From Noisy Traces to Root Causes: Structural Trajectory Analysis and Causal Extraction for Agent Optimization

2026-07-08 · Source: Computation and Language · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, quick

Summary

STRACE (Structural TRajectory Analysis and Causal Extraction) is a novel framework designed to optimize long-horizon agents by constructing high signal-noise optimization contexts. It addresses challenges in using real execution traces for reflection-based LLM optimizers, which often suffer from redundancy, heterogeneity, and irrelevant steps. STRACE operates on two levels: at the batch level, it identifies and filters redundant traces by mining failure patterns, retaining only representative failures. Within selected traces, it employs causal localization on a textual dependency graph to eliminate non-causal steps and pinpoint the precise root-cause module for optimization. Empirical evaluations demonstrate STRACE's superior performance over standard context-filtering baselines. Notably, on the challenging VeruSAGE-Bench formal verification task, STRACE improved success rates for human-expert designed agents by 1.4×, from 42.5% to 58.5%. The code was made available on 2026-07-08 at https://github.com/moomight/STRACE.

Key takeaway

For Machine Learning Engineers optimizing long-horizon agents with LLM reflection, you should consider integrating STRACE to refine your agent's learning process. This framework helps you filter redundant execution traces and precisely identify root-cause failures, significantly boosting optimization efficiency and agent success rates. Implementing STRACE can lead to more robust and effective agent policies, as demonstrated by its 1.4× success-rate improvement on complex tasks.

Key insights

STRACE improves LLM agent optimization by filtering noisy traces and causally identifying root causes, enhancing signal-to-noise.

Principles

• Redundant traces hinder LLM agent optimization.
• Causal localization identifies true failure root causes.
• High signal-noise contexts improve agent learning.

Method

STRACE constructs optimization contexts by first mining failure patterns to filter redundant traces, then performing causal localization on a textual dependency graph within selected traces to identify root-cause modules.

In practice

• Use STRACE to refine LLM agent training data.
• Apply causal localization for debugging agent failures.
• Filter redundant execution traces for efficiency.

Topics

• LLM Agent Optimization
• Causal Extraction
• Trajectory Analysis
• Formal Verification
• VeruSAGE-Bench
• Reflection Mechanisms

Code references

• moomight/STRACE

Best for: AI Scientist, Machine Learning Engineer

Related on AIssential

• Retrospective Harness Optimization: Improving LLM Agents via Self-Preference over Trajectory Rollouts — AI agents can self-improve their operational harness by retrospectively analyzing past unlabeled trajectories.
• STT-Arena: A More Realistic Environment for Tool-Using with Spatio-Temporal Dynamics — LLMs struggle with adaptive replanning under spatio-temporal dynamics, even advanced models.
• What Resolve Rate Hides: Trajectory Structure Diagnostics for Coding Agents — TraceProbe diagnoses coding agent behavior by normalizing trajectories into canonical actions and detecting anti-patterns or divergences.
• Your Agent Has a Genome: Sequence-Level Behavioral Analysis and Runtime Governance of LLM-Powered Autonomous Agents — Encoding agent actions into XEPV sequences enables quantitative behavioral analysis and runtime governance.
• Open-SWE-Traces: Advancing Dual-Mode Multilingual Distillation for Software Engineering Agents — Open-SWE-Traces provides a large, dual-mode multilingual dataset to train agentic LLMs for autonomous software engineering.
• ToolChoiceConfusion: Causal Minimal Tool Filtering for Reliable LLM Agents — Reliable LLM agent tool selection hinges on causal necessity, not just semantic relevance or executability.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computation and Language.