OpenAI says its models produced a Lean-verified proof that a widely believed speed limit for solving logic problems is false, and outside experts will now check it. For builders, GPT-6.1 Terra halves API prices while solving 50.3% of Terminal-Bench 4.0 tasks.
Mirendil released Fern 0.1, its first model, pretrained from scratch with a context window of up to 10M tokens. The context window is how much text a model can read in one go; here that is about 23 MB of code, or half a million lines. Fern is built for agent work on AI projects, such as debugging a training run from its logs or auditing training data. Access is listed as coming soon.
OpenAI announced three new discoveries by its models and asked outside experts to check them. The biggest says the Exponential Time Hypothesis, a widely believed limit on how fast Boolean formula satisfiability can be solved is false; the proof was verified in Lean, a machine proof checker. A group led by Russell Impagliazzo, Ramamohan Paturi and Ryan Williams will verify it. A physics group led by David Gross will examine a claimed existence proof for pure Yang, Mills theory.
Δτ-Bench tests agents in tool ecosystems that keep evolving: familiar tools disappear or get replaced while more efficient alternatives appear. It has 1,588 multi-turn tasks across 42 domains and six successive ecosystem states. The strongest model or harness reached only 42.3% task success, and gains from existing skill-learning methods did not hold as tools changed. The authors' Steward manager, which maintains skills and tools during tasks, kept success higher.
TraceTrial-Bench grades agent harnesses, the software around a model that handles context, tools and checks, separately from the model itself. Across 4 harnesses, 9 models and 95 tasks, or 3,420 runs, harness scores ranged from 59.8% to 72.0%. In 31.1% of runs the agent finished the task while failing at least one harness criterion, so a passing result can hide an unreliable setup.
AutoTavern, a team of LLM research agents, rebuilt Hearthstone Battlegrounds as a simulator in 20 days and trained 23 versions of a game-playing agent. On the live ladder its rating rose from 6,000 to about 15,000, and its last 20 records average a placement of 1.60 with 80% first places, better than the server's ten highest-rated players. On play tasks it answers 91%, against at most 63% for frontier LLMs.
Rubric-based rewards let reinforcement learning train on tasks without one right answer, but they mostly reward covering more rubric points, so trained models grow verbose and add wrong content. The FORK reward also scores precision: each part of an answer must be necessary and factually correct. Against the recall-only reward it kept similar coverage with shorter answers, and its win rate over the base model rose from 22.1% to 74.1% on HealthBench.
Continual Learning Exam tests whether agents improve from experience. It has 120 expert-written tasks, each a five-level ladder, across 89 professional roles, and later levels build on earlier work. Across 17 harness-model setups, feedback after a failure plus one revision added a median 5.3 percentage points. The best system cleared 47.7% of levels and finished all five levels in just 23.3% of tasks.
Entire launched a Git forge built for agents and humans working together, available to every developer today. Instead of reading a diff, developers see Trails: custom evals that agents score on every push, and code can merge once all evals are satisfied. Entire says pushes are up to 89x faster than GitHub, and its codebase graph saves up to 71% on tokens. Private repositories can be hosted in the US, EU, India or Australia.
Riverside launched Riverside Agents: a Producer, Editor, Clipper and Marketer that cover a show from prep before recording to publishing. When recording ends, the Clipper flags moments worth sharing and the Editor drafts a plan for the cut; finished clips move to the Marketer with a suggestion of where to post. Nothing goes out without the creator's sign-off. The agents are rolling out on web and iOS.
Hopper runs inference for real-time voice agents, where slow replies break a conversation. It trains models on a customer's production traces, then serves and improves them against live traffic. It lists 80 ms time to first token against 600 ms for GPT-4.1, and $0.30 per 1M input tokens against $2. Data is not retained by default.
The policy group ARI published a blueprint for federal AI governance limited to a handful of well-resourced frontier developers. Their published safety frameworks would have to meet minimum federal standards, with independent assurance that they do. They would file confidential quarterly disclosures on how much AI automates their research, with material-change disclosures filed within four business days. The plan sunsets in nine years unless Congress reauthorizes it.
Baidu released Unlimited OCR 2 on Hugging Face, a model that turns page images into text. Its new attention method, SR-SWA, squeezes the links between text and page images in very long inputs into a fixed set of working references, such as 512 tokens. That keeps per-step decoding speed and memory nearly constant however long the document gets. Baidu reports the top results on MPDocBench and the hundred-page MarathonDocBench.
You’re caught up.
That was all 12 stories in the 27 October issue.