PlanE: Meta Planning of Data, Tuning, and Inference for Extractive-based LLMs

# Meta Planning of Data, Tuning, and Inference for Extractive-based LLMs
Source: [https://arxiv.org/html/2607.20470](https://arxiv.org/html/2607.20470)
Jiacheng Wang♢,Weiyan Zhang♠11footnotemark:1,Guangya Yu♠ ♢Ant Group, Shanghai, China, ♠School of Information Science and Engineering, East China University of Science and Technology, Shanghai, China Correspondence:[xueqiao\.wjc@antgroup\.com](https://arxiv.org/html/2607.20470v1/[email protected]),[weiyanzhang@ecust\.edu\.cn](https://arxiv.org/html/2607.20470v1/[email protected])

###### Abstract

Enhancing the task\-specific capabilities of Large Language Models \(LLMs\) primarily requires substantial instruction\-tuning datasets\. However, the sheer volume of such data imposes a considerable annotation cost, and a lack of optimization methods for tailoring LLMs to specific tasks\. To address the above issues, we propose aPlanning framework for constructingExtractive\-based LLMs calledPlanE, which includes data decomposition, instruction tuning, and prompt inference\. Additionally, we introduce a Data\-Tuning\-Inference \(DTI\) planner, aimed at selecting the optimal base\-LLM and its DTI combinations for specific datasets to improve construction efficiency\. The experimental results demonstrate the effectiveness of our PlanE from two views: \(1\) across different datasets using the same base\-LLM, and \(2\) on the same dataset using different base\-LLMs\. Furthermore, we validate the generalizability of the proposed DTI planner under different optimization objectives\. The codes are publicly available at[https://github\.com/gugugu\-469/PlanE](https://github.com/gugugu-469/PlanE)\.

PlanE: Meta Planning of Data, Tuning, and Inference for Extractive\-based LLMs

Jiacheng Wang♢††thanks:Equal Contribution\., Weiyan Zhang♠11footnotemark:1††thanks:Corresponding Authors\., Guangya Yu♠♢Ant Group, Shanghai, China,♠School of Information Science and Engineering, East China Universityof Science and Technology, Shanghai, ChinaCorrespondence:[xueqiao\.wjc@antgroup\.com](https://arxiv.org/html/2607.20470v1/[email protected]),[weiyanzhang@ecust\.edu\.cn](https://arxiv.org/html/2607.20470v1/[email protected])

![Refer to caption](https://arxiv.org/html/2607.20470v1/x1.png)Figure 1:\(a\) Three key factors in building task\-specific LLMs\. \(b\) For the same base\-LLM, the optimal performance differs across different task datasets\. \(c\) On the same dataset, different base\-LLMs reach their best performance through different combinations of DTI\.## 1Introduction

LLMs have demonstrated prowess in producing human\-aligned responses to varied instructions\. A pivotal technique for enhancing the task\-specific capabilities of LLMs is instruction tuning, which aligns the model with human preferences using data in the form of instruction\-response pairs\.

Based on the stage of LLM optimization, these works can be categorized into three approaches\. First, data refinement employs data augmentation and data structure transformation to improve LLM performanceDinget al\.\([2024](https://arxiv.org/html/2607.20470#bib.bib44)\); Changet al\.\([2024](https://arxiv.org/html/2607.20470#bib.bib45)\)\. However, these methods are often constrained by data qualityLiet al\.\([2024a](https://arxiv.org/html/2607.20470#bib.bib43)\)\. Second, multi\-task tuning trains LLMs on multiple information extraction \(IE\) tasks simultaneously to improve generalizationAbdelazizet al\.\([2024](https://arxiv.org/html/2607.20470#bib.bib46)\)\. However, such methods do not significantly boost the performance of the specific task and come with high tuning costsBriefet al\.\([2024](https://arxiv.org/html/2607.20470#bib.bib47)\)\. Third, retrieval\-augmented generation integrates retrieval mechanisms with LLMs to enhance task performance through external knowledge bases \(KBs\)Wanget al\.\([2025](https://arxiv.org/html/2607.20470#bib.bib48)\); Suet al\.\([2024](https://arxiv.org/html/2607.20470#bib.bib49)\)\. However, this approach is limited by the challenges associated with building specialized KBsLiet al\.\([2024b](https://arxiv.org/html/2607.20470#bib.bib50)\)\.

Current research primarily focuses on the use of LLM technology, but there is a lack of studies on constructing the LLM for IE\-specific tasks\. When building an IE\-specific LLM, it is necessary to consider the optimization of the entire process, including data, tuning, and inference, as shown in Figure[1](https://arxiv.org/html/2607.20470#S0.F1)\(a\): 1\)Data\. Given the limited training data, constructing high\-quality instruction\-tuning data is a significant challenge\. 2\)Tuning\. For training data with different data structures, determining the best way to combine tuning strategies to maximize the potential of LLMs is key\. 3\)Inference\. For a tuned LLM, designing prompts to fully leverage its reasoning capabilities remains another critical issue\. Furthermore, the integration and joint optimization of these three stages into a comprehensive solution is also a major difficulty\. This is because, as shown in Figure[1](https://arxiv.org/html/2607.20470#S0.F1)\(b\), even when using the same three Data\-Tuning\-Inference \(DTI\) combinations, the optimal performance varies across different task datasets for the same base\-LLM\. Similarly, as shown in Figure[1](https://arxiv.org/html/2607.20470#S0.F1)\(c\), even on the same dataset, different base\-LLMs achieve their best performance with different DTI combinations, despite using the same three DTI combinations\.

In this paper, we propose a novel framework for selecting combinations to construct extractive\-based LLMs, calledPlanE\. The core idea is to design optimization schemes for the three key stages of the extractive\-based LLM construction process \(i\.e\., data, tuning, inference\), and propose a combination selection function to determine the optimal construction strategy\. Specifically, for the data, we introduce two data decompositions: Pipeline\-based and Bidirectional\. These schemes decompose the complex task into a combination of serialized atomic subtasks\. Then, for tuning, we adopt instruction tuning\. Based on the decomposed data structure of subtasks, we employ Supervised Fine\-Tuning \(SFT\), and a combination of SFT with Reinforcement Learning \(SFT\+RL\)\. Finally, for inference, we adopt prompt inference, including three inference strategies: direct inference, intersection inference, and union inference\. In addition, we design a DTI planner that predicts the optimal DTI combination for given task datasets and base\-LLMs through meta\-learning on empirical data, thereby optimizing extractive\-based LLM construction\.

The contributions of this paper are as follows:

- •We design a novel DTI planner that selects the optimal DTI combinations for the extractive\-based LLM\. Besides, this planner also performs multi\-objective optimization, balancing both the performance and efficiency of building the extractive\-based LLM\.
- •We propose the PlanE framework, which enhances high\-performance LLM construction through three key modules\. From a holistic optimization perspective, it determines the appropriate DTI combination based on the specified base\-LLM and dataset\.
- •We conduct extensive experiments on three task datasets\. The results indicate the effectiveness of our approach from two views: across different task datasets using the same base\-LLM, and on the same task dataset using different base\-LLMs\. In addition, we validate the generalizability of the proposed DTI planner under different optimization objectives\.

![Refer to caption](https://arxiv.org/html/2607.20470v1/x2.png)Figure 2:The framework of PlanE\.
## 2Related Work

Recent work in meta\-planning for task\-specific LLMs can be broadly categorized into several approaches: 1\) policy space design, 2\) meta\-policy learning, and 3\) optimization objectives\.

The first branch is policy space design, which defines the set of high\-level strategies or meta\-actions available to a meta\-planner\. For instance, DeepSpeed\-InferenceAminabadiet al\.\([2022](https://arxiv.org/html/2607.20470#bib.bib61)\)defines its policy space through kernel selection and parallelism configurations\. DyPlanParekhet al\.\([2025](https://arxiv.org/html/2607.20470#bib.bib63)\)constructs a dynamic policy space by composing reasoning paths \(e\.g\., Chain\-of\-Thought, Retrieval\-Augmented Generation\)\. In decentralized settings, MetaInfDuet al\.\([2025](https://arxiv.org/html/2607.20470#bib.bib64)\)leverages semantic embeddings and historical metadata to predict acceleration strategies, enabling zero\-shot generalization without online profiling\.

The second branch, meta\-policy learning, refers to constructing or learning the optimal mapping from context to actions within the policy space\. While MetaGPTHonget al\.\([2024](https://arxiv.org/html/2607.20470#bib.bib65)\)relies on handcrafted workflows, PromptAgentWanget al\.\([2024](https://arxiv.org/html/2607.20470#bib.bib67)\)employs reinforcement learning to discover dynamic reasoning and prompting strategies, effectively learning a meta\-policy\. Similarly, data\-level planning methods like DataRaterCalianet al\.\([2025](https://arxiv.org/html/2607.20470#bib.bib68)\)and GradMatchKillamsettyet al\.\([2021](https://arxiv.org/html/2607.20470#bib.bib69)\)for sample selection, and MoE\-ScalingTianet al\.\([2025](https://arxiv.org/html/2607.20470#bib.bib70)\)for expert configuration, can be viewed as specialized forms of meta\-policy learning for data and architecture management\.

The third branch, optimization objectives, defines the criteria or metrics that guide the selection and evaluation of policies within the policy space, such as latency, accuracy, cost, or efficiency\. DeepSpeed\-Inference balances latency and accuracy, while AdaCoTLouet al\.\([2025](https://arxiv.org/html/2607.20470#bib.bib66)\)uses Pareto search to navigate cost\-performance trade\-offs\. Meta\-R1Donget al\.\([2025](https://arxiv.org/html/2607.20470#bib.bib71)\)introduces root\-scaled efficiency for cross\-scale efficiency normalization, and ChinchillaHoffmannet al\.\([2022](https://arxiv.org/html/2607.20470#bib.bib72)\)provides a compute\-optimal scaling law that has been adopted in Llama 2Touvronet al\.\([2023](https://arxiv.org/html/2607.20470#bib.bib73)\)\.

Existing policy spaces are often limited to isolated aspects, lacking a unified representation that jointly covers data, training, and inference\. Meta\-policy learning remains fragmented, with some methods relying on handcrafted rules and others optimizing only local decisions\. Optimization objectives are either narrowly decoupled or based on idealized scaling laws that overlook practical multi\-objective trade\-offs\. To address these issues, this paper introduces a data\-driven planner that unifies the full configuration space into a single policy representation, learns an integrated meta\-policy via empirical performance modeling, and directly optimizes joint performance–efficiency objectives\.

## 3Overview

In this section, we first formally define three IE tasks\. Then, we outline our proposed PlanE\.

### 3\.1Task Formulation

Definition 1: Relation extraction task \(RE\)Given an input textC=\[c1,c2,…,cn\]C=\[c\_\{1\},c\_\{2\},\.\.\.,c\_\{n\}\]and a set of pre\-defined relationsℛ=\{r1,r2,…,rl\}\\mathcal\{R\}=\\\{r\_\{1\},r\_\{2\},\.\.\.,r\_\{l\}\\\}, wherennandlldenote the numbers of tokens inCCand relations inℛ\\mathcal\{R\}, respectively, the purpose of relation extraction task is to obtain a set of triplets𝒯=\{\(hi,ri,ti\)\}i=1m\\mathcal\{T\}=\\\{\(h\_\{i\},r\_\{i\},t\_\{i\}\)\\\}\_\{i=1\}^\{m\}, wherehi,tih\_\{i\},\\ t\_\{i\}represent the head and tail entities connected byrir\_\{i\}andmmdenotes the total number of extracted triplets\.

Definition 2: Event extraction task \(RE\)Given a textCe=\[c1,c2,…,cne\]C\_\{e\}=\[c\_\{1\},c\_\{2\},\.\.\.,c\_\{n\_\{e\}\}\]and the pre\-defined event typesℰ=\{e1,e2,…,ele\}\\mathcal\{E\}=\\\{e\_\{1\},e\_\{2\},\.\.\.,e\_\{l\_\{e\}\}\\\}, event extraction task aims to detect the mention event triggert​rtrand arguments𝒜=\{a1,a2,…,ame\}\\mathcal\{A\}=\\\{a\_\{1\},a\_\{2\},\.\.\.,a\_\{m\_\{e\}\}\\\}fromCeC\_\{e\}and the event typee∈ℰe\\in\\mathcal\{E\}\(e\.g\.,EAT,LOCATION, etc\)\.

