Research map / Learning theory and algorithms / Prediction and decision losses

Conformal Agent Error Attribution

NeurIPS 2026 · Learning theory and algorithms › Prediction and decision losses · approach: Conformal prediction sets · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 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 filtration-based conformal prediction (left, right, two-way filtration) that outputs contiguous prediction sets for locating the decisive error in failed LLM multi-agent trajectories, and uses those sets for automated rollbacks.

Problem

Multi-agent LLM systems fail on long traces, and point-prediction error attribution gives no uncertainty guarantee; the goal is a set of steps that contains the decisive error with finite-sample coverage, ideally contiguous for debugging and rollback.

Approach

Adapts conformal prediction to sequential data via filtration: left, right, and two-way filtering return the longest suffix, prefix, or their intersection whose monotone set-level score stays under a calibrated threshold. This yields contiguous sets with distribution-free coverage, and the score can be the normalized sum of step scores from any LLM judge.

Key results

• Right Filtering achieves the highest removal rate on Who&When (0.31 vs 0.20 for vanilla with naive gpt-4o-mini scores), matching its prefix-favoring error distribution.
• Fine-tuned Qwen3-1.7B scorer reaches AUROC 0.762 and accuracy 0.731 versus 0.519 AUROC for naive gpt-4o-mini.
• Right Filtering uses about 3.08 scoring calls per trajectory on Left Dense GSM8k versus 8.50 for vanilla and CRSVP.

Surprising findings

• Filtering algorithms are largely insensitive to the quality of the scoring function, while vanilla CP and CRSVP depend heavily on the fine-tuned scorer.
• Vanilla CP overcovers in the rollback experiment because its sets are not contiguous, so it yields no better success rate than Top-1.

Limitations

• Assumes exchangeable data and a single decisive error per failed trajectory.
• The best filtering algorithm must be matched to the calibration error-position distribution.
• Who&When has only 184 examples, and the fine-tuned scorer is trained only on synthetic error-injected GSM8k and MATH traces.

Experimental setup

Datasets: Who&When GSM8k MATH

Benchmarks: Who&When GSM8k Left Dense GSM8k Mid Dense GSM8k Right Dense MATH

Metrics: empirical coverage removal rate AUROC AUPRC accuracy success rate coverage cost NFE

Baselines: Vanilla conformal prediction CRSVP Top-1 prediction naive LLM-as-a-judge ECHO-style role-prompted scoring

Builds on: conformal prediction CRSVP ECHO AEGIS

Closest papers by meaning

• CORRECT: COndensed eRror RECognition via knowledge Transfer in multi-agent systems (ICML 2026 · similarity 0.90)
• Which Agent Causes Task Failures and When? On Automated Failure Attribution of LLM Multi-Agent Systems (ICML 2025 · similarity 0.87)
• DoVer: Intervention-Driven Auto Debugging for LLM Multi-Agent Systems (ICLR 2026 · similarity 0.87)
• Beyond LLM-Based Reasoning: Lightweight GNNs for Agent Failure Attribution (NeurIPS 2026 · similarity 0.87)
• AgentForesight: Online Auditing for Early Failure Prediction in Multi-Agent Systems (NeurIPS 2026 · similarity 0.86)
• AgenTracer: Who Is Inducing Failure in the LLM Agentic Systems? (ICLR 2026 · similarity 0.86)