@LijieyYang: Thrilled to share that our paper, “Thought-Level Beam Search for Reasoning,” has been accepted to COLM 2026! Different …

Thrilled to share that our paper, “Thought-Level Beam Search for Reasoning,” has been accepted to COLM 2026!

Different from Self-Consistency and Pruning-only approaches, Gambit actively reallocates test-time compute while generation is in flight — pruning low-scoring reasoning traces and branching from the strongest prefixes. See thread

[1/6] Test-time scaling drives much of the performance of large reasoning models: sample many long traces, then vote. Yet it is strikingly inefficient — most traces end in wrong answers, memory pressure inflates latency, and a single competition problem can occupy a GPU for hours.

[2/6] This shifts the central question from how much compute to spend to where to allocate it. Two paradigms fall short: parallel sampling treats traces as independent trials and saturates memory, while score-based pruning removes weak traces but leaves the freed capacity idle.

[3/6] Gambit formulates test-time reasoning as constrained compute allocation over partial trajectories. A lightweight scorer probes hidden states at reasoning-step boundaries; a periodic tournament prunes the lowest-scoring traces and branches from the highest-scoring prefixes.

[4/6] Children inherit the parent’s KV cache via prefix caching, so an alternative continuation costs only its new tokens.
Under identical B300 hardware budgets:

+6.7% on HMMT-24 and +3.3% on AIME-25 over pruning up to 68.5% fewer tokens than parallel sampling

[5/6] Paper: https://arxiv.org/abs/2608.08020 Code: https://github.com/Dao-AILab/gambit-parallel-reasoning…

[6/6] Thanks to compute support from @PrincetonPLI and Dao AI Lab. Glad to work with @lhyTHU (MIT CSAIL), @jiawzhao (Meta), and grateful for the amazing mentorship of @tri_dao and Ravi Netravali! #COLM2026 #LLM #Reasoning

We find that reinvesting the compute is the key. In our paper, we have used two types of scorers, one adapted from previous works and the other trained with our own recipe. To maintain the fair comparison, we use the first score for both pruning-based baselines and ours, shown in Table 1.

The first one is a gif to show score trajectories of different approaches. Feel free to click and view it-it should be working :)

Thank you so much Liane :)

Haha thanks bro

We further evaluated the ranking accuracy, to reflect calibration of scorers, and put it in Figure 9 and downstream accuracy in Table 2. Gambit improves the accuracy with either our trained history-aware scorers or off-the-shelf scorer from previous work.

Thanks for your interest! Our work finds out that reinvesting the compute is the key lever while scoring function is also important. In our paper, we have used two types of scorers, one adapted from previous works and the other trained with our own recipe. To maintain the fair comparison, we use the first score for both pruning-based baselines and ours, shown in Table 1.

We further evaluated the ranking accuracy, to reflect calibration of scorers, and put it in Figure 9 and downstream accuracy in Table 2. Gambit improves the accuracy with either our trained history-aware scorers or off-the-shelf scorer from previous work.