CriticGen: Generation-Aware Evaluation as Actionable Feedback

# CriticGen: Generation-Aware Evaluation as Actionable Feedback
Source: [https://arxiv.org/html/2609.05439](https://arxiv.org/html/2609.05439)
Zecheng Zuo11footnotemark:1Beijing University of Posts and Telecommunications, Beijing, ChinaSen WangBeijing University of Posts and Telecommunications, Beijing, ChinaChenghao FanUniversity of Science and Technology Beijing, Beijing, ChinaHaofen WangTongji University, Shanghai, ChinaYehui YangCorresponding author \(yangyehuisw@126\.com\)\. Also contactable via duhuifang123@gmail\.com\.AI Lab, Qifu Technology, Beijing, China

###### Abstract

Current evaluation methods for large language models are coarse\-grained and decoupled from generation, producing generic explanations that fail to provide actionable feedback for model improvement\. We propose CriticGen, a fine\-grained, generation\-aware evaluation framework that turns evaluation into actionable control for answer improvement\. CriticGen first generates sample\-specific evaluation dimensions and scoring criteria under high\-level categories such as subjective, objective, and self\-derived constraints\. These criteria then serve as a dynamic rubric for jointly producing a score, a reason, an executable refinement suggestion, and a refined answer\. This rubric\-conditioned refinement process enables models to diagnose flaws and perform targeted answer improvement\. Experimental results show that fine\-grained evaluation should be both instance\-specific and actionable\. CriticGen induces higher\-quality rubrics, improving relevance/coverage from 3\.33/4\.03 to 3\.97/4\.24\. CriticGen also achieves the best score correlations, with 0\.9556 Pearson and 0\.9560 Spearman, and raises the F1 of criterion\-grounded reasons and executable suggestions from 0\.6369/0\.5994 to 0\.7554/0\.7900\. Crucially, its feedback translates into reliable answer improvement, improving 73\.17% of answers with a 93\.28% non\-degradation rate\.

CriticGen: Generation\-Aware Evaluation as Actionable Feedback

## 1Introduction

![Refer to caption](https://arxiv.org/html/2609.05439v1/x1.png)Figure 1:Illustration of CriticGen’s core idea\. Existing methods often provide fixed or generic feedback that is weakly connected to generation\. CriticGen induces a sample\-specific rubric and uses it to score, diagnose, plan, and refine the answer, turning evaluation into actionable control for answer improvement\.Large language models \(LLMs\) are increasingly expected to produce not only correct final answers, but also faithful, consistent, and constraint\-aware reasoning processes\(Weiet al\.,[2022](https://arxiv.org/html/2609.05439#bib.bib1); Cobbeet al\.,[2021](https://arxiv.org/html/2609.05439#bib.bib2)\)\. For chain\-of\-thought \(CoT\) outputs\(Weiet al\.,[2022](https://arxiv.org/html/2609.05439#bib.bib1)\), evaluation is therefore no longer only a leaderboard metric\. It also serves as a diagnostic tool for identifying reasoning failures, curating training data, and guiding model improvement\(Ouyanget al\.,[2022](https://arxiv.org/html/2609.05439#bib.bib19)\)\. However, current evaluation methods still leave a gap between*judging*an answer and*improving*it\.

The first limitation is that many evaluation protocols are insufficiently adaptive to individual answers\. Outcome\-level metrics such as exact match or shallow text overlap only indicate whether the final answer matches a reference, but do not reveal which reasoning aspect is weak\(Papineniet al\.,[2002](https://arxiv.org/html/2609.05439#bib.bib22)\)\. Recent LLM\-as\-a\-judge and fine\-grained evaluation methods improve this by comparing responses, assigning ratings, or scoring outputs along multiple skills or criteria\(Zhenget al\.,[2023](https://arxiv.org/html/2609.05439#bib.bib7); Liuet al\.,[2023](https://arxiv.org/html/2609.05439#bib.bib9); Yeet al\.,[2024](https://arxiv.org/html/2609.05439#bib.bib10); Kimet al\.,[2024](https://arxiv.org/html/2609.05439#bib.bib11)\)\. Generative judges further enhance interpretability by producing natural\-language critiques together with judgments\(Liet al\.,[2024b](https://arxiv.org/html/2609.05439#bib.bib15)\)\. Nevertheless, their rubrics are often fixed at the task or scenario level\. Such criteria may cover general quality dimensions, but they may fail to capture the specific omissions, contradictions, or constraint violations that appear in a particular question–answer pair\.

The second limitation is that evaluation signals are often not directly executable for generation\. Evaluator models such as Prometheus and LLaVA\-Critic can provide fine\-grained scores or feedback, and these signals can be used for reward\-guided selection or later policy optimization\(Kimet al\.,[2024](https://arxiv.org/html/2609.05439#bib.bib11); Xionget al\.,[2025](https://arxiv.org/html/2609.05439#bib.bib12); Schulmanet al\.,[2017](https://arxiv.org/html/2609.05439#bib.bib20)\)\. This paradigm is effective, but it typically separates evaluation from generation: the evaluator is trained first, and its signal affects generation only through an additional optimization or selection stage\. Recent critic\-based generation methods show that critique can improve model outputs, for example by refining erroneous reasoning steps or by training stronger policy behavior from critic data\(Zhenget al\.,[2025](https://arxiv.org/html/2609.05439#bib.bib13); Wanget al\.,[2025](https://arxiv.org/html/2609.05439#bib.bib14)\)\. Yet the critique is still not necessarily grounded in the same fine\-grained criteria used for scoring, nor is it explicitly converted into a concrete edit plan for revising the same answer\.

We ask whether evaluation can be made both*sample\-specific*and*actionable*\. To this end, we proposeCriticGen, a generation\-aware evaluation framework that converts fine\-grained evaluation into controllable answer refinement\. CriticGen contains two separately trained SFT components coupled by a rubric interface\. First, a rubric generator induces a sample\-specific rubricR​\(q,a\)=\{\(di,Ci\)\}i=1mR\(q,a\)=\\\{\(d\_\{i\},C\_\{i\}\)\\\}\_\{i=1\}^\{m\}, where each dimensiondid\_\{i\}is paired with a graded scoring criterionCiC\_\{i\}\. The rubric is organized under high\-level constraint families, including subjective quality, objective correctness, and self\-derived reasoning constraints, while remaining specific to the current question and answer\. Second, a rubric\-conditioned evaluation\-refinement model takes each\(q,a,di,Ci\)\(q,a,d\_\{i\},C\_\{i\}\)as input and generates a structured trajectory\(si,ri,ui,ai′\)\(s\_\{i\},r\_\{i\},u\_\{i\},a^\{\\prime\}\_\{i\}\): a score, a criterion\-grounded reason, an executable refinement suggestion, and a refined answer\. In this trajectory, the reason explains*why*the answer receives the score, while the suggestion specifies*what*should be changed before rewriting\. Thus, CriticGen uses the same rubric to connect what to evaluate, why the answer is flawed, how it should be revised, and how the revised answer should be generated\.

We construct two supervision datasets to support this framework\.RubricDatatrains the rubric generator to produce answer\-specific dimensions and graded criteria\.RefineDatatrains the evaluation\-refinement model to produce the complete score–reason–suggestion–rewrite trajectory under each rubric item\. Unlike reward\-modeling or preference\-optimization pipelines, CriticGen does not rely on a separate reinforcement\-learning stage\. Instead, it studies whether structured supervised trajectories can directly teach models to transform evaluation into an executable refinement signal\.

Experiments validate the central hypothesis of CriticGen: fine\-grained evaluation should be both instance\-specific and actionable\. Human evaluation shows that CriticGen induces higher\-quality rubrics than a static checklist, improving relevance/coverage from 3\.33/4\.03 to 3\.97/4\.24\. OnRefineData, CriticGen\-Qwen3\.5\-9B achieves the best score correlations, reaching 0\.9556 Pearson and 0\.9560 Spearman, outperforming both the strongest open\-weight baseline and GPT\-5\. It also improves the generative components of evaluation, raising the best baseline F1 of criterion\-grounded reasons and executable suggestions from 0\.6369/0\.5994 to 0\.7554/0\.7900\. More importantly, CriticGen turns feedback into reliable answer refinement, improving 73\.17% of answers while maintaining a 93\.28% non\-degradation rate\. Out\-of\-domain results on Feedback Bench further show consistent transfer of rubric\-conditioned evaluation ability\.

We summarize our contributions as follows:

- •We propose sample\-specific rubric\-conditioned evaluation and refinement, which induces fine\-grained rubrics for each question–answer pair and uses them to guide a score–reason–suggestion–rewrite trajectory, turning evaluation from descriptive judgment into actionable control for answer improvement\.
- •We construct RubricData and RefineData to train the two SFT components of CriticGen, enabling rubric induction and rubric\-conditioned refinement without a separate reward\-optimization stage\.
- •We demonstrate strong empirical performance across rubric quality, score correlation, reason and suggestion generation, answer refinement, and ablation studies, validating the effectiveness of rubric\-conditioned refinement\.

## 2Data Construction

CriticGen requires supervision for two related but distinct capabilities: generating instance\-specific rubrics and refining answers under a given rubric\. We therefore construct the training data in two layers, both starting from a diverse pool of candidate answers\.

