TRACE is an open-source hierarchical memory system for language model agents, achieving 82.5% on MemoryAgentBench's EventQA task using gpt-oss-20B.
Research methodology for improving weak model performance through on-policy distillation from stronger models without access to strong model internals.
Research paper on PixWorld unifying 3D scene generation and reconstruction using pixel-space representations. Received 32 upvotes on HuggingFace.
Research advancing vision-language-action models to work without camera calibration, enabling deployment in more varied robotic and embodied AI contexts.
Research paper presenting LLM-as-a-Verifier, a framework for using language models as general-purpose verification systems across multiple domains. Also appeared on HuggingFace with 5 upvotes.