LLM judges are blind to reward-model-optimized readability decay in writing evals
sam_paech · x · 2026-09-07
Responding to whether creative-writing benchmarks can still be trusted, sampaech says any LLM-judged writing eval deserves a pinch of salt—read the samples yourself.
The core issue: LLM judges are largely blind to the readability problems caused by optimizing on reward-model preferences, and this contamination keeps getting worse.
More from Research
- EMNLP Oral: training-free Nexus Sampling cuts KV-cache memory 10x; OpenResearcher tops BrowseComp-Plus at 54.8% — yuz9yuz · 2026-10-27
- ICML 2026 position paper: LLM 'thinking traces' are a harmful anthropomorphism — rao2z · 2026-10-27
- Forking gap follows a power law in context frequency; masking rare n-grams shrinks it — ZimingLiu11 · 2026-10-27
- How forking happens: n-gram tables split logits into non-interacting subspaces — ZimingLiu11 · 2026-10-27
- The best LoRA explainer video yet: rank constraints meet intrinsic dimension — sophiamyang · 2026-10-27
- No π0 weights changed: Motoniq layer lifts real-robot precision fitting from 0/10 to 9/10 — hbouammar · 2026-10-27