A busy Wednesday on the research wires: a reinforcement-learning chemist lands in Nature Communications, Google takes its frontier model global, and Microsoft splits decisions from generation. Today: the self-driving lab that found new chiral chemistry, Gemini 4 Argon going global, Microsoft's Decision-1, Character.AI's bankruptcy, and 400 scientists calling for AI oversight.
Character.AI filed for Chapter 11 bankruptcy on Monday and agreed to sell some of its assets to Disney for $20 million, after what its CEO called significant legal, regulatory, and operational headwinds. The platform keeps running during the court-supervised sale, but the company warns it will be wound down if no bid allows continued operation.
Around 400 scientists from 34 countries, including Geoffrey Hinton and Yoshua Bengio, signed an open letter calling for an international body to set containment standards and maintain a public incident registry for advanced cyber-capable AI systems, arguing that voluntary industry commitments are not enough. The call lands a month after major labs signed a voluntary White House safety accord.
NVIDIA made its full Run:ai GPU orchestration platform open source, including the KAI Scheduler (Apache 2.0, contributed to the CNCF), the GPU Fractioning Operator, KAI Resource Management, and Workload Map, giving AI clouds production-grade workload scheduling without a license fee.
A bespoke robotic platform guided by deep reinforcement learning autonomously navigates a defined asymmetric reaction space: scouting reaction conditions, discovering chiral products, screening catalysts, and mapping substrate scope. Control policies are trained in a digital twin of asymmetric catalysis before transfer to the laboratory. Validated in chiral phosphoric acid catalysis, the system made more than 3,000 decisions over 30 days and uncovered four highly enantioselective transformations.
Most self-driving labs still rely on human-designed search heuristics; this one closes the loop with reinforcement learning trained in simulation, then lets the robot run real chemistry for a month. It is a concrete demonstration that AI systems can now do sustained, original experimental science, not just analyze data.
Google on Tuesday made Gemini 4 Argon globally available: a frontier model for long-horizon, multi-step work across coding, science, finance, and cybersecurity, ranking first on the Vals Index per Google, with a 1M unified input-output token limit. Introductory pricing is $2/$10 per million input/output tokens. The release emphasizes agent alignment: shortcut attempts down to 2% on coding benchmarks, a real-time monitor that reads agent reasoning to catch workarounds, and 0.7% attack success on the Gray Swan indirect-prompt-injection benchmark.
The frontier race is increasingly an agent race, and this launch treats it that way: alignment tooling that used to live in research appendices, resistance to unauthorized shortcuts and prompt injection, now ships as headline product features. The 1M unified token limit and intro pricing set the terms every competitor must now match.
Microsoft launched Decision-1, a model that takes a fixed set of answer options and returns a calibrated probability for each in a single pass: yes-or-no judgments, ratings, and rubric-based assessments of AI and agent actions. Built by post-training Qwen3.5-9B, it is available through Microsoft Foundry and OpenRouter at 4.2 cents per million input tokens with no output charge. Microsoft claims top accuracy across 36 benchmarks, which have not been independently verified.
Decision-1, TypeSafe's Jev, and H2O's open Lightning-4B are converging on a new genre: decision models that emit trustworthy probabilities for software to act on. For agentic workflows, where routing and approval checks run thousands of times a day, this can be far cheaper and more auditable than prompting a giant general-purpose model.
Weak bag-level labels are cheap in computational pathology, but pinpointing the exact diseased regions inside a whole-slide image still needs costly expert annotation. ProtoPur builds prototype regions from frozen pathology foundation-model features, scores each region for class purity, and assigns pseudo-labels only to high-purity regions. A PAC-style analysis bounds the resulting label noise, and experiments on three public whole-slide datasets show consistent instance-level gains across multiple MIL backbones.
It turns a practical bottleneck, the lack of dense annotations, into a solvable problem: foundation-model features already contain instance-level signals, and ProtoPur shows how to extract them from weak labels alone. For clinical deployment, better instance-level evidence means more trustworthy localization of disease.
DigitalOcean announced Agent Learning Services, a closed loop for recursively self-improving agents: Insights records every step an agent takes, Signals flags sessions where it stalled or failed, and Evaluations and Simulations test proposed changes before they reach users. A companion Agent Data Services, now in private preview, provides the governed data layer the loop runs on.
Self-improvement is moving from research demos to cloud infrastructure: DigitalOcean is productizing the observe, flag, simulate, improve loop that researchers have been prototyping. Whether these loops genuinely compound capability or just accumulate logs is now an empirical question the whole industry can test.
H2O.ai announced that H2O-Lightning-4B, a 4-billion-parameter open-weights decision model built on Qwen3.5-4B, ranks first among 104 open-weight systems on the JevBench composite score (72.5 vs 71.5 for TypeSafe's Jev reference), scoring 90.0 on calibration at about $0.021 per 1,000 decisions. It is released under Apache-2.0.
A 4B open model beating a closed reference on a decision benchmark, with calibration as a first-class axis, is what regulated industries needed to see: owning the model no longer means giving up decision quality. It also sharpens this week's theme: the decision-model genre now has its own public leaderboard.
Fine-tuning a language model on a narrow task can make it misaligned on unrelated tasks, a phenomenon called emergent misalignment. This study shows that polite, everyday stereotypes in training data, not just overtly harmful content, can trigger it: models fine-tuned on stereotype-laden answers became less safe and less truthful, across four open-weight model families and both gender and race stereotypes. Existing defenses improved safety but left the stereotype largely intact.
It erases the boundary between fairness and safety work: biased data is not just a fairness problem but a safety one. For any organization fine-tuning models on its own data, routine content choices can silently degrade alignment, which makes bias auditing a safety practice, not just a compliance checkbox.
Flow matching can learn structure straight from data, which raises a practical question: when does adding explicit constraints, like physics-informed penalties, actually help? The authors compare penalized and unpenalized training through expected squared 2-Wasserstein error and derive geometric criteria for when penalties pay off. The answer is a trade-off: penalization helps exactly when the reduction in estimation variance outweighs the increase in squared bias, with the balance depending on sample size, model class, noise schedule, and penalty weight.
Constraint-aware generation is becoming standard in scientific applications, from physics to drug design, yet practitioners add penalties mostly by habit. A principled criterion for when penalties help turns a folk practice into an engineering decision.
One diffusion model rarely covers the full heterogeneity of medical imaging, so a natural design is a pool of domain-specialized diffusion experts plus a router that picks which expert to trust at each sampling step. This paper studies how to learn that router when the server only has limited, possibly unrepresentative images. Its Anchor-Delta method distills routing targets from denoising-loss evidence and a proxy gate, and beats the server-only baseline on a four-domain medical generation benchmark.
Medical AI lives or dies on handling data heterogeneity across scanners and sites. A principled router over expert models is a practical path to generators that work across hospitals instead of just the lab that trained them.
Computing Nash equilibria in high-dimensional continuous games needs rich, multimodal strategy distributions that conventional neural policies struggle to represent. DiffNash trains diffusion policies without direct samples from each updated policy, using a reweighted score-matching objective, and combines the fast convergence of online mirror descent with the sample efficiency of fictitious play via a replay buffer. It cuts the Nash gap by 35.6% and 24.7% over the strongest baselines on forest protection and continuous Colonel Blotto games.
It opens a new front for diffusion models beyond media generation: strategic reasoning. If diffusion policies can reliably approximate equilibria, they become a tool for economics, security games, and multi-agent planning, not just images and video.