NeurIPS paper: hidden measurement error in LLM pipelines distorts benchmarks and annotation
IanArawjo · x · 2026-11-18
A new NeurIPS paper, 'Hidden Measurement Error in LLM Pipelines Distorts Annotation, Evaluation, and Benchmarking,' argues that evals drive deployment, safety standards, and research, yet noise from prompts, judge models, and agent harnesses is ignored — systematically underestimating error in benchmarking and annotation, with the problem worsening as data grows. Commentator srchvrs calls it a timely revisit of statistical testing in generative AI: new methods may not be needed, but different sources of variance and dependence must be properly accounted for. The author's thread also covers implications for confidence intervals and how noise makes benchmark hacking easier.
More from Research
- sokrypton shows models can reason about which beta sheets interact — sokrypton · 2026-11-18
- OpenAI study: persona-style CoT training drops monitorability from 0.94 to 0.22 — dair_ai · 2026-11-18
- LLM protein-folding test on fresh PDB entry lands near-random structures — MoAlQuraishi · 2026-11-18
- Lean-PEP turns worst-case optimization bounds into formally verified Lean proofs — prof_grimmer · 2026-11-18
- Two arms, two token streams, one shared vocabulary: new robot manipulation work open-sourced — DJiafei · 2026-11-18
- NUS's BiCAT action tokenizer for bimanual robots: 113/150 real-robot wins, 2.7x lower latency than OAT — DJiafei · 2026-11-18