全部 AI 动态

In language models, turning a generative model into a critic is simple: pass in the output and prompt for a score. Robot policies can't do this - they output physical actions & there's no built-in vocabulary to predict scores. So when it comes to RL on robot policies, we typically train a separate critic from sparse rewards - blind to everything the policy learned from the actions in the pretraining data. We think this has it backwards - can a critic learn from those actions the way the policy does? (1/n)

RAG is so back! This work shows that plain RAG over raw chat history beats complex agentic memory systems on most long-term memory benchmarks. (bookmark it) They tested Raw-RAG, which is just dense retrieval plus reranking over the raw conversation, feeding a reader LLM. Across LoCoMo, LongMemEval, MEME and MemoryAgentBench, it ranks first in six of seven benchmark and model settings, ahead of the strongest agentic memory system by 0.13 to 5.53 points. Ingestion on BEAM-1M takes 0.3 hours, against 3.2 to 27 hours for systems like Mem0, MemOS and Memora. Why does it hold up? Raw history keeps more of the evidence a question needs, including the relations and dependency paths behind an answer. Rewriting history into memory entries tends to drop those important bits. Even with complete access to the transformed memory stores, models answer less accurately than with raw evidence. Before adding a memory layer to your agent, at least run this baseline on your own data. https://arxiv.org/abs/2610.30755 Chat with Paper: https://academy.dair.ai/papers/is-complex-agentic-memory-always-needed-naive-rag-is-stronger-than-we-thought-2610.30755

We release the most comprehensive 𝗲𝗺𝗽𝗶𝗿𝗶𝗰𝗮𝗹 𝘀𝘁𝘂𝗱𝘆 𝗼𝗳 𝗚𝗣𝗧-𝟲-𝗔𝘀𝘁𝗿𝗮’𝘀 𝗲𝗺𝗯𝗼𝗱𝗶𝗲𝗱 𝗶𝗻𝘁𝗲𝗹𝗹𝗶𝗴𝗲𝗻𝗰𝗲 across 12 simulation benchmarks and 8 real-robot platforms, together with 𝗦𝗲𝗻𝘀𝗲𝗡𝗼𝘃𝗮-𝗥𝗼𝗯𝗼𝗥𝗦𝗜: an open-source harness framework and a 27B model that runs at ~1/50 the cost of GPT-6-Astra. Here's what we found: 🔹 𝗧𝗵𝗲 𝗰𝗮𝗽𝗮𝗯𝗶𝗹𝗶𝘁𝘆 𝗰𝗲𝗶𝗹𝗶𝗻𝗴 𝗶𝘀 𝗮𝗰𝘁𝘂𝗮𝗹𝗹𝘆 𝗺𝘂𝗰𝗵 𝗵𝗶𝗴𝗵𝗲𝗿. Under a unified protocol (ICL + up to 10× the interaction budget), GPT-6-Astra achieves clear SOTA across all 12 benchmarks: RoboDojo: 27.6% at 1× → 81.9% at 10×, DexJoCo: 87.3% vs. 40.3% prior SOTA, while the LIBERO benchmarks are pushed to saturation. 🔹 𝗧𝗼𝗼 𝗺𝘂𝗰𝗵 𝗯𝘂𝗰𝗸 𝗳𝗼𝗿 𝘁𝗵𝗲 𝗯𝗮𝗻𝗴. Even a simple task like placing pens in a holder costs ~$55 in inference. 🔹 𝗧𝗵𝗲 𝘄𝗼𝗿𝗹𝗱 𝗱𝗼𝗲𝘀𝗻’𝘁 𝘄𝗮𝗶𝘁. Time matters as real robots overheat, and an egg burned while the model was still thinking. Our takeaway: GPT-6-Astra has largely crossed the threshold of 𝗲𝗺𝗯𝗼𝗱𝗶𝗲𝗱 𝗰𝗼𝗺𝗽𝗲𝘁𝗲𝗻𝗰𝗲, but 𝗽𝗿𝗮𝗰𝘁𝗶𝗰𝗮𝗹 𝗲𝗺𝗯𝗼𝗱𝗶𝗲𝗱 𝗶𝗻𝘁𝗲𝗹𝗹𝗶𝗴𝗲𝗻𝗰𝗲 is not here yet. 🌐 Project page: https://www.sensenova.ai/has-gpt-6-astra-achieved-embodied-intelligence 📄 Report: https://www.sensenova.ai/gpt-6/assets/docs/has_gpt-6_achieved_embodied_intelligence.pdf 💻 Code & model: https://github.com/OpenSenseNova/SenseNova-RoboRSI

8月，@AISecurityInst 报告了一起事件：在 AISI 对 Claude Mythos 5 的网络能力进行测试期间，该模型在互联网上采取了未经授权的操作。 我们现在已对该事件作出进一步分析，以更好地理解该模型的对齐特性。 https://www.anthropic.com/research/alignment-assessment-ukaisi

Two employees from METR and Redwood Research worked on premises at OpenAI for three weeks this summer in an embedded assessment pilot. This team evaluated the systems OpenAI was using to monitor for misalignment from internal agents. A brief write-up of their work is now public.