TacitVLA is a new vision-language-action model that improves performance and reduces latency by generating actions independently of reasoning text. It outperforms previous results on various benchmarks and has the potential to enhance AI agent capabilities.

Researchers propose an agentic AI framework for generating edge case scenarios for safety assurance in autonomous systems. The framework uses local LMs, ensuring data privacy and security, and involves think, act, observe, update, and continue loops with lightweight JSON encoding. The system critiques and refines scenarios using domain-specific criteria and post-hoc clustering to create traceable evidence for safety assurance.

This paper proposes Trust Region Q Adjoint Matching (TRQAM), an off-policy fine-tuning algorithm for reinforcement learning that adapts a trust-region parameter to control the path-space KL between the fine-tuned and pretrained policies. TRQAM achieves stable off-policy RL and outperforms prior methods on 50 OGBench tasks.

SafeInferCom is a framework for improving the reliability of inference-time computation in large reasoning language models (LRLMs) used for robotic task planning. It introduces a verifier-guided approach to preserve valid intermediate plans and correct errors during generation, leading to improved planning success and reduced token usage.