Definition 3: Aspect\-based sentiment analysis task \(ABSA\)Given a textCa=\[c1,c2,…,cna\]C\_\{a\}=\[c\_\{1\},c\_\{2\},\.\.\.,c\_\{n\_\{a\}\}\], a set of pre\-defined sentiment polarities𝒮=\{s1,s2,…,sla\}\\mathcal\{S\}=\\\{s\_\{1\},s\_\{2\},\.\.\.,s\_\{l\_\{a\}\}\\\}\(e\.g\.,\{POSITIVE,NEGATIVE,NEUTRAL\}\\\{\\texttt\{POSITIVE\},\\texttt\{NEGATIVE\},\\texttt\{NEUTRAL\}\\\}\), wherenan\_\{a\}andlal\_\{a\}is the number of tokens inCaC\_\{a\}sentiment polarities in𝒮\\mathcal\{S\}, respectively, the ABSA task aims to extract a set of aspect\-sentiment triplets𝒴=\{\(a​tj,oj,sj\)\}j=1k\\mathcal\{Y\}=\\\{\(at\_\{j\},o\_\{j\},s\_\{j\}\)\\\}\_\{j=1\}^\{k\}fromCaC\_\{a\}\.kkis the number of aspect\-sentiment triplets anda​tiat\_\{i\},oio\_\{i\}denote the aspect term and opinion term inCaC\_\{a\}\.

### 3\.2Framework

As shown in Figure[2](https://arxiv.org/html/2607.20470#S1.F2), our solution is to construct an extractive\-based LLM \(e\.g\., RE\-LLM\) in two parts\.1\) Planning of Data, Tuning, and Inference\.Specifically, for the data, given the training set, we introduce two Data Decompositions: Pipeline\-based and Bidirectional\. Then, for tuning, based on the decomposed data structure of subtasks, we employ two Instruction Tuning: SFT and SFT\+RL\. Finally, for inference, based on tuned LLMs, we adopt three Prompt Inference strategies: direct inference, intersection inference, and union inference\.2\) Planner Construction\.To achieve automatic prediction of the optimal DTI combination, we design a DTI Planner\. This planner predicts the optimal DTI combination for given datasets and base\-LLMs through meta\-learning on empirical data, thereby optimizing extractive\-based LLM construction\.

## 4PlanE

This section details PlanE, including: \(1\) Planning of Data \(data decomposition\), Tuning \(instruction tuning\), and Inference \(prompt inference\), \(2\) Planner Construction, and \(3\) Planner Usage\.

### 4\.1Planning of Data, Tuning, and Inference

Data Decomposition\.This section presents a task decomposition method aimed at enhancing the performance of language models on complex tasks through data augmentation\. By decomposing a task into subtasks, we improve the precision of LLMs at each step and reduce error propagation\. We propose two decomposition strategies: pipeline\-based and bidirectional\. In pipeline\-based decomposition, the task is broken down into sequential subtasks, where the output of one subtask serves as the input for the next\. For example, in RE task, this can be implemented in two ways: first, performing entity recognition \(ER\) to extract head\-tail entity pairs, followed by relation classification \(RC\) to generate triplets; second, performing RC to determine the relation, then ER to extract the corresponding head\-tail entity pairs and output triplets\. In bidirectional decomposition, the task is approached from two directions to improve precision\. For example, in RE, first extracting the head entity, generating the relation, and then extracting the tail entity, or starting with the tail entity and proceeding accordingly\.

Instruction Tuning\.To align the model with diverse task structures and subtask objectives, we employ two tuning strategies: 1\) SFT\. All data is formatted into structured input\-output instruction pairs, and a single model is fine\-tuned to handle the RE task and its subtasks\. 2\) SFT \+ RL\. It is a two\-stage instruction tuning strategy\. First, the model is trained with SFT using the majority of the data, where all samples are formatted into structured input\-output instruction pairs to enable general task\-solving capability\. Then, RL is applied on the remaining subset to further optimize model performance through feedback\-based reward mechanisms\. Specifically, we adopt three RL\-based optimization methods, namely Group Relative Policy Optimization \(GRPO\)Shaoet al\.\([2024](https://arxiv.org/html/2607.20470#bib.bib60)\), Direct Preference Optimization \(DPO\)Rafailovet al\.\([2023](https://arxiv.org/html/2607.20470#bib.bib1)\), and Kahneman\-Tversky Optimization \(KTO\)Ethayarajhet al\.\([2024](https://arxiv.org/html/2607.20470#bib.bib2)\)\. Detailed information is listed in Appendix[H](https://arxiv.org/html/2607.20470#A8)\.

Prompt Inference\.To leverage the strengths of different decomposition strategies during inference, we propose the following methods: 1\) Direct Inference\. The model directly processes the input and generates the final output; 2\) Intersection Inference\. Aggregates results by retaining only outputs consistent across different inference chains\. 3\) Union Inference\. Combines results from multiple inference chains \(e\.g\., pipeline\-based or bidirectional\) by taking their union\. For example, the RE task is based on pipeline\-based data decomposition\. For direct inference, we directly output relation triplets from the given text\. For union/intersection inference, we first perform ER to extract head\-tail entity pairs, followed by RC to generate triplets, then perform RC to determine the relation, followed by ER to extract entity pairs\. Finally, we aggregate results at the triplet level using union or intersection\. Detailed prompts are listed in Appendix[I](https://arxiv.org/html/2607.20470#A9)\.

### 4\.2Planner Construction

The construction of the DTI planner for various objectives \(e\.g\., improve performance\) consists of five key steps:

1\.Planner definition:We define a planner that selects the optimal plan based on the task context\. The planner is trained on historical execution data, enabling it to automatically identify and select the best\-performing plan in specific scenarios\. Formally, given the historical execution data,

𝒟=\{\(tk\(i\),a\(i\),r\(i\)\)\}i=1N,\\mathcal\{D\}=\\\{\(t\_\{k\}^\{\(i\)\},a^\{\(i\)\},r^\{\(i\)\}\)\\\}\_\{i=1\}^\{N\},\(1\)wheretk\(i\)∈𝒯t\_\{k\}^\{\(i\)\}\\in\\mathcal\{T\}denotes thekk\-th task,a\(i\)∈𝒜a^\{\(i\)\}\\in\\mathcal\{A\}is the selected plan, andr\(i\)∈ℝr^\{\(i\)\}\\in\\mathbb\{R\}represents the performance feedback \(e\.g\., the F1\-score, which provides a more comprehensive measure for selecting the best combination as the harmonic mean of precision and recall\) of the plan on tasktk\(i\)t\_\{k\}^\{\(i\)\}\. The planner𝒫x\\mathcal\{P\}\_\{x\}is implemented by a parameterized ranking modelfx:𝒯×𝒜→ℝf\_\{x\}:\\mathcal\{T\}\\times\\mathcal\{A\}\\rightarrow\\mathbb\{R\}, which satisfies,

𝒫x​\(t\)=arg⁡maxa∈𝒜⁡fx​\(t,a\),\\mathcal\{P\}\_\{x\}\(t\)=\\arg\\max\_\{a\\in\\mathcal\{A\}\}f\_\{x\}\(t,a\),\(2\)meaning that it selects the plan with the highest F1 among all candidates\. The scoring functionfxf\_\{x\}is trained on𝒟\\mathcal\{D\}using a learning\-to\-max objective\.

2\.Element definition:In this paper, we explore two scenarios: planners across different datasets and planners across different models\. Accordingly,tk\(i\)t\_\{k\}^\{\(i\)\}represents either a model or a dataset\.a\(i\)a^\{\(i\)\}denotes different DTI combinations, with a total ofZZcandidate DTI combinationsC​OCO\.r\(i\)r^\{\(i\)\}represents the performance metric \(e\.g\. F1\) achieved by the model undertk\(i\)t\_\{k\}^\{\(i\)\}anda\(i\)a^\{\(i\)\}\.

3\.Identify optimal combination:We identify the combination that yields the highest F1\. This ensures that subsequent optimization is grounded in the best observed result\. Specifically, we determine the optimal combinationc​om​a​xco\_\{max\}as:

c​om​a​x=arg⁡max1≤j≤n​u​m​\(C​O\)⁡\(F​1j\)\.co\_\{max\}=\\arg\\max\_\{1\\leq j\\leq num\(CO\)\}\\;\(F1\_\{j\}\)\.\(3\)
4\.Build inequality constraints:To ensure that our planner correctly ranks the optimal DTI combination above all other candidates, we construct a set of inequality constraints comparing each non\-optimal configuration with the optimal one\. Besides, since each DTI combination consists of qualitative labels, we encode them into quantitative variables to enable numerical comparison and parameterized modeling within the DTI planner\. Specifically, we define the value ranges of three integer variables \(e\.g\., ranging from 1 to 10\)\.v​dvd,v​tvt, andv​ivirepresent Data Decompositionc​odco\_\{d\}, Instruction Tuningc​otco\_\{t\}, and Prompt Inferencec​oico\_\{i\}\. The detailed calculation methods for these variables are provided in Appendix[C](https://arxiv.org/html/2607.20470#A3)\.

In constructing the planner, we adopt a quadratic polynomial function with coefficientsso,m=\{a,b,c,d,e,f,g,h,i\}s\_\{o,m\}=\\\{a,b,c,d,e,f,g,h,i\\\}to model the nonlinear relationships between the DTI variables \(i\.e\.,v​dvd,v​tvt, andv​tvt\) and the F1\.yydenotes the predicted score of the model under a given combination, reflecting the relative effectiveness of the DTI combination:

y​\(c​o\)=a⋅v​d2\+b⋅v​t2\+c⋅v​i2\+d⋅vd⋅vt\+e⋅vd⋅vi\+f⋅v​t⋅v​i\+g⋅v​d\+h⋅v​t\+i⋅v​i\.\\begin\{split\}y\(co\)=&a\\cdot vd^\{2\}\+b\\cdot vt^\{2\}\+c\\cdot vi^\{2\}\+\\\\ &d\\cdot vd\\cdot vt\+e\\cdot vd\\cdot vi\+f\\cdot\\\\ &vt\\cdot vi\+g\\cdot vd\+h\\cdot vt\+i\\cdot vi\.\\end\{split\}\(4\)
For each non\-optimal combination, the inequality constraint is formulated as:

y​\(v​dj,v​tj,v​ij\)≤y​\(v​dm​a​x,v​tm​a​x,v​im​a​x\)\.y\(vd\_\{j\},vt\_\{j\},vi\_\{j\}\)\\leq y\(vd\_\{max\},vt\_\{max\},vi\_\{max\}\)\.\(5\)
These constraints ensure that the polynomial function assigns a higher score to the best\-performing combination during optimization, thereby guiding the model to learn a consistent and interpretable representation ofDTI effectiveness\.

5\.Solve coefficients:For eachtt\(representing a model or a dataset\), we can obtain its corresponding set of constraints based on the observed F1 performance under different DTI configurations\. We then integrate all constraints across scenarios to jointly solve for the coefficients that define a unified planner suited for the current model or dataset\. We optimize the coefficientsso,m=\{a,b,c,d,e,f,g,h,i\}s\_\{o,m\}=\\\{a,b,c,d,e,f,g,h,i\\\}of the quadratic model using gradient descent\. We define a differentiable loss function that penalizes violations of the optimality constraints\. Formally, the objective is written as:

ℒp​l​a​n​n​e​r=minso,m∑j≠maxReLU\(y\(coj\)−y\(com​a​x\)\+ε\),\\begin\{split\}\\mathcal\{L\}\_\{planner\}=&\\min\_\{s\_\{o,m\}\}\\sum\_\{j\\neq\\mathrm\{max\}\}\\mathrm\{ReLU\}\\\!\\big\(y\(co\_\{j\}\)\\\\ &\-y\(co\_\{max\}\)\+\\varepsilon\\big\),\\end\{split\}\(6\)whereε=10−10\\varepsilon=10^\{\-10\}is a small margin\. This gradient\-based method ensures the learned polynomial ranks the optimal DTI combination higher than all others\.

