DeepMind's Recirculation Cuts Perplexity by 23% With Zero Weight Changes
daniel_mac8 · x · 2026-08-21
Google DeepMind's paper 'Recirculation' introduces a method to give LLMs dynamic working memory. It feeds a slice of deep-layer activations back into shallow layers at each step, analogous to installing a mail chute in an office building so the ground floor knows what the top floor just learned.
Key Results on Gemma3:
- 23% lower perplexity
- +21% on GSM8k pass@1
- Model weights remain untouched
- No added latency during generation
The researchers discovered this capability was already latent in the model; no parameters were added, just a feedback loop.
Related event: DeepMind's Recirculation boosts LLM reasoning without retraining(10 posts)→
More from Research
- UCSD Researchers Propose Dataset Watermarking Detectable via Black-Box API Access — yuxiangw_cs · 2026-10-16
- EmbodiedEvo tops RoboDojo as agents inherit reusable skills and lessons — Aiden_Tech_Ai · 2026-10-16
- CHI reviewer laments: papers' Discussion sections now cite nothing — IanArawjo · 2026-10-16
- Casper: Understanding 90% of neurons ≠ understanding the network — StephenLCasper · 2026-10-16
- NVIDIA Launches cuPotential in CUDA-X to Bring AI Agents to Semiconductor Materials Simulation — tomaszbednarz · 2026-10-16
- Agentic Search Systems Are Rewarded Only for Final Answers, Raising Faithfulness Concerns — burkov · 2026-10-16