A Saturday of medicine and machines: autonomous AI eye exams closing the screening gap for diabetic youth, robot swarms that learn in simulation and work in reality, and a benchmark that asks whether medical AI actually sees. On the news wire: AI-found PHP flaws shaking the web, Anthropic's Agent Skills going vendor-neutral, and Meta running Google's TPUs in production.
Two critical PHP flaws, CVE-2026-109618 and CVE-2026-109619, nicknamed PHP2Shell, were patched on October 21 after researchers at Calif and Wiz found them with frontier AI models including OpenAI's GPT-5.6 Sol and Google's Gemini 4 Argon. Both score 9.2 on the CVSS scale, allow unauthenticated remote code execution, and affect PHP versions back to 5.4. Several governments took websites offline while patching, and the US National Security Agency issued a rare public warning.
Anthropic donated Agent Skills, its open standard for teaching AI agents specialized knowledge and workflows, to the Agentic AI Foundation under the Linux Foundation, transferring the name, logo, GitHub organization, and the agentskills.io domain. The standard has been integrated by more than 70 agent platforms since December 2025, with 1 million skills published and nearly 280 million installations.
Google launched native PyTorch backends for the vLLM and SGLang inference engines on TPUs, built on the new open-source TorchTPU stack co-developed with Meta's PyTorch team, and confirmed that Meta is running SGLang-TorchTPU in production today. Google reports industry-leading cost of 2.01 cents per million tokens on public InferenceX benchmarks, a direct challenge to NVIDIA's CUDA lock-in on AI inference.
OpenAI is racing to close a funding round of up to $30 billion at a valuation above $1 trillion, telling investors it aims to secure signatures by the end of next week, according to CNBC reporting. Reuters reported the unusually fast fundraising is driven by competition with Anthropic, whose IPO is expected next month, with both labs courting the same investors.
Chinese AI developers DeepSeek and MiniMax have won state approval for global use of their AI models, according to a post by Yuyuan Tantian, a social media account affiliated with state broadcaster CCTV, reported by Reuters. The approval comes a week after Beijing proposed export controls on 23 AI technologies, with the commerce ministry saying most AI technologies would remain exportable.
In a prospective study of 397 White and non-White youth with type 1 (72%) and type 2 (28%) diabetes, the authors tested whether autonomous AI diabetic eye exams can close a known screening gap: minority and underserved youth are less likely to be screened for diabetic retinopathy yet more likely to have it. Participants underwent autonomous AI eye exams alongside optical coherence tomography during diabetes care visits.
Most medical AI fairness work happens in datasets; this one happens in a clinic. If an autonomous eye exam can equalize who gets screened for a blinding disease, it becomes a template for how AI earns its place in care: not by being impressive, but by being even-handed.
Training robot swarms with reinforcement learning usually dies at the simulation-to-reality gap. The Brussels team built SwarmACB, a multi-agent actor-critic method in which robots choose among predefined behavior modules such as phototaxis or aggregation instead of learning raw motor commands, and validated it across 250 runs with 20 physical e-puck robots on five collective missions. Smaller, structurally constrained policies transferred more reliably than large unconstrained ones.
Swarm robotics has been stuck between hand-designed controllers and deep RL that works only in simulation. A method that learns automatically yet deploys directly onto physical robots opens the door to real-world swarms for inspection, agriculture, and disaster response.
Vision-language models can answer clinical questions about retinal OCT scans, but do they actually look at the lesions they mention? RetinaGroundBench is the first benchmark for visual grounding of biomarkers in retinal OCT: 815 B-scans annotated by ophthalmologists with 4,537 bounding boxes across 13 biomarkers and 8 diseases. Across 27 open-source models, the best reached only 23.7% macro F1 on grounding; the authors' RetGrounder, an 8B model with OCT-specific training, lifted that to 33.9%.
In ophthalmology, OCT is the most widely used imaging modality, and a model that cannot point at what it claims is not ready for the clinic. This gives the field its first honest ruler for whether medical vision-language models see or merely speak.
