Recirculation: Training-free inference architecture boosts model performance
chaumian · x · 2026-08-20
The paper proposes "Recirculation," an inference-time architectural enhancement for off-the-shelf foundation models that significantly reduces perplexity and boosts accuracy across generation and reasoning tasks.
Key Features:
- Inference-time enhancement: Incurs essentially no additional latency during generation, though requires serial processing in the prefill phase.
- Introduces recurrence: Allows the model to act as a dynamical system and track belief states, overcoming the depth limitation of feedforward transformers.
- Adaptive variant: Requires only light hyperparameter tuning while freezing the original model weights.
Results:
On the Gemma3 family, adaptive recirculation achieves a 23% reduction in perplexity on a suite of datasets, a 21% increase in accuracy on GSM8k, and reliable improvements on other downstream tasks.
Related event: DeepMind's Recirculation boosts LLM reasoning without retraining(10 posts)→
More from Infra
- Jensen Huang: Musk's AI buildout moved from K's to M's, now a CSP-tier NVIDIA customer — elonmusk · 2026-10-16
- EpochAI pegs GPT-6 training at up to 1.8e27 FLOP; scaling01 calls param estimates way too high — scaling01 · 2026-10-16
- Meta's Muse Spark Open scores 39 vs 48 for closed Spark 1.3 on AA Intelligence Index — EAccelerate_42 · 2026-10-16
- LLM spots useless 95GB-per-frame memcpy in ffmpeg dblur, cutting runtime from 2452ms to 49ms — IgorCarron · 2026-10-16
- DeepSeek-V4-Flash on 4x Radeon AI PRO R9700: the 'refused' 4-card config hits ~79 tok/s — brainchillzZ · 2026-10-16
- Tree packing lands in TRL's AsyncGRPO, cutting agentic RL training compute by over 26x — joecole · 2026-10-16