Divergence Decoding fuses LLM capabilities without retraining

Researchers have introduced Divergence Decoding, a novel training-free framework designed to fuse the capabilities of specialized scientific language models with generalist models. This method uses Jensen-Shannon divergence to monitor distributional disagreements between models, dynamically routing to a generalist model when a specialist shows significant divergence. Tested on Qwen and Llama series models across benchmarks like GPQA and ChemBench, Divergence Decoding demonstrated superior performance compared to single-model baselines, suggesting a new paradigm for adaptive inference-time collaboration between LLMs. AI

IMPACT This method could improve the performance of specialized AI models by allowing them to leverage general reasoning capabilities without requiring additional training.

RANK_REASON The cluster contains an academic paper detailing a new method for LLM capability fusion. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• arXiv
• ChemCoTBench
• Divergence Decoding
• GPQA: A Graduate-Level Google-Proof Q&A Benchmark
• Jensen-Shannon divergence
• Llama
• Qwen

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →