DigiNews

Tech Watch by Johan Denoyer

← Back to articles

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

Quality: 8/10 Relevance: 9/10

Summary

Shows a GitHub-based approach to duplicating small blocks of transformer layers (3 layers) in 24B-class LLMs to boost reasoning and logical deduction without training. Results on Devstral-24B and Qwen2.5-32B indicate meaningful gains in BBH logical deduction and other benchmarks; includes a reproducible quick-start workflow and codebase for layer duplication and evaluation. Useful for researchers and practitioners exploring model architecture tricks and inference-time optimizations.

🚀 Service construit par Johan Denoyer