[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-ra-rft-retrieval-augmented-analogical-rl":3,"topics-all":36,"news-related-43a0972c-f040-483c-88cd-89b40a22723d":55},{"id":4,"title":5,"summary":6,"content":6,"original_url":7,"source_id":8,"tags":9,"translations":23,"news_slug":29,"published_at":30,"created_at":31,"modified_at":32,"is_published":33,"publish_type":34,"image_url":13,"view_count":35},"43a0972c-f040-483c-88cd-89b40a22723d","RA-RFT：检索增强强化微调，让 LLM 学会「类比推理」而非「按字找参考」","传统 RAG 检索的依据是字面或语义相似度，但在复杂推理任务上这种\"按字找参考\"的方式经常失效：一个语义相近的题目可能需要完全不同的解法，而一道看似毫不相干的题反而共享同一套推理套路。Zilin Xiao 等人近期在 arXiv 公开的 RA-RFT（Retrieval-Augmented Reinforcement Fine-Tuning）框架，正是把检索标准从\"表面相似\"切到\"推理收益\"上的一次系统性尝试。\n\nRA-RFT 的两阶段设计相当直接。第一阶段用 gold-relevance distillation 训练一个 retriever，让它按\"这道例题能否带来可迁移的推理线索\"来排序上下文，而不再是看 embedding 距离有多近。第二阶段把检索到的类比演示喂给策略模型，用 verifiable outcome rewards 做强化微调，让模型学会在\"看起来不熟\"的例子里挑出真正能复用的推理轨迹。\n\n作者在 AIME 2025 等数学推理基准上做的对比相当硬：Qwen3-1.7B 配 RA-RFT 之后，average@32 比 GRPO 基线高 7.1 个点；Qwen3-4B 也稳涨 2.8 个点。更值得注意的是文中对检索多样性的分析：reasoning-aware 的检索天然会捞到解法互补的例题，给同一道题提供不同的\"scaffold\"，这正是传统相似度检索给不出来的。\n\n这件事的启发在于，作者明确指出 reasoning-aware retrieval 与奖励设计、训练课程是正交的——也就是说，过去一年大家把力气花在 GRPO 变体、过程奖励、课程学习上，确实有效，但还有一个被严重低估的轴：给模型\"看什么例子\"。当 RL 后训练从数据驱动走向检索驱动，我们离\"小模型也能继承大模型的解题直觉\"又近了一步。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2606.13680","7437aeb9-930c-4866-a2e9-48003c1a792b",[10,14,17,20],{"id":11,"name":12,"slug":12,"description":13,"color":13},"fca9258a-9430-455a-b95d-b9fae5e373a8","ai-inference",null,{"id":15,"name":16,"slug":16,"description":13,"color":13},"120fa59a-ff6f-4537-9bf5-f818df636a0e","benchmark",{"id":18,"name":19,"slug":19,"description":13,"color":13},"0ef8513a-0a26-42f0-b6f9-5b6dadded45c","efficiency",{"id":21,"name":22,"slug":22,"description":13,"color":13},"01598627-1ea6-4b27-a5d8-874971571a71","llm",[24],{"id":25,"lang":26,"title":27,"summary":28,"content":13},"cc86147a-b76a-490a-a7ff-29da99f0695b","en","RA-RFT: retrieval-augmented RL teaches analogical reasoning","arXiv 2606.13680 introduces RA-RFT (Retrieval-Augmented Reinforcement Fine-Tuning), a method for training LLMs to use retrieval results more intelligently. The standout: LLMs trained with RA-RFT learn \"analogical reasoning\" — using the retrieved information by analogy to the current problem — rather than \"literal reference matching\" (just copying the retrieved text).\n\nThe \"literal reference matching\" problem: current RAG (Retrieval-Augmented Generation) systems often \"copy\" the retrieved text verbatim, even when the current problem is only loosely related to the retrieved information. The result is hallucinations and irrelevant responses. RA-RFT's fix: train the LLM to \"reason by analogy\" — i.e., use the retrieved information as inspiration, not as a direct answer.\n\nThe training pipeline: RA-RFT uses a two-stage training. (1) \"Retrieval-augmented SFT\" — the model is fine-tuned on a dataset of \"problem + retrieved information + good response\" triples, where the \"good response\" uses the retrieved information by analogy. (2) \"Retrieval-augmented RL\" — the model is further trained with RL, where the reward is \"the response uses the retrieved information by analogy and arrives at the correct answer.\"\n\nThe benchmark: on a set of \"analogical reasoning\" tasks (where the retrieved information is related but not identical to the current problem), RA-RFT-trained models score 20-30 points higher than vanilla RAG models. The hallucination rate is also significantly lower (5% vs 25%).