Tag
DeepSeek Harness has been released, featuring a trace mode that visualizes Agent's thinking steps, tool calls, and return results, and uses colors to distinguish different types of content. It can be installed via npm and is suitable for developers to observe Agent work progress.
This paper studies whether single-turn uncertainty quantification methods transfer to interactive LLM agent trajectories, evaluating white-box, black-box, and reflexive scorers across five LLMs and four tool-use datasets. Results show that transfer is uneven, with black-box self-consistency often strongest, and recommend revalidating UQ methods at the trajectory level.
Introduces AD-MCQ and DEFT-RLVR, a method for verifiable reasoning in autonomous driving VLMs that defers future trajectory exposure to post-decision verification, improving reasoning faithfulness while reducing hallucinations.
AgentLens is a new open-source benchmark for evaluating coding agents that assesses the full trajectory of interactions, including instruction following, tool use, error recovery, and more, using formal verification and LLM-written reviews.
This paper proposes TIE, a knowledge fusion framework for masked diffusion language models that tracks confidence dynamics to identify reliable decoding trajectories and iteratively transfers partially denoised sequences between models, improving generation quality on reasoning tasks.
The author shares insights after trying various Agent Memory implementations, concluding that only strictly length-limited entry-level memory (like Hermes) and skill evolution based on trajectory precipitation are somewhat useful, while other graph-based or card-based methods are ineffective.
Nvidia released a 2.6B open-source world model that can generate controllable worlds from a single image, text prompt, and trajectory, running on a single GPU.
The article discusses how traditional observability methods fail as AI agents run for days and hundreds of steps, proposes trajectory-based Observability-Driven Development (ODD) to monitor and debug long-running agents, and introduces related practices on the Honeycomb platform.