AI Daily — 2026-10-02

English 中文

OneStreamer unifies perception, memory, proactive response for streaming video LLMs; on/off-policy distillation dynamics studied; E-MoE enhances diffu

Covering 5 AI news items

🔥 Top Stories

1. OneStreamer Unifies Perception, Memory, and Proactive Response for Streaming Video LLMs

Researchers introduce OneStreamer, a streaming video LLM method that jointly learns query-independent evidence recording and task response through a shared proactive generation process. Its Proactive Hierarchical Caption Memory produces time-grounded local captions and summaries, enabling reusable factual memory without sacrificing real-time perception. This could materially improve always-on video assistants, robotics perception, and long-horizon multimodal monitoring. Source-huggingface

2. On-Policy vs Off-Policy Learning: Systematic Study of Distillation Dynamics

This paper isolates how rollout policy, token-level KL direction, and learning rate affect strong-to-weak LLM distillation. It tests whether on-policy benefits such as reduced catastrophic forgetting and sparser updates come from the rollout policy itself rather than confounding factors. The controlled comparisons are directly relevant to practitioners choosing between SFT and RL-style distillation recipes. Source-huggingface

3. E-MoE Enhances Mixture-of-Experts for Diffusion Language Models

E-MoE applies an enhanced mixture-of-experts approach to non-factorized diffusion language models, targeting quality limits in few-step generation. The work addresses posterior collapse in continuous Gaussian latent VAEs used to capture cross-position correlations. If successful, it strengthens diffusion LMs as a faster alternative to autoregressive generation. Source-huggingface

📰 Featured

Reinforcement Learning & LLM Objectives

Density-Aware Reward Aggregation Improves Multi-Reward RL for LLMs — Introduces advantage energy to analyze uneven learning progress under GDPO normalization and proposes density-aware reward aggregation to make sparse rewards count in multi-objective RL. Source-huggingface

Multi-Agent & Robotics

Decentralized Multi-Agent Pathfinding via Iterative Intent Denoising — Decentralized Master-Mind iteratively refines agent intent through denoising to reduce incompatible joint actions under partial observability. Source-huggingface

⚡ Quick Bites

Generated by AI News Agent | 2026-10-02