Hidden Measurement Error in LLM Pipelines Distorts Benchmarks, NeurIPS Paper Warns
IanArawjo · x · 2026-11-18
- A new NeurIPS paper, "Hidden Measurement Error in LLM Pipelines Distorts Annotation, Evaluation, and Benchmarking", argues that LLM evals drive deployment, safety standards and research, yet noise from prompts, judge models and agent harnesses is routinely ignored — underestimating error in benchmarking and annotation, with the problem worsening as data grows. The authors also cover correct confidence intervals and how this noise makes benchmark hacking easier.
- scychanbrains flags potential p-hacking-like behavior in labs' published LLM evals, echoing statistical concerns from the deep RL era (Henderson et al., "Deep RL that Matters") that the community never fully internalized.
- IanArawjo notes related prompt-sensitivity issues discussed in his paper "Machine Psychology".
More from Models
- PhilBench quizzes 97 LLMs with the 2020 PhilPapers Survey of 1,785 philosophers — pwlot · 2026-11-18
- GPT-6.2 Sol tops browser-use leaderboard at 82%, 35% cheaper and 2x faster than predecessor — charliermarsh · 2026-11-18
- Decision models beat LLMs at structured output: Liquid AI d1 runs 2.7-3.6x faster than LFM3 — mervenoyann · 2026-11-18
- Why Anthropic matters: competition, pricing and open source — dh7net · 2026-11-18
- GitHub to Retire Six Copilot Models on Dec 14, 2026, Auto-Switching Enterprise Users — rcleffke · 2026-11-18
- Dev's Deep Dive: GPT-6.2 Astra Improved but Still 'Directionally Very Bad' at Debugging — kimmonismus · 2026-11-18