Recirculation adds inference-time memory to LLMs without retraining, boosting Gemma3 accuracy by 21%
pbaylies · x · 2026-08-20
The paper introduces "Recirculation," an inference-time architectural enhancement that gives pretrained Transformers a form of working memory by feeding deep-layer information back into earlier layers.
Key Mechanics:
- Recirculates a portion of understood context to earlier layers during token processing.
- Enables the model to act as a dynamical system to track belief states without weight updates.
Results on Gemma3:
- 23% reduction in perplexity.
- 21% increase in GSM8K accuracy.
- Near-zero extra generation latency (requires serial processing in prefill).
The findings suggest that memory mechanisms may already exist within pretrained models, and recirculation simply provides a way to persist them.
Related event: DeepMind's Recirculation boosts LLM reasoning without retraining(10 posts)→
More from Research
- Dev critiques compiling fast programs into slow neural nets: fun, but not useful — ctjlewis · 2026-10-16
- New paper uses revealed-preference methods to unpack why AI agents cooperate — brwilder · 2026-10-16
- Logit lens on a Beat Saber finetune reveals it reuses language grammar circuits — ctjlewis · 2026-10-16
- 128×128 record stands: how intermediate-work experiments foreshadowed reasoning models — ctjlewis · 2026-10-16
- TeamFusion: preference-based AI proxies debate to preserve minority views in team decisions — burkov · 2026-10-16
- MIT's C-GMT gives humanoid robots whole-body compliance without sacrificing motion tracking — MIT_CSAIL · 2026-10-16