Flow and diffusion generative models have established themselves as widely adopted density estimators for simulation-based inference (SBI), extending naturally from neural posterior estimation to likelihood and joint density estimation. Their principled optimization objectives and freedom from architectural constraints have driven rapid adoption across the natural sciences. Yet the most widely used SBI libraries remain PyTorch-based, leaving researchers who develop their forward models and analysis pipelines in JAX without a native option. We present GenSBI, an open-source library that implements flow matching, score matching, and denoising diffusion entirely in JAX. The library offers three transformer-based architectures -- SimFormer, Flux1, and a novel Flux1Joint that extends gate-modulated transformer blocks to joint density estimation -- all interchangeable through a unified interface that decouples generative method, neural backbone, and inference mode. GenSBI provides an end-to-end workflow from training through posterior calibration (SBC, TARP, LC2ST) and supports custom architectures with domain-specific embedding networks.

Simone Bombari asked us whether the 1-bit quantized random vector Y = sgn(Wx) has subgaussian norm bounded by a universal constant. Here W is an n n random Gaussian matrix, and x is an independent standard normal random vector in Rn. The question is nontrivial since the coordinates of Y are not independent. We give a strong positive answer to this question - for any bounded map instead of sgn() - using AI: AIDiscovery and Generalization (Theorem 1): To handle coordinate dependence, Gemini 3.5 Flash1 proposed decomposing the Gaussian vector into independent parts, using one part to "smooth" the sign function, and then applying Gaussian concentration for Lipschitz functions.

Safety defenses for large language models (LLMs) are typically trained and evaluated on single-turn prompts, yet real attacks often unfold as indirect, multi-turn probing. To defend against this more nuanced form of deception, we present a unified pipeline that generates realistic multi-turn deceptive question sets via multi-objective genetic prompt optimization with co-evolving mutation operators. We validate this dataset through a human study, which also revealed that early generations yielded the most convincing deception and practical constraints such as adherence filtering and ordering effects. Using this data, we were able to detect deceptive attempts to access prohibited information using simple, explainable geometric signals in embedding space coupled with a lightweight feed-forward classifier. Three geometric features (angular coverage, distance ratio, and linearity) augmented with pairwise similarity statistics led to a compact predictive model that achieved consistently high recall (0.89) across base, reworded, and truncated (three-turn) scenarios, with test-time F1 ranging from 0.74-0.86. The results support a central hypothesis that multi-turn deceptive intent leaves a stable geometric footprint that enables lightweight, transparent screening without expensive end-to-end training. We further discuss responsible uses, limitations, and paths toward larger, more diverse human-evaluated datasets. The primary contribution to artificial intelligence is the multi-objective evolutionary framework for prompt generation, and the engineering application is the deployment of a lightweight geometric detection system for LLM safety infrastructure.

Existing training approaches for large language models learn a single set of parameters, based on large volumes of data, which is typically heterogeneous, conflicting and often outright contradictory. As a result, the model is forced to compress conflicting goals, and inherent uncertainties into a single, averaged pattern of behaviour. We propose an $α$-Rényi variational framework for learning distributions over post-training parameters, offering an uncertainty-aware alternative to deep ensemble approaches. The resulting variational objective interpolates between classical variational Bayes and predictively oriented posterior learning, balancing between globally plausible individual models against systems of complementary specialists. We identify local stability criteria, demonstrating how model misspecification can make non-degenerate posterior spread locally favourable, manifesting contradictory or conflicting data as epistemic uncertainty. We apply our framework to LLM post-training, learning an ensemble of LoRA adapters attached to a shared, frozen base model, providing a scalable training procedure for both supervised fine-tuning and preference optimisation. Our approach enables training examples to be softly routed across ensemble members, promoting model specialisation and providing actionable uncertainty estimates across different tasks.

The Chinese company is adapting to the demise of Moore's Law, which guides chip production. It could complicate US chip dominance. Tingbo He, president of Huawei's chip-design subsidiary HiSilicon, says her company's engineers have developed a novel way to optimize semiconductors--and she believes it will close the performance gap between Chinese and Western chips over the next few years. Huawei's method, in short, focuses on speeding up computations across chips, circuits, and entire computing systems, rather than squeezing ever-more components onto a single piece of silicon. "We found a new path," He said at the IEEE International Symposium on Circuits and Systems in Shanghai last weekend.

