reasoning

Google's Gemini 4 Carbon claims the #1 spot on the Artificial Analysis Intelligence Index v5 with a score of 59, leading in complex reasoning, knowledge, and scientific tasks while ranking second behind Claude Opus 5.5 on coding agent benchmarks. The model is rolling out to selected users with a 50% launch discount and a 1M-token context window.

ServiceNow AI presents Apriel Voice 27B Preview, an open speech-language model built on Gemma 4 26B-A4B that unifies speech understanding, reasoning, and tool use in a shared backbone, trained with a two-stage procedure and online distillation to reach 98.90 on Big Bench Audio and 81.13 on Speech-BFCL with 1.91s mean latency.

论文提出 conditional interaction 框架，通过 System-1 轻量预测器筛选高价值交互、仅让 System-2 执行被接纳的候选，从而以 1/ρ_u 因子降低执行计算成本，并主张以密度可测的 ρ_u 和 informative compute 更好地描述扩展效率。

提出一个可验证框架，将LLM在滑块拼图上的原生规划能力分解为成功加权规划效率（SPE=R×Q×E），通过轨迹回放与认证最优解量化执行可靠性、规划协调性与动作经济性，揭示长视野规划的局限。

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。

An open-source project implementing test-time compute scaling with agent swarms, reporting large accuracy gains on math and coding benchmarks when run on mid-size open models like Gemma and Qwen, spending up to hundreds of millions of tokens per problem.

A community member released the Groot Series, a family of four MoE models ranging from 19B to 280B parameters, touting strong benchmark results including a 19B model running 128K context on a 16GB GPU and a 280B model scoring 63.47 on SWE-bench Pro at low cost.

inclusionAI 发布 Ling-3.1-flash-1019，Ling-3.1-flash 的升级版本，拥有 560B 总参数/25B 激活参数和 1M token 上下文窗口，通过后训练优化在推理与 agentic 能力上显著提升（AA-Briefcase Elo 从约 1400 升至 1524，Artificial Analysis Intelligence Index 达 44）。

MORE 是一个自进化训练框架，通过两阶段范式与 MATE-PPO 联合学习推理模式选择和计算预算分配，使智能体在编码、数学和通用推理基准上以更短的响应长度和更低的推理时间取得一致的性能提升。

IBM researchers formalize LLM consistency as scoped semantic invariance over knowledge-graph-defined semantic domains, showing inconsistencies cluster in stable hotspot subgraphs and that selective, layered KG grounding improves cross-model agreement and reliability.

The paper introduces MACR, a relay-based multi-agent framework that dynamically transfers reasoning ownership between agents using confidence-guided triggering and a bidding mechanism, improving accuracy on benchmarks like MathVista (+3.3%) and MMStar (+3.9%) while cutting token overhead by up to 6.0x versus other multi-agent strategies.

论文提出了一种名为 DIVA（Decision-Informed Value-of-information Acquisition）的训练无关方法，使 LLM 基于证据的决策价值来决定何时检索信息以及何时停止，从而避免为不改变决策的证据付费，并在四个领域、20 个 LLM 上将最佳基线的遗憾值降低 75%。

This paper develops a direction-sensitive, readout-relative account of semantic fragility in language models, showing that heterogeneous task-preserving rewrites can share failure dynamics while lacking a common input-independent failure direction, validated on Qwen3-32B and Gemma 4 31B.

SOPO（Socratic Guided Policy Optimization）提出了一种多轮强化学习框架，通过苏格拉底式教师引导反馈和基于轮次的优势缩放来解决span级幻觉检测中的奖励稀疏与长度膨胀问题，在RAGTruth上取得SOTA，超越GPT-5等模型并开源了1.7B和4B模型。

This paper quantifies how hints attributed to sources of varying reliability (professor, friend, or dog) shift answers in reasoning LLMs like Qwen, Gemma, and SmolLM, finding the shift follows an exponential tilt pattern that persists even when the source is declared untrustworthy.

论文提出 SAGE——一个统一的10候选图像选择基准，将 VWSD、VerSe 和 WikiDiverse 重构为共享的词项中心协议，并提出训练无关的 Gloss-Augmented Reasoning (GAR) 框架，使用 Gemma3-12B 将 Hits@1 从 65.90 提升至 71.22。

This paper reexamines the assumption that constrained decoding degrades LLM reasoning, showing via theory and causal ablations across 11 open-source models that effects depend on constraint type—format, validity, and reasoning constraints—with validity constraints improving reliability and reasoning constraints potentially boosting accuracy by pruning invalid paths.

DynaFlow introduces a framework that lets a main agent synthesize executable workflows to coordinate sub-agents adaptively, achieving 88.08% accuracy and a 10.90-point gain over baselines with a Qwen3.5-27B backbone across six benchmarks.

Moonshot AI 推出 Kimi K3.1 Pro，目前可通过 API 使用，并承诺于 2026 年 11 月 4 日发布完整模型权重。其基准测试成绩接近 Claude Opus 5.5 和 GPT-6 Astra，是首个在 Artificial Analysis 榜单上如此接近前沿闭源模型的（即将开源权重的）模型，定价与 GPT-6.1 Sol 持平。

蚂蚁集团发布了 Ling-3.0-flash-Sante 的模型权重，这是一个开放的医疗推理模型，采用 124B 总参数 / 5.1B 激活参数的 MoE 架构，支持 256K 上下文，采用 MIT 许可证，具备 agentic 深度研究与工具使用能力。