A Friday of agents doing science: a multi-agent AI for drug discovery in Science, language models doing sequential diagnosis, and where ER triage still treats patients unequally. On the news wire: four new open agent projects, Meta's $25 billion privacy penalty, and OpenAI's move into drug development.
The Agentic AI Foundation announced four new hosted projects at AGNTCon + MCPCon North America: Agent Skills from Anthropic, Open Responses from OpenAI, OpenEnv from Hugging Face, and OpenSandbox from Alibaba. The additions take the foundation from three projects to ten in its first year, covering how agents gain skills, how apps talk to models across providers, how agents are trained and evaluated, and where they run safely, all under vendor-neutral governance.
A New Mexico judge ordered Meta to pay $25 billion after a jury found the company misled consumers about data privacy, in a case rooted in the Cambridge Analytica scandal. Judge Francis Mathew calculated that more than $219 billion was warranted under state law before reducing it to $25 billion, still one of the largest corporate penalties in US history. Meta says it is confident it will win on appeal.
OpenAI acquired Orchestra Bio, a startup whose AI agents help pharmaceutical teams coordinate drug development programs, folding its team and technology into OpenAI's Rosalind life sciences unit. The agents track studies, flag schedule risks, and map how setbacks ripple through a program, while scientists keep final decisions. Terms were not disclosed.
Self-driving truck company Einride will test NVIDIA's Alpamayo vision-language-action model in its vehicles, aiming to become the first trucking operator to deploy it. The model interprets driving environments, recommends actions, and explains its reasoning; initial testing runs in shadow mode, with live road evaluation planned for the second half of 2027.
A Stanford team built the Virtual Biotech, an organization of AI agents modeled on a drug-development company, with agentic divisions for target discovery, safety assessment, modality selection, and clinical development. In one demonstration, more than 37,000 agents annotated outcomes from 55,984 clinical trials and found drugs targeting cell type-specific genes were 48% more likely to reach market.
Drug development fails more often from fragmented evidence than from bad science. An agentic organization that integrates evidence across scales, and shows its work at real decision points, is the most credible vision yet of AI reshaping how medicines get made.
Microsoft researchers built an interactive benchmark for sequential diagnosis, the iterative way clinicians actually reason: given a brief presentation, the system must decide what information to gather next and when to commit to a diagnosis. Their AI system simulates a panel of physicians that continually revises the differential diagnosis while selecting cost-effective tests, reaching 83.2% diagnostic accuracy on the benchmark.
Most medical AI is tested on one-shot questions, but real diagnosis is a sequence of decisions under cost and uncertainty. A benchmark that grades the workup, not just the answer, moves evaluation closer to the clinic.
Analyzing 1.35 million encounters at four U.S. emergency departments, the authors compared nurse triage assignments against Emergency Severity Index criteria applied computationally. Patients from ethnoracial minority groups were systematically assigned lower urgency than the criteria warranted, with odds ratios of 0.75 to 0.90 for Hispanic versus non-Hispanic White patients, and the disparities were largest in lower-risk visits.
Triage decides who waits and who gets seen, and small disparities compound into very different care. Finding the bias concentrated in lower-risk visits tells hospitals exactly where to look, and where AI triage tools must be audited first.
OpenMAP-HIE localizes neonatal hypoxic-ischemic encephalopathy lesions on ADC images without any voxel-level lesion labels. A diffusion model learns the healthy neonatal ADC distribution while a ControlNet branch preserves patient-specific anatomy; lesion evidence comes straight from the gap between conditional and unconditional noise predictions. Pathology knowledge of the characteristic ADC reduction in acute HIE guides the design.
Neonatal brain injury needs fast, expert-free localization, and annotated scans are scarce. An unsupervised method that writes pathology knowledge directly into the model points to how medical diffusion models can work where labels cannot.
Pathology foundation models still stumble when a classifier trained on one hospital’s slides meets another hospital’s: the domain shift problem. AdaptMIL folds domain adaptation into multiple instance learning with a two-stage procedure that decouples representation alignment from MIL aggregation, after the authors found that naive joint training can actually hurt.
Hospital-to-hospital generalization is the quiet blocker between pathology AI papers and pathology AI products. A training recipe that respects how MIL actually learns is the kind of unglamorous fix that moves the field.
