Precise but Uncoupled: Reviewer Precision Does Not Guarantee Critique Uptake in Multi-Agent Math Reasoning

Yang, Chih-Hsuan; Jiang, Jingyan; Vasudevan, Vikram; Yang, Cheng-Hau; Zheng, Huihuo; Chen, Le; Huerta, Eliu A.; Vishwanath, Venkatram; Foster, Ian T.; Thakur, Rajeev

doi:10.48550/arXiv.2607.15388

Two panels. Left, a stacked bar chart of inner-loop failure modes over wrong-initial-candidate cases: PER splits 11% repair, 49% neglect, 40% try-but-fail, while broadcast splits 26% repair, 26% neglect, 48% try-but-fail — 2.3 times more neglect under PER. Right, a grouped bar chart of five metrics for PER and broadcast: precision 0.86 versus 0.64, recall 0.75 versus 0.87, CouplingRate 0.34 versus 0.93, guided repair 0.05 versus 0.29, misleading resistance 0.92 versus 0.71. A shaded box marks CouplingRate and guided repair as the coupling gap. — A more precise reviewer does not guarantee better critique uptake or repair.On Omni-MATH, PER has higher reviewer precision but lower uptake and reviewer-guided repair than broadcast. This protocol comparison does not isolate a causal mechanism.

When does correct critique lead to repair?

We compare multi-agent protocols on 4,181 verifier-grounded Omni-MATH problems with matched gpt-oss-120b actors. The planner–executor–reviewer pipeline has a more precise reviewer than broadcast deliberation, yet verified useful critique changes its next candidate much less often. Detection, uptake, and successful repair are empirically separable: evaluating a reviewer alone can miss whether the system acts on its feedback. Within-PER interventions provide directional evidence about feedback delivery, without establishing an isolated causal mechanism.

Reproduction & release notes ↗Dataset & terms ↗

Release note: the legacy token and model-call cost figures in arXiv v1 are invalid and withdrawn; accuracy and review-transition results are unaffected. See the repository for artifact coverage and reproduction limits.

Citation

Download BibTeX ↗

Chih-Hsuan Yang et al. Precise but Uncoupled. arXiv preprint, 2026.
doi:10.48550/arXiv.2607.15388

BibTeX

@misc{yang2026preciseuncoupled,
  title         = {Precise but Uncoupled: Reviewer Precision Does Not Guarantee
                   Critique Uptake in Multi-Agent Math Reasoning},
  author        = {Yang, Chih-Hsuan and Jiang, Jingyan and Vasudevan, Vikram and
                   Yang, Cheng-Hau and Zheng, Huihuo and Chen, Le and
                   Huerta, Eliu A. and Vishwanath, Venkatram and
                   Foster, Ian T. and Thakur, Rajeev},
  year          = {2026},
  eprint        = {2607.15388},
  archivePrefix = {arXiv},
  primaryClass  = {cs.AI},
  doi           = {10.48550/arXiv.2607.15388},
  url           = {https://arxiv.org/abs/2607.15388}
}