Thought-Level Beam Search for Reasoning

Hugging Face Daily Papers Papers
reasoning beam-search compute-allocation test-time-scaling inference-algorithm efficiency machine-learning

Summary

Gambit improves reasoning model efficiency by using thought-level beam search to dynamically allocate compute to promising reasoning traces under fixed hardware budgets, yielding significant accuracy and throughput gains.

Test-time compute scaling is a primary driver of performance in large reasoning models (LRMs), but extreme inefficiency bounds current approaches, shifting the critical question from how much compute to spend, to where to allocate it. We formalize test-time reasoning as a constrained compute allocation problem over partial trajectories. Under a fixed hardware budget, existing paradigms fail to actively allocate the compute to the most promising partial progress: traditional parallel sampling treats traces independently and induces severe memory bottlenecks, while subtractive pruning starves hardware and fails to actively and sufficiently shift the output distribution. To overcome this dichotomy, we introduce Gambit, an inference algorithm that executes thought-level beam search. By periodically pruning unpromising trajectories and immediately branching from high-quality prefixes, Gambit dynamically concentrates compute onto the most promising reasoning traces via a light-weight scorer probing hidden states while maintaining continuous high hardware utilization. Extensive evaluations across multiple models and benchmarks demonstrate that Gambit strictly dominates existing baselines. Under identical hardware constraints, our method yields up to a +6.7\% absolute accuracy gain on HMMT-24 and +3.3\% on AIME-25 over pruning baselines, delivers >2times higher throughput on trace completion, and reduces total token consumption by up to 68.5\% relative to standard parallel sampling.
Original Article
View Cached Full Text

Cached at: 08/14/26, 11:30 PM

Paper page - Thought-Level Beam Search for Reasoning

Source: https://huggingface.co/papers/2608.08020

Abstract

Gambit improves reasoning model efficiency by using thought-level beam search to dynamically allocate compute to promising reasoning traces under fixed hardware budgets.

Test-time compute scalingis a primary driver of performance inlarge reasoning models(LRMs), but extreme inefficiency bounds current approaches, shifting the critical question from how much compute to spend, to where to allocate it. We formalize test-time reasoning as aconstrained compute allocationproblem overpartial trajectories. Under a fixed hardware budget, existing paradigms fail to actively allocate the compute to the most promising partial progress: traditionalparallel samplingtreats traces independently and induces severe memory bottlenecks, whilesubtractive pruningstarves hardware and fails to actively and sufficiently shift the output distribution. To overcome this dichotomy, we introduceGambit, an inference algorithm that executesthought-level beam search. By periodically pruning unpromising trajectories and immediately branching from high-quality prefixes,Gambitdynamically concentrates compute onto the most promising reasoning traces via a light-weight scorer probinghidden stateswhile maintaining continuous highhardware utilization. Extensive evaluations across multiple models and benchmarks demonstrate thatGambitstrictly dominates existing baselines. Under identical hardware constraints, our method yields up to a +6.7\% absolute accuracy gain on HMMT-24 and +3.3\% on AIME-25 over pruning baselines, delivers >2times higher throughput on trace completion, and reduces total token consumption by up to 68.5\% relative to standardparallel sampling.

View arXiv pageView PDFProject pageGitHub3Add to collection

Get this paper in your agent:

hf papers read 2608\.08020

Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash

Models citing this paper0

No model linking this paper

Cite arxiv.org/abs/2608.08020 in a model README.md to link it from this page.

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2608.08020 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2608.08020 in a Space README.md to link it from this page.

Collections including this paper0

No Collection including this paper

Add this paper to acollectionto link it from this page.

Similar Articles

Stratagem: Learning Transferable Reasoning via Trajectory-Modulated Game Self-Play

Hugging Face Daily Papers

STRATAGEM is a new framework for improving reasoning transferability in language models by using game self-play with a Reasoning Transferability Coefficient and Reasoning Evolution Reward to reinforce abstract, domain-agnostic reasoning patterns over game-specific heuristics. Experiments show strong improvements on mathematical reasoning, general reasoning, and code generation benchmarks.

Efficient Reasoning with Balanced Thinking

Papers with Code Trending

This paper introduces ReBalance, a training-free, plug-and-play method that dynamically balances overthinking and underthinking in large reasoning models, improving efficiency and accuracy across multiple benchmarks.

@cerebras: https://x.com/cerebras/status/2067357992929153268

X AI KOLs Timeline

An analysis of the economics and performance impact of AI reasoning models, showing that enabling reasoning can improve accuracy by 10-20% but costs 5-10x more tokens, and discussing different reasoning types and their applications.

Agentic Systems as Boosting Weak Reasoning Models

arXiv cs.AI

This paper studies verifier-backed committee search as inference-time boosting for reasoning language models, showing that a committee of weak reasoning models can match the performance of much stronger models on code repair tasks like SWE-bench Verified.