New Paper Decomposes LLM Decision Errors into Belief Loss and Optimization Residual
JessicaHullman · x · 2026-10-23
A new paper from Jessica Hullman's team tackles LLMs used as decision assistants (e.g. risk estimation, treatment recommendations) by decomposing their total regret on decision problems into two parts:
- Belief loss: the cost of acting on wrong beliefs, measured against the true posterior
- Optimization residual: the cost of not choosing optimally given the reported belief
The authors argue that aligning LLMs with human goals requires diagnosing why a bad call happens: did the model fail to form correct beliefs, or fail to optimize its action under the utility? The decomposition offers a framework for evaluating and improving LLM decision-making.
More from Research
- Open 14B Carbon-X reads raw DNA at 67% benchmark score, ~1000x cheaper than GPT-6.1 Sol — Thom_Wolf · 2026-10-23
- MPI-Sec PhD Thesis on Vulnerability Detection Benchmarks Includes Pre-LLM Classic "Top Score on the Wrong Exam" — mboehme_ · 2026-10-23
- New paper makes 9 falsifiable predictions toward a major AI-enabled terror attack by 2027-2028 — StephenLCasper · 2026-10-23
- HumanoidPose recovers humanoid robots' full-body joint states from RGB video alone — cholz · 2026-10-23
- Sophont's Holon-S1 predicts genes and biomarkers from routine pathology slides — iScienceLuvr · 2026-10-23
- Glasgow study: raised eyebrows and slow blinks make speakers look confident — Limor_Raviv · 2026-10-23