#### Diverse answer pool\.

We collect questions from five reasoning\-oriented datasets: GSM8K\(Cobbeet al\.,[2021](https://arxiv.org/html/2609.05439#bib.bib2)\), LIMO\(Yeet al\.,[2025](https://arxiv.org/html/2609.05439#bib.bib3)\), NaturalReasoning\(Yuanet al\.,[2025](https://arxiv.org/html/2609.05439#bib.bib4)\), NuminaMath\-CoT\(Liet al\.,[2024a](https://arxiv.org/html/2609.05439#bib.bib5)\), and QwQ\-LongCoT\-130K\(Amphora,[2024](https://arxiv.org/html/2609.05439#bib.bib6)\)\. Together, these datasets cover arithmetic reasoning, symbolic/logical reasoning, open\-domain natural\-language reasoning, competition\-level math, and long\-form CoT generation, which helps the rubric and refinement models avoid overfitting to a single reasoning style\. For each question, we prompt generator models of different scales and capabilities, ranging from weaker open\-weight models to stronger teacher models, to produce multiple candidate answers\. This multi\-model generation is deliberate: using models with different error profiles leads to a more balanced distribution of data across the full 0–5 quality range\.

#### Rubric supervision generation\.

For each question–answer pair\(q,a\)\(q,a\), GPT\-5\(OpenAI,[2025b](https://arxiv.org/html/2609.05439#bib.bib36)\)generates a sample\-specific rubricR∗=\{\(di∗,Ci∗\)\}i=1mR^\{\*\}=\\\{\(d\_\{i\}^\{\*\},C\_\{i\}^\{\*\}\)\\\}\_\{i=1\}^\{m\}, where each item contains an evaluation dimension and its 0–5 scoring criterion\. The dimensions are organized under the three high\-level families used in our method: subjective quality, objective correctness, and self\-derived constraints\. We store each annotated instance as\(q,a,R∗\)\(q,a,R^\{\*\}\), which provides supervision for rubric induction\.

#### Evaluation\-refinement supervision generation\.

Given each rubric item\(di∗,Ci∗\)\(d\_\{i\}^\{\*\},C\_\{i\}^\{\*\}\), GPT\-5 further acts as a judge and refiner\. It annotates a scoresi∗s\_\{i\}^\{\*\}, a criterion\-grounded reasonri∗r\_\{i\}^\{\*\}, an executable refinement suggestionui∗u\_\{i\}^\{\*\}, and a refined answerai′∗\{a\_\{i\}^\{\\prime\}\}^\{\*\}\. Each resulting instance is stored as\(q,a,di∗,Ci∗,si∗,ri∗,ui∗,ai′∗\)\(q,a,d\_\{i\}^\{\*\},C\_\{i\}^\{\*\},s\_\{i\}^\{\*\},r\_\{i\}^\{\*\},u\_\{i\}^\{\*\},\{a\_\{i\}^\{\\prime\}\}^\{\*\}\), forming the supervision for rubric\-conditioned evaluation\-to\-refinement learning\. The score and reason capture how the answer satisfies the criterion, while the suggestion and refined answer instantiate how the evaluation is converted into a concrete improvement\.

#### Score coverage completion\.

To reduce score imbalance, we check whether the answers for each question cover the full 0–5 quality range\. If some score regions are missing, we prompt GPT\-5 to synthesize additional answers targeted at the missing quality levels\. For these synthesized answers, we repeat rubric supervision generation and evaluation\-to\-refinement supervision generation\. This step improves score coverage and increases answer diversity beyond what is obtained from the initial set of generator models\.

#### Quality filtering\.

We first apply a consensus filter to acquire a reliable dataset in the evaluation\-to\-refinement stage\. Given the same question, answer, evaluation dimension, and scoring criteria, we ask Claude Sonnet 4\.5\(Anthropic,[2025](https://arxiv.org/html/2609.05439#bib.bib38)\), Gemini 3\.1 Pro Preview\(Google,[2026](https://arxiv.org/html/2609.05439#bib.bib39); Google AI for Developers,[2026](https://arxiv.org/html/2609.05439#bib.bib40)\)to independently generate the supervision\. We keep a sample only when their scores agree with the GPT\-5 score, and discard samples with score disagreement\. We further apply a lexical\-diversity and near\-duplicate filter to avoid highly templated or repetitively phrased generations\. Specifically, we compute field\-levelnn\-gram diversity and sample\-levelnn\-gram overlap over questions, answers, rubrics, reasons, refinement suggestions, and refined responses, and remove samples whose supervision is near\-duplicated by previously retained instances\. Appendix[B](https://arxiv.org/html/2609.05439#A2)provides the filtering details\. The resulting data consist of two datasets:RubricDatafor dynamic rubric induction, andRefineDatafor rubric\-conditioned evaluation\-to\-refinement learning\. We splitRubricDataandRefineDatainto training and test sets with ratios of 5:1 and 50:1, respectively\. Table[1](https://arxiv.org/html/2609.05439#S2.T1)summarizes the final data statistics\.

Table 1:Statistics of the constructed datasets after quality filtering\. learning\.![Refer to caption](https://arxiv.org/html/2609.05439v1/x2.png)Figure 2:Overview of the CriticGen framework\. CriticGen first induces a sample\-specific dynamic rubric for each question–answer pair, and then uses each rubric item to condition a unified model that jointly produces a score, a reason, an executable refinement suggestion, and a refined answer\.

## 3Method

CriticGen is designed around the two limitations identified in the introduction: evaluation should be fine\-grained for each instance, and its output should directly guide generation\. Accordingly, we separate the framework into two functions:*rubric induction*, which specifies what should be evaluated, and*rubric\-conditioned refinement*, which uses the evaluation result to improve the same answer\.

### 3\.1Problem Formulation

Letqqdenote a question andaaa candidate reasoning answer\. Instead of learning only a scalar evaluatorf​\(q,a\)→sf\(q,a\)\\rightarrow s, CriticGen aims to produce a structured set of rubric\-conditioned evaluation and refinement outputs:

F​\(q,a\)→\{\(di,Ci,si,ri,ui,ai′\)\}i=1m\.F\(q,a\)\\rightarrow\\\{\(d\_\{i\},C\_\{i\},s\_\{i\},r\_\{i\},u\_\{i\},a^\{\\prime\}\_\{i\}\)\\\}\_\{i=1\}^\{m\}\.\(1\)Here,did\_\{i\}is an evaluation dimension,CiC\_\{i\}is its graded scoring criterion,sis\_\{i\}is the predicted score,rir\_\{i\}is the score explanation,uiu\_\{i\}is an executable refinement suggestion, andai′a^\{\\prime\}\_\{i\}is the refined answer\. This formulation makes evaluation not only descriptive, through\(si,ri\)\(s\_\{i\},r\_\{i\}\), but also operational, through\(ui,ai′\)\(u\_\{i\},a^\{\\prime\}\_\{i\}\)\.

### 3\.2Sample\-Specific Rubric Induction

The first component induces a dynamic rubric for each question–answer pair\. Given\(q,a\)\(q,a\), a rubric generator𝒢\\mathcal\{G\}produces a set of dimensions and criteria:

R​\(q,a\)=\{\(di,Ci\)\}i=1m\.R\(q,a\)=\\\{\(d\_\{i\},C\_\{i\}\)\\\}\_\{i=1\}^\{m\}\.\(2\)To ensure broad coverage while avoiding a fixed rubric, we organize dimensions under three high\-level families:

- •Subjective quality: clarity, coherence, fluency, and readability;
- •Objective correctness: factual accuracy, numerical correctness, and task\-constraint satisfaction;
- •Self\-derived constraints: consistency among reasoning steps, absence of contradictions, and whether the conclusion follows from previous steps\.

For each generated dimensiondid\_\{i\}, the model also produces a 0–5 criterionCiC\_\{i\}that describes what different score levels mean for the current question and answer\. This differs from task\-level or scenario\-level rubrics: the generated criteria are tied to the concrete failure modes that may appear in the current answer\.

### 3\.3Rubric\-Conditioned Evaluation and Refinement

The second component takes a single rubric item\(di,Ci\)\(d\_\{i\},C\_\{i\}\)and evaluates the candidate answer under that criterion\. For each\(q,a,di,Ci\)\(q,a,d\_\{i\},C\_\{i\}\), a rubric\-conditioned model𝒮\\mathcal\{S\}generates four outputs:

𝒮​\(q,a,di,Ci\)→\(si,ri,ui,ai′\)\.\\mathcal\{S\}\(q,a,d\_\{i\},C\_\{i\}\)\\rightarrow\(s\_\{i\},r\_\{i\},u\_\{i\},a^\{\\prime\}\_\{i\}\)\.\(3\)The scoresi∈\{0,…,5\}s\_\{i\}\\in\\\{0,\\ldots,5\\\}measures how well the answer satisfies the criterion\. The reasonrir\_\{i\}explains the score by identifying the relevant strength or weakness\. The refinement suggestionuiu\_\{i\}turns this explanation into an executable instruction, such as adding a missing reasoning step, correcting a calculation, enforcing a neglected constraint, or clarifying a justification\. Finally, the refined answerai′a^\{\\prime\}\_\{i\}is generated under the same criterion\.

