Research map / Large language models / Instruction and fine-tuning
Toward Effective Tool-Integrated Reasoning via Self-Evolved Preference Learning
Yifei Chen, Guanting Dong, Zhicheng Dou
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add pre-aligned DPO after SFT (ablation) | Tool-Light (Qwen2.5-7B-Instruct) · 10 reasoning tasks average | average score: 56.6 -> 57.4 | Improved |
| add self-evolved DPO alignment (2 loops) after pre-aligned DPO (ablation) | Tool-Light (Qwen2.5-7B-Instruct) · 10 reasoning tasks average | average score: 57.4 -> 58.0 | Improved |
| increase self-evolved loops from 2 to 3 (ablation) | Tool-Light · 10 reasoning tasks average | performance: 58.0 -> 56.1 | Hurt |
| increase self-evolved loops from 2 to 5 (ablation) | Tool-Light · 10 reasoning tasks average | performance: 58.0 -> 54.1 | Hurt |
| use 2 loops instead of 1 loop (ablation) | Tool-Light · 10 reasoning tasks average | performance: 57.9 -> 58.0 | No clear effect |
| random positive example selection instead of min-tool-call, low-entropy criterion (ablation) | Tool-Light · 10 reasoning tasks average | performance: 58.0 -> 53.6 | Hurt |
| random negative example selection (ablation) | Tool-Light · 10 reasoning tasks average | performance: 58.0 -> 53.9 | Hurt |
| mix vanilla and entropy-guided data 1:1 instead of the tuned ratio (ablation) | Tool-Light · 10 reasoning tasks average | performance: 58.0 -> 56.9 | Hurt |
| multi-tool-integrated training (Tool-Light/Tool-Star) rather than single-tool training | Qwen2.5-7B-Instruct · math and knowledge tasks | average score: 50.4 -> 58.0 | Improved |
| fine-tuning instead of prompt engineering for tool use | Qwen2.5-7B-Instruct · 10 tasks average | average score: 35.3 -> 56.6 | Improved |
| introduce external tools via prompting only | Qwen2.5-7B-Instruct · knowledge-intensive tasks | F1: 33.0 -> 35.3 | Mixed |
| entropy-guided sampling to learn low-entropy paths | Tool-Light · output entropy analysis | output entropy | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- ReTool — ReTool: Reinforcement Learning for Strategic Tool Use in LLMs (ICLR 2026) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Advancing Tool-Augmented Large Language Models: Integrating Insights from Errors in Inference Trees (NeurIPS 2024 · similarity 0.90)
- Nemotron-Research-Tool-N1: Exploring Tool-Using Language Models with Reinforced Reasoning (ICLR 2026 · similarity 0.90)
- Are Tools Always Beneficial? Learning to Invoke Tools Adaptively for Dual-Mode Multimodal LLM Reasoning (ICML 2026 · similarity 0.89)
- PORTool: Importance-Aware Policy Optimization with Rewarded Tree for Multi-Tool-Integrated Reasoning (NeurIPS 2026 · similarity 0.88)
- Joint Optimization of Tool Creation and Use for Large Language Model Agents (NeurIPS 2026 · similarity 0.88)
- Incentivizing Agentic Reasoning in LLM Judges via Tool-Integrated Reinforcement Learning (ICLR 2026 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
