mcts

The paper identifies why MCTS with critics fails to improve computer-use agents and proposes Probe-and-Resume, a policy that invokes a critic only on candidate final states with full probe access to a forked environment, boosting Terminal-Bench 2.1 success by 8.4 points at less than half the cost of Codex CLI.

介绍 IneqAgent，一个将 LLM 推理与符号计算结合的智能体框架，用于 Lean 4 中的不等式形式化证明，采用 lemma 级蒙特卡洛树搜索协调引理构造与符号后端选择，在四个基准上达到 94.09% 的成功率。

一项受控等预算实验研究比较了蒙特卡洛树搜索（MCTS）与随机重启列表调度在DAG工作流调度中的表现，发现MCTS在相同预算下反而比简单的随机重启方法差约2.88%。研究指出MCTS的劣势源于选择规则锁定在首个扩展子节点上，而换成标准UCT规则后差距基本消失。

A preprint introduces FLEET, an algorithm that makes Best-of-N sampling reward-aware by attributing external rewards to tokens, storing hidden states with reward metadata in a vector store, and using modified MCTS to reweight logits in subsequent iterations. Results show equivalent or better performance with far fewer sampling iterations on GSM8K and LiveCodeBench with Llama 3.2 3B, plus a reusable metadata store that can serve as a prior for other tasks or enrich SFT/RL.

MoFlow formulates agentic workflow generation as a multi-objective MDP and solves it with Convex-Hull Monte Carlo Tree Search with set-valued backups, covering the Pareto front so workflows can be retrieved for any preference without retraining. It outperforms six baselines on math, code, and QA benchmarks in terms of hypervolume.

Harness Router is a decision layer for AI coding agents that improves tool selection before execution using Jev for ambiguity and MCTS for multi-step consequences.

Introduces PrimeScientist, a method for strategic allocation of research effort in autonomous research agents using an adaptive MCTS-based policy to improve research quality and sample efficiency under resource constraints.

This paper proposes a three-role Monte Carlo Tree Search framework for formal theorem proving in Lean 4, using the compiler as a reward oracle to improve sample efficiency and revealing reward hacking issues through kernel-level proof auditing.

This paper proposes a Unified Sequential Composition Model (USCM) and Latent Expansion Monte Carlo Tree Search (LE-MCTS) for fashion outfit generation, achieving state-of-the-art performance on multiple datasets.

Sakana AI introduces AB-MCTS, an inference-time scaling algorithm that enables multiple frontier AI models (Gemini 2.5 Pro, o4-mini, DeepSeek-R1-0528) to cooperate, significantly outperforming individual models on the ARC-AGI-2 benchmark.

SciOrch presents an 8B vision-language model trained with MCTS to coordinate multiple expert LLMs for multimodal scientific reasoning, achieving superior performance while reducing API costs.

This paper introduces Multimodal Multi-Dimensional Scalarization Process Reward Modeling (MMS-PRM), which enforces the worst dimension's robustness in multimodal reasoning to prevent failures like visual hallucinations from being masked by strong text logic.

This paper introduces Program-of-Layers (PoLar), a method that allows LLMs to dynamically skip or loop pretrained layers per input, improving accuracy and efficiency over fixed-depth inference.

StepPRM-RTL is a novel framework combining stepwise trajectory modeling, process-reward modeling (PRM), and retrieval-augmented fine-tuning (RAFT) to improve LLM-based RTL code generation for Verilog and VHDL, outperforming prior methods by over 10% in functional correctness metrics.

Proposes COMPASS, a cognitive MCTS-guided process alignment framework to enhance safety in LLM-powered search agents by synthesizing attack trajectories and isolating risky actions, achieving a favorable safety-utility trade-off with less training data.

DRS-GUI proposes a training-free dynamic region search framework for GUI grounding, using a lightweight UI Perceptor with human-like perceptual actions and Monte Carlo Tree Search to progressively locate instruction-relevant elements. Experiments show a 14% improvement on ScreenSpot-Pro for both general and GUI-specific MLLMs.