The key design choice is thatuiu\_\{i\}is not merely another explanation:rir\_\{i\}diagnoses the problem, whileuiu\_\{i\}specifies how the refiner should act\. We calluiu\_\{i\}*executable*because it gives a concrete edit operation and target content, e\.g\., “insert a step before the final conclusion to apply the non\-negativity constraint\.” Thus, the model learns to move from “why this score” to “what should be changed” and then to an improved answer\.

### 3\.4Training Objectives

We train the two components with supervised fine\-tuning on their corresponding constructed supervision\. This design makes the source of the gains explicit: CriticGen is not optimized by a separate reward or preference stage, but learns from two forms of structured supervision—instance\-specific rubrics and rubric\-conditioned evaluation\-to\-refinement trajectories\.

#### SFT for rubric induction\.

Let𝒟R=\{\(q\(n\),a\(n\),R∗\(n\)\)\}n=1NR\\mathcal\{D\}\_\{R\}=\\\{\(q^\{\(n\)\},a^\{\(n\)\},R^\{\*\(n\)\}\)\\\}\_\{n=1\}^\{N\_\{R\}\}denoteRubricData, whereR∗=\{\(di∗,Ci∗\)\}i=1mR^\{\*\}=\\\{\(d\_\{i\}^\{\*\},C\_\{i\}^\{\*\}\)\\\}\_\{i=1\}^\{m\}is the constructed ground\-truth rubric for a question–answer pair\. The rubric generator𝒢θ\\mathcal\{G\}\_\{\\theta\}parameterizes a conditional distribution over variable\-length rubric items and is optimized with the empirical negative log\-likelihood:

Lrubric​\(θ\)=−𝔼𝒟R​log⁡p𝒢θ​\(R∗∣q,a\)\.L\_\{\\mathrm\{rubric\}\}\(\\theta\)=\-\\mathbb\{E\}\_\{\\mathcal\{D\}\_\{R\}\}\\log p\_\{\\mathcal\{G\}\_\{\\theta\}\}\(R^\{\*\}\\mid q,a\)\.\(4\)In practice, this objective is implemented as a token\-level cross\-entropy over the serialized dimensions and their 0–5 scoring criteria\.

#### SFT for rubric\-conditioned evaluation and refinement\.

GivenRefineData, each training example contains an inputxi=\(q,a,di,Ci\)x\_\{i\}=\(q,a,d\_\{i\},C\_\{i\}\)and a consensus\-filtered teacher outputyi∗=\(si∗,ri∗,ui∗,ai′⁣∗\)y\_\{i\}^\{\*\}=\(s\_\{i\}^\{\*\},r\_\{i\}^\{\*\},u\_\{i\}^\{\*\},a\_\{i\}^\{\\prime\*\}\), where the scoresi∗s\_\{i\}^\{\*\}is retained only when multiple teacher models agree\. We train the rubric\-conditioned model𝒮ϕ\\mathcal\{S\}\_\{\\phi\}with a sequence likelihood objective over the serialized teacher output:

Lrefine​\(ϕ\)=−𝔼RefineData​log⁡p𝒮ϕ​\(yi∗∣xi\)\.L\_\{\\mathrm\{refine\}\}\(\\phi\)=\-\\mathbb\{E\}\_\{\\textsc\{RefineData\}\}\\log p\_\{\\mathcal\{S\}\_\{\\phi\}\}\(y\_\{i\}^\{\*\}\\mid x\_\{i\}\)\.\(5\)This single objective teaches the model to produce the complete score–reason–revision\-suggestion–rewrite trajectory under the same rubric item\. The fixed field order also provides lightweight structural supervision, so the model learns not only the content of each field but also the interface through which evaluation is converted into an executable refinement signal\.

Although the two modules are trained independently, they are coupled at inference time through the induced rubric, which serves as the interface between evaluation target generation and actionable refinement\.

## 4Experiments

We structure our experiments around the two components of CriticGen: sample\-specific rubric induction and rubric\-conditioned evaluation and refinement\. For rubric induction, we examine whether the induced rubricR​\(q,a\)=\{\(di,Ci\)\}i=1mR\(q,a\)=\\\{\(d\_\{i\},C\_\{i\}\)\\\}\_\{i=1\}^\{m\}serves as a useful assessment plan for a specific question–answer pair, using human judgments of rubric quality\. For rubric\-conditioned evaluation and refinement, we test whether conditioning on a rubric item\(di,Ci\)\(d\_\{i\},C\_\{i\}\)supports reliable scoring, criterion\-grounded reasoning, actionable revision suggestions, and improved answer refinement\.

### 4\.1Setup

#### Held\-out data\.

For rubric induction quality, we sample 240\(q,a\)\(q,a\)pairs, 48 from each of GSM8K, LIMO, NaturalReasoning, NuminaMath\-CoT, and QwQ\-LongCoT\. These pairs are from the rubric\-induction test set ofRubricData\. ForOurs, we generate answer\-specific rubricsR​\(q,a\)=\{\(di,Ci\)\}R\(q,a\)=\\\{\(d\_\{i\},C\_\{i\}\)\\\}\(mean 8\.4 dimensions per pair; Table[2](https://arxiv.org/html/2609.05439#S4.T2)\)\.Staticuses the same 240\(q,a\)\(q,a\)pairs with nine fixed, instance\-agnostic dimensions \(three per family: subjective quality, objective correctness, self\-derived constraints\), keeping checklist length comparable toOurs\. This setup tests whether answer\-specific rubrics improve relevance and coverage over static ones with matched dimension counts\.

#### Human evaluation protocol\.

For each of the 240 held\-out\(q,a\)\(q,a\)pairs, raters compare two rubrics under the same question and answer\. Rubric identities are blinded as Rubric A/B\. Two independent raters with prior experience in LLM\-as\-a\-judge evaluation and rubric design assign five\-point Likert scores\(Likert,[1932](https://arxiv.org/html/2609.05439#bib.bib28)\)for two attributes:*relevance*, measuring how well the rubric dimensions and criteria fit the current\(q,a\)\(q,a\), and*coverage*, measuring whether salient failure modes for judging the answer are represented\. If the two raters differ by at least one point on either rubric for either dimension, a senior adjudicator re\-scores both rubrics; otherwise we average the two ratings\. Raters are not authors of this work\.

#### Baseline models\.

We select baselines to cover backbone, scale, family, and proprietary\-model comparisons\. Qwen2\.5\-7B\-Instruct, Qwen3\.5\-4B, and Qwen3\.5\-9B are the backbones of our CriticGen variants, enabling direct task\-specific fine\-tuning\. Qwen2\.5\-14B\-Instruct and Qwen2\.5\-32B\-Instruct test whether gains come from scale alone, while Mistral\-7B\-Instruct\-v0\.3 and Ministral\-3\-14B\-Instruct\-2512 provide non\-Qwen open\-weight comparisons\. GPT\-5 serves as a high\-capacity proprietary reference evaluator\. Details and citations are provided in Appendix[D](https://arxiv.org/html/2609.05439#A4)\.

#### Evaluation and Refinement Metric Protocol\.

For rubric\-conditioned evaluation and refinement, we assess the outputs\(si,ri,ui,ai′\)\(s\_\{i\},r\_\{i\},u\_\{i\},a^\{\\prime\}\_\{i\}\)under each rubric item\(di,Ci\)\(d\_\{i\},C\_\{i\}\)along four strands:

Scoressis\_\{i\}: we report Pearson and Spearman correlations with the reference scores inRefineData\.

Reasonsrir\_\{i\}andexecutable suggestionsuiu\_\{i\}: since both reasons and suggestions are free\-form texts, we evaluate them by reference\-based semantic\-unit alignment rather than exact matching\(Zhanget al\.,[2020](https://arxiv.org/html/2609.05439#bib.bib25); Sellamet al\.,[2020](https://arxiv.org/html/2609.05439#bib.bib26); Liet al\.,[2025](https://arxiv.org/html/2609.05439#bib.bib27)\)\. For a generated textxix\_\{i\}and its ground\-truth referencexi∗x\_\{i\}^\{\*\}, we split them into semantic unitsXi=\{xi,j\}j=1miX\_\{i\}=\\\{x\_\{i,j\}\\\}\_\{j=1\}^\{m\_\{i\}\}andGi=\{gi,k\}k=1niG\_\{i\}=\\\{g\_\{i,k\}\\\}\_\{k=1\}^\{n\_\{i\}\}, and compute the pairwise similarity matrix

Ak,j\(i\)=Sim​\(gi,k,xi,j\)\.A^\{\(i\)\}\_\{k,j\}=\\mathrm\{Sim\}\(g\_\{i,k\},x\_\{i,j\}\)\.\(6\)We then define recall\-style coverage, precision\-style consistency, and their harmonic mean as:

Ri\\displaystyle R\_\{i\}=1ni​∑k=1nimaxj⁡Ak,j\(i\),\\displaystyle=\\frac\{1\}\{n\_\{i\}\}\\sum\_\{k=1\}^\{n\_\{i\}\}\\max\_\{j\}A^\{\(i\)\}\_\{k,j\},\(7\)Pi\\displaystyle P\_\{i\}=1mi​∑j=1mimaxk⁡Ak,j\(i\),\\displaystyle=\\frac\{1\}\{m\_\{i\}\}\\sum\_\{j=1\}^\{m\_\{i\}\}\\max\_\{k\}A^\{\(i\)\}\_\{k,j\},withFi=2​Pi​Ri/\(Pi\+Ri\)F\_\{i\}=2P\_\{i\}R\_\{i\}/\(P\_\{i\}\+R\_\{i\}\)\. Here,RiR\_\{i\}measures how well the generated text covers the ground\-truth semantic units, whilePiP\_\{i\}measures whether the generated units are supported by the ground\-truth reference\. We report macro\-averaged Precision, Recall, and F1 over all examples\. This protocol is applied to reasons by setting\(xi,xi∗\)=\(ri,ri∗\)\(x\_\{i\},x\_\{i\}^\{\*\}\)=\(r\_\{i\},r\_\{i\}^\{\*\}\), and to executable suggestions by setting\(xi,xi∗\)=\(ui,ui∗\)\(x\_\{i\},x\_\{i\}^\{\*\}\)=\(u\_\{i\},u\_\{i\}^\{\*\}\)\.