### 4\.3Planner Usage

Once the DTI planner is obtained for a specific scenario, it can be applied to a new model or dataset\. For example, according to the parameter calculation method \(see Appendix[C](https://arxiv.org/html/2607.20470#A3)\), only a subset of 10 representative combinations among the 28 candidates is required to infer all variable values\. These values are then fed into the planner, and the optimal DTI combination is determined by maximizing the quadratic function:

c​om​a​x=arg⁡maxc​o∈C​O⁡y​\(v​d,v​t,v​i\)\.co\_\{max\}=\\arg\\max\_\{co\\in CO\}\\;y\(vd,vt,vi\)\.\(7\)
Table 1:Statistics of all datasets\. “\#” = sample number, “Lang\.” = language, “Dom\.” = domain, “Med\.” = medical, “Gen\.” = general, “SFT” = supervised fine\-tuning, and “RL” = reinforcement learning\. The preprocessed dataset \(denoted by “\*”\) contains 7,000 random training samples from the original 14,339 to mitigate overfitting\.The optimal DTI combination usage:After obtaining the optimal DTI combination, the workflow proceeds in three stages\. First, the data decompositionc​odco\_\{d\}determines how the dataset is processed for training\. Second, the instruction tuningc​otco\_\{t\}specifies the tuning strategy used to tune the LLM\. Finally, the prompt inferencec​oico\_\{i\}determines the reasoning method applied during prediction\. The tuned\-LLM then produces its final outputs according to this inference strategy, completing the full DTI optimization\.

Table 2:Training sample statistics after data decomposition on three task datasets\. “→\\rightarrow” denotes the model’s data flow, connecting the model inputs \(left of the arrow\) to the outputs \(right of the arrow\)\. The tokens\[s​1\]\[s1\]and\[s​2\]\[s2\]serve as separators between multiple input contents\. Training scale for instruction tuning shown in Appendix[B](https://arxiv.org/html/2607.20470#A2)\.Table 3:The coefficient of the DTI planner in different objectives, where the variables’ mapping range is 1 to 10\.

## 5Experiments

### 5\.1Experimental Setup

Datasets\.We conduct the experiments on three public datasets, including CMeIE\-V2111[https://tianchi\.aliyun\.com/dataset/95414](https://tianchi.aliyun.com/dataset/95414)Zhanget al\.\([2022](https://arxiv.org/html/2607.20470#bib.bib27)\), ACE05Walkeret al\.\([2006](https://arxiv.org/html/2607.20470#bib.bib58)\), and 14LapXuet al\.\([2020](https://arxiv.org/html/2607.20470#bib.bib4)\)\. The statistics of these datasets are reported in Table[1](https://arxiv.org/html/2607.20470#S4.T1), and the training sample statistics after task decomposition are listed in Table[2](https://arxiv.org/html/2607.20470#S4.T2)\. Besides, since the gold references for the CMeIE\-V2 test set are not publicly available, we evaluate models on the validation set\.

Models and hyperparameters\.We experiment on open\-source LLMs with various parameter sizes, including Qwen3\-8BTeam \([2025](https://arxiv.org/html/2607.20470#bib.bib7)\), InternLM3\-8BCaiet al\.\([2024](https://arxiv.org/html/2607.20470#bib.bib8)\), LLaMA3\.1\-8BDubeyet al\.\([2024](https://arxiv.org/html/2607.20470#bib.bib52)\), and GLM4\-9BGLMet al\.\([2024](https://arxiv.org/html/2607.20470#bib.bib13)\)\. The parameter settings for tuning the LLMs are detailed in Appendix[A](https://arxiv.org/html/2607.20470#A1)\. We also test closed\-source LLMs via API\. Based on three trials, the average costs per run are $0\.0036 for GPT\-5\.1, $0\.0016 for QwQ\-MAX, and $0\.00037 for DeepSeek\-R1\. The coefficients of the DTI planner are solved, as shown in Table[3](https://arxiv.org/html/2607.20470#S4.T3)\. The analysis of coefficient solution non\-uniqueness and the variable mapping range values are provided in Appendix[F](https://arxiv.org/html/2607.20470#A6)and Appendix[G](https://arxiv.org/html/2607.20470#A7), respectively\.

MethodTime\(s\)↓\\downarrowP\(%\)R\(%\)F1\(%\)↑\\uparrowSea\.\(s\)↓\\downarrowSingle\-objectiveRandom66,12253\.1649\.4551\.246Grid Search96,59365\.4146\.4254\.30742,090Greedy Search113,26263\.5845\.3652\.94430,599Hill\-climbing Search96,59365\.4146\.4254\.30736,082MetaGPTGPT\-5\.178,21262\.3243\.9651\.55164QwQ\-MAX90,35244\.7960\.8251\.59165Deepseek\-R1113,26263\.5845\.3652\.94164PlanE \(ours\)96,59365\.4146\.4254\.30187,257Multi\-objectiveRandom66,12253\.1649\.4551\.246Grid Search79,59363\.1046\.1053\.28742,090Greedy Search11,40550\.7651\.4151\.09187,257Hill\-climbing Search96,59365\.4146\.4254\.30736,082MetaGPTGPT\-5\.111,16750\.8350\.7050\.76182QwQ\-MAX96,59345\.9059\.3251\.75185Deepseek\-R168,22751\.3451\.2551\.29181PlanE \(ours\)79,59363\.1046\.1053\.28187,257

Table 4:Comparison of single\-objective \(Performance\) and multi\-objective \(Performance \+ Efficiency\) optimization the on Qwen3\-8B in CMeIE\-V2\. “Time” = model construction time, “Sea\.” = search time for the optimal DTI combination\.IDCombinationTime\(s\)P\(%\)R\(%\)F1\(%\)Variablef​\(x\)d​a​t​a​s​e​tf\(x\)\_\{dataset\}f​\(x\)m​o​d​e​lf\(x\)\_\{model\}DataTuningInferencev​dvdv​tvtv​ivi1SFTDirect11,40550\.7651\.4151\.098101143\-7692SFT \+ GRPODirect17,34249\.9550\.3650\.15811\-18\-3703SFT \+ DPODirect11,16750\.8350\.7050\.76871152\-4564NoneSFT \+ KTODirect13,93151\.6650\.1150\.87881156\-5405Direct68,22751\.3451\.2551\.2910101126\-8616Intersection70,02963\.3045\.1452\.70101010227\-57SFTUnion70,02945\.4958\.1051\.0310102154\-7048Direct111,47052\.3451\.9952\.171011\-73\-5869Intersection113,26263\.5845\.3652\.941011016\-54010SFT \+ GRPOUnion113,26246\.1358\.1151\.431012\-46\-52011Direct77,64152\.4250\.8951\.641071123\-58912Intersection79,59363\.1046\.1053\.2810710220\-413SFT \+ DPOUnion79,59344\.7859\.8351\.221072150\-46314Direct94,87951\.7851\.6351\.701081131\-66015Intersection96,59365\.4146\.4254\.30108102291616Pipeline\-basedSFT \+ KTOUnion96,59345\.9059\.3251\.751082158\-52317Direct65,18550\.3950\.3350\.36110191\-84218Intersection66,80460\.6146\.0552\.3411010142\-63419SFTUnion66,80445\.5958\.5851\.271102113\-75720Direct110,83651\.1150\.4850\.7911163\-921Intersection112,49060\.9545\.5752\.151110103\-61122SFT \+ GRPOUnion112,49045\.4558\.3851\.1111284\-1423Direct76,53950\.7250\.2250\.47171144\-38424Intersection78,21262\.3243\.9651\.551710192\-44625SFT \+ DPOUnion78,21243\.5258\.8150\.02172166\-33026Direct88,75952\.5850\.7551\.65181133\-51727Intersection90,35263\.3444\.9052\.551810182\-48928BidirectionalSFT \+ KTOUnion90,35244\.7960\.8251\.59182155\-452

Table 5:The prediction results of constructing the DTI planner on the Qwen3\-8B in CMeIE\-V2\.IDCombinationTime\(s\)P\(%\)R\(%\)F1\(%\)Variablef​\(x\)d​a​t​a​s​e​tf\(x\)\_\{dataset\}f​\(x\)m​o​d​e​lf\(x\)\_\{model\}DataTuningInferencev​dvdv​tvtv​ivi1SFTDirect11,40550\.7651\.4151\.098101167\.9179\.02SFT \+ GRPODirect17,34249\.9550\.3650\.15811162\.5802\.73SFT \+ DPODirect11,16750\.8350\.7050\.76871193\.1449\.94NoneSFT \+ KTODirect13,93151\.6650\.1150\.87881187\.7366\.65Direct68,22751\.3451\.2551\.2910101215\.9488\.26Intersection70,02963\.3045\.1452\.704101010395\.03,028\.07SFTUnion70,02945\.4958\.1051\.0310102265\.4700\.88Direct111,47052\.3451\.9952\.171011176\.31,160\.59Intersection113,26263\.5845\.3652\.94310110298\.73,044\.210SFT \+ GRPOUnion113,26246\.1358\.1151\.431012219\.51,300\.211Direct77,64152\.4250\.8951\.641071229\.7775\.312Intersection79,59363\.1046\.1053\.28210710389\.93,096\.413SFT \+ DPOUnion79,59344\.7859\.8351\.221072277\.1963\.614Direct94,87951\.7851\.6351\.701081228\.1686\.615Intersection96,59365\.4146\.4254\.30110810394\.63,080\.616Pipeline\-basedSFT \+ KTOUnion96,59345\.9059\.3251\.751082276\.2883\.017Direct65,18550\.3950\.3350\.361101\-82\.0\-405\.518Intersection66,80460\.6146\.0552\.3411010\-283\.61,340\.519SFTUnion66,80445\.5958\.5851\.271102\-74\.8\-281\.120Direct110,83651\.1150\.4850\.7911132\.348\.121Intersection112,49060\.9545\.5752\.151110\-226\.01,138\.022SFT \+ GRPOUnion112,49045\.4558\.3851\.1111233\.299\.623Direct76,53950\.7250\.2250\.47171\-16\.9\-191\.324Intersection78,21262\.3243\.9651\.551710\-237\.41,336\.025SFT \+ DPOUnion78,21243\.5258\.8150\.02172\-11\.8\-91\.226Direct88,75952\.5850\.7551\.65181\-35\.6\-255\.727Intersection90,35263\.3444\.9052\.551810\-249\.81,344\.528BidirectionalSFT \+ KTOUnion90,35244\.7960\.8251\.59182\-29\.8\-147\.5

Table 6:The prediction results of constructing the DTI planner on multi\-objective optimization\. Since Performance \(F1\) and Efficiency \(training time,TT\) operate on different scales, we apply min\-max normalization to both metrics\. The final scoreSSfor a DTI combination is the weighted sum of the normalized components:S=α⋅Norm​\(F1\)\+β⋅Norm​\(T\)S=\\alpha\\cdot\\text\{Norm\}\(F\_\{1\}\)\+\\beta\\cdot\\text\{Norm\}\(T\)\(In our default setup,α=β=0\.5\\alpha=\\beta=0\.5\)\. This joint scoreSSreplaces the single\-objective in the ranking stage\. All subsequent steps—including the search and the planner’s loss calculation—remain identical to the single\-objective pipeline described\. The red numerical labels indicate the best performance ranking\.Table 7:Ablation study \(F1%\)\. “Pipe\.”, “Bidi\.”, “S\.”, “G\.”, “D\.”, “K\.”, “Int\.”, and “Best” denote the “Pipeline\-based”, “Bidirectional”, “SFT”, “GRPO”, “DPO”, “KTO ”, “Intersection” and “the best performance”, respectively\.Evaluation metrics\.Following the previous worksZhanget al\.\([2024](https://arxiv.org/html/2607.20470#bib.bib29)\), we employ three standard evaluation metrics, i\.e\., Precision \(P\), Recall \(R\), and strict Micro\-F1 \(F1\)\. Notably, in RE, a triplet is considered correct only if the relation type, along with the types and the boundaries of the head\-tail entities are precisely determined\. For EE, we decompose events into \(trigger, event type, argument, argument role\) quadruples, requiring exact matching of all components\. For ABSA, a triplet is regarded as correct only when the aspect, opinion, and sentiment all coincide with the ground truth\.

