Skip to content

Who Broke the System? Failure Localization in LLM-Based Multi-Agent Systems

🕒 Published (v1): 2026-07-08 23:33 UTC · Source: Arxiv · Venue: COLM 2026 · link

Why this paper was selected

COLM 2026; systematic failure localization framework for multi-agent LLM pipelines

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

AgentLocate is a framework for attributing failures in LLM-based multi-agent systems to both the responsible agent and the earliest decisive error step. It combines an LLM judge with multi-perspective evaluator verification, confidence-weighted aggregation, and LoRA fine-tuning to produce reliable, causally grounded failure localization. On two benchmarks, it consistently outperforms existing baselines in both agent-level and step-level accuracy.

Problem

When an LLM-based multi-agent system fails, identifying which agent caused the failure and at what step the trajectory became irreversibly misdirected is hard: errors propagate across tightly coupled agents over long horizons, causal responsibility is diffuse, and prior methods are brittle. Counterfactual replay (AgenTracer) is unstable because altering one step cascades into changed prompts and tool outcomes; taxonomy-driven matching (AEGIS) misses emergent coordination failures; poisoning forensics methods assume detectable adversarial traces absent in organic multi-agent failures.

Method

AgentLocate formalizes the task as finding the pair \((i^\star(\tau), t^\star(\tau))\) — the agent index and earliest decisive step \(t^\star\) such that correcting the action at \(t^\star\) would reverse the system-level failure. It operates in three stages:

  1. Decisive Failure Hypothesis Generation. An LLM Judge reads the trajectory \(\tau\) and query \(q\) and proposes \((\hat{i}, \hat{t})\). Two protocols are supported: all-at-once (full trajectory exposed at once) and step-by-step (growing prefix revealed incrementally, halting when the decisive step is detected).

  2. Evaluator Verification. \(H\) independent Evaluators each receive \((\tau, q, \hat{i}, \hat{t})\) under distinct prompt styles (base, concise, evidence-focused) and independently produce \((i_h, t_h, r_h, c_h)\) — a candidate location, rationale, and self-reported confidence \(c_h \in [0,1]\). Outputs are consolidated via confidence-weighted voting: $\(A(i,t) = \sum_{h \in H} c_h \cdot \mathbf{1}[(i_h, t_h) = (i,t)], \quad (i_{\text{eval}}, t_{\text{eval}}) = \arg\max_{(i,t)} A(i,t).\)$

  3. Adaptive Judge Improvement. Each failed trajectory yields a fine-tuning instance \(z_q = \{\tau, q, \hat{i}, \hat{t}, \{i_h, t_h, r_h, c_h\}_{h \in H}, i_{\text{eval}}, t_{\text{eval}}\}\). The Judge is fine-tuned with LoRA (rank 64, \(\alpha=128\), dropout 0.05, 3 epochs) on dataset \(\mathcal{D} = \{z_q\}_{q \in Q}\) built from a held-out refinement split, yielding \(\text{Judge}_{\text{ft}}\) used at inference.

Key Contributions

  • Formal definition of the multi-agent failure localization problem as identifying \((i^\star, t^\star)\) — the decisive agent-step pair.
  • AgentLocate framework treating localization as a verifiable and improvable process rather than a one-shot decision.
  • Judge-Evaluator refinement mechanism combining diverse reasoning styles with confidence-weighted aggregation.
  • Lightweight LoRA-based adaptive improvement that feeds evaluator critiques back into the Judge as supervised fine-tuning signal.
  • Evaluation on two benchmarks (Who&When Algorithm-Generated/Hand-Crafted subsets and Aegis-Bench) across four LLMs, plus comparison against poisoning forensics baselines.

Results

  • On Who&When Algorithm-Generated (all-at-once, Qwen-7B): AgentLocate achieves 69.05% agent-level and 38.10% step-level accuracy, vs. best baseline AgenTracer at 45.24%/11.90%.
  • On Who&When Algorithm-Generated (step-by-step, Qwen-7B): AgentLocate achieves 69.05% agent-level and 33.33% step-level, vs. best baseline AgenTracer at 38.10%/7.14%.
  • On Who&When Algorithm-Generated (all-at-once, GPT-4o): 61.90% agent-level / 30.95% step-level, vs. AgenTracer 45.24%/11.90%.
  • Holds the best agent-level performance across all four models (Qwen-7B, Llama-8B, Mistral-7B, GPT-4o) in both protocols.
  • Step-level accuracy is more tightly concentrated within 0–2 steps of the true failure step compared to all baselines (granular deviation analysis).
  • Remains efficient in token usage and running time relative to baselines.
  • Neither using the Judge hypothesis alone nor the Evaluator aggregation alone achieves the accuracy of the full pipeline.

Limitations

  • Step-level accuracy remains low in absolute terms even for AgentLocate (e.g., 38.10% best case on Algorithm-Generated), indicating localization is far from solved.
  • Fine-tuning requires a labeled refinement split of failed trajectories, adding data-collection overhead before deployment.
  • The evaluator confidence scores are self-reported by the same underlying LLM as the Judge, introducing a circularity risk.
  • AEGIS baseline cannot produce step-level predictions, making step-level comparisons incomplete.
  • Experiments assume a turn-based single-actor-per-step protocol; concurrent or asynchronous multi-agent architectures are not addressed.
  • Evaluation on only two benchmark families; coverage of real-world deployment settings is untested.

Relevance to Agentic AI / LLM Agents

Failure localization is a foundational reliability primitive for deployed multi-agent systems: without knowing which agent when caused a failure, debugging, monitoring, and targeted remediation are impossible at scale. AgentLocate directly addresses this gap for the class of orchestrated LLM pipelines that dominate current agentic AI deployments (software engineering agents, web agents, scientific discovery pipelines). The Judge-Evaluator-FineTune loop is a general pattern applicable to any setting where a single LLM pass is brittle — it operationalizes a form of self-critique and ensemble verification that is increasingly relevant as agents take consequential, long-horizon actions. The work also clarifies the boundary between failure localization and poisoning forensics, which matters for practitioners designing safety and accountability layers around agentic systems.