Refined answersai′a^\{\\prime\}\_\{i\}: a GPT\-5 verifier scores both the original answeraaand the refined answerai′a^\{\\prime\}\_\{i\}under the same rubric item\(di,Ci\)\(d\_\{i\},C\_\{i\}\)\. We report the improvement rate, unchanged rate, degradation rate, and the non\-degradation rate, where non\-degradation means that the refined answer is rated no worse than the original answer\.

### 4\.2Rubric Induction Quality

#### Rubric quality\.

Table[2](https://arxiv.org/html/2609.05439#S4.T2)reports human judgments of rubric quality on 240 held\-out\(q,a\)\(q,a\)pairs\. The adjudication rate is 17\.1%, and the pre\-adjudication Cohen’sκ\\kappa\(Cohen,[1968](https://arxiv.org/html/2609.05439#bib.bib29)\)is 0\.61 for relevance and 0\.57 for coverage\. Under the same\(q,a\)\(q,a\), raters assign higher scores to*Ours*than to*Static*on both relevance and coverage\. The improvement is larger on relevance, increasing from 3\.33 to 3\.97, while coverage increases from 4\.03 to 4\.24\. This result suggests that induced rubrics mainly improve instance\-level targeting rather than simply covering more generic evaluation aspects\. Qualitatively, Static dimensions behave as templates aligned with broad failure families, so raters still see many salient checks covered and assign mid\-to\-high coverage\. Answer\-conditioned induction adds criteria that explicitly mark omissions the fixed nine dimensions never single out, which is why raters score*coverage*higher for*Ours*\.

Table 2:Rubric induction: human judgments of rubric quality on 240 held\-out\(q,a\)\(q,a\)pairs\. Both rubrics are evaluated under the same question and answer; only the rubric text differs\. Scores are post\-adjudication means on a five\-point Likert scale\.Appendix[C](https://arxiv.org/html/2609.05439#A3)further supports that the induced rubrics are answer\-sensitive, not only higher\-quality under human judgment\.

### 4\.3Rubric\-Conditioned Evaluation Capability

Table 3:Rubric\-conditioned score evaluation quality\. P\. and S\. denote Pearson and Spearman correlations with reference scores inRefineData\.Table 4:Rubric\-conditioned reason generation quality\. P\., R\., and F1 denote semantic\-unit precision, recall, and F1 against reference reasons\.Table 5:Rubric\-conditioned executable suggestion generation quality\. P\., R\., and F1 denote semantic\-unit precision, recall, and F1 against reference suggestions\.Tables[3](https://arxiv.org/html/2609.05439#S4.T3),[4](https://arxiv.org/html/2609.05439#S4.T4), and[5](https://arxiv.org/html/2609.05439#S4.T5)evaluate rubric\-conditioned evaluation in terms of score prediction, reason generation, and executable suggestion generation\. Overall, CriticGen consistently outperforms its backbone models and strong instruction\-following baselines, indicating that sample\-specific rubrics provide more effective evaluation signals than generic prompting\.

For score prediction, Table[3](https://arxiv.org/html/2609.05439#S4.T3)shows that CriticGen achieves much stronger agreement with reference scores\. CriticGen\-Qwen3\.5\-9B obtains the best Pearson/Spearman correlations of 0\.9556/0\.9560, surpassing the strongest open\-weight baseline Qwen3\.5\-9B \(0\.8342/0\.8328\) by 0\.1214/0\.1232 and GPT\-5 \(0\.9212/0\.9143\) by 0\.0344/0\.0417\. The gains are also evident for smaller models: CriticGen\-Qwen2\.5\-7B improves Qwen2\.5\-7B\-Instruct from 0\.6355/0\.6300 to 0\.9518/0\.9521, suggesting that the benefit comes from rubric\-conditioned evaluation rather than model scale alone\.

Tables[4](https://arxiv.org/html/2609.05439#S4.T4)and[5](https://arxiv.org/html/2609.05439#S4.T5)further show that CriticGen improves the generative side of evaluation\. For reasons, CriticGen\-Qwen3\.5\-9B increases the best baseline F1 from 0\.6369 to 0\.7554, with recall rising from 0\.6092 to 0\.7594\. For executable suggestions, the gain is larger, improving F1 from 0\.5994 to 0\.7900 and recall from 0\.5342 to 0\.7921\. Similar improvements appear on smaller backbones\. These results support our claim that evaluation should go beyond scalar scoring: by jointly modeling scores, reasons, and executable suggestions under dynamic rubrics, CriticGen turns evaluation into actionable control for answer refinement\.

### 4\.4Refinement Capability

Figure[3](https://arxiv.org/html/2609.05439#S4.F3)compares each rubric\-conditioned refined answerai′a^\{\\prime\}\_\{i\}with the original answeraaunder the same rubric item, and categorizes the outcome as improved, unchanged, or degraded\. The non\-degradation rate is the sum of improved and unchanged cases\.

Overall, most models improve many answers, but their reliability varies\. Mistral\-7B\-Instruct\-v0\.3 performs worst, with 56\.71% improved cases and 27\.42% degraded cases\. Stronger baselines are more stable: Qwen2\.5\-32B\-Instruct achieves the highest non\-degradation rate of 95\.68%, while Ministral\-3\-14B\-Instruct\-2512 obtains the highest baseline improvement rate of 73\.79%\. CriticGen achieves a strong balance, improving 73\.17% of cases, leaving 20\.11% unchanged, and degrading only 6\.72%, for a non\-degradation rate of 93\.28%\. Compared with its Qwen2\.5\-7B backbone, CriticGen improves both the improvement rate, from 69\.49% to 73\.17%, and the non\-degradation rate, from 89\.24% to 93\.28%\.

These results show that CriticGen converts fine\-grained evaluation into actionable refinement signals, producing revisions that improve answers while keeping degradation risk low\.

![Refer to caption](https://arxiv.org/html/2609.05439v1/x3.png)Figure 3:Rubric\-conditioned refinement outcomes\. Each bar shows the proportions of improved, unchanged, and degraded cases after refinement under the same rubric item\.
### 4\.5Ablation Study on Evaluation\-to\-Refinement

We ablate the structured score–reason–suggestion–rewrite trajectory to study its role in rubric\-conditioned refinement\. Since rubric induction is evaluated separately in Section[4\.2](https://arxiv.org/html/2609.05439#S4.SS2), all variants use the same rubric condition\(di,Ci\)\(d\_\{i\},C\_\{i\}\)and differ only in the second\-stage supervision fields\. All experiments are based on CriticGen\-Qwen3\.5\-4B\-SFT and evaluated on the held\-out split ofRefineData\.

Table[6](https://arxiv.org/html/2609.05439#S4.T6)shows that Full CriticGen performs best, achieving an improved rate of 73\.54\. Removing reason or suggestion supervision only mildly affects the remaining evaluation metrics, but reduces refinement success to 60\.48 and 61\.57, respectively\. The rewrite\-only variant also underperforms Full CriticGen \(62\.35\) and loses interpretable score, reason, and suggestion outputs\. These results indicate that criterion\-grounded diagnosis and executable edit planning are crucial for converting evaluation into effective and controllable refinement\.

Table 6:Ablation study on the second\-stage evaluation\-to\-refinement model\. All variants use the same rubric condition\(di,Ci\)\(d\_\{i\},C\_\{i\}\)\. Score P\. denotes Pearson correlation for score prediction; reason and suggestion quality are measured by semantic\-unit F1\. Improved denotes the proportion of refined answers rated better than the original answers\.

## 5Conclusion

We presented CriticGen, a generation\-aware evaluation framework that turns fine\-grained evaluation into actionable control for answer refinement\. CriticGen first induces sample\-specific rubrics for each question\-answer pair and then uses the induced criteria to jointly produce a score, a criterion\-grounded reason, an executable refinement suggestion, and a refined answer\. Experiments show that CriticGen improves rubric quality, score correlation, reason and suggestion generation, and answer refinement reliability\. These results suggest that evaluation can serve not only as a diagnostic signal but also as a direct mechanism for controllable generation improvement\.

