The Archive

Language-grounded robot group joining predicts socially compliant poses for robots to join group formations using natural language descriptions.

Introduces STLC complexity method for transductive learning with fixed-point bounds matching classical inductive Rademacher complexity.

SWE-Flux benchmark evaluates LLM execution reasoning on 480 repository-level Python instances with gold labels from instrumented test execution.

Studies nonequilibrium dynamics of minimal recurrent transformer with repulsive attention; discovers bifurcations, chaos, and emergent locality.

Analyzes whether reordering mathematical rules affects LLM internal representations; finds 1B–8B models encode ordering patterns inconsistently.

"So absurd and embarrassing ..."

Provides geometric characterization of minimal-norm two-layer univariate ReLU classifiers; resolves optimal solutions under bias regularization.

Sequential LoRA protocol transfers depression severity prediction from English PHQ-8 to Chinese HAMD-17 clinical scales using Qwen3 backbone.

Context-Continuous Preference Learning uses Gaussian processes to personalize exoskeleton assistance from limited user feedback across operating conditions.

Analyzes repairability of approximate solvers in recursive state estimation by characterizing correction subspaces and covariance response in Kalman models.

Agent-Editing World Model proposes replacing observation prediction with task-state correction to reduce contamination and improve LLM agent long-horizon planning.

Decode-Augmented Rollout Training (DART) fixes motion loss in frozen-latent world models by retraining flow dynamics with horizon-aware supervision.

Google releases Gemini 3.8 Flash TTS models with 2,000+ voices and custom voice cloning from 30-second samples; Willison builds playground interface.

Clifford-VAE learns hyperdimensional vector representations via Clifford algebras and geometric autoencoders on MNIST, FashionMNIST, CIFAR-10.

Differentiable Gaussian Dynamics learns population-level collective behavior from aggregate counts via heterogeneous response propensities and recurrent feedback.

Memory Attention replaces value projections with token-indexed memory combined with contextual keys, reducing inference cost via lookup-based value construction.

Fine-tuning LLMs for translation causes catastrophic forgetting of MT-specific instructions; general mitigation methods fail to preserve formality/gender/length control on Llama 3.1/3.2.

Quantum score matching extends generative learning to quantum systems by handling noncommuting density operators with efficient circuit implementations.

Pro and Plus users will be able to use the Work tab on their phones

Entropy-calibrated credit assignment unifies on-policy distillation and GRPO for mathematical reasoning by preserving task credit across token-level teacher reweighting.

ForgetMimic enables removal of specific motions from RL-trained humanoid policies to address safety, privacy, and copyright concerns under GDPR.

The report suggests that greater exposure will not resolve the unease around the technology, nor reduce public support for AI regulation.

Study traces how pricing cues mislead LLM shopping agents under acquisition costs, revealing susceptibility across eight commercial models.

AnchorReasoning dataset annotates 416K frames with visually grounded chain-of-thought reasoning for long-tail autonomous driving on WOD-E2E.

Simon Willison explains CSS shadow roots with interactive examples; off-topic for AI industry coverage.

LEAP-CBF introduces Least-Effort Adversarial Potentials to design less conservative safety filters for uncertain nonlinear dynamical systems.

MicroQonv optimizes tensor reshaping to reduce memory overhead when applying microscaling quantization to convolutional neural networks.

Linguistic study examines distribution of Standard Arabic vs. Colloquial Arabic across X and Facebook across seven discourse categories.

Sen. Bernie Sanders (I-VT) and Rep. Greg Casar (D-TX) have introduced new legislation that would ban anyone from developing artificial superintelligence - a technology the bill describes as capable of the "destruction or disempowerment of humanity," including by overthrowing the government. Under the Ban Artificial Superintelligence Act, AI leaders would face up to 20 years in prison for violating the law. The bill doesn't just ban the more powerful artificial superintelligence; it would also pause the development of advanced AI systems, which lawmakers classify as frontier models trained on ...

Mizar is a 159M-parameter audio-language model combining CED-Small encoder with SmolLM2, targeting low-memory audio understanding tasks.