Latest stories
PersonaDose Tunes AI Persona Intensity With Calibrated Activation Steering
A new Hugging Face paper introduces PersonaDose, a method that lets users request a persona by description and a target mean inten...
Hugging Face Paper: Scaling and Distilling Text Embeddings Boosts Diffusion Language Models
A new paper finds that scaling text embeddings within the T5 family sharply improves continuous diffusion language models, and tha...
Hugging Face Paper: DMM Fixes Decentralized Multi-Agent Pathfinding's Coordination Gap
A new Hugging Face paper introduces DMM (Decentralized Master-Mind), which replaces one-shot action sampling with iterative intent...
X-Tree: Tokenizing Reusable Experience to Boost Agent Generalization
Hugging Face paper X-Tree mines a reusable skill hierarchy from agent trajectories with zero LLM calls, then trains on it as data,...
Survey Maps Post-Training and Alignment Strategies for Video Generation Models
A new TMLR-published survey offers the first comprehensive review of post-training and alignment in video generation, organizing m...
GraphForge Trains Working Agents on Real-File Workspaces
Hugging Face paper introduces GraphForge, an evidence-graph framework that builds agent training tasks and their verifiers from re...
ActiveSaddler Turns Agent Harness Training Into an Adaptive Curriculum
A new paper proposes treating the choice of training scenarios as a learnable dimension of LLM agent harness optimization, using a...
E-MoE Turns MoE Routing Into a Shared Latent for Faster Diffusion Language Models
A new paper proposes E-MoE, which uses the expert-routing decisions of a Mixture-of-Experts backbone as a discrete shared latent v...
Tiny LoRA Unlocks Hidden Depth in Transformers, Study Finds
A new paper shows pretrained transformers stop following reference chains after only a few lines, but a rank-8 LoRA at a single ea...
PoS Framework Gives LLM Agents Explicit Belief States to Escape 'Belief Trapping'
A new inference-time framework called PoS maintains explicit belief states for long-horizon LLM agents, detecting and recovering f...
On-Policy vs Off-Policy Distillation: Hugging Face Study Finds Rollout Policy Isn't the Main Driver
A controlled strong-to-weak distillation study across Llama3 and Qwen2.5 finds that token-level KL direction shapes task performan...
OneStreamer: A 4B Model That Learns What to Remember and When to Respond in Streaming Video
Hugging Face researchers introduce OneStreamer, a 4B streaming video LLM that jointly learns query-independent evidence recording...