Research Papers
New HEAL Method Targets Hallucinations in Multimodal LLMs via 'Synergy Heads'
A new paper introduces HEAL, a head-level information disentanglement and calibration method that identifies and reduces hallucina...
Cal-OPD Filters Out Teacher Self-Deviation in On-Policy Distillation
A new paper argues that the teacher-student discrepancy learned in on-policy distillation mixes in the teacher's own deviations, a...
Hugging Face Paper Unveils Gricea, an Open-Science Platform for Conversational AI Research
Gricea packages conversational AI studies as configurable, deployable research artifacts that can be run, inspected, shared, and r...
MoME: Hugging Face Paper Routes Token Memory by Context, Not Surface Form
A new paper proposes Mixture-of-Memory Embeddings (MoME), a context-aware memory mechanism that replaces each token's single memor...
GraphSkillEvo Turns LLM Agent Skills into Evolvable Graphs
A new Hugging Face paper proposes representing LLM agent skills as graph-structured natural-language artifacts and optimizing them...
Hugging Face Paper: MintAct Unifies UI Grounding, Navigation, and Visual Tool Use
MintAct is a family of vision-language models at 2B, 4B, and 8B scales that unifies UI grounding, multi-step navigation across mob...
Hugging Face Paper Unveils BI-Bench and BI-Agent to Automate End-to-End Business Intelligence
Researchers introduce BI-Bench, the first benchmark for end-to-end business intelligence, and BI-Agent, a tool-augmented agent tha...
Designer-RSI: Graphic-Design Agent Grows Its Own Skills Without Weight Updates
Hugging Face researchers present Designer-RSI, a continual-adaptation framework in which a frozen frontier model drives profession...
OmniVChat: Hugging Face Paper Tackles Native Audio-Visual Dialogue Data and Evaluation Gaps
A new Hugging Face paper introduces OmniVChat, a task where omni models take raw audio and video directly and return text, along w...
IntBMoE Decouples Expert Participation, Compute, and Memory in Mixture-of-Experts
A new block-conditioned Mixture-of-Experts architecture, IntBMoE, achieves full expert participation while keeping execution spars...
OmniVBench and Omni-R2V Dataset Target Gaps in Reference-to-Video Generation
Hugging Face researchers introduce OmniVBench, a benchmark spanning 7 task families and 18 fine-grained tasks with 12,172 checklis...
Paint-Anything Brings Exact 24-Bit Hex Color Control to Image Generation and Editing
Hugging Face researchers introduce Paint-Anything, a unified hex-prompt interface that lets users specify an object's exact 24-bit...