HuggingFace

Ai2 introduces TutorMoments, a replay-based evaluation framework and dataset for measuring whether LLMs can balance when to help and when to hold back in one-on-one math tutoring. Preliminary results show models tend to over-help, and prompt engineering only partially closes the gap to human tutors.

A Hugging Face repository hosting MiniMax-H3 models converted for ComfyUI usage, along with a Lightx2v distill LoRA for faster inference.

Hugging Face page for the Minimax-h3-Turbo video generation model, with instructions for using it via Diffusers and Colab/Kaggle notebooks.

Third-party ComfyUI-compatible LoRA conversions for MiniMax-H3 Turbo 4-step audio-video generation, including further-trained checkpoint-500 variants and an example workflow.

This paper introduces KVAE, a family of tokenizers for audio, image, and video designed for text-conditioned generative models, claiming competitive or superior reconstruction and generation quality compared to existing open-source tokenizers. The code and training details are publicly released.

EffectLearner is a semantic-reasoning-enhanced framework for real-world video object removal, combining a VLM-based Object-Effect Reasoner with a DiT-based Video Eraser and introducing the EffectWorld dataset to handle complex object-induced effects.

This paper introduces MASS, a method for multiplayer world models that disentangles world dynamics from view rendering using an authoritative shared state, enabling scalable and consistent multi-agent simulation with up to 1,024 concurrent players.

This paper proposes a systems blueprint and six-level capability ladder for building Economic World Models (EWMs), where heterogeneous agents interact with markets and institutions to generate emergent economic dynamics. It surveys existing work and emphasizes gaps in self-evolving agents, endogenous institutions, and empirical alignment.

This paper surveys the evolution of continual learning from parameter-centric methods to system-level adaptation, proposing a tri-axial framework (When, How, Where) to characterize learning across pre-training, post-training, and inference stages.

DyPES-VLA is a cross-embodiment VLA model that learns shared dynamics priors via future prediction and uses an embodiment-specific Mixture-of-Experts action head to control robots in their native action spaces, achieving state-of-the-art results on LIBERO, RoboCasa, and RoboTwin benchmarks.

HarnessOpt-Bench is a benchmark for evaluating LLMs' ability to optimize the harness—the prompts, tools, control flow, memory, and orchestration code—around a target agent, using a fixed evaluation budget. Experiments with five frontier LLMs show that optimizer models separate more than the coding harnesses they act through, with substantial room for improvement.

ChronoVision is a multimodal framework that improves temporal reasoning in vision-language models by aligning visual logic with latent imagery, using a Reconstructive Visual Head, ROI Attention Locating module, and reinforcement learning. It introduces the Vbvr-VQA dataset and achieves SOTA accuracy on temporal tracking benchmarks.

This paper introduces UniME-R1, an embedder-adviser framework for unified multimodal retrieval that generates Retrieval-Centric Chain-of-Thought (RC-CoT) conditioned on retrieval feedback, improving retrieval performance by learning from hard negatives.

This paper studies On-Policy Delta Distillation (OPD^2) for multilingual math reasoning in English, Korean, and Japanese, showing consistent improvements over standard OPD and narrowing language gaps.

GST-Bench is a new VQA benchmark for evaluating global spatial awareness in video understanding, testing whether VLMs can build coherent global scene representations from long-horizon egocentric video. Evaluation of 22 state-of-the-art VLMs shows a large gap versus humans, with the best model scoring 42.68 vs 79.08.

AgentOPSD proposes a critic-free recursive method for turn-level credit assignment in agentic RL, reweighting outcomes via Bayesian belief updates in log-odds space. It achieves 89.1% success on ALFWorld with Qwen2.5-7B, outperforming GRPO and self-distillation baselines.

CalibForge is an autonomous terminal-task synthesis system that uses adversarial solver calibration to create learnable tasks for training terminal agents. It constructs 5,431 calibrated tasks and improves agent performance on Terminal-Bench2.0, SWE-bench Pro, and Doc2Repo.

EnvACE introduces world rehearsal, an agentic reinforcement learning method that replaces external environment interaction by having the policy rehearse environment responses internally, achieving strong performance across multiple benchmarks.

PaDoc introduces a layout-grounded parallel decoding method for end-to-end document parsers, decoupling layout and content decoding to reduce decoding depth and improve throughput. It achieves state-of-the-art results on OmniDocBenchFull and significantly speeds up inference compared to sequential baselines.

Hugging Face announces Baseten as a supported Inference Provider, enabling serverless access to popular open-weight LLMs like Kimi K3, DeepSeek V4 Flash, and GLM-5.2 directly from model pages and SDKs.