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!
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.
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.
Why it matters
TRACE lets a smaller model reach 73.2% on SWE-benchThe 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.Full definition → by training on its own diagnosed capability gaps, beating larger models like Codex and GLM with a quarter of the training rollouts other RL methods need.
articleOpenDiscoveryTrace: Process Traces for Evaluating AI Scientist WorkflowsAayam Bansal, Keertan Balaji
articleFantastic Adaptive Taxonomies and How to Use ThemMert Cemri, Andrei Cojocaru, Melissa Pan, Shu Liu, Shubham Agarwal, Alexander Krentsel, Jay Tang, Kannan Ramchandran, Joseph E. Gonzalez, Matei Zaharia, Alex Dimakis, Ion Stoica
postTraining models to explain their own failures, then rewarding the explanationWRITER