Recirculation Mechanism Verified: +17% GSM8K with Zero Weight Changes
savvyRL · x · 2026-08-20
A developer independently verified Google's inference-time recurrence mechanism, which allows frozen off-the-shelf LLMs to act as dynamical systems by leaking deep representations back to shallow layers. Reproduction on Llama 3.2 1B (M4 MacBook) showed a +16.95% improvement on GSM8K-Platinum with zero weight changes. Reproduction code and optimized mlx prefill kernel are available.
More from Research
- LightOn releases mLateOn: SOTA multilingual retrieval with just 115M parameters — antoine_chaffin · 2026-08-20
- Berkeley & Princeton release MLS-Bench to test if AI can invent new ML methods — jiqizhixin · 2026-08-20
- ISMIR 2026 Conference Opens Call for Sponsors in Abu Dhabi — umpedronosapato · 2026-08-20
- Reasoning Model Underperforms in Law Write-on Competitions — neuranna · 2026-08-20
- Context Windows Are Not Memory: A Student Builds an Open-Source LLM Memory Framework — Cohere_Labs · 2026-08-20
- Study Reveals Compute-Optimal Scaling is Skill-Dependent: Memory Needs Params, Reasoning Needs Data — ml_perception · 2026-08-20