With architecture and data held fixed, a checkpoint-only reconstruction attack recovers training images best from Muon-trained models and worst from Adam-trained models, making the optimizer choice a control axis for training-data privacy.
A parallel diffusion drafter emits a whole block in one forward pass but each slot never sees which tokens earlier slots actually chose; a lightweight corrector that reads only the drafter's existing hidden states and logits scores adjacent candidates in parallel and re-selects left to right, giving the highest average throughput across four verifier backbones and ten benchmarks and staying ahead of the uncorrected drafter at 32 concurrent streams.
Assortment optimization under the Mallows model previously required either stylized structural simplifications or restrictive parametric assumptions, or else a highly involved quasi-polynomial-time scheme. The change here is that Mallows choice probabilities can be reproduced simultaneously for all offered sets by a single collection of mutually independent utilities. That exact equivalence, combined with a finite-support reduction, yields an -time approximation scheme allowing arbitrary product prices, concentration parameters, and modal rankings.
The work introduces TellTail, a query-only fingerprinting attack that identifies the embedding model behind a black-box retrieval system: it perfectly identifies the deployed retriever from full rankings, achieves 94.3% success from unordered top-3 results, and 92.5% from language-model-generated answers alone, and still identifies 44 of 53 retrievers in an end-to-end OpenWebUI deployment.
The work introduces cordial (correlated and distributed) learning for distributed learning tasks where agents' data are correlated: an agent's label depends on other agents' inputs for the same sample, and those inputs are themselves correlated; the method shares only low-dimensional outputs between agents while training local models to extract informative signals from peers, inducing a game in which each agent's loss depends on others' models, and the authors prove convergence with probability one to a globally optimal solution under a linear model despite the nonconvex global objective, with experiments on structured multi-digit MNIST tasks showing the method remains highly effective even in highly nonlinear settings.
The authors propose Local Support Learning (LSL), which frames catastrophic forgetting as a geometric problem in each weight matrix's input space and uses a Gaussian Mixture Model gate to enable a weight adapter only on the support of the current phase's training distribution, so that LLMs of up to 7 billion parameters can be finetuned across multiple phases without access to prior data while learning new tasks at full capacity and retaining pretrained and previously finetuned capabilities.
The work introduces Self-Play Pretraining with Zero Data as an initial proof-of-concept: starting from random initialization, a generator proposes programs interpreted by a universal Turing machine to produce byte sequences while a learner autoregressively predicts those byte sequences with standard cross-entropy, and the generator is trained with reinforcement learning to produce sequences at the frontier of the learner's capabilities, yielding an adaptive curriculum; because neither generator nor learner is trained on natural data, zero-shot performance on natural data serves as a clean test of transfer, and across several natural datasets zero-shot loss exhibits predictable scaling in compute, with the models also exhibiting in-context learning and discovering recognizable mathematical
This retrospective study analyzed 1,754 labeled public cystoscopy images to test Direct, Book-based, and Optimized prompts across GPT-5.2, GPT-5, GPT-5-Mini, and GPT-5-Nano for benign-versus-malignant classification, finding that GPT-5 and GPT-5-Mini with the optimized prompt reached accuracies of 86.7% and 89.2%, and that GPT-5 with the optimized prompt achieved 98.1% accuracy, 94.6% specificity, and 99.1% sensitivity in high-confidence triage at 62.0% image coverage, while prompt engineering improved calibration and triage without statistically significant performance gains.
In routine use of a surgical AI platform, the study automatically assigned Parkland Grading Scale scores across 249 consecutive laparoscopic cholecystectomies, grouped them into Low (PGS 1-2, n=78) and High (PGS 3-5, n=171) severity, compared surgical outcomes, and evaluated model F1, discrimination (AUC 0.932 and 0.896) and calibration against two surgeons' independent double reviews of a 75-video sample, finding that the High group had older age, higher ASA scores, more cholecystitis and urgent surgery, longer operative duration, more intraoperative events and bailouts, longer hospital stay, and more 90-day major complications and readmissions, with operative duration and hemorrhage-related events remaining significant after case-control matching (n=84).
You’re all caught up