### 5\.2Exp\-I: Main results

For a fair comparison, we build an LLM\-Planner by prompting MetaGPT with PlanE’s empirical data and redefining its operating procedure to select DTI combinations\. As listed in Table[4](https://arxiv.org/html/2607.20470#S5.T4), we choose three base\-LLMs \(GPT\-5\.1, QwQ\-MAX, DeepSeek\-R1\) as the LLM\-Planner\. We observe that: 1\) PlanE consistently outperforms all baselines under both single\- and multi\-objective optimization\. 2\) In both single\-objective and multi\-objective optimization settings, PlanE achieves performance comparable to Grid Search, while reducing search time by 554,833 seconds\.

### 5\.3Exp\-II: How is the generalization ability of the DTI planner?

View 1: Same task dataset, different base\-LLMs\.To verify the generalization of the proposed PlanE, we design the View 1 experiment: fitting the DTI function on several base\-LLMs using the same task dataset\. The plannerf​\(x\)d​a​t​a​s​e​tf\(x\)\_\{dataset\}is fitted on CMeIE\-V2 with three base\-LLMs: InternLM3\-8B, LLaMA\-3\.1\-8B, and GLM4\-9B \(detailed fitted data are shown in Appendix[D](https://arxiv.org/html/2607.20470#A4)\)\. The results are reported in Table[3](https://arxiv.org/html/2607.20470#S4.T3)\. To further assess the planner’s validity, we evaluate it on Qwen3\-8B, with the results shown in Table[5](https://arxiv.org/html/2607.20470#S5.T5)\. The DTI planner identifies the following best DTI combination: pipeline\-based data decomposition, SFT \+ KTO tuning, and intersection inference\. The resulting extractive\-based LLM achieves the best performance among the 28 candidate DTI combinations under single\-objective optimization focused solely on performance\.

View 2: Same base\-LLM, different task datasets\.To evaluate the effectiveness of PlanE from the View 2 experiment, we conduct an experiment in which the DTI planer is fitted on multiple task datasets using the same base\-LLM\. The plannerf​\(x\)m​o​d​e​lf\(x\)\_\{model\}is fitted using the results from EE and an ABSA task dataset based on Qwen3\-8B \(detailed fitted data are shown in Appendix[E](https://arxiv.org/html/2607.20470#A5)\), and the corresponding coefficient results are reported in Table[3](https://arxiv.org/html/2607.20470#S4.T3)\. To verify the planner’s generalization, we further test it on the RE dataset CMeIE\-V2, with the results listed in Table[5](https://arxiv.org/html/2607.20470#S5.T5)\. The planner identifies the following best DTI combination: pipeline\-based data decomposition, SFT \+ KTO tuning, and intersection inference\. The resulting extractive\-based LLM \(ID=15\) outperforms the other 27 candidate DTI combinations under single\-objective optimization focused on performance\.

### 5\.4Exp\-III: Can the DTI planner be used for multi\-objective optimization?

To evaluate the generalization capability of the DTI planner, we further apply it to multi\-objective optimization\. When constructing extractive\-based LLMs, we sometimes need to balance both performance and efficiency, measuring efficiency by the combined training and inference time\. The DTI planner is still constructed from two views, and its coefficients are listed in Table[3](https://arxiv.org/html/2607.20470#S4.T3)\. The prediction results of the planner are shown in Table[6](https://arxiv.org/html/2607.20470#S5.T6)\. Experimental results show that the DTI planner effectively achieves multi\-objective optimization\. Specifically, from the view of the same datasetf​\(x\)d​a​t​a​s​e​tf\(x\)\_\{dataset\}, the optimal choice for multi\-objective optimization is ID=6 \(highlighted in yellow in Table[6](https://arxiv.org/html/2607.20470#S5.T6)\); from the view of the same base\-LLMf​\(x\)m​o​d​e​lf\(x\)\_\{model\}, the optimal choice is ID=12 \(highlighted in blue in Table[6](https://arxiv.org/html/2607.20470#S5.T6)\)\.

### 5\.5Exp\-IV: Does DTI impact the performance of extractive\-based LLMs?

In this subsection, we conduct ablation experiments to examine how the three impact elements: Data, Tuning, and Inference, each contribute to building an extractive\-based LLM\. The ablation study spans three tasks, including: RE task \(e\.g\., CMeIE\-V2\), EE task \(e\.g\., ACE05\), and ABSA task \(e\.g\., 14Lap\), and across four base\-LLMs\. The results, summarized in Table[7](https://arxiv.org/html/2607.20470#S5.T7), reveal that each factor affects various base\-LLMs to varying extents, and their effectiveness across datasets shows no pattern\.

## 6Conclusion

In this paper, we propose PlanE, a framework that is the first to focus on extractive\-based LLM construction\. We design a DTI planner to select the optimal DTI combination for specific datasets/base\-LLMs to improve construction efficiency\. We conducted experimental analyses of PlanE’s generalizability from two views: \(1\) across different datasets using the same base\-LLM, and \(2\) on the same dataset using different base\-LLMs\. Moreover, results show that PlanE improves the F1\-score over MetaGPT by 1\.36% \(single\-objective\), and matches the performance of Grid Search under single\- and multi\-objective optimization while saving 554,833 seconds of search time\.

## Limitations

While PlanE offers a valuable solution for planning of Data, Tuning, and Inference for extractive\-based LLMs, it has three main limitations:

1\) The DTI planner is a discrete function and cannot be directly optimized using gradient information, which can be addressed through reinforcement learning methodsKaelblinget al\.\([1996](https://arxiv.org/html/2607.20470#bib.bib59)\)\.

2\) We do not consider all influencing factors, particularly practical constraints such as computational resources, or additional DTI combinations involving more complex and diverse data, tuning, and inference\. Therefore, we leave this exploration for future work\.

3\) We currently employ scalar encoding for the DTI planner\. Although it carries less information than vectorized alternatives, empirical results show it provides sufficient structural cues to stably reach the optimal performance \(ID 15 in Table[5](https://arxiv.org/html/2607.20470#S5.T5)\)\. Thus, the simplified representation does not bottleneck the current model’s efficacy\. Nevertheless, exploring vectorized representations remains a key future work to improve search efficiency and minimize the overhead of computing initial DTI combinations\.

## Acknowledgments

This work is supported by the Shanghai Natural Science Foundation Project under Grant 25ZR1402116\.

