Johns Hopkins benchmark finds accurate LLM agents fail at epistemic humility
JohnsHopkins · hf · 2026-10-09
Johns Hopkins researchers introduce a benchmark for epistemic humility (EH) in LLM agents: whether agents revise answers, acknowledge uncertainty, or persist with wrong conclusions when retrieved evidence conflicts with prior beliefs.
- EH is operationalized through three trajectory-level dimensions: Identify, Solve, and Escalate (ISE), tested under controlled knowledge conflicts and naturally occurring conflicts during multi-step execution, each with matched no-conflict controls.
- Evaluating four agents, higher task accuracy does not imply greater EH: some high-accuracy configurations recognize conflicts mid-execution but fail to communicate unresolved uncertainty in their final answers.
- Agents often detect conflicts early but fail to maintain or resolve them later; model-level interventions improve EH but often at the cost of task accuracy, suggesting EH emerges from interactions among backbone model, agent harness, and environment.
More from Research
- Embedding arithmetic demo: cartoon − animation − comic = diagram — mayfer · 2026-10-23
- Omnii Genome: Single Generalist Model Outperforms AlphaGenome Across All Benchmarks — JosephJacks_ · 2026-10-23
- Domingos Cites New Result Claiming RL Is 'Basically Useless' vs Supervised Learning — pmddomingos · 2026-10-23
- Word Arithmetic on Embeddings: "Bathroom Minus Toilet Equals Kitchen" Beats SAE Probing — mayfer · 2026-10-23
- Wolfram Opens Rule 30 Prizes to AIs: $30,000 for Three Open Problems — cephaloform · 2026-10-23
- Stanford and Databricks use AI agents to reproduce medical findings and test analytical robustness — ChengleiSi · 2026-10-23