PCWorld reports that Claude AI has developed a persistent habit of interrupting users to suggest they go to sleep during work sessions. This ongoing bug affects multiple Claude models including Sonnet 4.6 and Opus 4.7, with Anthropic acknowledging it as a troublesome character tic. Users can attempt to reduce this nagging behavior through custom instructions in Claude's settings while Anthropic works on a permanent fix. Claude has developed an unusual habit over the last couple of months: urging its users to stop what they're doing and get some rest. Just this week, yet another Claude user said that the chatbot tried to end a late-night coding session because "it's late" and "your work will be better after some sleep." "There's something deeply irritating about your primary work tool developing a personality that includes unsolicited bedtime enforcement," the user complained, which sparked a lengthy discussion about how Claude had bugged users about how should get some rest, too.

Trajectory is betting the rapid iteration cycle that supercharged vibe-coding can help all kinds of companies build AI products that learn continuously. Trajectory founders, Ronak Malde (left), Michael Elabd(center), and Arjun Karanam (right). A group of AI researchers who previously worked at Google DeepMind, Apple, OpenAI, and Meta Superintelligence Labs announced on Wednesday they're launching a new startup called Trajectory, which aims to help companies regularly improve their AI products by training on real-world user interactions. Trajectory wants to build a platform for AI that can learn continuously, a capability that researchers have long held up as a major barrier to further AI progress. OpenAI, Google, and Anthropic have found success training increasingly capable versions of AI models, especially for domains such as coding, math, and science.

Former Giants manager's daughter consulted ChatGPT before reporting altercation The 18-year-old daughter of former Yomiuri Giants manager Shinnosuke Abe said she consulted ChatGPT before reporting an alleged altercation with her father to a child guidance center. The 18-year-old daughter of former Yomiuri Giants manager Shinnosuke Abe said in a letter released on Tuesday that she had consulted ChatGPT before reporting an alleged physical altercation with her father to the child guidance center. Abe resigned from his position on Tuesday following his arrest on suspicion of physically assaulting his daughter . He has been released from police custody. According to reports, two of his daughters had been involved in an argument the previous day.

A striking geometric disparity has long persisted in the practice of deep learning. While modern neural network architectures naturally exhibit rich symmetry and equivariance properties, popular optimizers such as Adam and its variants operate inherently coordinate-wise, rendering them unable to respect the equivariance structures of the parameter space. We address this disparity by introducing a symmetry-compatible principle for optimizer design: the gradient update rule should be equivariant under the symmetry group acting on the corresponding weight block. Following this principle, we first provide a unified perspective on bi-orthogonally equivariant updates for general matrix layers, as employed by stochastic spectral descent, Muon, Scion, and polar gradient methods. More importantly, by moving from orthogonal groups to permutation and shared-shift symmetries, we derive symmetry-compatible optimizers for parameter blocks whose symmetries differ from those of general matrix layers: embedding and LM head matrices, SwiGLU MLP projections, and MoE router matrices. These constructions include one-sided spectral, row-norm, hybrid row-norm/spectral, row-aware, column-aware, centered row-norm, and left-spectral updates. They yield an end-to-end layerwise optimizer stack in which each major matrix-valued parameter class is assigned an update whose equivariance matches its symmetry group. We corroborate this principle through pre-training experiments on dense and sparse MoE language models, including Qwen3-0.6B-style, Gemma 3 1B-style, OLMoE-1B-7B-style, and downsized gpt-oss architectures. Across these experiments, symmetry-compatible update rules consistently improve final validation loss, reduce load imbalance in sparse MoE models, and in several cases improve training stability over the corresponding AdamW updates.

Large-scale search, recommendation, and retrieval-augmented generation (RAG) systems typically employ a two-stage architecture: an early-stage ranker (ESR) generates a candidate set, which is subsequently re-ranked by a late-stage ranker (LSR). While there are many reinforcement learning (RL) methods for training the LSR, end-to-end training of the ESR has proven challenging. In particular, naive application of "vanilla" policy gradient (V-PG) is not scalable for candidate-set sizes relevant for practical use due to exploding variance. This issue arises because V-PG propagates the gradient to the joint probability of the candidate sets, ignoring the contribution of each specific item in the candidate set to the reward. To mitigate this issue, we propose a novel "credit-assigned" policy gradient (CA-PG), which computes gradients with respect to the probability that the target item is chosen in any candidate set, i.e. marginalizing over all candidate sets that contain it. Our theoretical analysis reveals that CA-PG significantly reduces the variance of V-PG by marginalizing over the specific composition of the candidate set, while preserving the ability to learn the correct ranking of items under a reasonably aligned LSR policy. Experiments on both synthetic and real-world data demonstrate that CA-PG improves the convergence speed and training stability for ESRs utilizing the canonical Plackett-Luce model, especially when the candidate-set size is large.