@Azaliamirh: Check out TRACE, a new self-improvement approach where the agent identifies the missing capabilities behind its own fai…
Summary
TRACE is a new self-improvement approach where an AI agent identifies the missing capabilities behind its own failures and trains itself to address them. TRACE-trained Qwen3.6-27B achieves 73.2% on SWE-bench Verified, outperforming much larger models with fewer training rollouts.
View Cached Full Text
Cached at: 07/10/26, 12:11 PM
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.
By contrasting successful and failed trajectories, TRACE identifies its own weaknesses (such as bug localization or retrieval of the correct doc) and creates new synthetic environments to fix them.
The result is a transferable and sample-efficient synthetic env / data generation + fine-tuning pipeline for agentic tasks.
Great work led by @TarunSures41845 and @hangoo_kang!
Hangoo Kang @ ICML ✈️ (@hangoo_kang): “TRACE: Capability-Targeted Agentic Training” got Spotlight @ ICML AIWILD 🎉
Beats direct RL, GEPA, & synthetic-agent data on SWE-Bench Verified and τ²-Bench.
TRACE-Qwen3.6-27B tops GPT-5.2-Codex, GLM 5, & Claude 4.5 Sonnet on SWE-Bench.
Co-led with @TarunSures41845. Thanks to
Similar Articles
@ADarmouni: https://arxiv.org/pdf/2607.13988 Good RL work from Microsoft research that manages to improve the Qwen3 small MoE bette…
This paper introduces TRACE, a dense credit-assignment method for reinforcement learning of long-horizon agents that significantly improves Qwen3 small MoE models on agentic benchmarks without additional critic models.
qwen agentworld can self-correct in reasoning traces
A user experiments with Qwen AgentWorld and finds a system prompt that enables self-correction in reasoning traces, as demonstrated by the classic car wash test.
@Vtrivedy10: https://x.com/Vtrivedy10/status/2066571435871551655
A joint study by LangChain Labs and Fireworks AI demonstrates fine-tuning an open Qwen model to create a trace judge that detects 'perceived error' in production traces, achieving frontier performance at up to 100x lower cost. The model is evaluated on two internal datasets and shows generality across applications.
Trace: A Taxonomy-Guided Environment for Multidomain Visual Reasoning
TRACE is a taxonomy-guided environment with 1,000 visual reasoning tasks across 11 domains. Training Qwen2.5-VL-3B and Qwen2.5-VL-7B on 64,000 TRACE instances improves their macro-average performance across 24 external benchmarks by 3.51 and 4.06 percentage points respectively.
@heyshrutimishra: 90% of AI progress is an illusion built on benchmarks that reward memorization over thinking. TRACES strips that away w…
TRACES is a new AI benchmark by Apodex designed to measure discoverative AI by focusing on the problem-solving process rather than memorized answers, challenging traditional benchmarks.