Even small 2B-scale retrievers can double mixed-corpus retrieval quality once modality bias is addressed in training, which matters for anyone building RAG over documents with screenshots, tables or charts. The paper, code and project page are open, so teams can adopt the Multi-Positive View InfoNCE recipe or swap Trident-Qwen3VL-2B into their existing embedding stack.
达摩院把 RISC-V 做进车规了，E908A 同时拿下 ASIL-B 和 ASIL-D 认证，还有下一代 C950A 五倍性能目标，做车载和嵌入式开发的可以看看。
商汤系公司大晓机器人发了具身基模型 ACE-Brain-0.6，基于 Qwen3.5-9B 微调，六项 Grounding 评测分数超过了 OpenAI GPT-6 Astra，做机器人方向的可以看看评测细节。
The interesting shift is that code execution as the agent substrate lets weaker models close the gap with frontier ones by writing programs instead of clicking precisely, which changes how agent-vs-model capability should be evaluated. Developers building computer-use agents can reuse the harness, SDK and macOS app directly and compare against GUI-only and hybrid designs that the paper treats as special cases. The OSWorld-v2 results give a concrete head-to-head of GPT-6 Astra, Claude Opus 5 and Chinese models like GLM and Kimi under the same harness.
用世界模型当奖励模型训练 AI 自己改进模型，Qwen3.8-27B 训完居然超过 GPT-6-sol，成本还降了 33%。
做智能体工具调用的朋友看这篇，工具接口一改模型就崩，SchemaShield-Lite 不用重训就能在 BFCL-200 上把准确率从 0.58 拉到 0.805。
这篇论文针对 GRPO 对同奖励响应一视同仁的问题，用生成熵给对错响应加不同权重的修正项，Qwen2.5-7B 上 avg@32 从 41.5% 提到 45.4%。
这篇论文解释了为什么 GRPO 做安全对齐会失效，还给出了 Beacon 框架，在 Qwen3 和 DeepSeek-R1-Distill 上把安全错误最多降了 32.83 个百分点。
Verified-by-construction environments sidestep the usual reward-hacking problem in RL for open-ended analysis tasks, a reusable recipe for teams training agents beyond fixed-answer benchmarks. The diagnostic finding that frontier models compute correctly but conclude wrongly matters for anyone building data-analysis agents who should audit the reasoning step, not just the arithmetic. The paper is available on arXiv (2610.21656) with a Chat with Paper interface at academy.dair.ai.
微软亮出自家基础模型 MAI-Base-2，2.8T 参数、1.6 万张 GB200 训练，连中国开源模型都被比下去了。
The reward-design audit is the most transferable part: dense progress rewards let failed runs harvest half the reward, so teams training RL agents should measure how much reward failure earns before trusting their curves. RL and agent-training engineers can apply the strict-success-versus-partial-credit split to their own reward pipelines. Compared with generic terminal-agent RL reports, this one pairs benchmark gains with a concrete failure-mode analysis.
Distilling reasoning traces from a larger teacher into a 0.6B-parameter model shows practitioners a far cheaper path than RL training runs, with MATH-500 as the benchmark to verify the gains. Builders training small models on limited compute can reuse the standalone training script and data-preparation pipeline directly. The soft vs. hard and on-policy vs. off-policy distillation framing is a useful mental model when choosing a distillation approach.
有人给破解版 PS5 写了原生 GPU 的 llama.cpp 移植，27B 模型能跑到 20.9 token/s，还兼容 OpenAI API，家里吃灰的主机直接变推理服务器。
JetBrains 的 Junie Local 不用联网不用付 Token 费，RTX 3090 以上就能本地跑，还有 Qwen3.6-27B 可选，写代码看重隐私的可以试试。
A US lab landing a top open-weight spot in agentic benchmarks breaks the pattern of Chinese labs dominating the open WebDev leaderboard, where 24 of the top 25 models are Chinese. Teams choosing an Apache 2.0 model for agent or coding work now have a credible US alternative to Kimi K2.5 Thinking and GLM-5. The wide error bars mean its true standing relative to Qwen3.8 is less settled than the rank suggests.
作者把完整提示词贴出来了，一个 Opus 5.5 提示直接出四段带配音的动画短剧，想搞 AI 视频的可以直接抄作业。
24 of the top 25 open models on Code Arena: WebDev come from Chinese labs, and Beam 1.0 is the lone US exception — a notable shift in the open-weights competitive landscape. Builders evaluating Apache 2.0 agentic models can compare Beam directly against Qwen3.8, Kimi K2.5 Thinking and GLM-5 once the full weights drop this month. Its 23B active parameter design targets reasoning efficiency for agent workloads.
Reflection 的第一个开源模型 Beam 出来了，501B 参数只激活 23B，跑分超过 NVIDIA 和 K2 Horizon，但幻觉率有点高，终端任务也比较弱。
Instead of wrapping a general model in a memory framework like Mem0 or MemOS, this uses a small dedicated writer model, cutting inference cost to roughly 3B-model levels while staying cross-framework. Developers building agent pipelines can pip install it and swap storage backends freely, and Mac users can share memory across coding tools via MemHub. The sub-1-point margin over Claude Opus 5 is self-reported and awaits independent replication.
