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

Consistency as a Testable Property: Statistical Methods to Evaluate AI Agent Reliability

NeurIPS 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.

Proposes a U-statistic and kernel (MMD/GAK) framework to measure AI agent consistency under semantically preserving perturbations, with hypothesis tests, and validates it on SWE-bench Verified, Spider2-DBT, and BFCL with GPT-5-mini (Codex) and Kimi-K2 (OpenHands).

Problem

Output-only and pass@1 metrics do not reveal whether AI agents behave consistently under semantically equivalent prompt or task-level perturbations, and existing consistency metrics ignore execution trajectories.

Approach

Treats agent consistency as a statistical quantity: U-statistic estimators with hypothesis tests for output consistency, and MMD with a Global Alignment Kernel (approximating Levenshtein distance) for trajectory consistency, plus position-weighted kernels to localize where deviations occur.

Key results

• On SWE-bench Verified, the linear_mcp perturbation drops OpenHands+Kimi K2 accuracy to 0.253 with output consistency 0.605, while action composition JSD is 0.931 and ordering Levenshtein is 0.549.
• On BFCL, the xml-format perturbation drops OpenHands+Kimi K2 accuracy to 0.001 with output consistency 0.345, traced to malformed XML generation.
• Prompt-level perturbations keep output consistency at about 0.75-0.90, while task-level perturbations degrade it.

Surprising findings

• Kimi-K2 under XML output format fails almost completely (accuracy 0.001) while succeeding with JSON, suggesting JSON-specific training bias.
• Dead-code injection density 30 exceeds density 10 on both accuracy and consistency, with density 75 returning to density-10 levels.

Limitations

• Perturbation semantic equivalence is assumed and hard to validate empirically.
• Trajectory analysis treats action sequences as the unit and ignores intermediate reasoning or tool outputs.
• Theoretical assumptions (distributional invariance, homogeneous task distribution) may be violated in practice.

Experimental setup

Datasets: SWE-bench Verified Spider2-DBT BFCL

Benchmarks: SWE-bench Verified Spider2-DBT BFCL

Metrics: pass@1 output consistency (U-statistic) MMD Jensen-Shannon distance normalized Levenshtein distance

Baselines: pass@1 Rabanser et al. reliability metrics

Builds on: Rabanser et al. reliability metrics Global Alignment Kernel Maximum Mean Discrepancy Hoeffding decomposition

Closest papers by meaning

• Towards a Science of AI Agent Reliability (ICML 2026 · similarity 0.95)
• Agentic Confidence Calibration (ICML 2026 · similarity 0.90)
• Terminal-Bench: Benchmarking Agents on Hard, Realistic Tasks in Command Line Interfaces (ICLR 2026 · similarity 0.88)
• Establishing Best Practices in Building Rigorous Agentic Benchmarks (NeurIPS 2025 · similarity 0.88)
• Claw-Eval: Towards Trustworthy Evaluation of Autonomous Agents (NeurIPS 2026 · similarity 0.87)
• AgentHijack: Benchmarking Computer Use Agent Robustness to Common Environment Corruptions (ICML 2026 · similarity 0.87)