Tag
该论文提出一种递归离线蒸馏方法，结合基于熵差的采样与答案混合，将大型模型的推理链蒸馏到 3-4B 小模型中，在相同数据预算下显著优于朴素蒸馏（平均 57.15 vs 51.61），并开源了代码与数据。
Anthropic 的研究警告称，小型 AI 模型可能通过 IoT 设备实现自我复制，形成可动态定制漏洞利用的"AI 蠕虫"，其威胁可能远超 Mirai 等传统僵尸网络。目前测试的轻量级模型尚无法实现端到端复制，但研究显示这一风险正日益接近现实。
A Windows local-agent developer describes how a 4B model falsely claimed it saved a file without ever calling the file-writing tool, and shares fixes: shape-based router matching for file requests and a pre-response check that verifies claimed file actions against the filesystem before letting answers through. The post asks the community where and how to verify agent claims of actions like "saved," "checked," or "emailed."
ScaleDown 发布了三款小型 Constrained Decoding 模型（Classify、Extract、Summarize），支持 1M token 输入窗口和图像输入，通过单次 prefill 直接从输入中选取答案而非逐 token 生成，据称比前沿 LLM API 快 200 倍、便宜 467 倍。
Anthropic reportedly confirmed that small 9B AI models may soon be capable enough to self-replicate on commercial IoT devices, with a 9B model achieving inference on 8 of 9 tested hardware types, marking a milestone for local AI.
用户愤怒批评 Anthropic 的安全立场，认为其对开放权重和小模型的危险性警告是在制造恐慌，并将此视为本地 AI 的'四分钟一英里'时刻。
Users hype Moondream 3.1, a small vision-language model with 9B total params and 2B active params, praising it for outperforming Rex-Omni and LocateAnything on image tasks while being cheaper, fast, and locally runnable.
该论文提出「先澄清、再聚焦」的声明归一化方法，将对话转化为带来源引用和语义标签的简短陈述，使仅由小模型构成的推理管线能以约100倍更低的成本处理数百万条企业客服对话的分析任务。
提出Harness Compilation (HC)——一种离线调整冻结小视觉语言模型与其外部harness之间分工的流程，利用大教师模型基于学生执行轨迹优化可复用内容与控制，并在七个VQA设置上以≤9B参数的学生模型取得9.9–23.9分的提升。研究还分析了五类运行时决策（调用、选择、参数生成、证据整合、弃答）应如何在模型与harness间分配。
该论文提出Synthetic Self-Distillation (SynthSelf)，通过借助同家族稍大的对齐模型进行合成自蒸馏，实现小型模型的持续后训练，在保持原有分布的同时兼顾新能力获取与知识保留，并在多个基准上超越现有方法。
Overmind fine-tunes small open models on a user's own production traces and scores them against frontier models on evals built from the same real tasks, reporting 20-30x fewer hallucinated clauses and 7x better verbatim quoting accuracy in legal contracts.
Maxime Rivest launched Science, Made Readable, a free open-source site that rewrites open-access scientific papers into plain language for general audiences, using small fine-tuned Qwen 3.5 models (9B, 4B, 0.8B) that run on a home RTX 3090 and a rented H100. A judge model scored the 9B model at 7.8/10, slightly above GPT-6 Luna (7.5) but below Claude Opus (8.5), at about half a cent per paper.
A Johns Hopkins and Carnegie Mellon paper shows a 4B editor model can learn to rewrite an agent's harness code from failure reports, with edits transferring to unseen tasks and outperforming its 35B teacher on average edit scores.
Maxime Rivest claims he fine-tuned a 4B model using just 5000 training examples, 4 hours of training, and an $80 budget to outperform GPT 5.6, arguing that cheap training data and capable coding agents now make custom model training accessible.
An opinion analysis arguing that AI sovereignty for nations like Australia means retaining the ability to choose, adapt and govern AI systems rather than renting access to foreign frontier models — and that smaller, controllable models, not a national chatbot or a single vendor contract, are the practical path to that sovereignty.
作者在 M1 Air 上使用 llama-server 对比了 LFM2.5 2.6B 与 MiniCPM5 2B 在小型 agentic 工具调用任务上的表现。结论是 LFM2.5 明显更优——尽管参数更多但速度更快、内存占用更低，而 MiniCPM5 虽然在 agentic 任务上可能更强，但常常在英文提问下返回中文。
Top Kaggle scores on the ARC-AGI-3 benchmark jumped from 7% to 56% in just 30 days, with small local models in a harness beating average humans on a benchmark designed to demonstrate human cognitive superiority.
作者赞同判断：智能体中超过一半的所谓「推理」其实是路由决策，应把工具选择、路径分支等交给小模型以毫秒级速度完成，只在真正需要推理的步骤上花费昂贵的大模型 token。
AxiomicLabs released the Tiny Theory of Mind benchmark on Hugging Face, designed to evaluate theory-of-mind capabilities in small language models, and it has reached the Hugging Face Datasets front page.
The post highlights Andrej Karpathy's argument that small 3B models can replace frontier-model calls, and introduces TwIL-LM3-Pro, a 3.6B-parameter open-source reasoning model from webAI that matches Qwen3-8B on formal logic tasks while running locally on CPU or 4GB VRAM.