📄 Summary: This paper introduces controllable evaluation environments to systematically measure and distinguish exploration versus exploitation behaviors in language model agents without access to their internal policies. The approach uses partially observable 2D grids and task DAGs that can be programmatically adjusted to emphasize different decision-making challenges, enabling policy-agnostic performance metrics for embodied AI scenarios.
💡 Key Insight: You can objectively measure whether an AI agent is failing at exploring new options or exploiting known solutions by testing it in specially designed environments.
📄 Summary: LangFlow is the first continuous diffusion model for language that achieves performance comparable to discrete diffusion models by connecting embedding-space diffusion to Flow Matching via Bregman divergence. The approach introduces a novel ODE-based evaluation method and a learnable Gumbel-based noise scheduler that adapts to the information content of tokens at each step.
💡 Key Insight: Continuous diffusion (proven successful for images) can now match discrete methods for text by adapting the noise schedule to how much information each token carries.
📄 Summary: TREX is a multi-agent system that automates the complete LLM training lifecycle by orchestrating collaboration between Researcher and Executor modules to handle requirement analysis, literature review, strategy formulation, data preparation, and model training. The system models the experimental process as a search tree to efficiently explore training configurations and reuse previous results.
💡 Key Insight: AI agents can now automatically manage the entire process of improving language models by treating different training experiments as paths in a searchable tree.
📄 Summary: Seedance 2.0 is a unified multi-modal audio-video generation model supporting four input modalities (text, image, audio, video) with comprehensive content reference and editing capabilities. The model delivers significant improvements across video and audio generation quality, achieving performance on par with leading systems in both expert and user evaluations.
💡 Key Insight: Modern video generation models now handle multiple input types seamlessly and can edit both audio and video simultaneously in a single unified system.