The Model Knows, the Decoder Finds: Future Value Guided Particle Power Sampling

Summary

Introduces Auxiliary Particle Power Sampling (APPS): training-free blockwise parallel search over partial solutions with proposal-corrected reweighting and future-value-guided resampling, redistributing finite compute over competing prefixes. Also studies a learned selection-head variant.

Keywords

decoding · inference-time compute · particle sampling

Topics

decoding, inference-time compute, particle sampling

Research notes

• Discovery: Shared in #random-papers as an X post by Haitham Bou Ammar; resolved to arXiv 2605.02427.
• Method: APPS decoding: blockwise parallel partial-solution search, proposal-corrected importance reweighting, future-value-guided particle resampling; optional learned selection head.
• Submitted 2026-05-04, revised 2026-06-13. Benchmark numbers not captured; read back from arXiv if needed.