## References

- I\. Abdelaziz, K\. Basu, M\. Agarwal, S\. Kumaravel, M\. Stallone, R\. Panda, Y\. Rizk, G\. P\. S\. Bhargav, M\. Crouse, C\. Gunasekara, S\. Ikbal, S\. Joshi, H\. Karanam, V\. Kumar, A\. Munawar, S\. Neelam, D\. Raghu, U\. Sharma, A\. M\. Soria, D\. Sreedhar, P\. Venkateswaran, M\. Unuvar, D\. D\. Cox, S\. Roukos, L\. A\. Lastras, and P\. Kapanipathi \(2024\)Granite\-function calling model: introducing function calling abilities via multi\-task learning of granular tasks\.InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing: Industry Track,F\. Dernoncourt, D\. Preoţiuc\-Pietro, and A\. Shimorina \(Eds\.\),Miami, Florida, US,pp\. 1131–1139\.External Links:[Link](https://aclanthology.org/2024.emnlp-industry.85/),[Document](https://dx.doi.org/10.18653/v1/2024.emnlp-industry.85)Cited by:[§1](https://arxiv.org/html/2607.20470#S1.p2.1)\.
- R\. Y\. Aminabadi, S\. Rajbhandari, A\. A\. Awan, C\. Li, D\. Li, E\. Zheng, O\. Ruwase, S\. Smith, M\. Zhang, J\. Rasley,et al\.\(2022\)Deepspeed\-inference: enabling efficient inference of transformer models at unprecedented scale\.InSC22: International Conference for High Performance Computing, Networking, Storage and Analysis,pp\. 1–15\.Cited by:[§2](https://arxiv.org/html/2607.20470#S2.p2.1)\.
- M\. Brief, O\. Ovadia, G\. Shenderovitz, N\. BenYoash, R\. Lemberg, and E\. Sheetrit \(2024\)Mixing it up: the cocktail effect of multi\-task fine\-tuning on LLM performance \- a case study in finance\.External Links:[Link](https://openreview.net/forum?id=VoHJTTA2MB)Cited by:[§1](https://arxiv.org/html/2607.20470#S1.p2.1)\.
- Z\. Cai, M\. Cao, H\. Chen, K\. Chen, K\. Chen, X\. Chen, X\. Chen, Z\. Chen, Z\. Chen, P\. Chu,et al\.\(2024\)Internlm2 technical report\.arXiv preprint arXiv:2403\.17297\.Cited by:[§5\.1](https://arxiv.org/html/2607.20470#S5.SS1.p2.1)\.
- D\. A\. Calian, G\. Farquhar, I\. Kemaev, L\. M\. Zintgraf, M\. Hessel, J\. Shar, J\. Oh, A\. György, T\. Schaul, J\. Dean,et al\.\(2025\)DataRater: meta\-learned dataset curation\.arXiv preprint arXiv:2505\.17895\.Cited by:[§2](https://arxiv.org/html/2607.20470#S2.p3.1)\.
- K\. Chang, K\. Wang, N\. Yang, Y\. Wang, D\. Jin, W\. Zhu, Z\. Chen, C\. Li, H\. Yan, Y\. Zhou,et al\.\(2024\)Data is all you need: finetuning llms for chip design via an automated design\-data augmentation framework\.InProceedings of the 61st ACM/IEEE Design Automation Conference,pp\. 1–6\.Cited by:[§1](https://arxiv.org/html/2607.20470#S1.p2.1)\.
- B\. Ding, C\. Qin, R\. Zhao, T\. Luo, X\. Li, G\. Chen, W\. Xia, J\. Hu, A\. T\. Luu, and S\. Joty \(2024\)Data augmentation using LLMs: data perspectives, learning paradigms and challenges\.InFindings of the Association for Computational Linguistics: ACL 2024,L\. Ku, A\. Martins, and V\. Srikumar \(Eds\.\),Bangkok, Thailand,pp\. 1679–1705\.External Links:[Link](https://aclanthology.org/2024.findings-acl.97/),[Document](https://dx.doi.org/10.18653/v1/2024.findings-acl.97)Cited by:[§1](https://arxiv.org/html/2607.20470#S1.p2.1)\.
- H\. Dong, H\. Ye, W\. Zhu, K\. Jiang, and G\. Song \(2025\)Meta\-r1: empowering large reasoning models with metacognition\.arXiv preprint arXiv:2508\.17291\.Cited by:[§2](https://arxiv.org/html/2607.20470#S2.p4.1)\.
- Y\. Du, Z\. Wang, A\. Farhan, C\. Angione, H\. Yang, F\. Johnston, J\. P\. Buban, Y\. Zhao, and Y\. Yang \(2025\)Meta\-learning for speeding up large model inference in decentralized environments\.InSecond Conference on Language Modeling,External Links:[Link](https://openreview.net/forum?id=eLWn2XVMHA)Cited by:[§2](https://arxiv.org/html/2607.20470#S2.p2.1)\.
- A\. Dubey, A\. Jauhri, A\. Pandey, A\. Kadian, A\. Al\-Dahle, A\. Letman, A\. Mathur, A\. Schelten, A\. Yang, A\. Fan,et al\.\(2024\)The llama 3 herd of models\.arXiv preprint arXiv:2407\.21783\.Cited by:[§5\.1](https://arxiv.org/html/2607.20470#S5.SS1.p2.1)\.
- K\. Ethayarajh, W\. Xu, N\. Muennighoff, D\. Jurafsky, and D\. Kiela \(2024\)Kto: model alignment as prospect theoretic optimization\.arXiv preprint arXiv:2402\.01306\.Cited by:[Appendix H](https://arxiv.org/html/2607.20470#A8.p6.1),[§4\.1](https://arxiv.org/html/2607.20470#S4.SS1.p2.1)\.
- T\. GLM, A\. Zeng, B\. Xu, B\. Wang, C\. Zhang, D\. Yin, D\. Zhang, D\. Rojas, G\. Feng, H\. Zhao,et al\.\(2024\)Chatglm: a family of large language models from glm\-130b to glm\-4 all tools\.arXiv preprint arXiv:2406\.12793\.Cited by:[§5\.1](https://arxiv.org/html/2607.20470#S5.SS1.p2.1)\.
- J\. Hoffmann, S\. Borgeaud, A\. Mensch, E\. Buchatskaya, T\. Cai, E\. Rutherford, D\. de Las Casas, L\. A\. Hendricks, J\. Welbl, A\. Clark, T\. Hennigan, E\. Noland, K\. Millican, G\. van den Driessche, B\. Damoc, A\. Guy, S\. Osindero, K\. Simonyan, E\. Elsen, O\. Vinyals, J\. W\. Rae, and L\. Sifre \(2022\)Training compute\-optimal large language models\.InProceedings of the 36th International Conference on Neural Information Processing Systems,NIPS ’22,Red Hook, NY, USA\.External Links:ISBN 9781713871088Cited by:[§2](https://arxiv.org/html/2607.20470#S2.p4.1)\.
- S\. Hong, M\. Zhuge, J\. Chen, X\. Zheng, Y\. Cheng, J\. Wang, C\. Zhang, Z\. Wang, S\. K\. S\. Yau, Z\. Lin, L\. Zhou, C\. Ran, L\. Xiao, C\. Wu, and J\. Schmidhuber \(2024\)MetaGPT: meta programming for a multi\-agent collaborative framework\.InThe Twelfth International Conference on Learning Representations,External Links:[Link](https://openreview.net/forum?id=VtmBAGCN7o)Cited by:[§2](https://arxiv.org/html/2607.20470#S2.p3.1)\.
- L\. P\. Kaelbling, M\. L\. Littman, and A\. W\. Moore \(1996\)Reinforcement learning: a survey\.Journal of artificial intelligence research4,pp\. 237–285\.Cited by:[Limitations](https://arxiv.org/html/2607.20470#Sx1.p2.1)\.
- K\. Killamsetty, S\. Durga, G\. Ramakrishnan, A\. De, and R\. Iyer \(2021\)Grad\-match: gradient matching based data subset selection for efficient deep model training\.InInternational Conference on Machine Learning,pp\. 5464–5474\.Cited by:[§2](https://arxiv.org/html/2607.20470#S2.p3.1)\.
- Y\. Li, K\. Ding, J\. Wang, and K\. Lee \(2024a\)Empowering large language models for textual data augmentation\.InFindings of the Association for Computational Linguistics: ACL 2024,L\. Ku, A\. Martins, and V\. Srikumar \(Eds\.\),Bangkok, Thailand,pp\. 12734–12751\.External Links:[Link](https://aclanthology.org/2024.findings-acl.756/),[Document](https://dx.doi.org/10.18653/v1/2024.findings-acl.756)Cited by:[§1](https://arxiv.org/html/2607.20470#S1.p2.1)\.
- Z\. Li, C\. Li, M\. Zhang, Q\. Mei, and M\. Bendersky \(2024b\)Retrieval augmented generation or long\-context LLMs? a comprehensive study and hybrid approach\.InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing: Industry Track,F\. Dernoncourt, D\. Preoţiuc\-Pietro, and A\. Shimorina \(Eds\.\),Miami, Florida, US,pp\. 881–893\.External Links:[Link](https://aclanthology.org/2024.emnlp-industry.66/),[Document](https://dx.doi.org/10.18653/v1/2024.emnlp-industry.66)Cited by:[§1](https://arxiv.org/html/2607.20470#S1.p2.1)\.
- C\. Lou, Z\. Sun, X\. Liang, M\. Qu, W\. Shen, W\. Wang, Y\. Li, Q\. Yang, and S\. Wu \(2025\)AdaCoT: pareto\-optimal adaptive chain\-of\-thought triggering via reinforcement learning\.arXiv preprint arXiv:2505\.11896\.Cited by:[§2](https://arxiv.org/html/2607.20470#S2.p4.1)\.
- T\. Parekh, P\. Prakash, A\. Radovic, A\. Shekher, and D\. Savenkov \(2025\)Dynamic strategy planning for efficient question answering with large language models\.InFindings of the Association for Computational Linguistics: NAACL 2025,pp\. 6038–6059\.Cited by:[§2](https://arxiv.org/html/2607.20470#S2.p2.1)\.
- R\. Rafailov, A\. Sharma, E\. Mitchell, C\. D\. Manning, S\. Ermon, and C\. Finn \(2023\)Direct preference optimization: your language model is secretly a reward model\.InThirty\-seventh Conference on Neural Information Processing Systems,External Links:[Link](https://openreview.net/forum?id=HPuSIXJaa9)Cited by:[Appendix H](https://arxiv.org/html/2607.20470#A8.p3.2),[§4\.1](https://arxiv.org/html/2607.20470#S4.SS1.p2.1)\.
- Z\. Shao, P\. Wang, Q\. Zhu, R\. Xu, J\. Song, X\. Bi, H\. Zhang, M\. Zhang, Y\. Li, Y\. Wu,et al\.\(2024\)Deepseekmath: pushing the limits of mathematical reasoning in open language models\.arXiv preprint arXiv:2402\.03300\.Cited by:[Appendix H](https://arxiv.org/html/2607.20470#A8.p2.22),[§4\.1](https://arxiv.org/html/2607.20470#S4.SS1.p2.1)\.
- W\. Su, Y\. Tang, Q\. Ai, Z\. Wu, and Y\. Liu \(2024\)DRAGIN: dynamic retrieval augmented generation based on the real\-time information needs of large language models\.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\. 12991–13013\.External Links:[Link](https://aclanthology.org/2024.acl-long.702/),[Document](https://dx.doi.org/10.18653/v1/2024.acl-long.702)Cited by:[§1](https://arxiv.org/html/2607.20470#S1.p2.1)\.
- Q\. Team \(2025\)Qwen3: think deeper, act faster\.External Links:[Link](https://qwenlm.github.io/blog/qwen3/)Cited by:[§5\.1](https://arxiv.org/html/2607.20470#S5.SS1.p2.1)\.
- C\. Tian, K\. Chen, J\. Liu, Z\. Liu, Z\. Zhang, and J\. Zhou \(2025\)Towards greater leverage: scaling laws for efficient mixture\-of\-experts language models\.arXiv preprint arXiv:2507\.17702\.Cited by:[§2](https://arxiv.org/html/2607.20470#S2.p3.1)\.
- H\. Touvron, L\. Martin, K\. Stone, P\. Albert, A\. Almahairi, Y\. Babaei, N\. Bashlykov, S\. Batra, P\. Bhargava, S\. Bhosale,et al\.\(2023\)Llama 2: open foundation and fine\-tuned chat models\.arXiv preprint arXiv:2307\.09288\.Cited by:[§2](https://arxiv.org/html/2607.20470#S2.p4.1)\.
- C\. Walker, S\. Strassel, J\. Medero, and K\. Maeda \(2006\)ACE 2005 multilingual training corpus\.Linguistic Data Consortium, Philadelphia57,pp\. 45\.Cited by:[§5\.1](https://arxiv.org/html/2607.20470#S5.SS1.p1.1)\.
- K\. Wang, F\. Duan, P\. Li, S\. Wang, and X\. Cai \(2025\)LLMs know what they need: leveraging a missing information guided framework to empower retrieval\-augmented generation\.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\. 2379–2400\.External Links:[Link](https://aclanthology.org/2025.coling-main.163/)Cited by:[§1](https://arxiv.org/html/2607.20470#S1.p2.1)\.
- X\. Wang, C\. Li, Z\. Wang, F\. Bai, H\. Luo, J\. Zhang, N\. Jojic, E\. Xing, and Z\. Hu \(2024\)PromptAgent: strategic planning with language models enables expert\-level prompt optimization\.InThe Twelfth International Conference on Learning Representations,External Links:[Link](https://openreview.net/forum?id=22pyNMuIoa)Cited by:[§2](https://arxiv.org/html/2607.20470#S2.p3.1)\.
- L\. Xu, H\. Li, W\. Lu, and L\. Bing \(2020\)Position\-aware tagging for aspect sentiment triplet extraction\.InProceedings of the 2020 Conference on Empirical Methods in Natural Language Processing \(EMNLP\),pp\. 2339–2349\.Cited by:[§5\.1](https://arxiv.org/html/2607.20470#S5.SS1.p1.1)\.
- N\. Zhang, M\. Chen, Z\. Bi, X\. Liang, L\. Li, X\. Shang, K\. Yin, C\. Tan, J\. Xu, F\. Huang,et al\.\(2022\)CBLUE: a chinese biomedical language understanding evaluation benchmark\.InACL,pp\. 7888–7915\.Cited by:[§5\.1](https://arxiv.org/html/2607.20470#S5.SS1.p1.1)\.
- W\. Zhang, J\. Wang, C\. Chen, W\. Lu, W\. Du, H\. Wang, J\. Liu, and T\. Ruan \(2024\)A bidirectional extraction\-then\-evaluation framework for complex relation extraction\.IEEE Transactions on Knowledge and Data Engineering\.Cited by:[§5\.1](https://arxiv.org/html/2607.20470#S5.SS1.p3.1)\.

## Appendix AParameter Settings

For model fine\-tuning, we utilize Low\-Rank Adaptation \(LoRA\) with the following configuration: a batch size of 32, a learning rate of1×10−41\\times 10^\{\-4\}, and 4 epochs\. To optimize computational efficiency, we employ the bfloat16 data type\. The LoRA\-specific hyperparameters are set as follows: the rank of the low\-rank matrices \(l​o​r​a​\_​rlora\\\_r\) is 16, and the scaling factor \(l​o​r​a​\_​αlora\\\_\\alpha\) is 32\.

For model reinforcement learning, we use the GRPO, DPO, and KTO methods and utilize Low\-Rank Adaptation \(LoRA\) with the following configuration: a batch size of 8, a generation number of 8, a learning rate of1×10−61\\times 10^\{\-6\}, and 2 epochs\. To optimize computational efficiency, we employ the bfloat16 data type\. The LoRA\-specific hyperparameters are set as follows: the rank of the low\-rank matrices \(l​o​r​a​\_​rlora\\\_r\) is 16, and the scaling factor \(l​o​r​a​\_​αlora\\\_\\alpha\) is 32\.

Our experiments are conducted on a workstation running Ubuntu 20\.04\.6 LTS, with two Intel\(R\) Xeon\(R\) Platinum 8336C CPUs, four NVIDIA A800 GPUs, and 1\.0TiB of memory\.

During inference, we configure thet​e​m​p​e​r​a​t​u​r​etemperatureto 0\.3 and the top\-p sampling parameter \(t​o​p​\_​ptop\\\_p\) to 0\.9, ensuring a balance between diversity and coherence in the generated outputs\.

## Appendix BTraining Scale for Different Instruction Tunings

This section presents the number of training samples for different instruction tuning strategies across all three datasets\. We provide a detailed breakdown of the sample sizes for each decomposition approach, offering insight into the distribution of data used for training under various configurations, as shown in Table[A1](https://arxiv.org/html/2607.20470#A2.T1)\.

