losses under multi\-task learning\([Sener and Koltun, 2018](https://arxiv.org/html/2610.19206#bib.bib31)\)\. PCGrad projects conflicting gradients, CAGrad regularizes updates through worst\-task local improvement, Nash\-MTL negotiates a joint update, and FAMO adaptively balances loss reduction\([Yu et al\., 2020](https://arxiv.org/html/2610.19206#bib.bib32);[Liu et al\., 2021](https://arxiv.org/html/2610.19206#bib.bib33);[Navon et al\., 2022](https://arxiv.org/html/2610.19206#bib.bib34);[Liu et al\., 2023a](https://arxiv.org/html/2610.19206#bib.bib35)\)\. FairGrad controls fairness of task progress, while ExcessMTL addresses excessive weighting of noisy tasks\([Ban and Ji, 2024](https://arxiv.org/html/2610.19206#bib.bib40);[He et al\., 2024](https://arxiv.org/html/2610.19206#bib.bib41)\)\. Recent work also studies aligned objectives, selective task\-group updates, and task\-affinity prediction\([Efroni et al\., 2025](https://arxiv.org/html/2610.19206#bib.bib37);[Jeong and Yoon, 2025](https://arxiv.org/html/2610.19206#bib.bib36);[Ayman et al\., 2026](https://arxiv.org/html/2610.19206#bib.bib38)\)\.

Some MOO works cater to text regression\. SaMRL trains an autoregressive multi\-trait essay scorer with QWK\-based rewards and an MSE penalty, learning weights for multiple reward\-specific PPO losses\([Do et al\., 2024b](https://arxiv.org/html/2610.19206#bib.bib42)\)\. LMUnit combines preference, direct\-rating, and rationale supervision for output\-conditioned evaluation\([Saad\-Falcon et al\., 2025](https://arxiv.org/html/2610.19206#bib.bib43)\)\. Multi\-objective reward models fit attribute scores\([Wang et al\., 2024](https://arxiv.org/html/2610.19206#bib.bib4)\), and recent work combines Bradley–Terry preference and multi\-attribute regression objectives in a shared embedding space\([Zhang et al\., 2026b](https://arxiv.org/html/2610.19206#bib.bib45)\)\. Thus, these MOO methods are compatible withDualView’s evidence design by balancing optimization across outputs, whereas our two\-view design specifies the evidence available to each prediction\.

## Appendix HDataset Prompt Templates

Tables[22](https://arxiv.org/html/2610.19206#A8.T22)–[26](https://arxiv.org/html/2610.19206#A8.T26)specify the prompts used for each dataset\. Within a dataset, prompts across LLMs share the same task instructions, descriptions for all output targets, and input fields\. The broad prompt ends after these components\. The focused prompts append a short target\-specific instruction to this shared prefix\. Specifically, the templates reflectPib=\[s⁡\(𝒬\);xi\]P\_\{i\}^\{b\}=\[s\(\\mathcal\{Q\}\);x\_\{i\}\]andPi​jf=\[s⁡\(𝒬\);xi;q⁡\(dj\)\]P\_\{ij\}^\{f\}=\[s\(\\mathcal\{Q\}\);x\_\{i\};q\(d\_\{j\}\)\]\(§[3\.2](https://arxiv.org/html/2610.19206#S3.SS2)\), enabling shared\-prefix caching\. Braces indicate input fields and target\-specific substitutions\.

For the LLM\-as\-a\-Judge settings, the broad prompt is the zero\-shot query\.Table[21](https://arxiv.org/html/2610.19206#A8.T21)specifies the template for appending demonstrations drawn from the training split within each seed\.

