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: Recirculation: Training-Free Architectural Enhancement for LLM Reasoning(7 posts)→
More from Research
- Nearly 10% of Cancer Papers Flagged as Potentially Fake — rohanpaul_ai · 2026-08-20
- vLLM precision gap prevents GRPO convergence — SergioPaniego · 2026-08-20
- Aurora-80K releases: A modern tiny LLM with 80K params — Tall_Abrocoma_3533 · 2026-08-20
- Mini Kimi-K3 Replicated Under $250 Beats GPT-2 Benchmark — OtherRaisin3426 · 2026-08-20
- llama.cpp PR Uses AVX2 to Speed Up Large Batch IQ Quantization — pmttyji · 2026-08-20
- Qwen 2.5 72B Aces ACT Exam with Perfect Reading Score — on_line187 · 2026-08-20