Source note
Yet Even Less Is Even Better For Agentic, Reasoning, and Coding LLMs
Summary
This paper argues that software-engineering agents can improve more from a small set of filtered, high-value trajectories than from a large pool of raw trajectories. It introduces STITCH, a trajectory curation method, plus SandForge, a task construction and evaluation pipeline, and reports gains on SWE-bench-style coding tasks across Python, Java, and ArkTS.
Problem
- Training coding and agentic LLMs often needs many full trajectories, which are expensive to collect and clean.
- Raw agent traces contain a lot of low-value tokens, repeated actions, and broken context segments, which weakens supervised fine-tuning.
- Prior work focused more on scaling data volume than on finding which parts of trajectories actually carry the training signal, especially for larger models and multilingual software tasks.
Approach
- The paper builds an end-to-end pipeline called SandForge that turns real software repair records from GitHub into executable tasks, runs agents on them, and stores trajectories, patches, verifier outputs, rewards, and metadata.
- It proposes STITCH (
Sliding-memory Trajectory Inference and Task Chunking Heuristic), a two-stage filter for trajectory data. - In the first stage, STITCH uses automatically discovered trajectory features plus logistic regression to screen out weak trajectories based on signals such as code edits, tool usage, efficiency, and recovery behavior.
- In the second stage, it uses an LLM judge to split long trajectories into semantically safe chunks, carries forward a compressed memory summary between chunks, scores local segments, and keeps decision-critical sub-trajectories even when the full run is mediocre.
- The core idea is simple: keep the steps where the agent makes useful decisions or code changes, drop noisy parts, and fine-tune on that smaller but cleaner dataset.
Results
- On SWE-bench Verified, models trained with STITCH show up to 63.16% relative improvement over their base models.
- On Multi-SWE-bench (Java), MiniMax-M2.5-STITCH reaches 43.75% with the CodeArts Agent scaffold, a gain of +16.67% over the comparison stated in the paper.
- On HarmonyOS (ArkTS), GLM-4.7-STITCH raises the compilation pass rate to 61.31%, with an improvement of +43.34%.
- The paper states that these gains hold across agent frameworks such as mini-SWE-agent and MSWE-agent, model sizes from 30B to 355B, and multilingual settings.
- For the ArkTS setting, the paper says the model used less than 1K training trajectories.
- The excerpt does not provide full benchmark tables, variance, or detailed baseline names for every experiment, so the strongest concrete claims are the percentages above.