Use the labeled examples to help answer the query\.Labeled examples:Example 1 input:\{broad prompt for training example 1\}Example 1 answer JSON:\{gold\-score JSON object for training example 1\}Repeat the example block for examples 2, 3, and 4\.Query:\{broad prompt for the query example\}

Table 21:Shared template for few\-shot prompting\.Table 22:Broad and focused prompts for CreativeEval\.The broad prompt contains the shared task instructions and input fields\. Each focused prompt reuses that entire prefix and appends the target\-specific suffix, instantiated once per output\. Braces mark substitutions with actual data\.Table 23:Broad and focused prompts for MSumBench\.The broad prompt contains the shared task instructions and input fields\. Each focused prompt reuses that entire prefix and appends the target\-specific suffix, instantiated once per output\. Braces mark substitutions with actual data\.Table 24:Broad and focused prompts for RATE\.The broad prompt contains the shared task instructions and input fields\. Each focused prompt reuses that entire prefix and appends the target\-specific suffix, instantiated once per output\. Braces mark substitutions with actual data\.Table 25:Broad and focused prompts for PRISM\.The broad prompt contains the shared task instructions and input fields\. Each focused prompt reuses that entire prefix and appends the target\-specific suffix, instantiated once per output\. Braces mark substitutions with actual data\.Table 26:Broad and focused prompts for DSCodeBench\.The broad prompt contains the shared task instructions and input fields\. Each focused prompt reuses that entire prefix and appends the target\-specific suffix, instantiated once per output\. Braces mark substitutions with actual data\.
## Appendix IOutput\-Level Performance Tables

Tables[27](https://arxiv.org/html/2610.19206#A9.T27)–[34](https://arxiv.org/html/2610.19206#A9.T34),[35](https://arxiv.org/html/2610.19206#A9.T35)–[42](https://arxiv.org/html/2610.19206#A9.T42),[43](https://arxiv.org/html/2610.19206#A9.T43)–[50](https://arxiv.org/html/2610.19206#A9.T50),[51](https://arxiv.org/html/2610.19206#A9.T51)–[58](https://arxiv.org/html/2610.19206#A9.T58), and[59](https://arxiv.org/html/2610.19206#A9.T59)–[66](https://arxiv.org/html/2610.19206#A9.T66)report output\-level NRMSE, Pearson, and Spearman for CreativeEval, MSumBench, RATE, PRISM, and DSCodeBench, respectively, across all eight LLM backbones\. Values are averaged across three training seeds with sample standard deviation \(±\\pm\)\.

