All Topics
All Topics
Technology
Technology
AI
AI
Business
Business
Entertainment
Entertainment
News
News
Programming
Programming
Science
Science
Design
Design
Environment
Environment
Finance
Finance
Crypto
Crypto
Politics
Politics
Sports
Sports
Education
Education
Gaming
Gaming
Art
Art
Music
Music
Health
Health
Security
Security
Books
Books
Food
Food
Travel
Travel
Personal
Personal
Bluesky
Twitter

TRACE: A New System for Capability-Targeted Agentic Training via Automated Failure Analysis

By

[Submitted on 7 Apr 2026 (v1), last revised 2 Jul 2026 (this version, v2)]

2h ago· 2 min readenInsight

Summary

TRACE (Turning Recurrent Agent failures into Capability-targeted training Environments) is an end-to-end system for improving AI agent performance by automatically identifying missing capabilities from failed trajectories, synthesizing targeted training environments, and training LoRA adapters via reinforcement learning. The system uses a mixture-of-experts approach over capability adapters and demonstrates significant improvements on benchmarks: +15.3 points on τ²-Bench (customer-service) and +15.0 points on SWE-Bench Verified (software engineering), outperforming baselines like GEPA and SWE-RL while being more sample-efficient.

Source

bskyTRACE: A New System for Capability-Targeted Agentic Training via Automated Failure Analysisarxiv.org

Key quotes

· 5 pulled
Models often fail to complete agentic tasks because they lack core capabilities required by the target environment.
TRACE contrasts successful and failed trajectories to automatically identify missing capabilities, synthesizes a targeted training environment for each capability that rewards whether the capability is exercised.
TRACE can be effectively applied across different environments, improving over the base agent by +15.3 points on τ²-Bench, a customer-service agent benchmark, and by +15.0 points Pass@1 on SWE-Bench Verified, a software-engineering benchmark.
TRACE outperforms the strongest external baselines, GEPA and SWE-RL, by +8.6 points and +8.4 points, respectively.
Using fewer than one-fourth the number of rollouts, TRACE outperforms the best-performing baselines, GRPO and GEPA, and achieves higher final accuracy by +10.4 and +8.6 points on τ²-Bench.
Snippet from the RSS feed
Models often fail to complete agentic tasks because they lack core capabilities required by the target environment. However, mainstream approaches for addressing these failures typically either fine-tune directly on target environments or generate synthet

You might also wanna read

Comments

Sign in to join the conversation.

No comments yet. Be the first.