ended3월 19일· 1 sources

Show HN: Duplicate 3 layers in a 24B LLM, logical deduction .22→.76. No training

24B LLM에서 레이어 3개만 복제하면 논리 추론 0.22→0.76 — 학습 없이

Why it matters

By duplicating 3 specific contiguous layers in transformer LLMs using David Ng's RYS method, logical deduction scores jump from 0.22 to 0.76 on BBH benchmarks without any training or weight modification. The technique exploits functional 'reasoning circuits' — multi-layer cognitive units that benefit from a second forward pass — and was validated on Devstral-24B and Qwen2.5-32B using two consumer AMD GPUs in a single evening.

1
Sources
+0
24h
Growth
181d
Active
RYSlayer duplicationQwen2.5-32BDevstral-24Breasoning circuitsBBH

Sources

Related Issues