## Limitations

This work has several limitations\. First, CriticGen is trained with supervised fine\-tuning, which depends on the quality and coverage of the constructed supervision data\. Although our results show strong performance, the framework may still inherit biases or blind spots from the generated rubrics and refinement trajectories\. Second, our current experiments focus on text reasoning tasks, and further validation is needed on broader domains such as multimodal reasoning, long\-form generation, and interactive decision\-making\. Third, while CriticGen directly converts evaluation into refinement through a structured SFT trajectory, it does not yet optimize the model with reinforcement learning\. Future work can further use reinforcement learning or preference optimization to strengthen the alignment between rubric\-conditioned feedback and actual answer improvement\.

## References

- Amphora \(2024\)QwQ\-longcot\-130k\.Note:Hugging Face datasetExternal Links:[Link](https://huggingface.co/datasets/amphora/QwQ-LongCoT-130K)Cited by:[§2](https://arxiv.org/html/2609.05439#S2.SS0.SSS0.Px1.p1.1)\.
- Anthropic \(2025\)Introducing claude sonnet 4\.5\.Note:[https://www\.anthropic\.com/news/claude\-sonnet\-4\-5](https://www.anthropic.com/news/claude-sonnet-4-5)Accessed: 2026\-05\-26Cited by:[§2](https://arxiv.org/html/2609.05439#S2.SS0.SSS0.Px5.p1.2)\.
- K\. Cobbe, V\. Kosaraju, M\. Bavarian, M\. Chen, H\. Jun, L\. Kaiser, M\. Plappert, J\. Tworek, J\. Hilton, R\. Nakano, C\. Hesse, and J\. Schulman \(2021\)Training verifiers to solve math word problems\.InNeurIPS,Cited by:[Appendix A](https://arxiv.org/html/2609.05439#A1.SS0.SSS0.Px1.p1.1),[§1](https://arxiv.org/html/2609.05439#S1.p1.1),[§2](https://arxiv.org/html/2609.05439#S2.SS0.SSS0.Px1.p1.1)\.
- J\. Cohen \(1968\)Weighted kappa: nominal scale agreement with provision for scaled disagreement or partial credit\.Psychological Bulletin70\(4\),pp\. 213–220\.Cited by:[§4\.2](https://arxiv.org/html/2609.05439#S4.SS2.SSS0.Px1.p1.3)\.
- Google AI for Developers \(2026\)Gemini api models\.Note:[https://ai\.google\.dev/gemini\-api/docs/models](https://ai.google.dev/gemini-api/docs/models)Accessed: 2026\-05\-26Cited by:[§2](https://arxiv.org/html/2609.05439#S2.SS0.SSS0.Px5.p1.2)\.
- Google \(2026\)Gemini 3\.1 pro: a smarter model for your most complex tasks\.Note:[https://blog\.google/innovation\-and\-ai/models\-and\-research/gemini\-models/gemini\-3\-1\-pro/](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-1-pro/)Accessed: 2026\-05\-26Cited by:[§2](https://arxiv.org/html/2609.05439#S2.SS0.SSS0.Px5.p1.2)\.
- A\. Q\. Jiang, A\. Sablayrolles, A\. Mensch, C\. Bamford, D\. S\. Chaplot, D\. de las Casas, F\. Bressand, G\. Lengyel, G\. Lample, L\. Saulnier, L\. Renard Lavaud, M\. Lachaux, P\. Stock, T\. Le Scao, T\. Lavril, T\. Wang, T\. Lacroix, and W\. El Sayed \(2023\)Mistral 7b\.External Links:2310\.06825,[Link](https://arxiv.org/abs/2310.06825)Cited by:[6th item](https://arxiv.org/html/2609.05439#A4.I1.i6.p1.1)\.
- S\. Kim, J\. Shin, Y\. Cho, J\. Jang, S\. Longpre, H\. Lee, S\. Yun, S\. Shin, S\. Kim, J\. Thorne, and M\. Seo \(2024\)Prometheus: inducing fine\-grained evaluation capability in language models\.InICLR,Cited by:[Appendix A](https://arxiv.org/html/2609.05439#A1.SS0.SSS0.Px2.p1.1),[Appendix A](https://arxiv.org/html/2609.05439#A1.SS0.SSS0.Px3.p1.1),[Appendix F](https://arxiv.org/html/2609.05439#A6.p1.1),[§1](https://arxiv.org/html/2609.05439#S1.p2.1),[§1](https://arxiv.org/html/2609.05439#S1.p3.1)\.
- J\. Li, E\. Beeching, L\. Tunstall, B\. Lipkin, R\. Soletskyi, S\. C\. Huang, K\. Rasul, L\. Yu, A\. Jiang, Z\. Shen, Z\. Qin, B\. Dong, L\. Zhou, Y\. Fleureau, G\. Lample, and S\. Polu \(2024a\)NuminaMath\.Note:Hugging Face repositoryExternal Links:[Link](https://huggingface.co/AI-MO/NuminaMath-CoT)Cited by:[§2](https://arxiv.org/html/2609.05439#S2.SS0.SSS0.Px1.p1.1)\.
- J\. Li, S\. Sun, W\. Yuan, R\. Fan, H\. Zhao, and P\. Liu \(2024b\)Generative judge for evaluating alignment\.InICLR,Cited by:[Appendix A](https://arxiv.org/html/2609.05439#A1.SS0.SSS0.Px2.p1.1),[§1](https://arxiv.org/html/2609.05439#S1.p2.1)\.
- S\. Li, J\. Li, Y\. Qu, X\. Shi, Y\. Guo, Z\. He, Y\. Wang, and W\. Tan \(2025\)Semantic\-Eval: a semantic comprehension evaluation framework for large language models generation without training\.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\. 9675–9690\.External Links:[Document](https://dx.doi.org/10.18653/v1/2025.acl-long.477),[Link](https://aclanthology.org/2025.acl-long.477/),ISBN 979\-8\-89176\-251\-0Cited by:[§4\.1](https://arxiv.org/html/2609.05439#S4.SS1.SSS0.Px4.p3.6)\.
- R\. Likert \(1932\)A technique for the measurement of attitudes\.Archives of Psychology,New York\.Cited by:[§4\.1](https://arxiv.org/html/2609.05439#S4.SS1.SSS0.Px2.p1.2)\.
- Y\. Liu, D\. Iter, Y\. Xu, S\. Wang, R\. Xu, and C\. Zhu \(2023\)G\-Eval: NLG evaluation using GPT\-4 with better human alignment\.InEMNLP,Cited by:[Appendix A](https://arxiv.org/html/2609.05439#A1.SS0.SSS0.Px2.p1.1),[§1](https://arxiv.org/html/2609.05439#S1.p2.1)\.
- A\. Madaan, N\. Tandon, P\. Gupta, S\. Hallinan, L\. Gao, S\. Wiegreffe, U\. Alon, N\. Dziri, S\. Prabhumoye, Y\. Yang, S\. Gupta, B\. P\. Maurer, P\. Clark, S\. Welleck, and Y\. Yang \(2023\)Self\-refine: iterative refinement with self\-feedback\.InNeurIPS,Cited by:[Appendix A](https://arxiv.org/html/2609.05439#A1.SS0.SSS0.Px4.p1.1)\.
- Mistral AI \(2024\)Mistral\-7b\-instruct\-v0\.3\.Note:[https://huggingface\.co/mistralai/Mistral\-7B\-Instruct\-v0\.3](https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.3)Hugging Face model card\. Accessed: 2026\-05\-26Cited by:[6th item](https://arxiv.org/html/2609.05439#A4.I1.i6.p1.1)\.
- Mistral AI \(2025\)Ministral 3 14b\.Note:[https://docs\.mistral\.ai/models/model\-cards/ministral\-3\-14b\-25\-12](https://docs.mistral.ai/models/model-cards/ministral-3-14b-25-12)Official model card\. Accessed: 2026\-05\-26Cited by:[7th item](https://arxiv.org/html/2609.05439#A4.I1.i7.p1.1)\.
- OpenAI \(2025a\)GPT\-5 system card\.Note:[https://openai\.com/index/gpt\-5\-system\-card/](https://openai.com/index/gpt-5-system-card/)Accessed: 2026\-05\-26Cited by:[8th item](https://arxiv.org/html/2609.05439#A4.I1.i8.p1.1)\.
- OpenAI \(2025b\)Introducing gpt\-5\.Note:[https://openai\.com/index/introducing\-gpt\-5/](https://openai.com/index/introducing-gpt-5/)Accessed: 2026\-05\-26Cited by:[8th item](https://arxiv.org/html/2609.05439#A4.I1.i8.p1.1),[§2](https://arxiv.org/html/2609.05439#S2.SS0.SSS0.Px2.p1.3)\.
