Tag
论文提出 Agentic World Modeling（AWM），通过只读接口与真实环境交互式获取证据来模拟动作结果，实现无需环境副本的并行 agent 测试时扩展，在 τ-bench、AppWorld 和 OfficeBench 上显著优于单次模拟与训练式世界模型基线。
SelfMorph 让冻结的大模型在推理时根据输入自适应选择内部计算拓扑（原始前向、层重复、局部并行分支），通过轻量选择器统一管理这些计算形态，在多个基准上取得显著提升，提出测试时扩展的新维度。
Rubricon is a pipeline from Handshake AI that automatically generates and refines discriminative rubrics for LLM evaluation and post-training without human labels or fine-tuning, achieving 9% better alignment with expert rubrics than prior synthetic methods and 4.5% better HealthBench performance when training Qwen3.5-9B.
LARC introduces a latent reasoning framework in continuous space for unified multimodal models, replacing discrete token reasoning with continuous hidden-state reasoning to improve image generation quality and enable test-time scaling, achieving a 7% GenEval gain over BAGEL.
该论文将多重重要性采样(MIS)引入Guided Speculative Inference(GSI)框架，使用多个小型LLM作为草稿提案模型并由大型模型与过程奖励模型验证，证明了balance heuristic的最优性，并给出KL散度与奖励退化的新界。
MLE-World: a world model finetuned from Qwen3.5-4B that predicts ML experiment outcomes (error traces, logs, validation scores) without execution, enabling affordable test-time scaling for AI research agents. It matches real terminal outputs on 77.9% of held-out programs and boosts mean performance by up to 22.1% on MLE-bench Lite tasks.
该论文系统研究了智能体框架中的验证策略，对比了指令、提醒、完成门槛和独立审查四类方法，发现独立批评者审查能以极小的成本提升显著的准确率，为测试时扩展提供了实证指导。
TEX 是一个基于执行交叉验证的协作式测试时扩展框架，多个智能体并行生成测试与补丁并交叉执行形成反馈矩阵，在 SWT-Bench Verified 上达到 86.4% 的新 SOTA，比单智能体基线提升 9.0%。
提出Uncertainty-Aware Allocation (UAA)框架，通过兄弟节点语义分散度估计不确定性来动态分配LLM树搜索的每步扩展预算，在保持精度的同时将token使用量减少最多90%、延迟降低最多9倍，且可跨四种树搜索算法和开源/闭源模型泛化。
提出了Harness-Delta Attribution（HDA）评估协议，将agent harness演化带来的性能增益分解为过拟合、测试时计算扩展和可泛化改进三部分，发现11个executor模型和3个框架中过拟合与测试时扩展平均解释了54%和14%的增益，搜索集收益往往无法迁移到held-out任务。
ATEST: Adaptive Resource-Aware Test-Time Scaling for LLM Inference 提出了一种资源感知的测试时缩放框架，联合考虑额外采样的预期价值与当前 GPU 成本，在 GPU 负载变化下动态决定采样何时停止，从而在满足能耗预算的同时保持准确率，并显著减少 GPU 时间与能耗。
SHELDON 是一个科学推理框架，通过在假设空间和观测空间中进行联合且独立的搜索来提升 LLM 的推理能力，在诊断、抽象推理和优化基准上较基线最高提升 26.7 分，并生成多达 4.5 倍的新颖解决方案。
ANTE 是一个 agentic 强化学习框架，通过执行验证的动作前预测为编码智能体提供逐步信用分配，显著提升了 Qwen3-32B 在 SWE-bench Verified 上的性能（25.0% 提升至 57.2% pass@1）。其核心组件 BAPO 将预测锚定为逐步优势信号，ANTE-TTO 则在推理时无需外部评分器即可选择候选动作。
Opera is a verbal critic framework for long-horizon coding agents that treats each correction as a persistent note, tracking agent actions to distinguish mere compliance from actual resolution. It improves resolver rates by up to 15 percentage points across benchmarks and provides on-policy training data that boosts Qwen3.5-9B's SWE-Bench Pro performance without needing a critic at inference time.
提出了 PersonTTS，一个通过需求匹配的控制器初始化和源蒸馏程序化指导，实现个性化测试时缩放（多维用户需求联合满足）的摊销式智能体策略发现框架，在 AIME 和 HMMT 上显著优于现有 TTS 基线。
CLIFT 利用共形自验证(conformal self-verification)将稀疏的judge反馈转化为可复用的训练奖励与无需外部judge的测试时轨迹选择，在WebArena Infinity、VisualWebArena和Online Mind2Web上取得了开源web agent的最优表现。
Meta Superintelligence Labs finds that RL post-training narrows solution coverage of LLM agents (the 'Sharpening Tax'): base models with a light harness can solve more distinct agentic tasks at large K, while post-trained models win at pass@1. They propose PTGS, an adaptive-temperature RL method that reduces this tax while improving pass@1.
The paper provides theoretical guarantees for beam-search-based test-time scaling in LLM reasoning, proving a lower bound on vanilla beam search sample complexity and proposing confidence-filtered beam search (CF-Beam) that reduces coverage dependence from quadratic to nearly linear, outperforming sequence-level methods like Best-of-N on long-horizon tasks.
The paper introduces Hermes, a family of configurable harnesses that let models control context allocation across inference-time compute, and Hermes-Learn, a two-stage framework that trains models to make these contextual reasoning decisions adaptively, closing the gap for smaller open-source models.
The paper introduces GapFT, a fine-tuning method that selects training examples based on a model's Pass@K–Pass@1 gap, showing on LogiQA 2.0 and ReClor that it raises single-sample accuracy dramatically while using far less data than uniform verified RFT.