Weekly Trending

A local multi-agent coding harness that makes roles, handoffs, queues, recovery, and reviewable work state explicit.

A document index for hierarchy-preserving, reasoning-based retrieval without relying solely on vector chunks.

A real-time audiovisual model that combines persistent world context with an event stream for streaming interaction.

A multimodal memory system for continuous video that organizes perception around persistent entities across short-term, episodic, semantic, and procedural memory.

An agentic benchmark that tests whether AI manager agents pressure, threaten, or deceive subordinate agents after a task refusal.

A method for localized multimodal control in diffusion transformers, addressing object identity, materials, and spatial regions instead of relying only on a global text prompt.

An open-source toolkit that traces LLM-agent failures, attributes root causes across trajectories, and supports recovery and rerun loops.

A benchmark for multimodal language models that tests repeated observation, reasoning, and visual evidence seeking instead of a single static glance.

A benchmark for autonomous agents performing complex document operations with deterministic verification of state tracking, semantic correctness, and metadata edits.

A framework that models agents as Python objects: methods are actions, fields are state, docstrings are prompts, and type annotations become contracts.

This paper tests language models in multi-turn conversations where a user's goal is revealed, revised, and redirected over time. It finds that static benchmark strength does not transfer cleanly when an agent must keep track of changing user intent.

StateAct gives a primary computer-use agent code-level access to files, the DOM, and application state, while a GUI subagent handles only irreducibly visual steps.