Reproducing Recirculation: Gemma 3 PPL Drops 23% Without Fine-tuning
CatAstro_Piyush · x · 2026-08-28
An independent researcher reproduced the Recirculation paper, confirming that injecting source layer 11 into destination layer 4 effectively reduces perplexity on the Gemma 3 1B model. The original paper proposes this as an inference-time architectural enhancement that boosts accuracy in generation and reasoning tasks without training. Tests on the Gemma3 family show a 23% reduction in PPL and a 21% accuracy increase on GSM8k.
Related event: Independent Reproduction Confirms Recirculation Cuts Gemma 3 Perplexity(2 posts)→
More from Research
- False Collatz Proof via a Lean Bug: Talk Probes Proof Assistants as LLM-Found Bug Targets — marc_lelarge · 2026-10-16
- Goodfire's Nick Cammarata argues neural networks are 'already clean' in new interpretability essay — ctjlewis · 2026-10-16
- NEJM AI: prompt tuning cuts LLM clinical test-order cost ratio to 1.4x across 19 systems — dvir_a · 2026-10-16
- AI Recommends 3.2x Costlier Care; Models Add Tests in 2/3 of Visits Physicians Ordered None — dvir_a · 2026-10-16
- Prompting cuts LLM over-testing costs across 19 systems, but diagnosis agreement drops — dvir_a · 2026-10-16
- 24 AI Systems Recommend $314 Per Visit in Care vs Physicians' $97, a 3.2x Gap — dvir_a · 2026-10-16