DAAF addresses the gap between failure localization and repair in LLM agents by attributing failures to editable system assets like routing rules, knowledge segments, or prompt instructions. It learns the expected effect of replacing each asset from controlled replays evaluated by task outcomes. On held-out Telecom tasks, DAAF achieves 80.72% attribute Hit@1, recovering 62.65% of failed executions while limiting clean-task regression to 3.23%, and reaching 71.93% overall task success. This shows that intervention-grounded attribution can connect localization to executable system repair without requiring counterfactual replay at diagnosis time.
A method that links failure localization to repair by identifying which editable system asset to change
The Detection-Aware Attribution Framework learns from controlled replays to predict asset edits that improve task outcomes
Chinese Tech
Xiaoyang Yuan · Qi Liu · Yubin Ruan · Xinyi Mou · Zhuomeng Zhang · Wenjin Wang · +6 more
Tencent
Research Digest··3 min read
In one line
DAAF learns the effects of replacing system assets to decide which edit fixes LLM agent failures.
Before you read this
Two ideas this paper builds onfailure attributionIdentifying which component or step in a system caused an observed failure, often using structured trace representations.146 papers →counterfactual reasoningComparing actual outcomes with imagined alternatives to estimate the effects of different actions or conditions.114 papers →
Read it as
The authors propose DAAF, a framework that moves beyond localizing where a failure occurs in an LLM agent to determining which persistent, editable system asset (e.g., a routing rule or knowledge segment) should be changed.
The authors formalize the problem of component-attribute : given a failed execution, identify which versioned, addressable system asset (called an attribute) should be modified.
How it works
before and afterThe Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.