Graph Learning · LLM × Graph · Multi-Agent · Science
Showing 0 papers for 2026-09-08
Dr. Claw offers an auditable, human-in-the-loop workspace that wraps existing coding-agent executors, avoiding a new autonomous agent. It provides persistent state objects and a reusable skill library to coordinate multiple agents and connect human decisions with AI execution in research workflows.
We model orchestrator-worker interaction as a bilevel coordination game for multi-agent LLM systems, linking coordination, memory, and verification. Under bounded coupling, the workers’ local-update game is an approximate potential game whose equilibrium slack is controlled by decomposition quality, and reflection is analyzed as stochastic movement over semantic memory.
Iris-mini and Iris-pro are search agents trained on large-scale models, with a data pipeline that builds tasks from web hyperlink graphs. They construct multi-hop question chains over an entity graph, rewrite non-answer entities into descriptive references to avoid string matching, and only admit questions that a reference model cannot solve closed-book but can with supporting evidence.
Motion-Omni presents an end-to-end framework for joint speech and full-body motion in spoken dialogue, removing the traditional cascade where speech is generated first and motion is produced afterwards. The spoken-dialogue model natively outputs both spoken content and motion-driven actions, enabling joint optimization and a more natural avatar performance.
The Attention Triangle analyzes cross-modal attention in audio-video diffusion models to study semantic leakage. It identifies three cross-attention edges among text, audio, and video and shows that information can flow bidirectionally between audio and video, revealing how modalities influence each other during generation.