HuggingFace

CyberSecQwen-4B is a small, specialized 4B parameter model fine-tuned for defensive cybersecurity tasks, designed to run locally on a single GPU, addressing privacy, cost, and air-gapped deployment needs.

Allen AI releases EMO, a mixture-of-experts model where modular structure emerges naturally from data, enabling use of just 12.5% of experts for a task while maintaining near full-model performance.

HiDream-ai has open-sourced HiDream-O1-Image (8B), a unified image generative foundation model built on a Pixel-level Unified Transformer (UiT) that natively handles text-to-image, image editing, and subject-driven personalization at up to 2048×2048 resolution without external VAEs or disjoint text encoders. It debuted at #8 in the Artificial Analysis Text to Image Arena and is positioned as a leading open-weights text-to-image model.

A tutorial and project demonstrating LoRA fine-tuning of Qwen3-1.7B on AMD MI300X using ROCm for clinical question answering, providing a CUDA-free alternative for medical AI development.

EMO is a Mixture-of-Experts model that enables modular deployment by grouping similar domain tokens with shared experts, achieving performance comparable to standard MoEs while allowing significant expert pruning (25% experts retain 99% performance) without performance degradation.

PianoCoRe is a large-scale piano MIDI dataset unifying and refining open-source corpora with 250,046 performances of 5,625 pieces by 483 composers, featuring note-level alignments for music information retrieval and including a MIDI quality classifier and alignment refinement pipeline.

GeoStack introduces a geometric framework to compose independently trained domain experts in Vision-Language Models without catastrophic forgetting, achieving constant-time inference and a 10x reduction in geometric error.

StraTA proposes strategic trajectory abstraction for long-horizon LLM agents, using hierarchical GRPO-style rollout with diverse strategy sampling and critical self-judgment to improve sample efficiency and final performance over frontier models and prior RL baselines.

This paper introduces a framework for validating comparative LLM safety scoring without ground-truth labels, using an 'instrumental-validity chain' to establish deployment evidence. It demonstrates the method using a local-first tool called SimpleAudit on Norwegian safety packs and compares models like Borealis and Gemma 3.

This paper introduces Sparkle, a new dataset and benchmark for instruction-guided video background replacement, addressing the lack of high-quality training data in this domain. It proposes a scalable pipeline with decoupled guidance to generate realistic foreground-background interactions.

This research paper investigates how Large Language Models encode social role granularity as a structured latent dimension. It demonstrates that this 'Granularity Axis' is consistent across architectures like Qwen3 and Llama-3, and can be causally manipulated via activation steering.

This paper introduces an auto-research framework using specialist agents to iteratively refine training recipes through an empirical loop of code execution and feedback. The system autonomously improves performance on tasks like Parameter Golf and NanoChat without human intervention by leveraging lineage feedback.

This paper introduces MARBLE, a gradient-space optimization framework for multi-reward reinforcement learning fine-tuning of diffusion models, which harmonizes policy gradients without manual weighting.

This paper introduces ScaleLogic, a framework demonstrating that RL training compute scales as a power law with reasoning depth in LLMs. It highlights that logical expressiveness is key to improving downstream transfer and training efficiency.

This paper introduces the AI Co-Mathematician, a workbench that uses agentic AI to support mathematicians in open-ended research tasks like ideation and theorem proving. Early tests show the system achieving state-of-the-art results on hard problem-solving benchmarks, including a 48% score on FrontierMath Tier 4.

Skill1 is a unified framework that trains a single policy to co-evolve skill selection, utilization, and distillation using a shared task-outcome objective. Experiments on ALFWorld and WebShop show it outperforms existing baselines in complex task environments.

This paper introduces SkillOS, a reinforcement learning framework that enables LLM agents to learn long-term skill curation policies for self-evolution, improving performance and generalization across tasks.

Cola DLM is a hierarchical latent diffusion language model that uses text-to-latent mapping and conditional decoding to achieve efficient, non-autoregressive text generation.

UniPool introduces a shared expert pool architecture for Mixture-of-Experts models, reducing parameter growth with depth while improving efficiency and performance over standard MoE baselines.

This paper introduces DeScore, a video reward model that decouples reasoning and scoring processes to improve training efficiency and generalization. It addresses the limitations of existing discriminative and generative reward models by using a 'think-then-score' paradigm with multimodal large language models.