ResearchNovember 10, 2026
-17
❄️
Free Uncertainty from Particle Decoding
A study shows that in-context learning can work without feature learning.
#uncertainty#SMC#LLM-reasoning#decoding
🔥 What happened
A team led by Erfan Baghaei Potraghloo shows that Sequential Monte Carlo decoding already computes the normalizing constant of the power distribution — which hands you sequence-level Rényi entropy for free, no extra forward passes.
💡 Why it matters
Until now, whole-completion uncertainty meant either burning many extra samples (self-consistency) or settling for token-local log-probs. The order-1/2 variant beats log-prob as a selective-prediction signal in 10 of 13 settings, gaining 0.022–0.044 AUROC on MATH500. An 8-particle probe matches 64-particle accuracy with 34.2% fewer particles.
⚡ Our take
A "free lunch" that actually earns the name — if you're already running particle decoding, the uncertainty readout is sitting right there. The catch: most teams will have to rebuild their decoding stack to cash in.
The title, summary and analysis of this item were produced automatically by an AI system and have not been editorially reviewed. They may contain errors, bias or omissions — when in doubt, read the linked original source.