OpenCoF framework and dataset released for Chain-of-Frame reasoning in video generation
Researchers introduced OpenCoF, a framework comprising the OpenCoF-17K dataset, designed to enable Chain-of-Frame (CoF) reasoning in video generation models. This approach uses temporally connected frames as a reasoning path, distinct from traditional Chain-of-Thought (CoT). The work addresses the lack of dedicated supervision for CoF reasoning in existing video generators.
Entities
Related
ARDY: Autoregressive Diffusion with Hybrid Representation for Interactive Human Motion Generation
Researchers introduced ARDY, a streaming generation framework that enables real-time synthesis of 3D human motions with text and kinematic control, bridging the gap between offline precision and online speed. The method addresses limitations in existing online approaches regarding controllability and complex text semantics.
PRX team publishes Part 4 detailing data strategy with VLM re-captioning and streamable corpus
The PRX team released Part 4 of their series, outlining their data strategy for assembling training data. They mix public and internal datasets, re-caption images using a VLM, and convert the result into a streamable corpus for training PRX. The post details guiding principles for diverse pre-training data.
↑ Updated Mon, Jul 6, 2026, 03:30 PM — PRX Part 4 details data strategy: public/internal datasets, VLM re-captioning, streamable corpus
Vox Director: open-source agent skill automates Vox-style explainer videos from one topic
A new open-source agent skill called Vox Director automates the creation of Vox-style paper-collage explainer videos from a single topic. It runs on Atlas Cloud API and local ffmpeg, handling script, keyframes, motion, voice-over, music, and captions. The project was released on GitHub on July 10, 2026.
DynaKRAG: A unified framework for learnable evidence control in multi-hop RAG
Researchers propose DynaKRAG, a framework that learns a state-conditioned policy to dynamically choose among evidence operations (retrieval, reformulation, critique, sufficiency checking) in multi-hop retrieval-augmented generation. This moves beyond fixed pipelines, potentially improving accuracy and flexibility in complex QA tasks.
VAORA reward design addresses VLM failures in interactive physical reasoning
A new paper on arXiv introduces VAORA (Visual Action Outcome Reasoning Alignment), a reward design that targets two key failure modes in vision-language models: hallucinated chain-of-thought reasoning and misalignment between reasoning and actions. VAORA uses a Visual Alignment Reward to anchor reasoning to visual context, aiming to improve generalization in unseen interactive physical reasoning tasks.