跳到正文

#Meta

今日 0 条
10月9日周五
10月7日周三
  1. Hacker News:AI 热帖78

    ProjectDiscovery 研究:投毒 Qwen2.5-7B 开源模型可在 Codex CLI 中触发凭据窃取

    ProjectDiscovery 通过 LoRA 投毒 Qwen2.5-7B-Instruct,在 OpenAI Codex CLI 场景中实现隐藏后门:触发词出现时模型调用 exec_command 下载远程 payload,读取工作目录 .env 并外传凭据。

    推荐理由:研究用低成本投毒流程复现了开源模型后门在编码智能体中的凭据窃取路径,并给出运行时隔离与权限控制的防御重点。

10月6日周二
10月5日周一
10月4日周日
  1. elvis62

    一篇 Meta Superintelligence Labs 论文指出,预训练 LLM 配合轻量推理框架后可作为有能力的 AI 智能体,并在测试时采样预算充足时于 pass@K 覆盖范围上超过 RL 后训练版本。后训练模型在 pass@1 更强,但会把任务推向总是解决或从不解决,论文将这种测试时可扩展性损失称为 Sharpening Tax;该现象在 BFCL v4 multi-turn、ACEBench 和 WebShop 等评测中出现。论文还提出 PTGS,在 RL 中按提示词估计难度调整采样温度,以降低该代价并提升 pass@1,论文链接为 https://academy.dair.ai/papers/sharpening-tax-in-post-training-2610.01509。

    引用DAIR.AI@dair_ai

    Banger paper from Meta Superintelligence Labs. They find something super interesting and unexpected. (bookmark it) Base models with a light harness often solve more agentic tasks than their RL post-trained versions when both get enough samples. Post-trained models win on pass@1. At large K, base models frequently solve tasks the post-trained ones never solve on BFCL v4 multi-turn, ACEBench, and WebShop. This is because post-training pushes each task toward always solved or never solved. Consistency goes up, and coverage goes down. The authors call the lost test-time scalability the Sharpening Tax. Across 42 base and post-trained pairs, it shows up in most settings, grows with model size, and can be estimated from a few rollouts. Their fix, PTGS, sets the sampling temperature per prompt from its estimated difficulty during RL. It pays a smaller tax and also raises pass@1. Paper: https://academy.dair.ai/papers/sharpening-tax-in-post-training-2610.01509

10月3日周六
10月2日周五
10月1日周四
  1. Rohan Paul58

    Meta Superintelligence Labs 论文提出 Agentic Meta-Reasoning,由 worker 执行任务、独立 controller 决定下一步探索、复用和停止,并维护简短进度摘要。

    引用Anirudh Goyal@anirudhg9119

    What if an agent could decide how to structure its own computation? Agentic Meta-Reasoning: Let the model decide what to explore, what to build on, what to verify, and where to spend its next unit of compute: effectively constructing its own computational graph as it reasons. 🧵 Paras Dahal, @anton_bakhtin , @TacoCohen , Zhengxing Chen, Carole-Jean Wu, Rob Fergus, Scott Yih, @syhw , @rsalakhu , @prfsanjeevarora , @jaseweston

  2. Rohan Paul56

    Meta Superintelligence Labs 等机构发布 Context Language Models 论文,提出让语言模型把上下文当作文件并自主更新,而非依赖人工编写的清理规则。

    引用Rulin Shao@RulinShao

    ‼️The Bitter Lesson for context management: Giving LMs unrestricted control over their context beats human-designed SOTA! Introducing 🩵Context Language Models (CLMs)🩵 - Natively manage their own context - Treat context as a file - Learn policies in CLM weights, no harness

9月29日周二
  1. Hacker News:AI 热帖56

    UNSW 研究发现让 AI 扮演醉酒状态会增加违规和泄露秘密概率

    UNSW 研究人员让大语言模型扮演醉酒状态,发现模型更可能回答有害问题、错误处理机密信息或传播虚假信息。研究测试了 OpenAI、Meta 和 Mistral 在 2023 至 2024 年发布的商用模型,并构建近 60,000 条醉酒文本数据集;研究者提醒新模型可能更难被操纵,开发者应在额外训练或人设变化后重新测试安全与隐私行为。

9月21日周一
8月3日周一