July 17, 2026

Proposes SearchOS, a multi-agent framework for web-based information-seeking agents that converts fragile, implicit search progress into explicit, persistent, shared state via schema-completion-with-citations, a search-oriented context manager tracking task frontiers and evidence graphs, pipeline-parallel scheduling, and a hierarchical skill system. Shows measurable gains over single- and multi-agent baselines on benchmark datasets.

SEED tackles the sparse-reward problem in outcome-based RL for agentic LLMs by having the policy extract reusable natural-language 'skills' (workflows and failure-avoidance patterns) from its own completed trajectories, then using the action-probability gap between skill-augmented and unaugmented rollouts as a dense, token-level on-policy distillation signal. Reports gains on both text and vision agentic tasks plus better generalization to unseen scenarios.

Introduces BadWAM, exposing adversarial vulnerabilities in world-action models (WAMs) that couple future-state prediction with robotic action generation. Small visual perturbations induce 'World-Action Drift Attacks': an action-only variant collapses task success from 96.5% to 43.1%, while a stealthier 'imagination-preserving' variant keeps the model's predicted future looking clean while still forcing harmful actions, showing that moderate regularization can mask the drift entirely.

A fully open (weights, data, and training recipe) 4B-parameter video multimodal LLM using an Inflated 3D Vision Transformer with adaptive frame resolution, paired with a scalable video-data synthesis pipeline producing general, long-form, and streaming-video datasets. Claims to surpass prior open-source models with equal or larger parameter counts across multiple video benchmarks.