Reward-model ensembles from different pretraining seeds resist reward hacking better
vxnuaj · x · 2026-10-29
vxnuaj notes that reward-model ensembles built from different pretraining seeds resist reward-hacking better than ensembles that differ only in fine-tuning seed—diversity at the pretraining stage yields more genuinely decorrelated ensemble members.
More from Research
- GPT-6.1 Terra scores 28.1% on agentic WeirdML v3, matching Claude Fable 5.1 at lower cost — scaling01 · 2026-10-29
- Ex-OpenAI Researcher's 26,000-Word Post Says Neural Networks May Not Be Black Boxes — CurieuxExplorer · 2026-10-29
- visloc-rs: Rust library for SfM, VI-SLAM and 3D Gaussian Splatting hits 150 stars — rsasaki0109 · 2026-10-29
- Anthropic's GRAFT fights reward hacking by deleting parameters, not penalizing behavior — dl_weekly · 2026-10-29
- Ex-Google Researcher's Elorian Launches Glimpse: Frontier Models Fail 50% on Visual Counting Benchmark — Latent Space · 2026-10-29
- TMLR Reviewers Quit Over Aggressive Desk Rejections Amid Rising AI Slop — miniapeur · 2026-10-29