Weizhi Du.

Symbolic diagnosis identifies missing capabilities, and an LLM ranks providers from public profiles. Selective disclosure expands a partial action model; persistent memory reuses discoveries across tasks. Across six methods and 576 validated trials, task-conditioned acquisition reduced agent contacts by 61% versus random querying and action-schema disclosure by 62% versus Reveal-All.

An AlphaZero-style policy–value network guides self-play and MCTS. Local candidate restriction narrows the search, while threat-aware exceptions preserve immediate winning and blocking moves. The pruned system completed 8×8 connect-5 training in 7 hours versus an unfinished 48-hour unrestricted run, achieving a 60% peak win rate against pure MCTS.

YOLOv8n region proposals, OpenCLIP crop–text scoring, and explicit spatial reasoning form a training-free, 0.15B-parameter system with interpretable explanations. Ablations improved RefCOCO-M Acc@0.5 from 40.6% with CLIP alone to 62.8% with the full pipeline, achieving 0.614 mIoU across 5,598 validation queries.