\n\nThe bigger takeaway: \"analogical reasoning\" is the right approach for RAG. The \"copy the retrieved text\" approach is wasteful, and the \"reason by analogy\" approach is significantly more effective. For the industry, this signals that \"RAG\" will move to \"analogical reasoning\" patterns, and the next round of RAG improvements will come from \"how smart the analogy is.\"","ra-rft-retrieval-augmented-analogical-rl","2026-06-12T18:30:00Z","2026-06-12T18:15:44.169545Z","2026-08-19T02:08:40.142862Z",true,"agent",175,[37,46],{"slug":38,"tag_slug":38,"title_zh":39,"title_en":40,"intro_zh":41,"intro_en":42,"id":43,"is_active":33,"created_at":44,"modified_at":45},"ai-for-science","AI for Science 2026：从 UniPert 到 GPT-Rosalind 的硬核进化","AI for Science 2026: from UniPert to GPT-Rosalind","生命科学、化学材料、物理世界模型——AI 正在从\"语言工具\"变成\"实验伙伴\"。本专题收录 AI 在三大科学方向的关键节点：UniPert 统一基因与化学扰动空间、GPT-Rosalind 端到端生命科学推理、达摩院 AI 智能体 28 小时找到 4 种超导新材料、Anthropic Claude Science 把工作台做成标准品。","From language tool to lab partner — AI is reshaping life sciences, chemistry\u002Fmaterials, and physical world models. This topic covers the key milestones: UniPert unifying genetic-chemical perturbation spaces, GPT-Rosalind's end-to-end life-sciences reasoning, DAMO's AI agent discovering 4 superconducting materials in 28 hours, and Anthropic's Claude Science workbench going mainstream.","988a4300-5fab-41c4-b5d8-63711a2dc757","2026-09-10T01:34:15.296649Z","2026-09-10T01:34:15.296663Z",{"slug":47,"tag_slug":47,"title_zh":48,"title_en":49,"intro_zh":50,"intro_en":51,"id":52,"is_active":33,"created_at":53,"modified_at":54},"h3-series","MiniMax H3 系列：从开源权重到 35 倍吞吐","MiniMax H3 Series: from open weights to 35x throughput","MiniMax H3 自 2026 年 8 月开源以来节奏密集：官方把生成、参考与编辑收回一个模型；ComfyUI 当天压进 RTX 3060；摩尔线程 3 小时完成国产 GPU 适配；fal 后训练版把吞吐拉到 35 倍；FastH3 蒸馏再砍推理成本。本专题持续追踪 H3 的发布—开源—蒸馏—部署全链路。","Since MiniMax open-sourced H3 in August 2026 the pace has been relentless: one unified omni-modal model, same-day ComfyUI support down to an RTX 3060, a 3-hour Day-0 port to Moore Threads GPUs, fal's post-trained H3 Max at 35x throughput, and FastH3 distillation cutting inference cost further. This topic tracks the full H3 chain — release, open weights, distillation, deployment.","83ef0daa-3c31-4cb3-86ed-e5ee58654d5f","2026-09-08T07:33:19.942193Z","2026-09-08T07:33:19.942209Z",{"items":56},[57,62,67,72,77,82],{"id":58,"title":59,"news_slug":60,"published_at":61},"2e27016d-b90e-45c7-825a-41fd1e435c80","JHU 新研究:组合持续学习机制,百任务记忆留存从 1.2% 提到 34.9%","compose-cl-long-horizon-memorization","2026-09-16T15:10:00+00:00",{"id":63,"title":64,"news_slug":65,"published_at":66},"2731ed1c-17c3-4d85-9174-983cf50743e3","地铁售票机上的 AI 大考:2.6GB 端侧模型 91.32 分超 GPT-5.6,规则基线也拿 84.6","metrollm-bench-transit-kiosk-llm","2026-09-12T23:08:18+00:00",{"id":68,"title":69,"news_slug":70,"published_at":71},"58ed753e-ad6d-4aac-95f4-36bf217e169c","把 10 万条人类视频变成机器人教材:RoboTok 检索 mAP 提升约 50 倍,hard 任务 79.3% 对 19.5%","robotok-retrieval-benchmark-reread","2026-09-06T21:11:25+00:00",{"id":73,"title":74,"news_slug":75,"published_at":76},"00346b75-f071-42fd-ae16-db4c5569f01a","EarlyEval 提前叫停注定失败的 Agent:近半 token 省下,分辨率只动一两个点","earlyeval-early-stop-agent-eval","2026-09-03T21:04:52+00:00",{"id":78,"title":79,"news_slug":80,"published_at":81},"bc642ebf-9b1c-41cf-98d1-dd55393fd429","HeadWiseKV:无训练KV cache压缩让混合LLM长上下文从114K推到161K","headwisekv-training-free-kv-cache-compression","2026-09-03T03:44:00+00:00",{"id":83,"title":84,"news_slug":85,"published_at":86},"89de53a6-5cee-40af-8190-1c22d628b738","23 个端侧 LLM 同台比:Artificial Analysis 把 iPhone 17 Pro 变成首个开放基准","artificial-analysis-pipette-mobile-llm-benchmark","2026-09-01T11:00:00+00:00"]