Offline eval scores dropped but quality improved: what Qwant's experience reveals about benchmark overfitting
antoine_chaffin · x · 2026-10-29
- @rdeveaud (Qwant): repeatedly trained ranking models whose offline metrics degraded while users perceived better overall search quality — a strong sign of overfitting the eval, meaning negative eval results don't always mean the model got worse.
- @antoinechaffin: after debiasing his decision model's preference for option length, its Habermas Machine score fell from 60 to 45, below even the "pick the longest" heuristic (50%, vs 24% for random). Yet he argues this prevents the model from collapsing when token counts shift — so he doesn't care if popular benchmarks score it lower.
Together: benchmark scores and real quality can diverge, and negative eval results need contextual interpretation.
More from Research
- Nature paper releases TenK10K phase 1: single-cell RNA-seq from ~2,000 people maps how genetic variants shape immune gene expression — mo_lotfollahi · 2026-10-29
- Axiom's agents are formalizing CFSG in 43M lines of code, possibly the largest open-source project ever — SucceededMind · 2026-10-29
- MIT's 'AI telepathy' KV-cache transfer beats text by 40% accuracy at 4x speed, across model vendors — LucaAmb · 2026-10-29
- Coding agents fail to reimplement models from papers, sparking doubts about how papers are written — charlieharris01 · 2026-10-29
- Information-theoretic scaling laws for diffusion models predict when memorization begins — davidad · 2026-10-29
- Amandine Flachs launches side project orbench — IgorCarron · 2026-10-29