ToMAS converts diagnosed multi-agent coordination failures into training items, then reports honestly that its own training run did nothing

arXiv·low signalXBlueskyLinkedInCopy link

ToMAS applies four explicit convertibility criteria to diagnosed execution traces, converting inter-agent misalignment cases (labelled FC2 in MAST-Data) into partner-state reasoning items; a full pass over 242 eligible non-AG2 traces produced 39 clean items, with an 18-trace reliability pilot reaching 94.4% raw agreement and Cohen's kappa 0.92. The GRPO feasibility experiment on Qwen2.5-1.5B is a documented null: every condition cleared the ROUGE-L threshold on the same 2 of 28 held-out Magentic GAIA items, and post-hoc adapter checks showed the LoRA update stayed numerically negligible at max abs delta W around 7e-6, so all conditions decode identically to the untrained checkpoint. The authors name the two blockers any conclusive study must fix — a provenance gap between training and evaluation items, and lexical-overlap scoring.