Source record

Published by arXiv cs.MA (T1 source). The original is at https://arxiv.org/abs/2609.16986.

Pipeline notes

The summary and note below are generated by the signal pipeline — they are Beyond Desk’s reading, not quotations from the source.

ToMAS converts 242 diagnosed multi-agent LLM coordination failures into 39 clean Theory-of-Mind benchmark items, with 94.4% annotator agreement. A GRPO training pilot on Qwen2.5-1.5B failed to show learning effects because LoRA updates were numerically negligible (max ΔW ≈ 7e-6).

Unusually candid about its own experimental null result; the paper's value is the conversion pipeline and the two matched-domain evaluation requirements it surfaces for future work.

Cited by

No citations on record.