AI Paper Atlas

Research map / Multi-agent systems and games / LLM agents

Scope Delineation Before Localization: A Two-Stage Framework for Enhancing Failure Attribution in Multi-Agent Systems

Kai Sun, Wenqiang Li, Bo Dong, Yuxin Lin, Jingyao Zhang, Bin Shi

AAAI 2026 · Multi-agent systems and games › LLM agents · approach: Self-evolving language agent learning · 2 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes a two-stage framework that first delineates the scope of likely failure in multi-agent system logs before localizing the responsible agent and step, to improve LLM-based failure attribution.

Problem

Failure attribution in LLM-based multi-agent systems, where existing LLM methods struggle with lengthy interaction logs and lack expert debugging knowledge.

Approach

Drawing on human debugging strategies, it narrows the search scope within long logs before localizing the failure, rather than attributing directly over the full trace.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add two-stage scope delineation with Stepwise Scope Delineation before localizationSTEP BY STEP · Who&When Hand-Crafted, GPT-4ostep accuracy (%): 8.77 -> 25.93Improved
add Expertise-Assisted Scope DelineationSDBL · Who&When Hand-Crafted, GPT-4ostep accuracy (%): 25.93 -> 27.78Improved
use delineated scope (SSD) instead of random scope (RSD) (ablation)SDBL · Who&When Hand-Crafted, GPT-4ostep accuracy (%): 14.81 -> 25.93Improved
use delineated scope (SSD) instead of random scope (RSD) (ablation)SDBL · Who&When Algorithm-Generated, GPT-4ostep accuracy (%): 31.75 -> 32.54No clear effect
add Overstep paradigm to expertise-assisted scope (ablation)SDBL (EASD) · Who&When Hand-Crafted, GPT-4ostep accuracy (%): 18.52 -> 27.78Improved
add Loop paradigm to expertise-assisted scope (ablation)SDBL (EASD) · Who&When Hand-Crafted, GPT-4ostep accuracy (%): 12.96 -> 27.78Improved
apply SDBL (EASD) with DeepSeek-R1 backboneALL AT ONCE · Who&When Hand-Crafted, DeepSeek-R1agent accuracy (%): 55.56 -> 50.00Hurt
apply SDBL (EASD) with DeepSeek-V3 backboneALL AT ONCE · Who&When Algorithm-Generated, DeepSeek-V3step accuracy (%): 19.84 -> 39.68Improved
use SDBL (SSD) stepwise scope delineation instead of single-passSDBL · Who&When Hand-Crafted, GPT-4otoken consumption: 17106 -> 111310Hurt
use expertise-assisted scope in a single passSDBL (EASD) · Who&When Hand-Crafted, GPT-4otoken consumption: 17106 -> 24657Mixed
prompt LLM directly for K failure-prone steps (ALL AT ONCE*) instead of delineation (ablation)scope delineation · Who&When Hand-Crafted Hit@KHit@K (%): 9.26 -> 25.93 (Hit@3 vs EASD)Hurt

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy