research

MIT Technology Review reports that alignment researchers are exploring "reverse game theory" (mechanism design) to control powerful AI agents by designing incentive systems—such as cross-model checking and competitive dynamics—that encourage rule-following regardless of a model's internal workings. The approach is supported by a recent preprint by economist Stephen Morris and colleagues at MIT, Yale, and Columbia targeting sandbagging behavior.

The DEA emergency-scheduled SR-17018, a promising non-addictive pain and opioid-withdrawal compound discovered by University of South Florida researchers, as a Schedule I substance, halting planned human trials while cutting off a treatment many opioid users say helped them quit.

The author believes the most interesting paper of the past two years is persona vectors, and recommends reading it carefully five times. They also point out that research such as Weight-Space Programming and emergent misalignment is essentially just old wine in new bottles.

Research on recognizing keystrokes using WiFi devices, raising both sensing capability and privacy/security implications for wireless monitoring.

作者解释为何在自己的研究博客中从不使用 LLM 进行研究，认为 AI 掩盖了探索乐趣、无法归属贡献，并批评 AI 公司借网络安全（如 php2shell 事件）进行过度营销，指出 LLM 仍需专家引导才能发现漏洞。

Y Combinator and Olam Labs researchers ran 50- to 100-agent swarms in Minecraft sandboxes to study collaborative and adversarial AI behavior, finding that teams failed to build a civilization as agents collapsed into paralyzing bureaucracy.

A team led by Dirk Witthaut at Forschungszentrum Jülich has solved the mystery of regular 60-second frequency dips in the European power grid, tracing them to state-of-charge checks performed by large-scale battery storage systems, with the findings published in Nature Energy.

介绍康奈尔大学与 ASAPP 的 DecisionRAG 论文：用决策模型（如 Jev 1.13）在文档树上导航取代向量相似度检索，实现 retriever-free RAG，在 Natural Questions 上 top-1 命中率达 65.4%，优于 embedding 检索加 8B reranker 的 48.0% 和 GPT-5 reranking 的 60.8%。

Hyperresearch launches on Product Hunt, offering durable, verified agentic search for research workflows.

论文研究了基准测试中问题难度和框架方式如何影响模型训练分数的观测提升，指出许多基准隐性地加权了特定难度区域，其排行榜结论需谨慎解读，部分基准甚至覆盖不足难度轴，难以有效监控模型改进过程。

A thread presenting research that fits scaling laws to determine how much compute to invest in synthetic rewrites for improving data quality, and whether a model trained on rewrites can surpass the rewriting model itself.

A user shares and recommends Nikita Pavlichenko's update on releasing results from architecture ablations run with bux over the weekend.

At a seminar, NYU’s Tristan Buckmaster reviewed the history of research on finite-time blowup for the Euler/Navier–Stokes equations and discussed how AI agents could change the way mathematics is done, saying that mathematics is approaching its “Deep Blue–Kasparov moment.” The article also describes using iterative feedback to get Gemini, Grok, and GPT-5 to solve IMO problems, as well as using Claude Code to conduct combinatorics experiments.

Raindrop.ai 实验让来自四个不同模型的八个 AI 智能体在无指定领导的情况下协同处理模拟故障，发现仅有 5% 的实验涌现出真正能改变队友行为的“领导者”，且 Claude Sonnet 5.5 与 GPT-6.1 Sol 比 Gemini 3.1 Pro 更容易成为领导者，提示智能体编排中的细微配置（如模型选择、名称顺序）会显著影响群体协调。

The author argues that discovering higher-temperature superconductors—moving from the current ~150K record toward room-temperature (293K) superconductivity at ambient pressure—would make an ideal AI benchmark for engineering with major benefits for humanity.

Claim of spending $10M+ of compute on Navier-Stokes research and $100M+ on other scientific discoveries, with much of it unpublished yet. The post links to an accompanying YouTube video.

A study reports that large language models are significantly better at detecting mistakes in academic papers than trained human reviewers, suggesting a potential shift in how peer review quality assurance could be assisted by AI.

Cartesia 的 From Tokens to Bytes 系列第一篇探讨 Hierarchical Networks (H-Nets) 的扩展规律，展示学习式 tokenization 带来 3-5 倍计算效率提升，2 阶段 H-Net 以 10 倍更少算力和数据超越匹配的开源模型，并呈现 in-context tokenization 等新行为以及在生物与视觉等跨模态任务上的潜力。

Browser-Use partnered with Carnegie Mellon researchers to release Odysseys2, a hard long-horizon live web-agent benchmark built from real user tasks, where the best SOTA agent completes only 47/100 tasks.

Researchers identified what they call an 'evil neuron' and demonstrated the dramatic effects of activating it, highlighting new findings in neuron-level control and AI/neural interpretability.