CombinationData StructureTrain SamplesDataTuningCMeIE\-V2NoneSFTStep1 SFT:c→\(h,r,t\)c\\rightarrow\(h,r,t\);7,000SFT \+ RLStep1 SFT:c→\(h,r,t\)c\\rightarrow\(h,r,t\);6,000Step2 RL:c→\(h,r,t\)c\\rightarrow\(h,r,t\);1,000Pipeline\-basedSFTStep1 SFT:c→\(h,r,t\)c\\rightarrow\(h,r,t\);c→\(h,t\)c\\rightarrow\(h,t\);c→rc\\rightarrow r;r​\[s​1\]​c→\(h,t\)r\{\[\}s1\{\]\}c\\rightarrow\(h,t\);h​\[s​1\]​t​\[s​2\]​c→rh\{\[\}s1\{\]\}t\{\[\}s2\{\]\}c\\rightarrow r;69,287SFT \+ RLStep1 SFT:c→\(h,r,t\)c\\rightarrow\(h,r,t\);c→\(h,t\)c\\rightarrow\(h,t\);c→rc\\rightarrow r;r​\[s​1\]​c→\(h,t\)r\{\[\}s1\{\]\}c\\rightarrow\(h,t\);h​\[s​1\]​t​\[s​2\]​c→rh\{\[\}s1\{\]\}t\{\[\}s2\{\]\}c\\rightarrow r;59,287Step2 RL:c→\(h,r,t\)c\\rightarrow\(h,r,t\);c→\(h,t\)c\\rightarrow\(h,t\);c→rc\\rightarrow r;r​\[s​1\]​c→\(h,t\)r\{\[\}s1\{\]\}c\\rightarrow\(h,t\);h​\[s​1\]​t​\[s​2\]​c→rh\{\[\}s1\{\]\}t\{\[\}s2\{\]\}c\\rightarrow r;10,000BidirectionalSFTStep1 SFT:c→\(h,r,t\)c\\rightarrow\(h,r,t\);c→\(h\)c\\rightarrow\(h\);c→\(t\)c\\rightarrow\(t\);h​\[s​1\]​c→\(t,r\)h\{\[\}s1\{\]\}c\\rightarrow\(t,r\);t​\[s​1\]​c→\(h,r\)t\{\[\}s1\{\]\}c\\rightarrow\(h,r\);62,887SFT \+ RLStep1 SFT:c→\(h,r,t\)c\\rightarrow\(h,r,t\);c→\(h\)c\\rightarrow\(h\);c→\(t\)c\\rightarrow\(t\);h​\[s​1\]​c→\(t,r\)h\{\[\}s1\{\]\}c\\rightarrow\(t,r\);t​\[s​1\]​c→\(h,r\)t\{\[\}s1\{\]\}c\\rightarrow\(h,r\);52,887Step2 RL:c→\(h,r,t\)c\\rightarrow\(h,r,t\);c→\(h\)c\\rightarrow\(h\);c→\(t\)c\\rightarrow\(t\);h​\[s​1\]​c→\(t,r\)h\{\[\}s1\{\]\}c\\rightarrow\(t,r\);t​\[s​1\]​c→\(h,r\)t\{\[\}s1\{\]\}c\\rightarrow\(h,r\);10,000ACE05NoneSFTStep1 SFT:c→\(t​r,e,a\)c\\rightarrow\(tr,e,a\);3,342SFT \+ RLStep1 SFT:c→\(t​r,e,a\)c\\rightarrow\(tr,e,a\);2,842Step2 RL:c→\(t​r,e,a\)c\\rightarrow\(tr,e,a\);500Pipeline\-basedSFTStep1 SFT:c→\(t​r,e,a\)c\\rightarrow\(tr,e,a\);c→\(t​r,a\)c\\rightarrow\(tr,a\);c→ec\\rightarrow e;e​\[s​1\]​c→\(t​r,a\)e\{\[\}s1\{\]\}c\\rightarrow\(tr,a\);t​r​\[s​1\]​a​\[s​2\]​c→etr\{\[\}s1\{\]\}a\{\[\}s2\{\]\}c\\rightarrow e;18,794SFT \+ RLStep1 SFT:c→\(t​r,e,a\)c\\rightarrow\(tr,e,a\);c→\(t​r,a\)c\\rightarrow\(tr,a\);c→ec\\rightarrow e;e​\[s​1\]​c→\(t​r,a\)e\{\[\}s1\{\]\}c\\rightarrow\(tr,a\);t​r​\[s​1\]​a​\[s​2\]​c→etr\{\[\}s1\{\]\}a\{\[\}s2\{\]\}c\\rightarrow e;16,094Step2 RL:c→\(t​r,e,a\)c\\rightarrow\(tr,e,a\);c→\(t​r,a\)c\\rightarrow\(tr,a\);c→ec\\rightarrow e;e​\[s​1\]​c→\(t​r,a\)e\{\[\}s1\{\]\}c\\rightarrow\(tr,a\);t​r​\[s​1\]​a​\[s​2\]​c→etr\{\[\}s1\{\]\}a\{\[\}s2\{\]\}c\\rightarrow e;2,700BidirectionalSFTStep1 SFT:c→\(t​r,e,a\)c\\rightarrow\(tr,e,a\);c→\(t​r\)c\\rightarrow\(tr\);c→\(a\)c\\rightarrow\(a\);t​r​\[s​1\]​c→\(e,a\)tr\{\[\}s1\{\]\}c\\rightarrow\(e,a\);a​\[s​1\]​c→\(t​r,e\)a\{\[\}s1\{\]\}c\\rightarrow\(tr,e\);18,794SFT \+ RLStep1 SFT:c→\(t​r,e,a\)c\\rightarrow\(tr,e,a\);c→\(t​r\)c\\rightarrow\(tr\);c→\(a\)c\\rightarrow\(a\);t​r​\[s​1\]​c→\(e,a\)tr\{\[\}s1\{\]\}c\\rightarrow\(e,a\);a​\[s​1\]​c→\(t​r,e\)a\{\[\}s1\{\]\}c\\rightarrow\(tr,e\);16,094Step2 RL:c→\(t​r,e,a\)c\\rightarrow\(tr,e,a\);c→\(t​r\)c\\rightarrow\(tr\);c→\(a\)c\\rightarrow\(a\);t​r​\[s​1\]​c→\(e,a\)tr\{\[\}s1\{\]\}c\\rightarrow\(e,a\);a​\[s​1\]​c→\(t​r,e\)a\{\[\}s1\{\]\}c\\rightarrow\(tr,e\);2,70014lapNoneSFTStep1 SFT:c→\(a​t,o,s\)c\\rightarrow\(at,o,s\);906SFT \+ RLStep1 SFT:c→\(a​t,o,s\)c\\rightarrow\(at,o,s\);756Step2 RL:c→\(a​t,o,s\)c\\rightarrow\(at,o,s\);150Pipeline\-basedSFTStep1 SFT:c→\(a​t,o,s\)c\\rightarrow\(at,o,s\);c→\(a​t,o\)c\\rightarrow\(at,o\);c→sc\\rightarrow s;s​\[s​1\]​c→\(a​t,o\)s\{\[\}s1\{\]\}c\\rightarrow\(at,o\);a​t​\[s​1\]​o​\[s​2\]​c→sat\{\[\}s1\{\]\}o\{\[\}s2\{\]\}c\\rightarrow s;5,127SFT \+ RLStep1 SFT:c→\(a​t,o,s\)c\\rightarrow\(at,o,s\);c→\(a​t,o\)c\\rightarrow\(at,o\);c→sc\\rightarrow s;s​\[s​1\]​c→\(a​t,o\)s\{\[\}s1\{\]\}c\\rightarrow\(at,o\);a​t​\[s​1\]​o​\[s​2\]​c→sat\{\[\}s1\{\]\}o\{\[\}s2\{\]\}c\\rightarrow s;4,327Step2 RL:c→\(a​t,o,s\)c\\rightarrow\(at,o,s\);c→\(a​t,o\)c\\rightarrow\(at,o\);c→sc\\rightarrow s;s​\[s​1\]​c→\(a​t,o\)s\{\[\}s1\{\]\}c\\rightarrow\(at,o\);a​t​\[s​1\]​o​\[s​2\]​c→sat\{\[\}s1\{\]\}o\{\[\}s2\{\]\}c\\rightarrow s;800BidirectionalSFTStep1 SFT:c→\(a​t,o,s\)c\\rightarrow\(at,o,s\);c→\(a​t\)c\\rightarrow\(at\);c→\(o\)c\\rightarrow\(o\);a​t​\[s​1\]​c→\(o,s\)at\{\[\}s1\{\]\}c\\rightarrow\(o,s\);o​\[s​1\]​c→\(a​t,s\)o\{\[\}s1\{\]\}c\\rightarrow\(at,s\);5,248SFT \+ RLStep1 SFT:c→\(a​t,o,s\)c\\rightarrow\(at,o,s\);c→\(a​t\)c\\rightarrow\(at\);c→\(o\)c\\rightarrow\(o\);a​t​\[s​1\]​c→\(o,s\)at\{\[\}s1\{\]\}c\\rightarrow\(o,s\);o​\[s​1\]​c→\(a​t,s\)o\{\[\}s1\{\]\}c\\rightarrow\(at,s\);4,448Step2 RL:c→\(a​t,o,s\)c\\rightarrow\(at,o,s\);c→\(a​t\)c\\rightarrow\(at\);c→\(o\)c\\rightarrow\(o\);a​t​\[s​1\]​c→\(o,s\)at\{\[\}s1\{\]\}c\\rightarrow\(o,s\);o​\[s​1\]​c→\(a​t,s\)o\{\[\}s1\{\]\}c\\rightarrow\(at,s\);800

Table A1:Training samples for different data decomposition and tuning strategies for each dataset\.VariableCombination ElementValue Assignment Rulev​dvdNoneF1 \[None, SFT, Direct\]Pipeline\-basedF1 \[Pipeline\-based, SFT, Direct\]BidirectionalF1 \[Bidirectional, SFT, Direct\]v​tvtSFTF1 \[None, SFT, Direct\]SFT \+ GRPOF1 \[None, SFT \+ GRPO, Direct\]SFT \+ DPOF1 \[None, SFT \+ DPO, Direct\]SFT \+ KTOF1 \[None, SFT \+ KTO, Direct\]v​iviDirect\( F1 \[None, SFT, Direct\] \+ F1 \[Pipeline\-based, SFT, Direct\] \+ F1 \[Bidirectional, SFT, Direct\] \) / 3Intersection\( F1 \[Pipeline\-based, SFT, Intersection\] \+ F1 \[Bidirectional, SFT, Intersection\] \) / 2Union\( F1 \[Pipeline\-based, SFT, Union\] \+ F1 \[Bidirectional, SFT, Union\] \) / 2

Table A2:The specific principles for mapping DTI strategies to numerical variables\. “F1\[\]” represents the F1 score corresponding to the specific DTI combination\.VariableStrategyCombinationF1\(%\)Avg\. F1\(%\)Mapping IntervalDataTuningInference1≤v≤101\\leq v\\leq 101≤v≤201\\leq v\\leq 20v​dvdwithout DecompositionNoneSFTDirect51\.0951\.09816Pipeline\-based DecompositionPipeline\-basedSFTDirect51\.2951\.291020Bidirectional DecompositionBidirectionalSFTDirect50\.3650\.3611v​tvtSFTNoneSFTDirect51\.0951\.091020SFT \+ GRPONoneSFT \+ GRPODirect50\.1550\.1511SFT \+ DPONoneSFT \+ DPODirect50\.7650\.76713SFT \+ KTONoneSFT \+ KTODirect50\.8750\.87816v​iviDirect InferenceNoneSFTDirect51\.0950\.9111Pipeline\-basedSFTDirect51\.29BidirectionalSFTDirect50\.36Intersection InferencePipeline\-basedSFTIntersection52\.7052\.521020BidirectionalSFTIntersection52\.34Union InferencePipeline\-basedSFTUnion51\.0351\.1524BidirectionalSFTUnion51\.27

Table A3:Calculation example of the three variables based on Qwen3\-8B in the CMeIE\-V2\.v​d,v​t,v​i∈ℤvd,vt,vi\\in\\mathbb\{Z\}\.IDCombinationGLM4\-9BInternLM3\-8BLLaMA3\.1\-8BDataTuningInferenceTime\(s\)F1\(%\)VariableTime\(s\)F1\(%\)VariableTime\(s\)F1\(%\)Variablev​dvdv​tvtv​iviv​dvdv​tvtv​iviv​dvdv​tvtv​ivi1NoneSFTDirect10,77449\.991619,72548\.12110110,38251\.0671022SFT \+ GRPODirect17,20850\.59110122,06747\.4213119,32950\.567823SFT \+ DPODirect11,12549\.1511113,10147\.5014112,23149\.457424SFT \+ KTODirect15,23150\.3518116,84247\.2211115,55848\.707125Pipeline\-basedSFTDirect64,47353\.07106174,11849\.871010174,03250\.8711026Intersection65,49353\.551061075,17350\.6110101075,23251\.92110107Union65,49352\.59106475,17350\.771010975,23250\.9711018SFT \+ GRPODirect112,83452\.6110101142,38449\.771031125,73351\.401829Intersection113,89853\.08101010143,45649\.9010310126,93652\.05181010Union113,89852\.4710104143,45650\.421039126,93651\.1418111SFT \+ DPODirect76,28252\.18101194,26349\.79104184,49250\.4514212Intersection77,32252\.921011095,30850\.071041085,69251\.40141013Union77,32251\.90101495,30849\.74104985,69250\.0914114SFT \+ KTODirect94,82851\.391081120,05449\.231011101,84550\.7411215Intersection95,91451\.8910810121,09950\.0110110103,03151\.31111016Union95,91450\.901084121,09949\.591019103,03151\.4611117BidirectionalSFTDirect66,24151\.1446170,90949\.15610170,38851\.161010218Intersection67,72553\.08461072,43050\.286101072,03252\.3010101019Union67,72551\.6246472,43049\.73610972,03250\.891010120SFT \+ GRPODirect112,43151\.544101133,80349\.28631122,68050\.99108221Intersection113,88953\.1941010135,32350\.236310124,38452\.621081022Union113,88951\.654104135,32350\.48639124,38451\.30108123SFT \+ DPODirect77,12851\.2441190,98349\.6964183,87950\.76104224Intersection78,64552\.90411092,46950\.44641085,54551\.961041025Union78,64551\.3541492,46949\.8864985,54550\.52104126SFT \+ KTODirect95,75751\.97481118,32348\.47611106,65451\.12101227Intersection97,17654\.034810119,86549\.836110108,31651\.701011028Union97,17651\.85484119,86548\.15619108,31651\.201011

Table A4:Fitted data of the DTI plannerf​\(x\)d​a​t​a​s​e​tf\(x\)\_\{dataset\}on the CMeIE\-V2 across three base\-LLMs\. The range of the variables’ mapping is from 1 to 10\.Table A5:Fitted data of the DTI plannerf​\(x\)m​o​d​e​lf\(x\)\_\{model\}on Qwen3\-8B for two task datasets\. The range of the variables’ mapping is from 1 to 10\.
## Appendix CCalculation Method of the Three Variables

This section represents the calculation method of the three variables\. Each variable corresponds to a different stage in the model workflow\. Representative F1 results from the corresponding configurations are used as the values of each variable\. The specific principles for mapping DTI strategies to numerical variables are shown in Table[A2](https://arxiv.org/html/2607.20470#A2.T2)\.

