Research map / Multi-agent systems and games / LLM agents

FT-Dojo: Towards Autonomous LLM Fine-Tuning with Language Agents

ICML 2026 · Multi-agent systems and games › LLM agents · approach: Self-evolving language agent learning · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces FT-Dojo, an interactive benchmark of 13 LLM fine-tuning tasks across 5 domains, and FT-Agent, a fine-tuning-oriented agent with structured planning, fail-fast validation, and feedback analysis that beats OpenHands and other baselines on 10 of 13 tasks.

Problem

End-to-end LLM fine-tuning (data construction, training configuration, diagnosis from feedback) has not been systematically studied as an interactive agent task, and general-purpose coding agents perform poorly on it.

Approach

Frames autonomous fine-tuning as a sandboxed interactive environment where data strategy and training config are joint optimization targets, and proposes FT-Agent with a three-stage loop (structured strategy proposal, progressive fail-fast validation, multi-level feedback analysis) on top of a general LLM planner.

Key results

• FT-Agent achieves the best performance on 10 of 13 tasks against Base Model, Manual SFT, and tool-augmented OpenHands.
• On AIME 2025, FT-Agent reaches 11.11% accuracy (3-run mean) while the base model, OpenHands, and Manual SFT all score 0%.
• On the five-task controlled comparison, FT-Agent with GPT-5.2 reaches a 42.83 average versus 34.55 for OpenHands and 36.06 for Codex; Claude Code averages 39.86.

Surprising findings

• General-purpose OpenHands in some tasks falls below the un-tuned base model.
• Claude Code (Sonnet-4.6) scores higher than FT-Agent on average in the five-task comparison, and Claude Code leads on PI4PC and Visualization.

Limitations

• Manual SFT baseline is constrained to a two-pass protocol and is not an expert upper bound.
• Multi-run statistics only for FT-Agent; baselines are single runs; limited to Qwen2.5-7B and 3B target models.
• Agents show limited causal diagnosis and long-horizon planning (shotgun debugging on PI4PC).

Experimental setup

Datasets: DeepScaleR PANORAMA ChemCoTDataset FinanceIQ TableInstruct AIME 2025 TableBench ChemCoTBench

Benchmarks: AIME 2025 PANORAMA (PAR4PC, NOC4PC, PI4PC) TableBench (DA, FC, NR, Vis) FinanceIQ ChemCoTBench (Mol_Und, Mol_Edit, Mol_Opt, Reaction)

Metrics: Accuracy Macro-F1 Pass@1 MAE Tanimoto similarity success rate valid ratio BLEU FTS

Baselines: Base Model (Qwen2.5-7B-Instruct) Manual SFT OpenHands Codex Claude Code

Builds on: OpenHands LlamaFactory OpenCompass MLE-bench

Closest papers by meaning

• MLE-Dojo: Interactive Environments for Empowering LLM Agents in Machine Learning Engineering (NeurIPS 2025 · similarity 0.91)
• Offline Training of Language Model Agents with Functions as Learnable Weights (ICML 2024 · similarity 0.90)
• AgentRefine: Enhancing Agent Generalization through Refinement Tuning (ICLR 2025 · similarity 0.89)
• Enhancing Decision-Making for LLM Agents via Step-Level Q-Value Models (AAAI 2025 · similarity 0.89)
• Reinforce LLM Reasoning through Multi-Agent Reflection (ICML 2025 · similarity 0.89)
• Coevolving with the Other You: Fine-Tuning LLM with Sequential Cooperative Multi-Agent Reinforcement Learning (NeurIPS 2024 · similarity 0.89)