Three new journal papers lead today: a head-to-head test of pathology foundation models on more than 100,000 prostate biopsies, an RL-guided editor that optimizes antibodies for several drug properties at once, and a study of how clinician bias shows up in cancer genetic counseling. In the news: OpenAI's GPT-6.1 Terra, OpenAI's claimed proofs in mathematics, computer science and physics, a DeepMind and Springer Nature peer review assistant, and iFLYTEK's Spark-X2.6.
OpenAI released GPT-6.1 Terra on October 26, a medium-sized model built for the computer work of an office day: reading screens, filling out forms, interpreting technical drawings and editing video. Its API price is half that of GPT-5.6 Terra, at $1 per million input tokens and $5 per million output tokens, and it is available in ChatGPT Work, Codex and the API. OpenAI says it matches GPT-6 Astra on Agents' Last Exam at one-tenth the token price and finishes tasks more than three times faster.
OpenAI says its internal models produced three research results on October 26: a proof of the local smoothing conjecture in every dimension, which would also settle the Kakeya and Stein restriction conjectures; a disproof of the Exponential Time Hypothesis, which OpenAI says was verified in Lean; and an existence proof for quantum chromodynamics and Yang-Mills theory. The mathematics and computer science materials were released for inspection, while the physics material is held by an expert group led by Nobel laureate David Gross, which will write it up itself. Independent confirmation is still pending, and OpenAI says it does not seek academic credit for the results.
Springer Nature and Google DeepMind announced a research partnership on October 26 to co-design an AI peer review assistant built on Gemini models. The assistant will help reviewers navigate long manuscripts, analyze text, charts and figures, and flag areas that need closer attention, running entirely inside Springer Nature's secure environment so author and reviewer material stays private. A prototype is already being tested, and both partners stress that expert human judgment stays at the center of peer review.
At its 1024 Global Developer Festival in Beijing on October 26, iFLYTEK unveiled Spark-X2.6, a model designed to work across devices and the cloud, alongside new developer tools. The company's developer competition drew 52,120 teams from 351 cities in 15 countries, a record for the event.
Foundation models trained by self-supervision are widely promoted as a universal solution for pathology, but their edge over task-specific models has rarely been measured at clinical scale. This study compares both approaches on more than 100,000 prostate biopsies from 6,268 patients across 15 sites. Foundation models help most when labeled data is scarce, yet with enough labels, task-specific models match them and sometimes surpass them, and all models reach a quadratic Cohen's kappa of 0.91 against pathologists' grading.
Hospitals weighing a generalist pathology model against training their own get a rare apples-to-apples answer, and the answer is that data volume, not model size, decides. That reframes a purchasing question as a data question, which is cheaper to act on.
Improving one property of a therapeutic antibody often worsens another, so AbOptimizer treats optimization as reinforcement-learning-guided sequence editing. An agent chooses editing sites and actions, keeping, substituting, inserting or deleting residues, while an antibody language model proposes plausible mutations and a Pareto-based reward balances affinity, specificity, immunogenicity, viscosity and clearance. Tested on emibetuzumab, trastuzumab and siltuximab, it improved the targeted properties in silico, and lab validation showed improved binding affinity in optimized variants of all three antibodies, with improved specificity for emibetuzumab and siltuximab.
Antibody optimization is usually a slow trade-off, handled one property at a time. A method that improves several properties together, and survives contact with wet-lab validation, could shorten one of the most expensive stretches of drug development.
Disparities in genetic counseling are well documented, but less is known about how a counselor's own biases reach the patient. In this cohort study of 185 patients and 15 genetic counselors and trainees at two comprehensive cancer centers, patients reported therapeutic alliance, satisfaction and decisional conflict after counseling sessions about hereditary cancer risk. Among Black patients, stronger counselor endorsement of explicit medical-context stereotypes favoring White over Black patients was associated with poorer alliance, lower satisfaction and greater decisional conflict, while no counselor bias measure was significantly associated with outcomes among White patients after adjustment.
The finding narrows a broad problem to a specific, trainable target: explicit stereotypes about medical encounters, not implicit attitudes in general. That gives clinics something concrete to measure and address in training, rather than another reason to worry in the abstract.
Today's new pathology paper is in the Selections above. No further standalone papers in this column today.
Today's new agentic AI paper is in the Selections above. No further standalone papers in this column today.
Today's new fairness paper is in the Selections above. No further standalone papers in this column today.
No new generative modeling papers in today's window. Recent issues carry related work on diffusion models and memory.