1
又快又聪明，阿里发布Qwen-Audio-3.0-Realtime：实时语音大模型四项功能升级
阿里实时语音大模型四项功能同步升级，智商与口语抗噪双突破，堪称又快又聪明。
2
小米开源 Xiaomi-Robotics-U0：具身领域首个“通吃”四类任务的统一生成模型
小米开源具身智能统一生成模型，可通吃四类任务，推理速度提升83倍，开源爱好者不容错过。
IT之家 7 月 15 日消息，小米今日发布 Xiaomi-Robotics-U0—— 一个拥有 380 亿参数 的多模态自回归具身生成基础模型，是 具身领域首个“通吃”四类任务的统一生成模型 ，打通了机器人图片和视频数据的生成与编辑链路。 具身场景生成（Scene Generation）—— 模型…
3
SlimPer: Make Personalization Model Slim and Smart
SlimPer提出让个性化模型精简且智能的新方法，兼顾性能与效率，值得关注。
arXiv:2607.12281v1 Announce Type: cross Abstract: Transformer-style architectures are increasingly adopted for industrial recommendation systems, yet …
4
Silent Failures in Quantized LLM Reasoning: A Taxonomy-Based Analysis of Hollow Convergence and Failure Mode Shifts
量化后的LLM看似准确率不变，但推理过程已悄然“空心化”，本文用六类故障图谱揭示这一隐患。
arXiv:2607.09999v1 Announce Type: cross Abstract: We show that post-training quantization can silently alter how large language models reason even whe…
5
AeroRAG: Structured Multimodal Retrieval-Augmented LLM for Fine-Grained Aerial Visual Reasoning
新型多模态检索增强框架让大模型精准理解航拍图像细节，解决细粒度视觉推理难题。
arXiv:2604.17889v2 Announce Type: replace Abstract: Despite recent progress in multimodal large language models (MLLMs), reliable visual question answ…
6
PiCSAR: Probabilistic Confidence Selection And Ranking for Reasoning Chains
提出概率置信度选择与排序方法，优化大模型推理链，提升复杂推理的准确性与可解释性。
arXiv:2508.21787v3 Announce Type: replace-cross Abstract: Best-of-n sampling improves the accuracy of large language models (LLMs) and large reasoning…
7
Agentic Context Learning with Self-Discovered Specification
让AI自动发现任务规范，在上下文中动态学习，实现更强自主推理能力。
arXiv:2607.09794v1 Announce Type: new Abstract: Context learning is an emerging inference-time task where LLMs must learn and apply novel, task-specif…
8
Show HN: Libargus:Low-latency local LLM runner via OpenJDK Panama FFM (Java 22)
低延迟的JVM本地LLM推理方案，利用OpenJDK Panama FFM避免传统串行化与GC开销。
Most existing approaches for running local LLM inference within the JVM ecosystem rely on spawning out-of-process daemons via REST sidecars (introduci…
9
Motion4Motion: Motion Transfer Across Subjects at Inference
无需额外训练，推理阶段即可实现跨主体的精准运动迁移，为视频生成与动画制作带来简洁高效的解决方案。
arXiv:2607.11644v1 Announce Type: new Abstract: This work explores the motion transfer from one video to another, which is crucial in animation for di…
10
FastTPS: An Optimized Method for LLM Token Phase for AI accelerators
针对LLM推理中token阶段低并行度瓶颈，FastTPS提出优化方法，提升AI加速器吞吐量。
arXiv:2607.11211v1 Announce Type: new Abstract: The popularity of large language models (LLMs) escalates an ongoing demand for effective inference. Ho…
11
Depth-Entropy Guided Sampling for Training-Free LLM Reasoning
无需训练的LLM推理新方法：深度熵引导采样，提升推理效率与质量。
arXiv:2607.09693v1 Announce Type: cross Abstract: Reinforcement learning (RL) has become the dominant paradigm for improving the reasoning capabilitie…
12
OpsMem: Dual-Memory Reasoning with Cross-Memory Resonance for Failure Diagnosis
新型双记忆推理框架，利用跨记忆共振技术精准诊断系统故障，提升运维智能化水平。
arXiv:2607.11357v1 Announce Type: new Abstract: Failure diagnosis in modern software systems requires iterative evidence acquisition and hypothesis re…
13
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…
14
OS-Pruner: Pruning Chains-of-Thought of Reasoning Models via Optimal Stopping
针对大模型CoT推理中“过度思考”的冗余步骤，用最优停止策略动态剪枝，降本增效不丢精度。
arXiv:2607.11089v1 Announce Type: new Abstract: Large Language Models (LLMs) have achieved remarkable success in complex reasoning tasks through Chain…
15
colibri – 在 25GB 内存电脑上运行 GLM-5.2 (744B MoE)
普通电脑也能跑744B参数大模型！colibri开源工具让25GB内存畅玩GLM-5.2。
colibri 是一个非常实用的开源项目，它能让普通电脑也能运行超大语言模型（GLM-5.2（744B），并且可以在无显卡的情况下，仅使用 CPU，但需要至少25G 内存。@Appinn 普通电脑跑 GLM-5.2（744B）模型 colibri 使用纯 C 语言，零依赖。可以按需从硬盘加载 Exp
16
MemDecay: Region-Aware KV Cache Eviction for Efficient LLM Agent Inference
新方法MemDecay基于区域感知的KV缓存淘汰，大幅提升LLM Agent推理效率，降低内存占用。
arXiv:2607.10582v1 Announce Type: cross Abstract: Large language model (LLM) agents accumulate heterogeneous context, including system instructions, p…
17
Evidence-Backed Video Question Answering
一篇关于视频问答的创新论文，强调利用证据推理来提升答案的准确性与可解释性。
arXiv:2607.11862v1 Announce Type: cross Abstract: Current Video Large Language Models (Video LLMs) excel in question answering (QA) but largely operat…
18
从分布式电力到分布式算力：Sunrun 在美启动边缘 AI 推理节点部署试点
美国家用储能巨头Sunrun试点将AI推理节点部署到用户微电网，分布式电力变身分布式算力，边缘AI落地新场景。
IT之家 7 月 14 日消息，自称美国最大家用电池储能、太阳能和家用并网电站供应商的 Sunrun 当地时间本月 8 日宣布启动一项边缘人工智能推理计算节点部署试点项目， 将分布式电力转化为分布式算力 。 Sunrun 表示其拥有超过 110 万现有客户，而这些客户可选择在其家用表后微电网中加装来…
19
Steam 社交推理游戏节开幕：《Among Us》国区六折 15 元，持续至 7 月 17 日
IT之家 7 月 14 日消息，Valve 现已开启 Steam 社交推理游戏节，平台推出大量推理类游戏折扣，包含谋杀悬疑、捉迷藏等多种依靠玩家判断力的游戏作品。 官方介绍，该类游戏的核心乐趣就在于玩家需要仔细分辨周围角色，判断对方究竟是 " 朋友还是敌人 "，过程中充满了猜忌与趣味性。 本次活动现…
20
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…