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.
More from Research
- Tencent Releases GameWAM: A Unified World-Action Model for Video Game Control — tencent · 2026-08-28
- Agentic Game Dev as Verifiable Trajectory Engine for Scaling World Models — NationalUniversityofSingapore · 2026-08-28
- Visual Intelligence Fails in Unwritten Cases: Plural Viewpoints First — HirokatuKataoka · 2026-08-28
- Survey on LLM Agent Evaluation: Taxonomy and Enterprise Challenges — kalyan_kpl · 2026-08-28
- Building a Sci-Fi Movie RAG Agent in ~60 Lines of TypeScript — mastra_ai · 2026-08-28
- 2026 AI System Danus Reproduces Complex Matroid Theory Proof — maier_ak · 2026-08-28