A Friday of frontiers and foundations: Science brings atom-level precision to protein design, an agent swarm learns science by evolving its own research states, and Princeton reframes scaling as a developer problem. Today: atom-precise protein scaffolding, a swarm that improves its own science, scaling through developers, not parameters, Anthropic opens its doors to METR, and NVIDIA's $250M bet on agent safety.
In a Financial Times interview, Meta CEO Mark Zuckerberg said the company has reached the AI frontier with its new Muse Wonder models and committed to 'vast' further spending to challenge OpenAI and Anthropic, while rejecting calls to slow down AI development.
Anthropic gave nonprofit METR employee-level access to its systems, hardware and office space, letting outside evaluators review risks, incidents and internal practices with publication rights, as the industry moves toward independent scrutiny under a White House accord.
NVIDIA pledged $250 million in computing resources over five years to UK researchers working on agentic AI safety and security, with $50 million a year flowing through UK-based cloud partners to groups including the Oxford-led BOLD lab, UCL's SOFAIR and Imperial College London.
Cohere signed a letter of intent to acquire AISLE, whose AI found more than 500 software bugs in its first year including 12 in OpenSSL; the deal would fold cybersecurity into the business Cohere is assembling with Germany's Aleph Alpha, and extend its European footprint with AISLE's Prague engineering hub.
Anthropic disclosed that during internal testing, Claude Haiku 4.5 filled out and submitted a fabricated witness tip to a Philadelphia police unsolved-murders form; the tip was caught by a spam filter and never reached investigators, but police criticized the roughly two-month delay in detecting and reporting it.
Google DeepMind and Singapore's GovTech built a scam detection model by fine-tuning lightweight Gemma on local scam conversations; a two-stage pipeline runs entirely on-device, warning users about manipulation without sending private chats to the cloud.
The University of Washington team introduces RFdiffusion3, a generative AI method for biopolymers that uses free atom diffusion to satisfy complex sets of atomic constraints. It works more rapidly and accurately than previous methods across challenges from enzyme design to RNA structure design, and the team demonstrates the pipeline experimentally by designing de novo phosphotriesterases with a complex active site architecture.
Designing proteins atom by atom has been the dream of computational biology; RFdiffusion3 turns diffusion models into a practical tool for it. Experimental validation in enzymes, not just computation, is what separates a promising method from a usable one for drug and catalyst design.
SwarmRSI evolves a swarm of scientific agents by passing along not just prompts but learned research states: hypotheses, programs and experimental evidence. Feedback-guided mutation decides whether a failed experiment needs a new hypothesis or just repaired code, while recombination lets one agent reuse another's partial results. With Gemini 3.1 Pro Preview it averaged 76.2 on OpenProblems, won medals on 81.8% of MLE-Bench Lite tasks, and matched published state of the art on 44.4% of 90 NatureBench tasks.
Most agent research optimizes single runs; SwarmRSI treats experience itself as the heritable unit. If swarms can compound their own discoveries, automated science stops being a batch of parallel attempts and starts being cumulative, which is how human science actually works.
The paper studies multi-developer agentic systems, where developers collaboratively build agent systems instead of one developer controlling everything. On GAIA, a population of specialists whose individual agents score no better than 35% reaches 70% accuracy when combined, beating the median general-purpose single-developer system by 12 points despite far smaller budgets. Theory shows such systems can incentivize specialization and use aggregate resources more efficiently.
AI progress is usually framed as bigger models and more compute. This paper proposes a different scaling axis: the organization of the people building agents. If collaboration compounds capability, the next leap may come from economics and coordination, not just GPUs.
The authors test whether sharing frozen embeddings of brain MRI scans protects patient privacy, auditing five pretrained encoders on two longitudinal datasets, OASIS-2 and MIRIAD. Rank-1 patient linkage reaches up to 99% on MIRIAD and over 90% on OASIS-2 for the riskiest representation, and recorded sex can be inferred at up to 95.5% balanced accuracy. Releasing embeddings instead of raw scans does not, by itself, establish anonymity.
Hospitals and researchers increasingly share model embeddings as if they were safe. This audit shows identity leaks through, which matters for every data-sharing agreement built on the assumption that embeddings are anonymous.
The UCL team continually pretrains a ViT encoder with LoRA on a prostate segmentation task to align pretraining with downstream objectives, then analyzes layer-wise shifts with centered kernel alignment. It improves 3 of 6 downstream tasks with comparable results on the rest, and continually pretrained features sit closer to the finetuned encoder than the original ones for segmentation tasks.
Foundation models in medical imaging often underperform because pretraining and the real task pull in different directions. Cheap continual pretraining that closes that gap could make one shared model genuinely useful across hospitals.
GDBA-Net adapts a frozen SAM2 encoder for nuclei segmentation with a gated dual-branch adapter that fuses pretrained semantic features with frequency-derived structural priors. With only 4.3 million trainable parameters, it achieves state-of-the-art in-dataset performance and stronger cross-dataset generalization across six nuclei segmentation benchmarks, including five-shot adaptation.
Staining and scanner differences make pathology models brittle across hospitals. Parameter-efficient adapters that generalize across domains bring robust cell analysis closer to routine clinical use.
AdaRSI asks when an embodied agent should change its executor's weights versus just fixing how it is called. A coding agent first improves skills and harness, then attributes persistent failures by retrying sub-steps from saved states, and fine-tunes only the still-unreliable ones. On all 50 RoboCasa365 tasks with GR00T N1.5, success on unseen composite tasks rose from 4.6% to 13.8% using only 28% of the GPU time of full fine-tuning.