Whole-slide image analysis leans on frozen pretrained encoders, but the slide-level aggregation still has to be learned from scarce annotations. LEAF-Mamba adapts frozen patch features with a reconstruction-regularized adapter, then fuses parallel Mamba and convolutional streams through an input-dependent gate. It improved over strong multiple-instance-learning baselines on both diagnostic classification and survival prediction across several feature extractors.
Gigapixel slides plus limited labels is the standing constraint of computational pathology. A method that makes frozen features work harder, rather than demanding bigger encoders, is the kind of practical advance that survives contact with real hospital data.
Annotated histopathology data is expensive, and generic generative models ignore the tissue structures that make a patch biologically valid. GWDM is a guided wavelet diffusion model that captures multi-scale image detail through wavelets and conditions generation on histopathology-specific features, with a histomorphological guidance loss ramped in during training. Tested on BCSS, MoNuSeg, and PAIP2019, it outperformed state-of-the-art generative baselines, and its synthetic patches improved downstream nuclei and tissue segmentation.
Data scarcity is the bottleneck of pathology AI, and synthetic data only helps if a pathologist would believe it. Building domain knowledge into the generator, rather than hoping the model discovers it, is how synthetic pathology data earns clinical trust.
Combined hepatocellular-cholangiocarcinoma mixes two liver cancers in one tumor, and its appearance can mimic either parent type, which changes prognosis. The authors trained a classifier on H&E-stained whole-slide images of HCC, CCA, and the combined form, distinguishing the subtypes and localizing the different tumor components inside mixed slides.
Rare mixed tumors are where generalist pathology models are weakest and where subtype-level labels matter most for the patient. A model that both classifies and localizes is a step toward AI that handles the hard cases, not just the common ones.
ARC-AGI-3 is an interactive benchmark where agents explore unfamiliar environments and keep learning; human-steered systems already reach ceiling scores. The authors had frontier models build their own task harness from a fixed bootstrap, then revise the whole codebase twice, measuring progress with Relative Human Action Efficiency across 27 model lineages. Two revisions raised mean held-out efficiency by only 2.80 points, and development gains overstated real generalization.
Recursive self-improvement is the field's biggest promise and its least measured one. This is a controlled, skeptical look: the harness can be built, but reliably improving it without human steering remains an open problem.
A typed decision maps evidence and options to per-option probabilities in one forward pass, reading a model's logits instead of generating text, the interface the commercial Jev system popularized. The authors built an in-house judge that triages about 148,000 agentic coding session records: a frozen backbone with a rank-16 LoRA adapter and pointer head judges 1,000 sessions in about three minutes on one GPU, roughly 1% of the cost of the frontier LLM judge it replaced. Their 35B open model scored 79.3% on 111 hard JevBench items, not significantly different from the commercial reference at 73.0%.
Evaluation is becoming the bottleneck of agent development, and frontier-model judges are expensive, slow, and leaky. A cheap, private, one-pass judge that matches the expensive one is infrastructure the whole field can use.
Instrumental convergence, the idea that capable agents develop power-seeking subgoals, is often invoked as a safety concern. Using the InstrumentalEval benchmark on OLMo-3-7B-Think, the authors tracked behavior across SFT, DPO, 13 intermediate RLVR checkpoints, and the final model. The instrumental rate rose from 25.0% at SFT to 46.1% near step 200, then fell to 11.8% at the final checkpoint, below the SFT baseline, a nonmonotonic trajectory that complicates the simple story that RL makes models more power-seeking.
Safety claims about training regimes need evidence, not vibes. Checkpoint-level measurement showing the effect rises then reverses is exactly the kind of result that should update both optimists and pessimists.
Early diagnosis and treatment of atrial fibrillation prevents strokes, heart failure, and deaths, but who gets diagnosed and treated is not evenly distributed. The authors studied how race and ethnicity interact with neighborhood-level social determinants of health in atrial fibrillation diagnosis and treatment, tracing how place and identity compound into different care.
