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

A new algorithm called Auxiliary Particle Power Sampling (APPS) offers a way to improve the reasoning of large language models at inference time without additional training, according to a paper posted on arXiv [1]. The method uses a bounded population of partial solutions to approximate a sequence-level power target, redistributing computation across competing prefixes rather than following a single path [2]. The work, submitted by Tu Nguyen on 4 May 2026 and revised on 13 Jun 2026, addresses a bottleneck in "reasoning without training" [1]. Base large language models—systems with many parameters trained on vast amounts of text—already assign non-trivial probability to correct multi-step solutions, but locating these modes efficiently during decoding remains difficult [2][11]. Power sampling biases decoding toward these modes by targeting a distribution proportional to p_theta(x)^alpha with alpha > 1, yet practical approximations must account for future-dependent correction factors that determine which prefixes stay promising [2]. APPS propagates hypotheses in parallel using proposal-corrected power reweighting and refines their survival through future-value-guided selection at resampling boundaries [2]. The algorithm provides a direct scaling knob in the particle count and predictable peak memory [1]. The future-value signal is instantiated with short-horizon rollouts, and the authors also study an amortized variant that replaces rollouts with a lightweight learned selection head [2]. The paper, hosted on the open-access repository arXiv, contributes to a growing body of research on inference-time techniques for large language models [9]. arXiv, founded in 1991, now receives about 24,000 submissions per month and has surpassed two million articles, serving as a primary distribution channel for preprints in computer science, physics, and related fields [9]. The APPS manuscript appeared in three versions, growing from 109 KB to 137 KB across revisions [1]. By improving the accuracy–runtime trade-off of training-free decoding, APPS supports the view that inference-time power approximation can recover gains often attributed to post-training [2]. The approach does not require modifying the underlying model weights, instead operating entirely at the decoding stage [1].