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

Agentic Confidence Calibration

ICML 2026 · Multi-agent systems and games › LLM agents · approach: Agentic benchmarks and orchestration · 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.

Introduces Holistic Trajectory Calibration (HTC), a 48-feature logistic calibrator over agent log-prob trajectories, plus a pretrained General Agent Calibrator (GAC) that gets lowest ECE on out-of-domain GAIA.

Problem

Agents are overconfident in failure, and existing calibration methods built for static single-turn outputs cannot handle compounding errors, tool-induced uncertainty, and opaque multi-step failure modes.

Approach

Extract process-level features (cross-step dynamics, intra-step stability, positional indicators, structural attributes) from token log-probability trajectories and fit a simple ridge- or lasso-regularized logistic calibrator; pretrain it on pooled datasets as a general calibrator.

Key results

• On HLE, HTC-reduced reaches ECE 0.031 and Brier 0.090 versus 0.686 ECE and 0.560 Brier for LastStep-TP.
• Pretrained GAC-reduced achieves the lowest GAIA validation ECE of 0.118 against 0.142 for DirectTrain (reduced).
• Calibrator trained on SimpleQA transfers to HotpotQA and outperforms direct training on that target across all metrics (reduced: ECE 0.070 vs 0.082).

Surprising findings

• Calibrator trained on MMLU-Pro transfers poorly to HLE (ECE 0.504), despite feature-space proximity, which the authors attribute to reasoning distribution shift.
• The GAC-full variant has higher GAIA Brier score and AUROC than DirectTrain (reduced), so the ECE win does not hold across all metrics.

Limitations

• Relies on token log-probabilities, so it requires logprobs access from the LLM provider.
• Transfer degrades under distribution shift across reasoning formats (e.g., MMLU-Pro to HLE, GAIA ECE still 0.118 to 0.128).

Experimental setup

Datasets: SimpleQA HotpotQA StrategyQA MATH500 GPQA MMLU-Pro HLE GAIA

Benchmarks: SimpleQA HotpotQA StrategyQA MATH500 GPQA MMLU-Pro HLE GAIA

Metrics: ECE Brier Score AUROC

Baselines: Verbalized Confidence LastStep-TP GlobalTrace-TP LastStep-TP + Temp GlobalTrace-TP + Temp LSTM Transformer MLP XGBoost Gaussian Process

Builds on: Temperature Scaling smolagents OAgents Deep Think with Confidence

Closest papers by meaning

• Consistency as a Testable Property: Statistical Methods to Evaluate AI Agent Reliability (NeurIPS 2026 · similarity 0.90)
• Towards a Science of AI Agent Reliability (ICML 2026 · similarity 0.90)
• Proper Scoring Rules for Agentic Uncertainty Quantification (NeurIPS 2026 · similarity 0.88)
• AgenTracer: Who Is Inducing Failure in the LLM Agentic Systems? (ICLR 2026 · similarity 0.87)
• Terminal-Bench: Benchmarking Agents on Hard, Realistic Tasks in Command Line Interfaces (ICLR 2026 · similarity 0.87)
• Claw-Eval: Towards Trustworthy Evaluation of Autonomous Agents (NeurIPS 2026 · similarity 0.87)