- L\. Ouyang, J\. Wu, X\. Jiang, D\. Almeida, C\. L\. Wainwright, P\. Mishkin, C\. Zhang, S\. Agarwal, K\. Slama, A\. Ray, J\. Schulman, J\. Hilton, F\. Kelton, L\. Miller, M\. Simens, A\. Askell, P\. Welinder, P\. Christiano, J\. Leike, and R\. Lowe \(2022\)Training language models to follow instructions with human feedback\.NeurIPS\.Cited by:[Appendix A](https://arxiv.org/html/2609.05439#A1.SS0.SSS0.Px3.p1.1),[§1](https://arxiv.org/html/2609.05439#S1.p1.1)\.
- K\. Papineni, S\. Roukos, T\. Ward, and W\. Zhu \(2002\)BLEU: a method for automatic evaluation of machine translation\.InACL,Cited by:[§1](https://arxiv.org/html/2609.05439#S1.p2.1)\.
- Qwen Team \(2026a\)Qwen3\.5\-4b\.Note:[https://huggingface\.co/Qwen/Qwen3\.5\-4B](https://huggingface.co/Qwen/Qwen3.5-4B)Hugging Face model card\. Accessed: 2026\-05\-26Cited by:[4th item](https://arxiv.org/html/2609.05439#A4.I1.i4.p1.1),[Appendix D](https://arxiv.org/html/2609.05439#A4.p3.1)\.
- Qwen Team \(2026b\)Qwen3\.5\-9b\.Note:[https://huggingface\.co/Qwen/Qwen3\.5\-9B](https://huggingface.co/Qwen/Qwen3.5-9B)Hugging Face model card\. Accessed: 2026\-05\-26Cited by:[5th item](https://arxiv.org/html/2609.05439#A4.I1.i5.p1.1),[Appendix D](https://arxiv.org/html/2609.05439#A4.p3.1)\.
- J\. Schulman, F\. Wolski, P\. Dhariwal, A\. Radford, and O\. Klimov \(2017\)Proximal policy optimization algorithms\.InICLR,Cited by:[Appendix A](https://arxiv.org/html/2609.05439#A1.SS0.SSS0.Px3.p1.1),[§1](https://arxiv.org/html/2609.05439#S1.p3.1)\.
- T\. Sellam, D\. Das, and A\. Parikh \(2020\)BLEURT: learning robust metrics for text generation\.InProceedings of the 58th Annual Meeting of the Association for Computational Linguistics,D\. Jurafsky, J\. Chai, N\. Schluter, and J\. Tetreault \(Eds\.\),Online,pp\. 7881–7892\.External Links:[Document](https://dx.doi.org/10.18653/v1/2020.acl-main.704),[Link](https://aclanthology.org/2020.acl-main.704/)Cited by:[§4\.1](https://arxiv.org/html/2609.05439#S4.SS1.SSS0.Px4.p3.6)\.
- N\. Shinn, F\. Cassano, A\. Gopinath, K\. Narasimhan, and S\. Yao \(2023\)Reflexion: language agents with verbal reinforcement learning\.InNeurIPS,Cited by:[Appendix A](https://arxiv.org/html/2609.05439#A1.SS0.SSS0.Px4.p1.1)\.
- T\. Wang, P\. Yu, X\. E\. Tan, S\. O’Brien, R\. Pasunuru, J\. Dwivedi\-Yu, O\. Golovneva, L\. Zettlemoyer, M\. Fazel\-Zarandi, and A\. Celikyilmaz \(2023\)Shepherd: a critic for language model generation\.arXiv preprint arXiv:2308\.04592\.Cited by:[Appendix A](https://arxiv.org/html/2609.05439#A1.SS0.SSS0.Px2.p1.1)\.
- X\. Wang, C\. Li, J\. Yang, K\. Zhang, B\. Liu, T\. Xiong, and F\. Huang \(2025\)LLaVA\-critic\-R1: your critic model is secretly a strong policy model\.arXiv preprint arXiv:2509\.00676\.Cited by:[Appendix A](https://arxiv.org/html/2609.05439#A1.SS0.SSS0.Px4.p1.1),[§1](https://arxiv.org/html/2609.05439#S1.p3.1)\.
- Y\. Wang, Z\. Yu, W\. Yao, Z\. Zeng, L\. Yang, C\. Wang, H\. Chen, C\. Jiang, R\. Xie, J\. Wang, X\. Xie, W\. Ye, S\. Zhang, and Y\. Zhang \(2024\)PandaLM: an automatic evaluation benchmark for LLM instruction tuning optimization\.InICLR,Cited by:[Appendix A](https://arxiv.org/html/2609.05439#A1.SS0.SSS0.Px2.p1.1)\.
- J\. Wei, X\. Wang, D\. Schuurmans, M\. Bosma, F\. Xia, E\. Chi, Q\. V\. Le, D\. Zhou,et al\.\(2022\)Chain\-of\-thought prompting elicits reasoning in large language models\.NeurIPS\.Cited by:[Appendix A](https://arxiv.org/html/2609.05439#A1.SS0.SSS0.Px1.p1.1),[§1](https://arxiv.org/html/2609.05439#S1.p1.1)\.
- T\. Xiong, X\. Wang, D\. Guo, Q\. Ye, H\. Fan, Q\. Gu, H\. Huang, and C\. Li \(2025\)LLaVA\-critic: learning to evaluate multimodal models\.InCVPR,Cited by:[Appendix A](https://arxiv.org/html/2609.05439#A1.SS0.SSS0.Px3.p1.1),[§1](https://arxiv.org/html/2609.05439#S1.p3.1)\.
- A\. Yang, B\. Yang, B\. Zhang, B\. Hui, B\. Zheng, B\. Yu, C\. Li, D\. Liu, F\. Huang, H\. Wei,et al\.\(2025\)Qwen2\.5 technical report\.External Links:2412\.15115,[Link](https://arxiv.org/abs/2412.15115)Cited by:[1st item](https://arxiv.org/html/2609.05439#A4.I1.i1.p1.1),[2nd item](https://arxiv.org/html/2609.05439#A4.I1.i2.p1.1),[3rd item](https://arxiv.org/html/2609.05439#A4.I1.i3.p1.1),[Appendix D](https://arxiv.org/html/2609.05439#A4.p3.1)\.
- S\. Ye, D\. Kim, S\. Kim, H\. Hwang, S\. Kim, Y\. Jo, J\. Thorne, J\. Kim, and M\. Seo \(2024\)FLASK: fine\-grained language model evaluation based on alignment skill sets\.InICLR,Cited by:[Appendix A](https://arxiv.org/html/2609.05439#A1.SS0.SSS0.Px2.p1.1),[§1](https://arxiv.org/html/2609.05439#S1.p2.1)\.
- Y\. Ye, Z\. Huang, Y\. Xiao, E\. Chern, S\. Xia, and P\. Liu \(2025\)LIMO: less is more for reasoning\.arXiv preprint arXiv:2502\.03387\.Cited by:[§2](https://arxiv.org/html/2609.05439#S2.SS0.SSS0.Px1.p1.1)\.
- W\. Yuan, J\. Yu, S\. Jiang, K\. Padthe, Y\. Li, D\. Wang, I\. Kulikov, K\. Cho, Y\. Tian, J\. E\. Weston, and X\. Li \(2025\)NaturalReasoning: reasoning in the wild with 2\.8m challenging questions\.arXiv preprint arXiv:2502\.13124\.Cited by:[§2](https://arxiv.org/html/2609.05439#S2.SS0.SSS0.Px1.p1.1)\.
- T\. Zhang, V\. Kishore, F\. Wu, K\. Q\. Weinberger, and Y\. Artzi \(2020\)BERTScore: evaluating text generation with BERT\.InInternational Conference on Learning Representations,External Links:[Link](https://openreview.net/forum?id=SkeHuCVFDr)Cited by:[§4\.1](https://arxiv.org/html/2609.05439#S4.SS1.SSS0.Px4.p3.6)\.
- L\. Zheng, W\. Chiang, Y\. Sheng, S\. Zhuang, Z\. Wu, Y\. Zhuang, Z\. Lin, Z\. Li, D\. Li, E\. P\. Xing, H\. Zhang, J\. E\. Gonzalez, and I\. Stoica \(2023\)Judging LLM\-as\-a\-judge with MT\-bench and chatbot arena\.InNeurIPS,Cited by:[Appendix A](https://arxiv.org/html/2609.05439#A1.SS0.SSS0.Px2.p1.1),[§1](https://arxiv.org/html/2609.05439#S1.p2.1)\.
- X\. Zheng, J\. Lou, B\. Cao, X\. Wen, Y\. Ji, H\. Lin, Y\. Lu, X\. Han, D\. Zhang, and L\. Sun \(2025\)Critic\-CoT: boosting the reasoning abilities of large language model via chain\-of\-thought critic\.InFindings of ACL,Cited by:[Appendix A](https://arxiv.org/html/2609.05439#A1.SS0.SSS0.Px4.p1.1),[§1](https://arxiv.org/html/2609.05439#S1.p3.1)\.
- D\. M\. Ziegler, N\. Stiennon, J\. Wu, T\. B\. Brown, A\. Radford, D\. Amodei, P\. Christiano, and G\. Irving \(2019\)Fine\-tuning language models from human preferences\.arXiv preprint arXiv:1909\.08593\.Cited by:[Appendix A](https://arxiv.org/html/2609.05439#A1.SS0.SSS0.Px3.p1.1)\.

## Appendix ARelated Work

#### Reasoning Evaluation Beyond Final Answers\.

Chain\-of\-thought prompting improves mathematical and multi\-step reasoning by eliciting intermediate steps\(Weiet al\.,[2022](https://arxiv.org/html/2609.05439#bib.bib1); Cobbeet al\.,[2021](https://arxiv.org/html/2609.05439#bib.bib2)\), but also complicates evaluation: correct answers may still contain incomplete, unclear, or constraint\-inconsistent reasoning\. Our work evaluates and revises reasoning under explicit criteria, rather than merely encouraging longer chains\.

