Intel / post

TRACE: An Agent That Trains Itself on Its Own Failure Modes

Source: x.com
Date: Published Jul 9, 2026

Check out TRACE, a new self-improvement approach where the agent identifies the missing capabilities behind its own failures and trains itself to address them. TRACE-trained Qwen3.6-27B reaches 73.2% on SWE-bench Verified, outperforming much larger models like Codex 5.2 and GLM 5, while beating GRPO and GEPA with <1/4 the training rollouts. Exciting work led by @hangoo_kang and @TarunSures41845!

Jul 9, 2026 View on X

Context

is a widely used test of whether an AI coding can fix real, previously reported bugs in real open-source codebases, so a strong score there is a proxy for practical bug-fixing ability. Stanford AI Lab reports a training method called TRACE that let a comparatively small open model beat several larger, more expensive ones on this benchmark.

TRACE works by having the agent compare its own successful and failed attempts at a task to identify which specific it is missing, such as locating the right file or reading the correct documentation, and then generating targeted practice environments that exercise just that missing skill, instead of running broad reinforcement learning across every task equally.

Trained this way, the Qwen3.6-27B model reached 73.2% pass rate on SWE-bench Verified, ahead of larger frontier systems including Codex 5.2 and GLM 5 on the public leaderboard, and it beat the GRPO and GEPA training methods using under a quarter of their training rollouts. These are Stanford's own reported figures on one benchmark; they show what TRACE achieved on this specific coding-fix test, not a general claim that the resulting model outperforms those systems on every task.

Terms in this piece · Glossary

SWE-bench — The standard benchmark for AI coding agents: real GitHub issues from real repositories, scored by whether the agent's patch passes the project's own tests.
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
agent skill — A reusable instruction file that teaches an agent how to do one job well — the procedure, the tools, and what counts as done.
open weights — A model whose trained parameters are published for anyone to download and run — unlike API-only models you can access but never possess.

Read the source x.com