- •The data decomposition variable \(vdv\_\{d\}\) represents how the dataset is structured, including None \(no decomposition\), Pipeline\-based decomposition, and Bidirectional decomposition\. For None \(no decomposition\), it is F1 \(None, SFT, Direct\)\. For Pipeline\-based decomposition, it is F1 \(Pipeline\-based, SFT, Direct\)\. For Bidirectional decomposition, it is F1 \(Bidirectional, SFT, Direct\)\.
- •The tuning variable \(vtv\_\{t\}\) describes the instruction tuning method, including SFT, SFT \+ GRPO, SFT \+ DPO, and SFT \+ KTO\. For SFT, it is F1 \(None, SFT, Direct\)\. For SFT \+ GRPO, it is F1 \(None, SFT \+ GRPO, Direct\)\. For SFT \+ DPO, it is F1 \(None, SFT \+ DPO, Direct\)\. For SFT \+ KTO, it is F1 \(None, SFT \+ KTO, Direct\)\.
- •The inference variable \(viv\_\{i\}\) denotes the inference strategy\. For Direct, the value is \(F1 \(None, SFT, Direct\) \+ F1 \(Pipeline\-based, SFT, Direct\) \+ F1 \(Bidirectional, SFT, Direct\)\) / 3\. For Intersection, it is \(F1 \(Pipeline\-based, SFT, Intersection\) \+ F1 \(Bidirectional, SFT, Intersection\)\) / 2, and for Union, it is \(F1 \(Pipeline\-based, SFT, Union\) \+ F1 \(Bidirectional, SFT, Union\)\) / 2\.

An example of the calculation for the three variables based on Qwen3\-8B on the CMeIE\-V2 dataset is illustrated in Table[A3](https://arxiv.org/html/2607.20470#A2.T3)\. Based on the value assignment rules in Table[A2](https://arxiv.org/html/2607.20470#A2.T2), the detailed process of calculatingv​dvdis as follows:

\(1\) Obtaining F1 scores from Table[5](https://arxiv.org/html/2607.20470#S5.T5)\.

- •None:F​1=51\.09F1=51\.09
- •Pipeline\-based:F​1=51\.29F1=51\.29
- •Bidirectional:F​1=50\.36F1=50\.36

\(2\) Determining the mapping range\.

According to the rules, the value range is the integers from 1 to 10\. Therefore, the method with the lowest F1 score is mapped to11, and the method with the highest F1 score is mapped to1010:

- •v​dB​i​d​i​r​e​c​t​i​o​n​a​l=1vd\_\{Bidirectional\}=1
- •v​dP​i​p​e​l​i​n​e−b​a​s​e​d=10vd\_\{Pipeline\-based\}=10

\(3\) Computing the mapped value\.

Using linear proportional mapping:

v​dN​o​n​e=51\.09−50\.36\(51\.29−50\.36\)/9\+1\.vd\_\{None\}=\\frac\{51\.09\-50\.36\}\{\(51\.29\-50\.36\)/9\}\+1\.After rounding to the nearest integer:

v​dN​o​n​e≈8\.vd\_\{None\}\\approx 8\.
v​tvtandv​ivi’s calculation method is the same asv​dvd’s\.

Table A6:The coefficient of the DTI planner in the objective of performance, where the range of the variables’ mapping is from 1 to 10\. “new” denotes a new set of coefficients compared to the Table[3](https://arxiv.org/html/2607.20470#S4.T3)\.IDCombinationP\(%\)R\(%\)F1\(%\)Variablef​\(x\)d​a​t​a​s​e​tf\(x\)\_\{dataset\}f​\(x\)m​o​d​e​lf\(x\)\_\{model\}f​\(x\)d​a​t​a​s​e​tn​e​wf\(x\)^\{new\}\_\{dataset\}f​\(x\)m​o​d​e​ln​e​wf\(x\)^\{new\}\_\{model\}DataTuningInferencev​dvdv​tvtv​ivi1NoneSFTDirect50\.7651\.4151\.098101143\-769\-170\-5872SFT \+ GRPODirect49\.9550\.3650\.15811\-18\-370\-78\-2053SFT \+ DPODirect50\.8350\.7050\.76871152\-456\-75\-3914SFT \+ KTODirect51\.6650\.1150\.87881156\-540\-99\-4495Pipeline\-basedSFTDirect51\.3451\.2551\.2910101126\-861\-174\-7086Intersection63\.3045\.1452\.70101010227\-5401,2897Union45\.4958\.1051\.0310102154\-704\-127\-5418SFT \+ GRPODirect52\.3451\.9952\.171011\-73\-586\-116\-3589Intersection63\.5845\.3652\.941011016\-540\-1741,09210Union46\.1358\.1151\.431012\-46\-520\-99\-25111SFT \+ DPODirect52\.4250\.8951\.641071123\-589\-91\-52312Intersection63\.1046\.1053\.2810710220\-4331,29213Union44\.7859\.8351\.221072150\-463\-53\-37614Direct51\.7851\.6351\.701081131\-660\-112\-57715Intersection65\.4146\.4254\.301081022916421,29916SFT \+ KTOUnion45\.9059\.3251\.751082158\-523\-71\-42417BidirectionalSFTDirect50\.3950\.3350\.36110191\-842\-193\-49118Intersection60\.6146\.0552\.3411010142\-634\-587019Union45\.5958\.5851\.271102113\-757\-148\-39420SFT \+ GRPODirect51\.1150\.4850\.7911163\-918121Intersection60\.9545\.5752\.151110103\-611\-6581522Union45\.4558\.3851\.1111284\-14333723SFT \+ DPODirect50\.7250\.2250\.47171144\-384\-58\-25924Intersection62\.3243\.9651\.551710192\-4463992025Union43\.5258\.8150\.02172166\-330\-24\-18226SFT \+ KTODirect52\.5850\.7551\.65181133\-517\-96\-32827Intersection63\.3444\.9052\.551810182\-4893291128Union44\.7960\.8251\.59182155\-452\-58\-245

Table A7:The prediction results of constructing the DTI planner on the Qwen3\-8B in CMeIE\-V2\. The range of the variables’ mapping is from 1 to 10\.
## Appendix DDetailed Fitted Data of the DTI Planner on CMeIE\-V2 Dataset

This section presents the results of three LLMs \(i\.e\., GLM4\-9B, InternLM3\-8B, and LLaMA3\.1\-8B\) on CMeIE\-V2, as reported in Table[A4](https://arxiv.org/html/2607.20470#A2.T4)\. These results are used to solve the coefficients of the DTI planner\. Specifically, 1\) The “F1” is used to determine the optimal combination for single\-objective optimization; “F1 and Time” is used to further refine the optimal combination for multi\-objective optimization\. 2\) For each DTI combination, the values of the variablesv​dvd,v​tvt, andv​iviare calculated and substituted into the inequality formula[5](https://arxiv.org/html/2607.20470#S4.E5)\. 3\) The inequality solving results from three base\-LLMs are combined to determine the coefficients of the DTI plannerf​\(x\)d​a​t​a​s​e​tf\(x\)\_\{dataset\}\.

## Appendix EDetailed Fitted Data of the DTI Planner on Qwen3\-8B Model

This section presents the results of Qwen3\-8B on two different task datasets, EE \(i\.e\., ACE05\) and ABSA \(i\.e\., 14Lap\), as listed in Table[A5](https://arxiv.org/html/2607.20470#A2.T5)\. These results are used to solve the coefficients of the DTI planner\. Specifically, 1\) The “F1” is used to determine the optimal combination for single\-objective optimization; “F1 and Time” is used to further refine the optimal combination for multi\-objective optimization\. 2\) For each DTI combination, the values of the variablesv​dvd,v​tvt, andv​iviare calculated and substituted into the inequality formula[5](https://arxiv.org/html/2607.20470#S4.E5)\. 3\) The inequality solving results from both datasets are combined to determine the coefficients of the DTI plannerf​\(x\)m​o​d​e​lf\(x\)\_\{model\}\.

## Appendix FAnalyze the Non\-uniqueness of the Coefficient Solutions

As shown in Table[A6](https://arxiv.org/html/2607.20470#A3.T6), under the same search range, we can still stably obtain the optimal DTI combination ID 15 \(as shown in Table[A7](https://arxiv.org/html/2607.20470#A3.T7)\) by changing the coefficient solution of another set of planners\. This result indicates that the information conveyed by the scalar encoding is adequate and effective for the current task, and does not impose a limitation on the final performance\. In other words, the coefficients of the DTI planner are not a unique solution\.

Table A8:The coefficient of the DTI planner, where the range of the variables’ mapping is from 1 to 20\.IDCombinationTime\(s\)P\(%\)R\(%\)F1\(%\)Variablef​\(x\)d​a​t​a​s​e​tf\(x\)\_\{dataset\}f​\(x\)m​o​d​e​lf\(x\)\_\{model\}DataTuningInferencev​dvdv​tvtv​ivi1SFTDirect11,40550\.7651\.4151\.0916201\-235\-22912SFT \+ GRPODirect17,34249\.9550\.3650\.151611\-74\-863SFT \+ DPODirect11,16750\.8350\.7050\.761613194\-8374NoneSFT \+ KTODirect13,93151\.6650\.1150\.8716161\-9\-13685Direct68,22751\.3451\.2551\.2920201\-200\-20876Intersection70,02963\.3045\.1452\.7020202025642047SFTUnion70,02945\.4958\.1051\.0320204\-56\-14608Direct111,47052\.3451\.9952\.172011\-128\-939Intersection113,26263\.5845\.3652\.9420120\-646258810SFT \+ GRPOUnion113,26246\.1358\.1151\.432014\-138\-3511Direct77,64152\.4250\.8951\.642013196\-71112Intersection79,59363\.1046\.1053\.28201320193425013SFT \+ DPOUnion79,59344\.7859\.8351\.2220134184\-29314Direct94,87951\.7851\.6351\.70201617\-120915Intersection96,59365\.4146\.4254\.30201620259432216Pipeline\-basedSFT \+ KTOUnion96,59345\.9059\.3251\.7520164120\-70117Direct65,18550\.3950\.3350\.361201\-452\-299818Intersection66,80460\.6146\.0552\.3412020156390319SFTUnion66,80445\.5958\.5851\.271204\-283\-227420Direct110,83651\.1150\.4850\.7911143\-421Intersection112,49060\.9545\.5752\.151120\-324328722SFT \+ GRPOUnion112,49045\.4558\.3851\.111145715023Direct76,53950\.7250\.2250\.4711310\-125324Intersection78,21262\.3243\.9651\.5511320249431825SFT \+ DPOUnion78,21243\.5258\.8150\.021134112\-73926Direct88,75952\.5850\.7551\.651161\-155\-190927Intersection90,35263\.3444\.9052\.5511620248423128BidirectionalSFT \+ KTOUnion90,35244\.7960\.8251\.591164\-19\-1306

Table A9:The prediction results of constructing the DTI planner on the Qwen3\-8B in CMeIE\-V2, where the range of the variables’ mapping is from 1 to 20\.
## Appendix GAnalyze the Values of the Variables’ Mapping Range

This section represents the analysis of the values of the variable mapping range\. To further examine the robustness of the proposed PlanE framework with respect to the variable range, we extend the value range of the three variables from 1–10 to 1–20 and repeat all experiments under the same settings\. The extended experimental results are presented in Table[A8](https://arxiv.org/html/2607.20470#A6.T8)and Table[A9](https://arxiv.org/html/2607.20470#A6.T9), corresponding to Table[3](https://arxiv.org/html/2607.20470#S4.T3)and Table[5](https://arxiv.org/html/2607.20470#S5.T5), respectively\. The comparison shows that the optimal DTI configurations and overall performance trends remain consistent across both ranges\. In both the 1–10 and 1–20 variable ranges, the validation phase consistently identifies the same optimal configuration \(ID = 15\)\. This indicates that enlarging the variable range does not affect the experimental outcomes, demonstrating the stability and reliability of PlanE in capturing the underlying relationships between data decomposition, tuning method, and inference strategy\.