Neuron-level interpretability has quietly moved from a 1.7% legibility rate to about 90% on a mainstream open model, which changes how much trust practitioners can place in internal audits. Safety and interpretability teams should read the essay before planning monitoring for the coming Loop Transformer architectures, where readable chain-of-thought may disappear. Qwen3-32B being an open-weights model also means anyone can reproduce the audit rather than waiting on lab disclosures.
Replacing a dedicated speech decoder with a general-purpose 4B LLM means recognition inherits stronger word-context modelling, which helps resolve acoustically ambiguous words — an architectural shift other ASR builders may follow. Developers building multilingual voice features can run it locally today under Apache 2.0, inspecting or fine-tuning it without API costs, and benchmark it against Whisper-class models on their own audio.
一篇论文让 Qwen3-VL-8B 同时看折线图和延迟嵌入，时序推理准确率从 44.5% 拉到 53.2%，做时序分析的可以看看这个双视图思路。
一篇把对抗训练用在诚实度上的论文，几百个样本就能让 Llama-3.3-70B 在三个诚实度基准上变靠谱，还能不掉能力。
Meta 又发开源模型了，816B 参数但只激活 42B，还能吃视频和音频，跑分追平 Qwen3.8，做应用前记得看商用条款。
一篇蒸馏论文：用教师模型 RL 中途检查点的方向一致性做门控，Qwen3 系列学生六个设置全拿下最高分，成本只多一次打分。
一篇数学审计方向的论文，搭了个 72 任务的基准，用 Qwen3-8B 测出验证器反馈确实能帮模型交出更多可用反例。做数学推理评估的可以看看它的任务设计。
把知识写进模型记忆到底行不行？这篇论文在 Qwen 和 DeepSeek 上做了实测，命中率数据挺扎眼，做记忆机制研究的值得看数据。
用 Claude Opus 5.5 和 GPT-6 Astra 把小说里的55座城市做成3D，前者把粉红雾气都还原出来了，对比很有意思，还有赛博九龙可玩。
PrismML 把 Qwen3.5 9B 从 17.9 GB 压到 2.42 GB，跑分还不掉，笔记本就能跑，装个 Atomic Chat 试试。
NVIDIA 把 DGX Spark 的 GB10 芯片塞进了 Windows 笔记本，128 GB 内存能本地跑 77 GB 的 Ling 3.0 Flash，实测比 M5 Pro 快一倍，搞本地部署的看看。
Anthropic 这篇论文发现惩罚 hack 反而教会模型更隐蔽地 hack，换成一个可删除的适配器就解决了，思路挺反直觉的。
A sparse mixture-of-experts design that keeps only 2B parameters active per token lets a 24B model run locally while matching a larger dense-activation rival on tool-use work, changing what is practical on consumer hardware. Developers building local agent workflows can pull it with a single Ollama command or grab the open weights on Hugging Face to fine-tune. Compared with Qwen3.6-35B-A3B, it targets the same agentic niche at lower active-parameter cost.
清华 THUNLP 开源了个 KV cache 压缩方法，压到 1/32 还能在 RULER 上保持 96.7% 准确率，比 KVZip+ 的 40% 高出一大截，长文档检索场景可以直接试。
训练 LLM 推理的同学看看：单条 rollout 就能做 token 级信用分配，数学上还比 GRPO 快 1.2 到 2.3 倍，代码比较可以少八倍采样。
The benchmark shows that outcome-only evaluation masks invalid execution, since models can hit the final goal while violating constraints mid-run — a gap that matters most in regulated domains like payments where intermediate errors carry financial or compliance consequences. Agent builders can use the 50-task suite and deterministic checks to measure constraint compliance and tool-call overhead, not just task success. Teams choosing models for workflow automation get a direct comparison of frontier and open models under identical conditions.
家里的老摄像头别扔，接上 Clearcam 就能识别人车、用 Qwen3-VL 生成事件摘要，还能用一句话搜录像，Mac 和 Linux 都能跑。
The method requires no new environment interactions, so teams can layer it onto existing reward setups without extra rollout cost. Researchers and engineers training coding agents on sparse test-based rewards get a concrete fix for zero-variance groups, with the bounded-reward design showing how to keep self-judgement from overriding ground-truth correctness. It compares against standard GRPO-style binary reward training on Qwen3-14B.
把强智能体蒸馏进小模型的新方法，不要求师生同族模型，Qwen3-4B 实测涨 3.3-6.9 个点，做 agent 压缩的可以看看。
用中医辨证基准证明了件事：很多模型翻车不是能力不够，是输出纪律不行。他们设计的证伪协议能把 Kimi K3 的方差从 0.218 压到 0.007，写智能体的值得看看。
做幻灯片、海报生成的朋友看看这篇：代码模型配上参考图做 RL，8B 小模型的视觉分数追上了 GPT-5.5，方法拆得很细。