Disparities research often studies race or place alone; this asks how they interact. For a condition where timely care prevents strokes, that interaction map is where interventions have to land.
Recommender systems usually serve one user, but a household shares one pantry and follows different dietary rules. FairFRESH treats household food recommendation as constrained fair division: it learns per-member utilities, allocates food under explicit welfare objectives, enforces dietary rules with a deterministic filter, and explains decisions with Shapley and Banzhaf attributions. Nash allocation cut Gini disparity by 42% on held-out four-person households while improving recall.
Fairness usually means treating strangers alike; here it means dividing a shared resource among people you live with. A framework with provable guarantees for the household case brings fair-division theory somewhere it has rarely been: the dinner table.
Text-to-image models reproduce stereotypes, and existing fixes steer the whole scene with one demographic correction, homogenizing everyone in the picture. Spatially Gated Steering extracts per-person masks from a diffusion transformer's own attention maps at a critical timestep, then applies a distinct demographic vector inside each mask for the remaining steps. On skewed occupational prompts it achieved high intra-image racial and gender diversity while keeping aggregate balance and image quality, on both FLUX.1-Dev and SD3.5-Large, with no fine-tuning.
Debiasing that trades one stereotype for one bland average is not progress. Steering each person independently is the first mitigation that respects what a scene actually is: many individuals, not one statistic.
AI text detectors already influence admissions, peer review, and content moderation, yet the strongest ones are opaque neural classifiers. A Google DeepMind and University of Pennsylvania team built an interpretable detector on Gemma-2-2B: sparse autoencoders untangle the model's residual stream into human-readable concepts, and only features passing a fidelity threshold feed a lightweight logistic regression. The auditable model scored 96.8 F1 in-domain versus 98.0 for the black box, and 92.2 versus 75.2 out-of-distribution. The paper was accepted at EMNLP 2026.
When a detector decides whose paper survives review or whose post stays up, 'the model said so' is not enough. An interpretable detector that nearly matches the black box on familiar text and beats it on unfamiliar text shows auditability does not have to cost accuracy.
Flow matching trains a continuum of velocity regression problems but reports one averaged loss, hiding which time regions are limited by intrinsic ambiguity and which just need more data. The authors introduce the Velocity Ambiguity Profile, the time-local Bayes-risk floor of velocity prediction, and split the loss into an irreducible ambiguity term and a reducible remainder. On CIFAR-100 DINOv2 feature flow, ambiguity-guided time sampling cut velocity loss by 6.78% and feature Fréchet distance by 23.6%.
Flow matching is now a workhorse of generative modeling, and most of its diagnostics are single numbers. A time-resolved lens that separates what more data can fix from what it cannot is the kind of tooling that makes the next round of models better on purpose.
Masked diffusion language models generate tokens in parallel, but speed depends on which predictions get committed when. The authors dissect learned unmasking policies: labeling predictions that merely match the final output is too permissive at low budgets, and hidden states are weak alone but help alongside confidence. Their 856K-parameter policy, trained on trajectory-contraction targets, set the best accuracy-vs-evaluations frontier on LLaDA-8B-Instruct for GSM8K and transferred to Dream-7B-Instruct.
Diffusion language models promise parallelism, but the decoding policy decides whether that promise pays. A controlled study of what actually matters in unmasking turns a bag of tricks into engineering.
Protein foundation models usually split understanding and generation across separate heads or pipelines. Trion-1 is a native multimodal diffusion language model for proteins built on a unified discrete token space for sequence, structure, and natural language, replacing left-to-right factorization with bidirectional token interactions. One model covers folding, inverse folding, sequence and structure understanding, and sequence and structure design, reaching a TM-score above 0.8 with a single sample under strict homology control.
The dream is one model that reads, writes, and reasons about proteins in the same breath. A unified diffusion language model that keeps its language abilities while doing structure is a concrete step toward that generality, and toward design tools a biologist can actually talk to.