![Refer to caption](https://arxiv.org/html/2607.20470v1/x3.png)Figure A1:The prompt for direct inference on the CMeIE\-V2\.![Refer to caption](https://arxiv.org/html/2607.20470v1/x4.png)Figure A2:The prompt for pipeline\-based inference on the CMeIE\-V2\.
## Appendix HDifferent Reinforcement Learning Methods

This section represents the three reinforcement learning methods used in this paper\.

GRPOShaoet al\.\([2024](https://arxiv.org/html/2607.20470#bib.bib60)\): This method optimizes the model based on group\-relative rewards instead of pairwise comparisons\. The core idea is to jointly evaluate a batch of model responses and use a static scoring function to assess the quality of each output from two perspectives — format\-based and value\-based rewards\. The policy is then updated relative to the mean performance within the group, thereby improving training stability and convergence efficiency\.

𝒥GRPO​\(θ\)=𝔼\\displaystyle\\mathcal\{J\}\_\{\\text\{GRPO\}\}\(\\theta\)=\\mathbb\{E\}\[q∼P​\(Q\),\{oi\}i=1G∼πθold\]\\displaystyle\\left\[q\\sim P\(Q\),\\\{o\_\{i\}\\\}\_\{i=1\}^\{G\}\\sim\\pi\_\{\\theta\_\{\\text\{old\}\}\}\\right\]\(8\)1G​∑i=1G\\displaystyle\\frac\{1\}\{G\}\\sum\_\{i=1\}^\{G\}\(min\(πθ​\(oi\|q\)πθold​\(oi\|q\)Ai,\\displaystyle\\left\(\\min\\left\(\\frac\{\\pi\_\{\\theta\}\(o\_\{i\}\|q\)\}\{\\pi\_\{\\theta\_\{\\text\{old\}\}\}\(o\_\{i\}\|q\)\}A\_\{i\},\\right\.\\right\.clip​\(πθ​\(oi\|q\)πθold​\(oi\|q\),1−ϵ,1\+ϵ\)​Ai\\displaystyle\\text\{clip\}\\left\(\\frac\{\\pi\_\{\\theta\}\(o\_\{i\}\|q\)\}\{\\pi\_\{\\theta\_\{\\text\{old\}\}\}\(o\_\{i\}\|q\)\},1\-\\epsilon,1\+\\epsilon\\right\)A\_\{i\}−βDK​L\(πθ∥πref\)\),\\displaystyle\\left\.\\vphantom\{\\frac\{\\pi\_\{\\theta\}\}\{\\pi\_\{\\theta\}\}\}\-\\beta D\_\{KL\}\(\\pi\_\{\\theta\}\\\|\\pi\_\{\\text\{ref\}\}\)\\right\),where𝔼​\[q∼P​\(Q\),\{oi\}i=1G∼πθold​\(O\|q\)\]\\mathbb\{E\}\\left\[q\\sim P\(Q\),\\\{o\_\{i\}\\\}\_\{i=1\}^\{G\}\\sim\\pi\_\{\\theta\_\{\\text\{old\}\}\}\(O\|q\)\\right\]denotes the expectation where the questionqqis sampled from the question distributionP​\(Q\)P\(Q\), and the output group\{oi\}i=1G\\\{o\_\{i\}\\\}\_\{i=1\}^\{G\}consists ofGGcandidate responses sampled from the old policyπθold\\pi\_\{\\theta\_\{\\text\{old\}\}\}\.GGrepresents the number of responses sampled for each question,πθ​\(oi\|q\)\\pi\_\{\\theta\}\(o\_\{i\}\|q\)denotes the probability of generating tokenoio\_\{i\}given questionqqunder the current policyπθ\\pi\_\{\\theta\}, andπθold​\(oi\|q\)\\pi\_\{\\theta\_\{\\text\{old\}\}\}\(o\_\{i\}\|q\)denotes the probability of generating tokenoio\_\{i\}given questionqqunder the old policyπθold\\pi\_\{\\theta\_\{\\text\{old\}\}\}\.ϵ\\epsilonis a clipping hyperparameter that limits the magnitude of policy updates,β\\betarepresents the coefficient of the KL divergence penalty term,πref\\pi\_\{\\text\{ref\}\}is the reference model, andDK​LD\_\{KL\}is the KL regularization term that constrains the policy from deviating excessively from the reference model\. Finally,Ai\{A\}\_\{i\}denotes the advantage estimate of theii\-th output, computed based on the group\-relative rewards\.

DPORafailovet al\.\([2023](https://arxiv.org/html/2607.20470#bib.bib1)\): This method simplifies RLHF by eliminating the need for a reward model\. Instead, it directly optimizes the policy toward preferred responses based on preference pairs\. The core idea is to align the model distributionπθ\\pi\_\{\\theta\}with the reference policyπref\\pi\_\{\\text\{ref\}\}using a logistic objective that favors higher\-probability preferred outputs\.

ℒDPO\\displaystyle\\mathcal\{L\}\_\{\\text\{DPO\}\}\(πθ;πref\)=−𝔼\(x,yw,yl\)∼𝒟\[logσ\(\\displaystyle\(\\pi\_\{\\theta\};\\pi\_\{\\text\{ref\}\}\)=\-\\mathbb\{E\}\_\{\(x,y\_\{w\},y\_\{l\}\)\\sim\\mathcal\{D\}\}\\Bigl\[\\log\\sigma\\Bigl\(\(9\)βlogπθ​\(yw\|x\)πref​\(yw\|x\)−βlogπθ​\(yl\|x\)πref​\(yl\|x\)\)\],\\displaystyle\\beta\\log\\frac\{\\pi\_\{\\theta\}\(y\_\{w\}\|x\)\}\{\\pi\_\{\\text\{ref\}\}\(y\_\{w\}\|x\)\}\-\\beta\\log\\frac\{\\pi\_\{\\theta\}\(y\_\{l\}\|x\)\}\{\\pi\_\{\\text\{ref\}\}\(y\_\{l\}\|x\)\}\\Bigr\)\\Bigr\],
whereπθ\\pi\_\{\\theta\}denotes the trainable policy model, andπref\\pi\_\{\\text\{ref\}\}represents the reference model that serves as a baseline policy\.𝒟\\mathcal\{D\}denotes the dataset of human preference pairs\(x,yw,yl\)\(x,y\_\{w\},y\_\{l\}\), wherexxis the given input \(e\.g\., a prompt or question\),ywy\_\{w\}is the preferred \(winning\) response, andyly\_\{l\}is the less\-preferred \(losing\) response\.β\\betais a temperature coefficient controlling the sharpness of the preference contrast\.σ​\(⋅\)\\sigma\(\\cdot\)represents the sigmoid function, converting the logit difference into a probability\.

KTOEthayarajhet al\.\([2024](https://arxiv.org/html/2607.20470#bib.bib2)\): This method is inspired by behavioral economics, particularlyKahneman–Tversky Prospect Theory, introducing asymmetric weighting for positive and negative feedback\. The core idea is to encourage the model to be risk\-averse for negative outcomes and risk\-seeking for positive outcomes, improving controllability and alignment stability\.

ℒKTO​\(πθ,πref\)=𝔼x,y∼𝒟​\[λy−v​\(x,y\)\],\\mathcal\{L\}\_\{\\text\{KTO\}\}\(\\pi\_\{\\theta\},\\pi\_\{\\text\{ref\}\}\)=\\mathbb\{E\}\_\{x,y\\sim\\mathcal\{D\}\}\\left\[\\lambda\_\{y\}\-v\(x,y\)\\right\],\(10\)
v​\(x,y\)=\{λD​σ​\(β​\(rθ​\(x,y\)−z0\)\),if​y∼yd\|xλU​σ​\(β​\(z0−rθ​\(x,y\)\)\),if​y∼yu\|xv\(x,y\)\\\!=\\\!\\begin\{cases\}\\lambda\_\{D\}\\sigma\(\\beta\(r\_\{\\theta\}\(x,y\)\\\!\-\\\!z\_\{0\}\)\),&\\text\{if \}y\\sim y\_\{\\text\{d\}\}\|x\\\\ \\lambda\_\{U\}\\sigma\(\\beta\(z\_\{0\}\\\!\-\\\!r\_\{\\theta\}\(x,y\)\)\),&\\text\{if \}y\\sim y\_\{\\text\{u\}\}\|x\\end\{cases\}\(11\)whereλD\\lambda\_\{D\}andλU\\lambda\_\{U\}are hyperparameters that weight desirable and undesirable outputs, respectively;σ​\(⋅\)\\sigma\(\\cdot\)denotes the sigmoid function;β\\betais the temperature coefficient controlling the scaling of the reward difference;rθ​\(x,y\)r\_\{\\theta\}\(x,y\)is the reward function, andz0z\_\{0\}is the KL divergence between the two policies\.ydy\_\{\\text\{d\}\}andyuy\_\{\\text\{u\}\}indicate the desirable and undesirable outputs for a given inputxx, respectively\.

## Appendix IPrompt Examples of Inference

This section represents the prompt examples of inference on three datasets\. Specifically, the prompt for direct inference on CMeIE\-V2 is shown in Figure[A1](https://arxiv.org/html/2607.20470#A7.F1), the pipeline\-based inference prompt in Figure[A2](https://arxiv.org/html/2607.20470#A7.F2), and the bidirectional inference prompt in Figure[A3](https://arxiv.org/html/2607.20470#A9.F3)\. For the ACE05 dataset, the prompts for direct, pipeline\-based, and bidirectional inference are shown in Figures[A4](https://arxiv.org/html/2607.20470#A9.F4),[A5](https://arxiv.org/html/2607.20470#A9.F5), and[A6](https://arxiv.org/html/2607.20470#A9.F6), respectively\. Similarly, for the 14Lap dataset, the prompts for direct, pipeline\-based, and bidirectional inference are presented in Figures[A7](https://arxiv.org/html/2607.20470#A9.F7),[A8](https://arxiv.org/html/2607.20470#A9.F8), and[A9](https://arxiv.org/html/2607.20470#A9.F9), respectively\.

![Refer to caption](https://arxiv.org/html/2607.20470v1/x5.png)Figure A3:The prompt for bidirectional inference on the CMeIE\-V2\.![Refer to caption](https://arxiv.org/html/2607.20470v1/x6.png)Figure A4:The prompt for direct inference on the ACE05\.![Refer to caption](https://arxiv.org/html/2607.20470v1/x7.png)Figure A5:The prompt for pipeline\-based inference on the ACE05\.![Refer to caption](https://arxiv.org/html/2607.20470v1/x8.png)Figure A6:The prompt for bidirectional inference on the ACE05\.![Refer to caption](https://arxiv.org/html/2607.20470v1/x9.png)Figure A7:The prompt for direct inference on the 14Lap\.![Refer to caption](https://arxiv.org/html/2607.20470v1/x10.png)Figure A8:The prompt for pipeline\-based inference on the 14Lap\.![Refer to caption](https://arxiv.org/html/2607.20470v1/x11.png)Figure A9:The prompt for bidirectional inference on the 14Lap\.