HyperThink open-sources hypernetwork that writes per-question weight updates to skip reasoning
udmrzn · x · 2026-11-20
Code and weights are out for HyperThink (COLM 2026, KAIST + NYU, MIT license).
The method trains a hypernetwork that turns each question into a query-conditioned parameter update for a frozen LLM, letting it answer without an explicit long thinking trace — "thinking in weights, not tokens," a step toward accurate System 1 reasoning.
- A vector-quantized decoder regularizes the predicted updates
- Training uses the base model's correct post-thinking answers as supervision
More from Research
- Mathematician claims Riemann Hypothesis proof with Claude: 17.3M lines of Lean, unverified — imjustnewatai · 2026-11-21
- Riemann Hypothesis proof claim released as Lean formalization on Papers with Code — Michael_D_Moor · 2026-11-21
- "Wright Brothers first flight for math": KOLs hail a new era after math breakthrough — jamesdouma · 2026-11-21
- Stanford-led paper: 27B open model rewrites its own weights, gains 6.6 points without training — rohanpaul_ai · 2026-11-21
- Language Models Can Rewrite Their Own Weights: 27B Model Self-Improves Without Gradients — Jiajun Fan · 2026-11-21
- Anthropic mathematician claims he and Claude proved the Riemann Hypothesis, formalized in Lean — NinaDSchick · 2026-11-21