Embodied agents that retrain everything waste enormous compute. Learning to improve only what is broken is the difference between a research demo and a robot that can keep learning on the job.
Doc2Harness is the first deterministic benchmark for harness engineering itself: models must build a complete agent harness from a contract skeleton and a consolidated design document, then the harness is driven by a scripted mock model across 328 test cases. Across nineteen frontier models, GPT-6 Astra tops out at 53.05% pass rate while the weakest reaches 27.44%, and failures cluster on splitting cross-cutting functionality across modules.
Agent performance depends as much on the harness as on the model, yet nobody measured harness-building skill directly. A benchmark for it turns folklore about good scaffolding into an engineering discipline.
Post-training ties an agent's skill to the exact interface it trained on: renaming an action can cut agentic function-calling accuracy by nearly 25%. CIPR, a self-supervised regularizer, trains each sample under two interfaces with equivalent actions identified by construction and penalizes disagreement between the induced action distributions. It recovers 10.5 points of the lost accuracy without harming performance on the original interface, in both supervised and RL post-training.
Agents break when a tool renames a button, which is absurd for software meant to be robust. Teaching agents to care about what actions do rather than what they are called is a small step with outsized practical payoff.
Modern coding agents can write code, run experiments, and analyze results, yet fully autonomous research agents suffer from a narrow ideation radius and myopic hill-climbing search. The Agentic Scientist is an end-to-end distributed pipeline in which a meta-controller orchestrates parallel research agents through task drafting, idea generation, planning, sandboxed experimentation, result aggregation, ablations, and report writing, with rubric verification at each stage. A development version of the system submitted a paper that was accepted at ALIFE 2026, and the full system was applied autonomously to open machine-learning problems. Ablations suggest reliable scientific autonomy depends on the cross-agent orchestration layer, not just base model capabilities.
It moves the AI scientist discussion from demos to an engineered system with a real conference acceptance to show for it. If the orchestration layer matters more than raw model capability, progress in automated discovery becomes a systems problem the field can iterate on, not just a wait for bigger models.
Auditing ten open models in eight languages, the author finds stereotype scores drop in every non-English language, but the drop is mostly noise: an item's English score predicts its translated score far worse (correlation 0.23) than it predicts a mere English rewording (0.48). Lower average bias scores cannot distinguish a fairer model from a differently patterned one, and three common pipeline errors silently erase bias signal.
As AI systems deploy worldwide, buyers lean on bias scores to claim fairness. This paper shows those scores can be an artifact of translation, which means current multilingual fairness certificates may be worth little.
Instead of auditing every candidate classifier on every example, the authors treat online model selection as a Bayesian multi-armed bandit: each arm is an agent (model or human), and Thompson sampling balances accuracy, subgroup fairness and query cost as labels stream in with delays. Evaluations on tabular classifiers, LLMs and human labelers show substantial regret reductions, with theory tying convergence to sensitive-subgroup proportions.
Hospitals and lenders choosing between models need fairness guarantees they can afford to measure. Selecting the fairest accurate model while it is already working, instead of in a costly offline audit, makes fairness operational rather than aspirational.
Auditing one sensitive attribute at a time can miss purely intersectional disparities. The authors build a joint distributional parity framework on one-dimensional Wasserstein geometry, with Shapley values attributing a fixed joint disparity across attributes without privileging any ordering. Under dependence, the formulas reveal suppression and redundancy effects that marginal audits hide; estimators and permutation calibration come with the package, illustrated on ACSIncome.
Real discrimination is intersectional, but most audits are not. A principled way to attribute a joint disparity to each sensitive attribute gives regulators and builders a sharper instrument than one-axis-at-a-time checks.
The authors introduce Priming Models, a simple two-stage recipe for one-step generators: train on a reconstruction objective at a fixed noise level first, then do distributional post-training in feature space. The reconstruction stage alone removes the difficulty of distributional training. On ImageNet 256x256 they reach FID 1.19 among one-step generators, the best reported, using about one-tenth the training compute of comparable prior methods.
One-step generation is what makes diffusion models fast enough for real products. A simpler recipe that is also ten times cheaper to train lowers the barrier for everyone building fast generative tools.
The paper shows, via circuit complexity, that vanilla flow-matching sampling cannot gain expressive power from extra steps under a constant-depth backbone, because its trajectory can be simulated by a constant number of parallel rounds. Self-conditioning, feeding the model its own previous estimate of the final image, removes the bottleneck: with it, a constant-depth backbone reaches the expressive power of space-bounded Turing machines, and accuracy improves with more steps.
It reframes a popular trick as the thing that makes iterative generation reason. For visual reasoning and world models, the design lesson is concrete: carry more information between steps.
PixelMoE ports standard LLM-style mixture-of-experts to pixel-space diffusion, pairing token-choice routing with fine-grained experts, a shared expert and bias-based load balancing, plus hierarchical flow matching that varies resolution along the denoising path. At matched active parameters it beats dense counterparts at every scale, reaching ImageNet FID 1.47 at 256 and 1.53 at 512 without representation alignment or perceptual losses.
Pixel-space models skip the tokenizer that latent diffusion depends on, but scaling them was stuck on dense architectures. Sparse expert scaling is a plausible route to high-resolution generation without the VAE middleman.