#### LLM Judges and Fine\-Grained Criteria\.

LLM\-as\-a\-judge methods scale automatic evaluation by prompting or training models to compare responses and assign ratings\(Zhenget al\.,[2023](https://arxiv.org/html/2609.05439#bib.bib7); Liuet al\.,[2023](https://arxiv.org/html/2609.05439#bib.bib9); Wanget al\.,[2024](https://arxiv.org/html/2609.05439#bib.bib8)\)\. Fine\-grained evaluators decompose quality into skills or criteria\(Yeet al\.,[2024](https://arxiv.org/html/2609.05439#bib.bib10); Kimet al\.,[2024](https://arxiv.org/html/2609.05439#bib.bib11)\), and generative judges further produce critiques or feedback\(Liet al\.,[2024b](https://arxiv.org/html/2609.05439#bib.bib15); Wanget al\.,[2023](https://arxiv.org/html/2609.05439#bib.bib16)\)\. However, their criteria are often task\- or scenario\-level, and their outputs usually remain scores, critiques, or rewards\. CriticGen instead induces a sample\-specific rubric for each question–answer pair and conditions both evaluation and refinement on it\.

#### Evaluation as Reward for Model Improvement\.

Evaluators are also used as reward models or selection functions for improving separate policy models\. Preference modeling and RLHF\-style methods learn reward signals from human or model feedback\(Ziegleret al\.,[2019](https://arxiv.org/html/2609.05439#bib.bib23); Ouyanget al\.,[2022](https://arxiv.org/html/2609.05439#bib.bib19); Schulmanet al\.,[2017](https://arxiv.org/html/2609.05439#bib.bib20)\), while Prometheus and LLaVA\-Critic provide fine\-grained feedback for reward\-guided selection or reinforcement learning\(Kimet al\.,[2024](https://arxiv.org/html/2609.05439#bib.bib11); Xionget al\.,[2025](https://arxiv.org/html/2609.05439#bib.bib12)\)\. In these methods, evaluation and generation are typically separated\. CriticGen studies a more direct path: the model scores the current answer, diagnoses weaknesses, generates an executable refinement suggestion, and rewrites the answer under the same rubric\.

#### Self\-Critique and Critic\-Based Generation\.

Self\-refinement and Reflexion show that model\-generated feedback can improve later outputs through critique and revision\(Madaanet al\.,[2023](https://arxiv.org/html/2609.05439#bib.bib17); Shinnet al\.,[2023](https://arxiv.org/html/2609.05439#bib.bib18)\)\. Critic\-CoT trains step\-wise critiques for reasoning errors\(Zhenget al\.,[2025](https://arxiv.org/html/2609.05439#bib.bib13)\), and LLaVA\-Critic\-R1 shows that critic training can yield strong policy behavior after reinforcement learning on critic data\(Wanget al\.,[2025](https://arxiv.org/html/2609.05439#bib.bib14)\)\. These works validate critic signals, while CriticGen focuses on converting dynamically induced, fine\-grained rubrics into directly executable refinement signals by grounding the score, reason, suggestion, and rewrite in the same instance\-specific criteria\.

## Appendix BData Filtering

To ensure that the constructed data are not dominated by templated or repetitive generations, we conduct lexical\-diversity analysis and near\-duplicate filtering during data construction\. This process is applied after score\-consensus filtering and before the final train/test split\.

#### Text normalization\.

We normalize all textual fields before computing lexical statistics\. Specifically, we lowercase English text, remove redundant whitespace, and tokenize each field into word\-level tokens\. For each instance, we consider both the input\-side fields and the generated fields in the constructed data\. The input\-side fields include the question and the original answer\. The generated fields include the induced rubric, the evaluation reason, the executable refinement suggestion, and the refined answer\. We analyze these fields separately because they play different roles in the constructed evaluation\-to\-refinement trajectory\.

#### Field\-level lexical diversity\.

We first measure field\-level lexical diversity using Distinct\-nn, defined as the ratio of uniquenn\-grams to the total number ofnn\-grams in a field collection:

Distinct​\-​n=\|UniqueNgramsn​\(𝒯\)\|\|AllNgramsn​\(𝒯\)\|,\\mathrm\{Distinct\}\\text\{\-\}n=\\frac\{\|\\mathrm\{UniqueNgrams\}\_\{n\}\(\\mathcal\{T\}\)\|\}\{\|\\mathrm\{AllNgrams\}\_\{n\}\(\\mathcal\{T\}\)\|\},\(8\)where𝒯\\mathcal\{T\}denotes all texts from a specific field\. Higher Distinct\-nnvalues indicate greater lexical diversity and lower template reuse\. We compute Distinct\-2, Distinct\-3, and Distinct\-4 for questions, answers, rubrics, reasons, executable suggestions, and refined answers\. Table[7](https://arxiv.org/html/2609.05439#A2.T7)reports the field\-level diversity statistics\.

Table 7:Field\-level lexical diversity of the constructed data\. The generated evaluation\-to\-refinement fields, especially reasons, executable suggestions, and refined answers, show substantial high\-ordernn\-gram diversity\. Questions exhibit lower diversity because they are more task\-driven and often follow similar query forms\. Overall, the results suggest that the constructed data are not dominated by repeated lexical templates\.
#### Sample\-level near\-duplicate filtering\.

Field\-level diversity measures the overall lexical variation of the dataset, but does not directly identify near\-duplicate instances\. We therefore apply a sample\-level overlap filter\. For each candidate instance, we concatenate its generated fields, including the rubric, reason, executable suggestion, and refined answer, into a single text sequence\. We then compute the 4\-gram Jaccard overlap between the candidate and previously retained instances:

J4​\(xi,xj\)=\|𝒢4​\(xi\)∩𝒢4​\(xj\)\|\|𝒢4​\(xi\)∪𝒢4​\(xj\)\|,J\_\{4\}\(x\_\{i\},x\_\{j\}\)=\\frac\{\|\\mathcal\{G\}\_\{4\}\(x\_\{i\}\)\\cap\\mathcal\{G\}\_\{4\}\(x\_\{j\}\)\|\}\{\|\\mathcal\{G\}\_\{4\}\(x\_\{i\}\)\\cup\\mathcal\{G\}\_\{4\}\(x\_\{j\}\)\|\},\(9\)where𝒢4​\(x\)\\mathcal\{G\}\_\{4\}\(x\)denotes the set of 4\-grams in instancexx\. A candidate instance is removed if its maximum 4\-gram Jaccard overlap with any retained instance exceedsτ=0\.75\\tau=0\.75\. In implementation, we use MinHash\-based locality\-sensitive hashing to retrieve high\-overlap candidates efficiently, followed by exact Jaccard computation for final filtering\.

Overall, Table[7](https://arxiv.org/html/2609.05439#A2.T7)shows that the constructed data retain meaningful lexical variation across both input\-side and generated fields\. Although the questions are relatively more standardized due to the task\-oriented data construction process, the generated reasons, executable suggestions, and refined answers exhibit high\-order diversity, indicating that the evaluation\-to\-refinement trajectories are not merely repeated templates\.

## Appendix CAnswer Sensitivity of Induced Rubrics

#### Data\.

We use 180 held\-out questions with 3–5 candidate answers each\. The questions are held out from training and are used only for the answer\-sensitivity analysis\.

#### Protocol\.

To measure whether rubrics are answer\-sensitive, we compare rubrics generated for different answers to the same question\. ForOurs, we compare answer\-specific rubricsR​\(q,a\)R\(q,a\)andR​\(q,a′\)R\(q,a^\{\\prime\}\)\. ForQ\-only, we use a shared question\-only rubricR​\(q\)R\(q\)for both answers, which cannot vary with the answer by construction\. ForQ\-only×\\times2, we compare two independently generated question\-only rubrics for the same question, providing a lexical\-noise control for repeated rubric generation\. Rubric change is measured by the lexical Jaccard shift over pooleddimension\_nametokens\.

#### Results\.

We evaluate whether induced rubrics change when the answer changes\. As a lightweight, reproducible proxy, we measure*lexical*turnover indimension\_namestrings only and exclude the 0–5 criterion text, which reduces sensitivity to long paraphrases in scoring anchors\. For a rubricR​\(q,a\)R\(q,a\), we take everydimension\_name, lowercase it, tokenize it into alphanumeric word*types*, and pool those types into a single setWaW\_\{a\}\(duplicate tokens across dimensions count once\)\. Fixing the same questionqq, we induceR​\(q,a\)R\(q,a\)andR​\(q,a′\)R\(q,a^\{\\prime\}\)from distinct answersaaanda′a^\{\\prime\}, constructWaW\_\{a\}andWa′W\_\{a^\{\\prime\}\}, and compute the Jaccard distance

DJaccard=1−\|Wa∩Wa′\|\|Wa∪Wa′\|\.D\_\{\\mathrm\{Jaccard\}\}=1\-\\frac\{\|W\_\{a\}\\cap W\_\{a^\{\\prime\}\}\|\}\{\|W\_\{a\}\\cup W\_\{a^\{\\prime\}\}\|\}\.\(10\)The distance is 0 if both token sets are empty\. LargerDJaccardD\_\{\\mathrm\{Jaccard\}\}means less overlap in the word types appearing anywhere in dimension titles when the conditioning answer changes\. Table[8](https://arxiv.org/html/2609.05439#A3.T8)summarizes means on 180 held\-out questions\.Q\-onlyuses the same question\-only rubric for both answers, so its shift is zero by design\.Q\-only×\\times2compares two independently generated question\-only rubrics and shows a small shift \(0\.09\), reflecting lexical variation from repeated generation\.OurscomparesR​\(q,a\)R\(q,a\)andR​\(q,a′\)R\(q,a^\{\\prime\}\)and achieves a much larger shift of 0\.37\. This confirms a necessary property of sample\-specific evaluation: the rubric changes with the answer, not just the question\. Together with the human evaluation in Section[4\.2](https://arxiv.org/html/2609.05439#S4.SS2), these results indicate that our rubrics are both more relevant to each\(q,a\)\(q,a\)and more sensitive to answer\-specific evaluation needs\.

