Latest stories
Hugging Face Paper Reframes Fine-Tuning as a Direction-Selection Problem
A new paper proposes drift-constrained optimization, arguing that only the direction of fine-tuning updates matters when staying w...
Register Tokens Let Diffusion Language Models Reason With a Fixed-Size State
A new Hugging Face paper proposes register tokens that carry reasoning progress across generation chunks in masked diffusion langu...
ModAR: Hugging Face Paper Shows Modality-Autoregressive World-Action Models Beat RGB-Only Prediction
A new Hugging Face paper introduces ModAR, the first world-action model that autoregressively denoises multiple future visual moda...
ModularRSI: A Modular Framework for Generalizable Agent Harness Self-Improvement
A new paper introduces ModularRSI, a benchmark-disjoint, contrastive, and modular framework that evolves agent harnesses by decomp...
Study: Child-Framed Prompts Push Frontier AI Toward Shared Architecture Blueprints
A new paper reports that framing prompts around a child audience led six frontier model types from OpenAI, Anthropic, xAI, and Goo...
Hugging Face Paper FLAT Unifies Multimodal Retrieval and Generation in One 1D Token Space
FLAT is a representation pre-training framework that jointly optimizes a shared multimodal encoder with text-to-image and image-to...
HarnessVLN Brings Training-Free Embodied Navigation Under One Agent Harness
A new Hugging Face paper introduces HarnessVLN, a zero-shot, training-free navigation framework whose Agent Harness coordinates pe...
Hugging Face Paper Proposes Unified Framework for Policy Improvement and Recursive Self-Improvement
A new paper introduces Generalized Agent Iteration (GAI), a formal framework that treats classical iterative policy improvement an...
StepAudio 3 Music Debuts With Explicit Planning and 5-Minute Song Generation
StepAudio 3 Music is a large-scale, long-form music generation model that combines a 50-Hz single-codebook tokenizer, a flow-match...
StepAudio 3 Realtime Brings Think-While-Speaking to Full-Duplex Voice AI
A new technical report introduces StepAudio 3 Realtime, an audio-language foundation model built around a continuous listen-conver...
New Study Maps How Transformer Representations Evolve, With Surprising Training Gains
Researchers decompose learned Transformer updates into parallel and perpendicular components, finding that direction-preserving up...
Hugging Face Paper Proposes Mind2Dialogue to Train Human-Aware AI Assistants
Mind2Dialogue simulates users' unspoken mental states and converts them into privileged supervision, teaching language models to u...