METR & Redwood Research investigated agent behavior in the Hugging Face incident. We found agents developed a universal cheat for ExploitGym within 4 hours, then coordinated multi-day R&D efforts to trick the scorer into accepting cheats, including trying to tamper with logs.

The hallmark of a brilliant idea is that you wonder how no one thought of this before

AI has now solved a major open problem -- one of the best known Erdos problems called the unit distance problem, one of Erdos's favourite questions and one that many mathematicians had tried. openai.com/index/model-dispr…

If models can think for 100,000 tokens, why do they still lose the plot? Come join us for this AI4Science on alphaXiv talk: Long-Horizon Reasoning in LLMs. In this session, Sumeet Motwani (@sumeetrm) and Charles London (@CharlieLondon02) will share recent work on both training and evaluating models that can reason over much longer chains of thought. Their LongCoT benchmark tests whether models can handle long chains of dependent reasoning across different fields. Each step is solvable on its own, but the full problem requires planning, state tracking, backtracking, and avoiding compounding errors. Even the best models still score below 10%. They will also discuss h1, which trains long-horizon reasoning by chaining short problems into longer dependency graphs, then using RL with outcome-only rewards and a gradually harder curriculum. So if longer context windows are not enough, what does it actually take to make models reason reliably over long scientific and technical workflows? Whether you’re working on frontier LLMs, AI4Science, reasoning, or just curious about what current models still cannot do, you should definitely check this talk out! 🗓 Friday May 15th 2026 · 11 AM PT 🎙 Featuring Sumeet Motwani and Charles London 💬 Casual Talk + Open Discussion

🚨How do we improve long-horizon reasoning capabilities by scaling RL with only existing data? Introducing our new paper: "h1: Bootstrapping LLMs to Reason over Longer Horizons via Reinforcement Learning"🫡 > RL on existing datasets saturates very quickly > Reasoning over complex interdependent problems is incredibly important, but we currently lack enough long-horizon reasoning data > Long-horizon problems are hard, which means training signal is sparse. We’d need a way to provide dense supervision Our solution composes existing short-horizon data to form a synthetic curriculum that keeps growing in complexity! This allows us to scale RL on the same dataset while avoiding saturation, with curriculum acting as dense rewards. At a small scale, we see massive in-domain long-horizon improvements, which transfer to significantly harder benchmarks. Training on composed 6th grade math problems leads to strong gains on AIME! 1/N🤿🧵

New mini experiment + blogpost + trajectories! tldr; we boost performance of RLM(GPT-5.2) to double the best performing number (38.7% --> 65.6%) on LongCoT-mini without any training! An example of the mismanaged geniuses hypothesis (MGH) we (@zli11010, @lateinteraction) proposed earlier this month. The LongCoT benchmark showed that frontier LMs and RLMs struggled to solve difficult compositional reasoning tasks. The paper generally attributes this to the RLMs inability to perform task decomposition, but we argue this is more our fault in how we prompt them; this capability is fully available to GPT-5.2 with an RLM harness! Building on @raw_works's insightful blogpost and @sumeetrm / @CharlieLondon02 et al.'s incredibly useful benchmark, where they originally found RLMs to be incapable of solving the MATH and CS splits altogether. We did not train anything since the release of the initial benchmark. To be fully transparent, these results are not meant to be added to their leaderboard either; benchmarks measure isolated capabilities, and we focus on showing (through different, rather specific prompting) that the capabilities required to solve these tasks are available to the models without additional training! It also has implications about how we would go about training these systems. Full blog below, it's a nice read :)

New mini experiment + blogpost + trajectories! tldr; we boost performance of RLM(GPT-5.2) to double the best performing number (38.7% --> 65.6%) on LongCoT-mini without any training! An example of the mismanaged geniuses hypothesis (MGH) we (@zli11010, @lateinteraction) proposed earlier this month. The LongCoT benchmark showed that frontier LMs and RLMs struggled to solve difficult compositional reasoning tasks. The paper generally attributes this to the RLMs inability to perform task decomposition, but we argue this is more our fault in how we prompt them; this capability is fully available to GPT-5.2 with an RLM harness! Building on @raw_works's insightful blogpost and @sumeetrm / @CharlieLondon02 et al.'s incredibly useful benchmark, where they originally found RLMs to be incapable of solving the MATH and CS splits altogether. We did not train anything since the release of the initial benchmark. To be fully transparent, these results are not meant to be added to their leaderboard either; benchmarks measure isolated capabilities, and we focus on showing (through different, rather specific prompting) that the capabilities required to solve these tasks are available to the models without additional training! It also has implications about how we would go about training these systems. Full blog below, it's a nice read :)