Table 8:Answer sensitivity of induced rubrics on 180 held\-out questions\.*Q\-only*uses a shared question\-level rubric,*Q\-only×\\times2*compares two independently generated question\-level rubrics, and*Ours*compares answer\-conditioned rubrics for different answers to the same question\.

## Appendix DBaseline Models and CriticGen Variants

We compare CriticGen with a diverse set of instruction\-following LLM baselines\. Unless otherwise specified, all baseline models are evaluated under the same rubric\-conditioned prompts and output format, and are not fine\-tuned on our constructed data\.

- •Qwen2\.5\-7B\-Instruct\.Qwen2\.5\-7B\-Instruct serves as the primary open\-weight backbone in our experiments\(Yanget al\.,[2025](https://arxiv.org/html/2609.05439#bib.bib30)\)\. It provides a strong 7B\-scale instruction\-following baseline and is used to assess the zero\-shot or prompt\-only capability of a model with the same architecture scale as our main CriticGen variant\.
- •Qwen2\.5\-14B\-Instruct\.We include Qwen2\.5\-14B\-Instruct as a medium\-scale Qwen2\.5 baseline\(Yanget al\.,[2025](https://arxiv.org/html/2609.05439#bib.bib30)\)\. Comparing it with Qwen2\.5\-7B\-Instruct and Qwen2\.5\-32B\-Instruct allows us to examine whether a larger model capacity alone improves rubric\-conditioned evaluation and refinement\.
- •Qwen2\.5\-32B\-Instruct\.Qwen2\.5\-32B\-Instruct is used as a larger Qwen2\.5\-family baseline\(Yanget al\.,[2025](https://arxiv.org/html/2609.05439#bib.bib30)\)\. It provides a scale\-controlled comparison for evaluating whether CriticGen’s gains can be attributed to task\-specific supervision rather than simply to increasing the number of model parameters\.
- •Qwen3\.5\-4B\.We include Qwen3\.5\-4B as a newer small\-scale Qwen\-series baseline\(Qwen Team,[2026a](https://arxiv.org/html/2609.05439#bib.bib31)\)\. It allows us to test whether the recent base\-model improvements can already support rubric\-conditioned evaluation without task\-specific fine\-tuning\.
- •Qwen3\.5\-9B\.Qwen3\.5\-9B is used as a newer medium\-scale Qwen\-series baseline\(Qwen Team,[2026b](https://arxiv.org/html/2609.05439#bib.bib32)\)\. Together with Qwen3\.5\-4B, it helps evaluate the robustness of CriticGen’s training strategy across different Qwen generations and parameter scales\.
- •Mistral\-7B\-Instruct\-v0\.3\.We use Mistral\-7B\-Instruct\-v0\.3 as a non\-Qwen open\-weight instruction\-following baseline\(Jianget al\.,[2023](https://arxiv.org/html/2609.05439#bib.bib33); Mistral AI,[2024](https://arxiv.org/html/2609.05439#bib.bib34)\)\. Its inclusion helps assess whether CriticGen’s advantages are specific to the Qwen model family or remain competitive against independently developed open\-weight instruction models\.
- •Ministral\-3\-14B\-Instruct\-2512\.We include Ministral\-3\-14B\-Instruct\-2512 as a recent non\-Qwen instruction\-following baseline\(Mistral AI,[2025](https://arxiv.org/html/2609.05439#bib.bib35)\)\. It provides an additional comparison point for evaluating CriticGen against contemporary open\-weight instruction models\.
- •GPT\-5\.GPT\-5 is included as a strong proprietary model baseline\(OpenAI,[2025b](https://arxiv.org/html/2609.05439#bib.bib36),[a](https://arxiv.org/html/2609.05439#bib.bib37)\)\. In addition to serving as a high\-capability comparison model, it is also used as a reference judge in parts of our data construction and verification pipeline\.

Our CriticGen variants are obtained by applying the same supervised fine\-tuning procedure to select Qwen backbones using our constructed training data\. Specifically, CriticGen\-Qwen2\.5\-7B\-SFT, CriticGen\-Qwen3\.5\-4B\-SFT, and CriticGen\-Qwen3\.5\-9B\-SFT are initialized from Qwen2\.5\-7B\-Instruct\(Yanget al\.,[2025](https://arxiv.org/html/2609.05439#bib.bib30)\), Qwen3\.5\-4B\(Qwen Team,[2026a](https://arxiv.org/html/2609.05439#bib.bib31)\), and Qwen3\.5\-9B\(Qwen Team,[2026b](https://arxiv.org/html/2609.05439#bib.bib32)\), respectively, and trained onRubricDataandRefineData\. This design isolates the effect of CriticGen’s data and learning objective from the choice of backbone: the base instruction models measure zero\-shot or prompt\-only rubric\-conditioned evaluation ability, while the CriticGen variants measure the effect of learning to generate sample\-specific rubrics and to produce the coupled score–reason–revision\-suggestion–rewrite trajectory from our supervision\.

## Appendix ETraining Details

We train CriticGen variants with supervised fine\-tuning \(SFT\)\. All training runs are initialized from the corresponding instruction\-tuned backbone and use the same serialized output format as described in Section[3](https://arxiv.org/html/2609.05439#S3)\. For rubric induction, the model is trained to generate the complete set of induced rubric items, including dimension names and 0–5 scoring criteria\. For rubric\-conditioned evaluation and refinement, the model is trained to generate the complete target sequence, including the score, reason, executable revision suggestion, and refined answer\. We optimize the standard next\-token prediction loss with AdamW, using a learning rate of2×10−52\\times 10^\{\-5\}, a cosine learning\-rate schedule, a warmup ratio of 0\.03, and weight decay of 0\.01\. Unless otherwise specified, we train for 3 epochs with bf16 precision, gradient checkpointing, and a maximum sequence length of 4096 tokens\. The effective batch size is set to 128 through gradient accumulation\.

We choose the SFT checkpoint as the main CriticGen model in all reported experiments\. This choice keeps the empirical comparison focused on the proposed data construction and evaluation\-to\-refinement formulation\. It also avoids introducing an additional preference\-optimization stage whose gains may depend on preference\-pair construction, verifier calibration, and reward\-model noise rather than on the core CriticGen framework\.

## Appendix FOut\-of\-Domain Evaluation on Feedback Bench

To further examine the out\-of\-domain generalization ability of CriticGen, we evaluate it on Feedback Bench from Prometheus\(Kimet al\.,[2024](https://arxiv.org/html/2609.05439#bib.bib11)\)\. Feedback Bench is designed for absolute grading with customized score rubrics\. Each instance provides an instruction, a response to be evaluated, a reference answer, and a fine\-grained scoring rubric, and the evaluator LM is required to assign a scalar score and feedback according to the given rubric\.

Different from our in\-domainRefineData, Feedback Bench is not used during training and follows an independently constructed evaluation format\. We therefore use it as an out\-of\-domain benchmark to test whether the evaluation capability learned by CriticGen transfers to external rubric\-conditioned scoring scenarios\. To ensure a fair comparison, we compare CriticGen\-Qwen2\.5\-7B\-SFT with evaluator LMs of a similar scale, including Qwen2\.5\-7B\-Instruct, Qwen3\.5\-4B, and Mistral\-7B\-Instruct\-v0\.3\.

Following the evaluation protocol of Prometheus, we measure the correlation between model\-predicted scores and reference scores using Pearson correlation, Spearman correlation, and Kendall’sτ\\tau\. The results are shown in Table[9](https://arxiv.org/html/2609.05439#A6.T9)\.

Table 9:Out\-of\-domain evaluation on Feedback Bench\. We report Pearson, Spearman, and Kendall’sτ\\taucorrelations between model\-predicted scores and reference scores\.CriticGen\-Qwen2\.5\-7B\-SFT achieves the best performance across all three correlation metrics\. Compared with its base model Qwen2\.5\-7B\-Instruct, CriticGen improves Pearson correlation from 0\.7520 to 0\.7556, Spearman correlation from 0\.7522 to 0\.7612, and Kendall’sτ\\taufrom 0\.6606 to 0\.6755\. Although the improvement is modest, the consistent gains across all metrics suggest that CriticGen’s rubric\-conditioned evaluation capability can transfer to out\-of\-domain customized scoring settings\.