Oxford-led NeurIPS paper: frontier LRMs mirror human rule discovery and brain activity
sreejan_kumar · x · 2026-09-25
A NeurIPS-accepted study from Oxford, Columbia, NYU, MIT and Harvard had 32 fMRI-scanned humans and frontier LRMs play VGDL grid games with no rules given. Top LRMs matched human learning curves, discovering rules in similar steps; model hidden states predicted human BOLD responses across cortical and subcortical regions — the first demonstration that LRM representations align with human brain activity during active learning. Ablations attribute alignment to in-context representations of game-state sequences.
More from Research
- Analog chip runs LLM attention 100x faster than H100 using 70,000x less power, Nature paper claims — anselm · 2026-09-25
- aaru publishes 2,993-question simulation eval with 7.62% mean TVD and 3.53% MAE — marcbhargava · 2026-09-25
- New research shows LLM leaderboards are less stable than you'd hope — beirmug · 2026-09-25
- ECCV 2026 paper studies which high-dimensional latents suit diffusion models — _akhaliq · 2026-09-25
- LeWAM: Lightweight World Action Model Hits 92.28% Success on RoboTwin 2.0 — udmrzn · 2026-09-25
- Sakana AI hires Jürgen Schmidhuber to lead its recursive self-improvement lab — The Decoder · 2026-09-25