Structural analysis filters noisy agent traces to pinpoint root causes

STRACE framework combines batch-level failure pattern mining and per-trajectory causal localization over textual dependency graphs to improve reflection-based agent optimization.

PaperBig Techcs.CLarXiv:2607.07702v1 ↗

Ying Chang · Jiahang Xu · Xuan Feng · Chenyuan Yang · Peng Cheng · Yuqing Yang

University of Chinese Academy of Sciences · Microsoft Research

Research Digest·9 July 2026·1 min read

Read the paper →

Thread:Credit Assignment in Agentic RL

The authors introduce STRACE, a framework that constructs high-signal optimization contexts for long-horizon agents by removing redundant and non-causal steps from execution traces. On a formal verification benchmark, STRACE boosted the success rate of human-expert designed agents from 42.5% to 58.5%, a 1.4× improvement over standard context-filtering baselines.

What they did

The authors developed STRACE to address the problem that real execution traces of long-horizon agents contain many irrelevant or redundant steps, making it difficult for LLM-based optimizers to diagnose failures effectively. STRACE operates at two levels: at the batch level, it mines failure patterns across many trajectories to filter out redundant traces and retain only representative failures; at the individual trajectory level, it builds a textual dependency graph over steps, then performs causal localization to remove non-causal steps and identify the true root-cause module for optimization.

The framework was evaluated on VeruSAGE-Bench, a challenging formal verification task, using human-expert designed agents. STRACE was compared against standard context filtering baselines, including naive truncation and sliding window approaches.

Analysis

This work addresses a practical bottleneck in the increasingly popular paradigm of using LLMs to critique and improve other LLM-based agents. By framing trace selection as a causal extraction problem, the authors move beyond naive context-reduction heuristics. The 1.4× gain on a human-expert baseline is notable, suggesting that even well-designed agents can benefit from better optimization signals. The open-source release of the code facilitates further research.