Efficient Long Sequence Decoding, Video Generation as Multimodal Reasoning, and Neuro-Symbolic Validation of Chain-of-Thought

Latest research summaries in ML, Robotics, CV, NLP and AI

Welcome to today’s edition of State of AI 👋

This edition covers a diverse range of topics, from breakthroughs in efficient inference for large language models, to new paradigms for multimodal reasoning, and advances in validating the logical consistency of model-generated explanations. Researchers are pushing the boundaries of what’s possible with AI systems, tackling challenges in scalable deployment, robustness, and interpretability.

Here’s what caught our attention:

• SnapStream: Efficient Long Sequence Decoding on Dataflow Accelerators - A novel method for compressing key-value caches in large language models, enabling 4x reduction in memory usage with minimal accuracy loss.

• Thinking with Video: Video Generation as a Promising Multimodal Reasoning Paradigm - A new benchmark and evaluation of using video generation as a unified framework for visual and textual reasoning.

• VeriCoT: Neuro-symbolic Chain-of-Thought Validation via Logical Consistency Checks - A system that automatically formalizes and verifies the logical validity of step-by-step explanations from language models.

• Logit-Entropy Adaptive Stopping Heuristic for Efficient Chain-of-Thought Reasoning - A training-free technique to adaptively halt the generation of chain-of-thought rationales, saving compute without sacrificing accuracy.

• X-Diffusion: Training Diffusion Policies on Cross-Embodiment Human Demonstrations - A framework that leverages human demonstrations to train robot control policies, while avoiding learning infeasible motions.

Let’s get into it 👇

Contents

1. SnapStream: Efficient Long Sequence Decoding on Dataflow Accelerators

2. Question the Questions: Auditing Representation in Online Deliberative Processes

3. PixCLIP: Achieving Fine-grained Visual Language Understanding via Any-granularity Pixel-Text Alignment Learning

4. Thinking with Video: Video Generation as a Promising Multimodal Reasoning Paradigm

5. Optimal Inference Schedules for Masked Diffusion Models

6. Forgetting is Everywhere

7. VeriCoT: Neuro-symbolic Chain-of-Thought Validation via Logical Consistency Checks

8. Logit-Entropy Adaptive Stopping Heuristic for Efficient Chain-of-Thought Reasoning

9. X-Diffusion: Training Diffusion Policies on Cross-Embodiment Human Demonstrations

10. Particle-Grid Neural Dynamics for Learning Deformable Object Models from RGB-D Videos

11. Learning to Control Self-Assembling Morphologies: A Study of Generalization via Modularity

12. Optimizing Sensor Placement in Urban Storm Sewers: A Data-Driven Sparse Sensing Approach

13. CREA: A Collaborative Multi-Agent Framework for Creative Image Editing and Generation

14. Revisiting Federated Fine-Tuning: A Single Communication Round is Enough for Foundation Models

15. When retrieval outperforms generation: Dense evidence retrieval for scalable fake news detection

SnapStream: Efficient Long Sequence Decoding on Dataflow Accelerators