Table 27:CreativeEval, Gemma E4B output\-level metrics\. Green highlights a winningDualViewcell and yellow highlights a winning baseline cell\.Table 28:CreativeEval, Qwen 9B output\-level metrics\. Green highlights a winningDualViewcell and yellow highlights a winning baseline cell\.Table 29:CreativeEval, Qwen 27B output\-level metrics\. Green highlights a winningDualViewcell and yellow highlights a winning baseline cell\.Table 30:CreativeEval, Gemma 31B output\-level metrics\. Green highlights a winningDualViewcell and yellow highlights a winning baseline cell\.Table 31:CreativeEval, Llama 70B output\-level metrics\. Green highlights a winningDualViewcell and yellow highlights a winning baseline cell\.Table 32:CreativeEval, Qwen 72B output\-level metrics\. Green highlights a winningDualViewcell and yellow highlights a winning baseline cell\.Table 33:CreativeEval, GPT 120B output\-level metrics\. Green highlights a winningDualViewcell and yellow highlights a winning baseline cell\.Table 34:CreativeEval, Qwen 122B output\-level metrics\. Green highlights a winningDualViewcell and yellow highlights a winning baseline cell\.Table 35:MSumBench, Gemma E4B output\-level metrics\. Green highlights a winningDualViewcell and yellow highlights a winning baseline cell\.Table 36:MSumBench, Qwen 9B output\-level metrics\. Green highlights a winningDualViewcell and yellow highlights a winning baseline cell\.Table 37:MSumBench, Qwen 27B output\-level metrics\. Green highlights a winningDualViewcell and yellow highlights a winning baseline cell\.Table 38:MSumBench, Gemma 31B output\-level metrics\. Green highlights a winningDualViewcell and yellow highlights a winning baseline cell\.Table 39:MSumBench, Llama 70B output\-level metrics\. Green highlights a winningDualViewcell and yellow highlights a winning baseline cell\.Table 40:MSumBench, Qwen 72B output\-level metrics\. Green highlights a winningDualViewcell and yellow highlights a winning baseline cell\.Table 41:MSumBench, GPT 120B output\-level metrics\. Green highlights a winningDualViewcell and yellow highlights a winning baseline cell\.Table 42:MSumBench, Qwen 122B output\-level metrics\. Green highlights a winningDualViewcell and yellow highlights a winning baseline cell\.Table 43:RATE, Gemma E4B output\-level metrics\. Green highlights a winningDualViewcell and yellow highlights a winning baseline cell\.Table 44:RATE, Qwen 9B output\-level metrics\. Green highlights a winningDualViewcell and yellow highlights a winning baseline cell\.Table 45:RATE, Qwen 27B output\-level metrics\. Green highlights a winningDualViewcell and yellow highlights a winning baseline cell\.Table 46:RATE, Gemma 31B output\-level metrics\. Green highlights a winningDualViewcell and yellow highlights a winning baseline cell\.Table 47:RATE, Llama 70B output\-level metrics\. Green highlights a winningDualViewcell and yellow highlights a winning baseline cell\.Table 48:RATE, Qwen 72B output\-level metrics\. Green highlights a winningDualViewcell and yellow highlights a winning baseline cell\.Table 49:RATE, GPT 120B output\-level metrics\. Green highlights a winningDualViewcell and yellow highlights a winning baseline cell\.Table 50:RATE, Qwen 122B output\-level metrics\. Green highlights a winningDualViewcell and yellow highlights a winning baseline cell\.Table 51:PRISM, Gemma E4B output\-level metrics\. Green highlights a winningDualViewcell and yellow highlights a winning baseline cell\.Table 52:PRISM, Qwen 9B output\-level metrics\. Green highlights a winningDualViewcell and yellow highlights a winning baseline cell\.Table 53:PRISM, Qwen 27B output\-level metrics\. Green highlights a winningDualViewcell and yellow highlights a winning baseline cell\.Table 54:PRISM, Gemma 31B output\-level metrics\. Green highlights a winningDualViewcell and yellow highlights a winning baseline cell\.Table 55:PRISM, Llama 70B output\-level metrics\. Green highlights a winningDualViewcell and yellow highlights a winning baseline cell\.Table 56:PRISM, Qwen 72B output\-level metrics\. Green highlights a winningDualViewcell and yellow highlights a winning baseline cell\.Table 57:PRISM, GPT 120B output\-level metrics\. Green highlights a winningDualViewcell and yellow highlights a winning baseline cell\.Table 58:PRISM, Qwen 122B output\-level metrics\. Green highlights a winningDualViewcell and yellow highlights a winning baseline cell\.Table 59:DSCodeBench, Gemma E4B output\-level metrics\. Green highlights a winningDualViewcell and yellow highlights a winning baseline cell\.Table 60:DSCodeBench, Qwen 9B output\-level metrics\. Green highlights a winningDualViewcell and yellow highlights a winning baseline cell\.Table 61:DSCodeBench, Qwen 27B output\-level metrics\. Green highlights a winningDualViewcell and yellow highlights a winning baseline cell\.Table 62:DSCodeBench, Gemma 31B output\-level metrics\. Green highlights a winningDualViewcell and yellow highlights a winning baseline cell\.Table 63:DSCodeBench, Llama 70B output\-level metrics\. Green highlights a winningDualViewcell and yellow highlights a winning baseline cell\.Table 64:DSCodeBench, Qwen 72B output\-level metrics\. Green highlights a winningDualViewcell and yellow highlights a winning baseline cell\.Table 65:DSCodeBench, GPT 120B output\-level metrics\. Green highlights a winningDualViewcell and yellow highlights a winning baseline cell\.Table 66:DSCodeBench, Qwen 122B output\-level metrics\. Green highlights a winningDualViewcell and yellow highlights a winning baseline cell\.