Chopthin-Consensus Power Sampling Boosts LLM Reasoning Accuracy Without Retraining
A new decoding technique called Chopthin-Consensus Power Sampling (CCPS) tackles a long‑standing weakness in Sequential Monte Carlo (SMC) methods for large language models. Traditional SMC resampling equalizes particle weights, which can discard low‑weight but potentially correct reasoning paths, reducing the diversity of the search space. CCPS replaces equal‑weight resampling with a Chopthin…
Key points
- CCPS uses a Chopthin resampler to limit weight ratios, preserving diverse reasoning paths during LLM decoding
- Semantic‑majority selection merges identical and semantically similar answers, boosting final answer accuracy
- Across three open‑weight models and five benchmarks, CCPS achieved up to 10.6 % absolute accuracy improvement
The authors pair this diversity‑preserving resampling with a semantic‑majority selection step that merges identical token sequences and clusters semantically equivalent answers. Tested on three open‑weight LLMs across five reasoning benchmarks, CCPS improved oracle coverage in 13 of 15 cases and matched or outperformed the prior Power‑SMC baseline in 14 of 15, delivering up to a 10.6‑point absolute accuracy gain. The code has been released publicly on GitHub, enabling researchers to apply the method without any model retraining.
These results suggest that enhancing particle diversity during inference can meaningfully boost reasoning performance, offering a training‑free avenue for improving LLM reliability on complex tasks.
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Research
All →- Anthropic research suggests small AI models may self-replicate on IoT devices · 7 src
- Researcher predicts AI engineering acceleration but not superintelligence · 1 src
- MIT researchers show neural networks can build own learning rules · 1 src
- Anthropic co-founder predicts full RSI possible by 2027 · 1 src
- Sakana AI’s MASS method improves model self-judgment but risks baking errors into training · 1 src
Comments
via GitHub Discussions