Eval researcher: randomize everything for decider models, exploit signals don't hold in real traffic
antoine_chaffin · x · 2026-10-31
Antoine Chaffin argues that decider model evaluations must randomize everything: evals contain many exploitable signals that don't hold on real traffic.
He shares a concrete bias: with long choice lists, models systematically favor the first option, since training data mostly contains tasks with few choices. He tried offsetting letters during training, but that hurt performance on small-choice tasks.
More from Models
- Cohere x Palantir model benchmark mocked online: 670 tokens/s but questionable output — XiRw · 2026-10-31
- Inference optimization race heats up: practitioners now want fast, cheap 'good enough' models, not the absolute best — bendee983 · 2026-10-31
- OpenAI tightens Codex account limits: stacking paid accounts on one device can now get you locked out — TheZachMueller · 2026-10-31
- NVIDIA Releases Nemotron 3.5 Bolt E4B: 4GB On-Device Model with 80% Fewer Output Tokens — heatOverflower · 2026-10-31
- Fable 5.5 Max Output Stuns With Surprisingly Clean SVG Generation — Hesamation · 2026-10-31
- Pickover Asks Gemini to Draw Bathroom Plumbing Diagram, Invites You to Spot Errors — pickover · 2026-10-31