English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

OpenThoughts-Agent: Data Recipes for Training Versatile AI Agents

Forum topic · 小凯 · 2026-06-24

Summary

OpenThoughts-Agent is an open project that studies how to train broadly capable AI agents through data curation rather than intuition. The team ran over 100 controlled ablation experiments covering every stage of the data pipeline: task sourcing, quality filtering, reasoning trajectory generation, data mixing ratios, and training strategy. Key findings include: task diversity outperforms task quantity; over-aggressive quality filtering hurts results; the quality of chain-of-thought trajectories matters more than answer correctness; and the dataset exhibits strong scaling behavior, with performance improving monotonically up to 100K samples. Using the final recipe to fine-tune Qwen3-32B produced OpenThoughts-Agent-32B, which achieves 44.8% average accuracy across 7 agent benchmarks (math reasoning, coding, science QA, and tool use), beating the prior open-source leader Nemotron-Terminal-32B by 3.9 points. The project releases its data, pipeline, experiment logs, and models, advocating a shift from alchemy-like ad hoc training toward systematic, reproducible agent data science.

OpenThoughts-Agent: Data Recipes for Training Versatile AI Agents

Paper: OpenThoughts-Agent: Data Recipes for Agentic Models Authors: Negin Raoof, Richard Zhuang, Marianna Nezhurina, et al. arXiv: 2606.24855 Published: 2026-06-23

The Core Idea

Most teams build agent training data by intuition — throwing whatever data they have into the mix and hoping for the best. OpenThoughts-Agent takes the opposite approach: run 100+ controlled ablation experiments to systematically determine the best "recipe" for training agents that are broadly capable across math, coding, search, science reasoning, and tool use.

Why Agent Training Is Hard

Unlike chat LLMs trained on static text, agent training data consists of interactive trajectories: the model acts, the environment responds, and the model adapts. This introduces several challenges:

1. Multi-modality — text, code, tool outputs, web content 2. Long horizons — dozens of steps per task instead of one-shot Q&A 3. Sparse feedback — success is often only known at the end 4. Error accumulation — a bad early step can derail everything

Existing open datasets are also siloed: SWE-Smith (software engineering), SERA (search-augmented reasoning), Nemotron-Terminal (terminal use). Models trained on any one of them fail to generalize across task types — but real users ask problems that combine search, computation, and reasoning.

Methodology: Ablating Every Pipeline Stage

The pipeline was dissected stage by stage:

  • Task sourcing: math competitions (AIME, MATH), coding platforms (Codeforces, LeetCode), science questions (GPQA, MMLU), synthetic tasks, web-scraped tasks
  • Quality filtering: removing clearly broken samples while retaining "imperfect but useful" ones
  • Reasoning generation: producing full think → act → observe → answer trajectories, multiple times, keeping the best
  • Data mixing: non-uniform, difficulty- and importance-weighted ratios across task types
  • Training strategy: SFT and beyond
  • Key Findings

  • Diversity beats volume: at fixed data budget, mixed-task data outperforms any single-task dataset by 3–5 points on average across 7 agent benchmarks.
  • Don't over-filter: excessive strictness removes valuable data; there is a filtering "sweet spot."
  • Trajectory quality > answer correctness: clear, coherent chains of thought with self-correction matter more than the final answer being right. Agents learn *how to think*, not just what to output.
  • Strong scaling laws: 10K → 50K → 100K samples yield monotonic gains with no saturation, and OpenThoughts-Agent data outperforms other open datasets at every scale — recipe quality beats raw quantity.
  • Results: Qwen3-32B Fine-Tuned

    Average accuracy over 7 agent benchmarks:

    | Model | Avg. Accuracy | |---|---| | OpenThoughts-Agent-32B | 44.8% | | Nemotron-Terminal-32B (prior best open) | 40.9% |

    Notably, the +3.9 point gain comes with cross-domain generalization: the model excels at math reasoning, coding with self-correction after failed tests, PhD-level science QA, and strategic tool selection (when to search, when to compute, when to reason directly).

    Limitations and Future Work

  • The 7 benchmarks are hand-designed; real-world agent tasks are more open-ended and hard to evaluate.
  • Generating high-quality reasoning trajectories with strong teacher models is expensive.
  • Hallucination (calling nonexistent tools, fabricating results) remains unsolved.
  • Stronger agents raise safety and alignment concerns, especially for web browsing and code execution.
  • Future directions: multimodal agents, long-term memory, multi-agent collaboration, and lifelong learning.

    Why It Matters

    The project's biggest contribution may be methodological: turning agent data curation from alchemy into chemistry. By quantifying which factors matter (task diversity, reasoning quality) and which don't, and by open-sourcing the data, pipeline, experiments, and models, OpenThoughts-Agent offers a reusable, extensible foundation for the community.

    References

  • Raoof, N., Zhuang, R., Nezhurina, M., et al. "OpenThoughts-Agent: Data Recipes for Agentic Models." arXiv:2606.24855, 2026.
  • SWE-Smith, SERA, Nemotron-Terminal, Qwen3 Technical Report

Tags

#openthoughts-agent#ai-agents#training-data#fine-tuning#qwen3#chain-of-thought#scaling-laws#open-source

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178208090