A question about a whole-slide image rarely names every morphological finding needed to answer it. BeyondPath uses a Preserve-Focus-Fuse design: it preserves slide-wide morphology in a Morphology Bank learned across questions, retrieves question-specific evidence from all foreground patches, then fuses the two into a compact visual prefix for the language model. It led all eight report-generation metrics on WSI-Bench with a single checkpoint.
Whole-slide question answering has been torn between looking everywhere and looking where asked. This shows the answer is both: keep the broad morphology, then aim the question at it.
Anthropic researchers built Agent4Sci, a long-horizon testbed that measures how far today’s frontier coding agents can improve published best models on 20 scientific ML benchmarks across 11 domains. Claude Opus improved the published best on every benchmark, by 10.4% on average, but the gains came almost entirely from ML engineering tricks rather than research-level insight; lightweight plug-in skills for ideation and experiment management nearly doubled the improvement.
The industry pitches coding agents as automated scientists; this is one of the first large-scale measurements of that claim, and it draws a clean line between engineering sweeps and genuine research insight.
As agent platforms start improving their own deployed agents with no human in the loop, the authors propose two measurable preconditions for letting a self-improvement step run unattended: verifier strength, how well an automated verifier agrees with a trusted reference, and recoverability, how cheaply a bad change can be undone. On their platform the rule lets prompt, tool, and skill edits run unattended while keeping weight updates under human sign-off.
Self-improvement is leaving the lab for production, and this gives operators a concrete, measurable safety gate instead of vibes.
Scientific work with LLMs, from literature synthesis to experiment design, keeps coming back to one need: updating beliefs as evidence accumulates. The Oxford team builds a principled evaluation of whether LLMs can do Bayesian belief updating in context, testing both internal probabilistic coherence and recovery of an externally determined posterior, and challenges the assumption that models cannot do it.
If agents are to help with science, they need to reason about uncertainty the way science does. This is a first rigorous check of whether in-context Bayesian updating is within reach.
Multimodal LLMs are everywhere, but their social biases are far less examined than text-only models’. Melange is a synthetic visual bias benchmark spanning seven bias categories, built by extending the language-only BBQ dataset with carefully generated visual scenarios so the target attributes are clearly visible. Generated images pass reliability checks to keep the measurement honest.
You cannot fix a bias you cannot see, and in multimodal models the bias hides in the pixels. A benchmark that puts the attribute visibly in the image is the prerequisite for honest evaluation.
When a small quantized model gives fewer biased answers, is it fairer or just more evasive? Testing Qwen3, Gemma-3, and SmolLM2 on the BBQ benchmark across sizes and quantization settings, the authors find no simple scaling trend for bias, and 20 of the 26 configurations with fewer biased responses also abstained more. Lower measured bias can partly mean higher abstention, not better fairness.
Edge deployment is pushing small quantized models into real decisions, and this is a cautionary footnote: the fairness number can improve for the wrong reason.
Public LoRA pools hold dozens of adapters per backbone, and combining them forces a choice: merge weights in one pass and lose quality, or run one pass per LoRA and pay linearly. LoRASpace observes that stacked LoRA factors need far fewer directions than their summed ranks, finds the minimal subspace keeping 99% of factor energy via SVD, and reuses it in a training-free static mode at merge cost and a dynamic routing mode that is over 6x faster than composite inference with six active LoRAs.
Personalized generation lives or dies on composing adapters cheaply. A shared subspace that makes multi-LoRA composition nearly free removes one of the main taxes on creative AI tooling.
Protein engineering needs sequences that trade off several properties at once, but experimental campaigns allow only a few batched measurements. MODS steers diffusion models at inference time toward different objective tradeoffs by sampling scalarization weights and surrogate ensemble members per proposal, updating the surrogates after each batch without retraining the generator. It pairs with different diffusion models and steering procedures.
Designing proteins is multi-objective by nature, and wet-lab feedback is precious. Inference-time steering that respects both could shorten the loop between computation and the bench.
Text-to-image systems return several images per prompt, but models optimized for single-image quality produce boringly similar sets. T-DIG is an iterative framework where users inspect default generations and request variations in natural language along whatever axis they care about, background, style, clothing, refining until satisfied.
Generation quality has outrun generation control. Putting diversity itself under text control turns the image model from a slot machine into a design partner.