1
TTHE: Test-Time Harness Evolution
测试时计算新范式：TTHE方法让模型在推理阶段自我进化，无需重新训练。
arXiv:2607.08124v1 Announce Type: cross Abstract: The behavior of an LLM agent is determined not only by the underlying model, but also by its harness…
2
Show HN: Getting GLM 5.2 running on my slow computer
在慢速电脑上运行744B参数的GLM-5.2模型，实测不同精度下多token预测头的性能表现。
A few days ago I found myself trying out GLM 5.2 and was really positively impressed. The capabilities and security I was getting from this LLM are si…
3
Nemotron-Labs-3-Puzzle-75B-A9B: Compressing Hybrid MoE LLMs
70多位作者联合发布Nemotron混合MoE大模型压缩方案，大幅降低推理成本且保持性能。
arXiv:2607.04371v1 Announce Type: new Abstract: We present Nemotron-Labs-3-Puzzle-75B-A9B, a compressed variant of Nemotron-3-Super optimized for inte…
4
MosaicKV: Serving Long-Context LLM with Dynamic Two-D KV Cache Compression
长上下文LLM推理提速新方法，MosaicKV通过动态二维KV缓存压缩，显著降低显存占用并保持精度。
arXiv:2607.00760v1 Announce Type: new Abstract: Long-context LLM services now sustain prompts with hundreds of thousands to millions of tokens, making…
5
英伟达刷新 DeepSeek V4 推理纪录：单 Token 成本降至 1/5，AI 吞吐量最高提升 20 倍
英伟达Blackwell平台将DeepSeek V4推理单Token成本砍至1/5，吞吐量暴增20倍，AI推理效率再破纪录。
IT之家 7 月 1 日消息，英伟达昨日（6 月 30 日）发布博文，宣布在英伟达 Blackwell 平台上，通过优化全栈推理，相比较 DeepSeek V4 模型 1 个月前上线初期， 单 Token 成本最多降至五分之一。 IT之家注：单 Token 成本（Cost Per Token）指模型…
6
Diagnosing and Mitigating Context Rot in Long-horizon Search
arxiv新研究：诊断并缓解长程搜索中的"上下文旋转"问题，优化AI推理过程
arXiv:2606.29718v1 Announce Type: cross Abstract: Extensive context has become the norm as Large Language Models (LLMs) are increasingly deployed in l…
7
Predict, Reuse, and Repair: Accelerating Dynamic Sparse Attention for Long-Context LLM Decoding
提出预测-重用-修复机制，动态稀疏注意力加速长上下文LLM解码，降低推理延迟。
arXiv:2606.30389v1 Announce Type: new Abstract: Dynamic sparse attention (DSA) accelerates long-context LLM decoding by attending to only the top-K KV…
8
英伟达在华启动机器人人才招聘，聚焦具身智能等四大方向
英伟达在华高薪招人，瞄准具身智能与人形机器人技术落地，布局四大方向加速AI计算平台应用。
IT之家 6 月 30 日消息，据《每日经济新闻》从英伟达处获悉，全球 AI 芯片巨头英伟达（NVIDIA）近日在中国启动大规模机器人人才招聘计划，围绕具身智能、仿真、部署及解决方案架构四大核心方向开放多个岗位，覆盖北京、上海、深圳三地。 据报道，本次招聘中，具身智能团队岗位数量最多，共开放 6 个…
9
W4A4 Quantization for Inference on Wan2.2-I2V-A14B
提出W4A4量化方案，在Wan2.2视频生成模型上实现高效推理，挑战低比特大模型量化极限
arXiv:2606.29337v1 Announce Type: new Abstract: We summarize our submission to Sub-Challenge 1: W4A4 Quantization for Inference (HiF4 / MXFP4) of the …
10
Ranking Before Serving: Low-Latency LLM Serving via Pairwise Learning-to-Rank
利用成对学习排序技术，在推理前优化请求排序，大幅降低大模型服务延迟。
arXiv:2510.03243v3 Announce Type: replace-cross Abstract: Efficient scheduling of large language model (LLM) inference tasks is critical for achieving…
11
Where Do CoT Training Gains Land in LLM based Agents?
揭秘思维链训练如何提升大模型智能体性能，实证分析增益落点与机制
arXiv:2606.26935v1 Announce Type: new Abstract: Chain-of-thought (CoT) reasoning is widely used in language-model agents, but prior work has shown tha…
12
ReNIO: Reweighting Negative Trajectory Importance for LLM On-Policy Distillation
针对LLM在线蒸馏中负轨迹重要性不均衡问题，提出重加权方法提升推理效果
arXiv:2606.23104v1 Announce Type: cross Abstract: On-policy distillation (OPD) improves LLM reasoning by training a student model on its own generated…
13
郭明錤：谷歌开发 TPU v9 芯片推理优化升级款，联发科接单
IT之家 6 月 22 日消息，证券分析师 郭明錤 今日下午表示，谷歌将在代号 "Humufish" 的 TPU v9 自有 AI 芯片基础上推出一个旨在提高推理能力的升级改款，并由联发科独家获得订单。 这款代号可能是 "Triggerfish" 的 TPU v9 升级版芯片 旨在缓解既有 AI A…
14
How to become an AI infrastructure engineer?
从SRE视角剖析AI推理基础设施，揭秘真实工作流与技能树，适合想转型该领域的工程师
Hi, I currently work on a GenAI platform for one of the largest local industrial companies. My daily work mostly involves building inference infrastru…
15
ALIGNBEAM : Inference-Time Alignment Transfer via Cross-Vocabulary Logit Mixing
提出推理时对齐迁移新方法，跨词汇混合logit无需额外训练即可转移对齐能力，为大模型高效部署提供新思路。
arXiv:2606.12342v1 Announce Type: cross Abstract: Domain fine-tuning degrades the safety of large language models: fine-tuned specialists readily comp…
16
RKSC: Reasoning-Aware KV Cache Sharing and Confident Early Exit for Multi-Step LLM Inference
多步LLM推理新突破：通过推理感知KV缓存共享与自信提前退出机制，大幅提升效率，已被ICML 2026 Workshop收录。
arXiv:2606.09937v1 Announce Type: cross Abstract: We introduce RKSC (Reasoning-Aware KV Cache Sharing), a training-free inference framework that elimi…
17
Learning Evidence Highlighting for Frozen LLMs
冻结大模型也能学会高亮关键证据？新方法提升LLM可解释性。
arXiv:2604.22565v2 Announce Type: replace-cross Abstract: Large Language Models (LLMs) can reason well, yet often miss decisive evidence when it is bu…
18
Hyperflux: Pruning Reveals Importance
提出Hyperflux，一种基于L0的新型网络剪枝方法，将剪枝建模为连续演化系统，揭示剪枝过程中神经元重要性变化。
arXiv:2504.05349v4 Announce Type: replace-cross Abstract: Network pruning is used to reduce inference latency and power consumption in large neural ne…
19
Stabilizing On-Policy Distillation for MLLM Reasoning with Global Normalization
用全局归一化稳定多模态大模型基于策略的蒸馏，提升推理性能与训练效率的创新方法。
arXiv:2606.09091v1 Announce Type: new Abstract: On-policy distillation (OPD) has recently emerged as an important post-training paradigm. By using a s…
20
与爱为舞亮相腾讯云AI产业应用大会，深耕教育大模型，打造下一代学习Agent
教育垂类大模型+学习Agent落地实践，推理性能提升3-4倍，看AI如何重塑下一代学习体验。