1
小米公开 MiMo-V3 核心架构 HySparse2：面向长程多轮 Agent 优化推理效率
IT之家 9 月 24 日消息，小米 MiMo 今天公布面向 MiMo-V3 的核心架构 HySparse2，瞄准下一代长上下文 Agent。新架构旨在让模型以更低成本处理网页、文件、代码和工具调用记录，并从不断增长的任务历史中更准确地检索关键信息。 IT之家附官方详细介绍如下： HySparse2…
2
Thinking at the Right Size: Amortized Distillation Across Post-Trained LLMs
不同尺寸模型间摊销蒸馏，让推理成本与思考深度灵活匹配，EMNLP 2026 前沿成果
arXiv:2608.22854v1 Announce Type: new Abstract: Practical deployment of large language models (LLMs) requires families of post-trained variants---inst…
3
Asymmetric Capacity Allocation in Self-Refinement Pipelines
探索自精炼流程中非对称容量分配策略，揭秘资源优化与推理效率的关键突破。
arXiv:2608.21345v1 Announce Type: new Abstract: Self-refinement, typically structured as generation, critique, and revision, is a widely adopted parad…
4
Forking Fast: Efficiently Estimating Uncertainty Dynamics in Text Generation
快速估计文本生成不确定性动态的新方法，让模型自我评估更高效精准。
arXiv:2608.19611v1 Announce Type: cross Abstract: LLM reasoning is stochastic, and so understanding a model requires grappling with the distribution o…
5
Second Thought: Reasoning in Parallel as LLM Agents Act and Observe
ReAct范式下智能体行动观察时推理停滞，新研究提出并行推理机制，让LLM在环境交互中持续思考，显著提升决策效率。
arXiv:2608.13667v1 Announce Type: new Abstract: LLM agents in the ReAct paradigm alternate between reasoning, acting, and observing, but deliberate re…
6
EASy: Towards Efficient LLM-Based Agentic System
EASy为LLM智能体系统提速增效，直击推理与行动协同的优化痛点
arXiv:2608.04588v1 Announce Type: cross Abstract: Agentic systems have emerged as a promising paradigm for solving complex tasks by coordinating speci…
7
DriveCode: Domain Specific Numerical Encoding for LLM-Based Autonomous Driving
DriveCode 提出领域专用数字编码，破解 LLM 在自动驾驶中数值推理不准、解码低效的难题，让大模型既算得准又跑得快。
arXiv:2603.00919v3 Announce Type: replace Abstract: Large language models (LLMs) have shown great promise for autonomous driving. However, discretizin…
8
Allocation Before Ranking: Decoupled Token Compression for OmniLLMs
让全模态大模型告别冗余Token，解耦压缩新思路，先分配后排序，推理效率可期！
arXiv:2608.01665v1 Announce Type: new Abstract: Token compression in OmniLLMs is typically posed as a single saliency-ranking problem: score each mult…
9
TrAC: Trace-Conditioned Answer Consistency for Efficient Uncertainty Quantification in LLMs
用追溯条件优化答案一致性，为LLM不确定性量化提供高效新方案，值得关注。
arXiv:2608.00422v1 Announce Type: new Abstract: Large language models (LLMs) can generate fluent reasoning traces that nevertheless lead to incorrect …
10
Knowing When to Quit: Diagnosing and Training LLMs to Abort Futile Reasoning
让大模型学会适时放弃无效推理，减少空耗、提升回答效率与准确率。
arXiv:2607.29211v1 Announce Type: new Abstract: Large language models generate computationally expensive yet semantically void reasoning on beyond-cap…
11
397B参数追平万亿模型，上海AI Lab发布科学智能体新基座 | WAIC 2026
上海AI Lab以397B参数实现万亿级模型性能，首创记忆与思考解耦架构，推理效率暴增4倍，改写大模型进化逻辑。
采用非Transformer架构“Mobius”，不做泛化通用问答
12
Valid $\ne$ Necessary: Diagnosing Latent Inefficiency in Chain-of-Thought
揭示CoT推理中有效步骤≠必要步骤，诊断大模型过度推理的冗余成本
arXiv:2607.11266v1 Announce Type: new Abstract: Chain-of-Thought (CoT) prompting has significantly advanced the reasoning capabilities of Large Langua…
13
Hierarchical Chain-of-Thought: Enhancing LLM Reasoning Performance and Efficiency
层次化思维链新方法，显著提升大模型推理性能与效率，值得关注的前沿研究。
arXiv:2604.00130v2 Announce Type: replace Abstract: Chain-of-Thought (CoT) prompting has significantly improved the reasoning capabilities of large la…
14
FPTQuant: Function-Preserving Transforms for LLM Quantization
新型LLM量化方法FPTQuant，通过功能保持变换解决离群值问题，高效节能不损性能。
arXiv:2506.04985v2 Announce Type: replace Abstract: Large language models (LLMs) require substantial compute, and thus energy, at inference time. Whil…
15
Making AI search smarter
揭秘Cloudflare如何优化AI搜索，让结果更精准，技术干货满满。
Search is how we find nearly everything on the web — creators, merchants, answers. AI is rewriting the rules, leaving creators caught between staying …
16
EquivPruner: Boosting Efficiency and Quality in LLM-Based Search via Action Pruning
用动作剪枝替代参数剪枝，LLM搜索不再“又慢又贵”，效率质量双赢的ACL 2026研究。
arXiv:2505.16312v2 Announce Type: replace Abstract: Large Language Models (LLMs) excel at complex reasoning through search algorithms, yet current str…
17
The Language-Energy Divide: Measuring Energy Costs of Multilingual LLM Inference
首次系统量化不同语言在LLM推理中的能耗差异，揭示语言鸿沟对AI可持续性的影响
arXiv:2606.21869v1 Announce Type: cross Abstract: Large language models (LLMs) are increasingly deployed in multilingual settings, yet the energy cost…
18
Stepwise Token Selection for Efficient Multimodal Large Language Models
多模态大模型推理成本主要来自视觉token，提出逐步选择方法大幅降低计算量，实现高效生成。
arXiv:2606.16067v1 Announce Type: new Abstract: In multimodal large language models (MLLMs), inference cost is largely dominated by the visual token p…
19
Learning When to Sample: Confidence-Aware Selective Sampling for Efficient Chain-of-Thought Reasoning
通过自适应置信度评估选择性采样推理路径，在保持CoT性能的同时大幅降低计算成本，为高效大模型推理提供新思路。
arXiv:2603.08999v3 Announce Type: replace Abstract: Large language models (LLMs) can achieve strong reasoning performance through chain-of-thought (Co…
20
Less is More: Improving LLM Reasoning with Minimal Test-Time Intervention
揭秘LLM推理效率新思路：最小干预即可显著提升性能，少即是多！
arXiv:2510.13940v4 Announce Type: replace-cross Abstract: Recent progress in large language models (LLMs) has focused on test-time scaling to improve …