Amazon paper finds rule-following AI benchmarks mark correct answers wrong; one scorer fix lifts LegalBench scores by 80+ points
rohanpaul_ai · x · 2026-11-09
An Amazon paper argues that benchmarks testing whether AI models apply tax, legal and business rules (RuleArena, LegalBench, etc.) often grade correct answers as wrong because their reference answers or scoring code contain bugs — so scores measure the benchmark's defects, not the model.
The team audited 9 public benchmarks via source tracing, rerunning reference programs, and controlled scorer tests, then rescored unchanged model responses after each fix. LegalBench's tax scorer deleted decimal points, reading $259,487.13 as 25,948,713; fixing it raised accepted answers from 8-9 to 87-91 of 96 for each of three models (80+ points). In RuleArena Tax, restoring a required zero floor changed 27 of 300 references.
Takeaway: before trusting a low score on a rule-following benchmark, read the scorer code and manually check a few failed answers. The paper also releases BAR (Benchmark Audit and Revision), a versioned collection of audit evidence and revisions.
Related event: Amazon Paper Finds Scoring Flaws in Rule-Based AI Benchmarks(2 posts)→
More from Models
- GLM matches Opus at mass-converting Windows drivers into LPE 0days, says security researcher — MalwareTechBlog · 2026-11-09
- Thread: OpenAI's own report shows models' 80% time horizon on AI R&D is just 15 minutes — tobyordoxford · 2026-11-09
- Opus 5.5 bypassed a YubiKey by pushing over HTTPS with a saved OAuth token — appenz · 2026-11-09
- Rumor: Nvidia to pour massive compute into Arete-27B model — generativist · 2026-11-09
- Gemini models keep blowing past length limits while Opus 5.5 over-refuses — jessi_cata · 2026-11-09
- Continual Learning Touted as December's Biggest AI Story as Third Moment Runs 10,000 Agents — haider1 · 2026-11-09