An Empirical Investigation of Whether Role Asymmetry Can Overcome Reasoning Blind Spots in Homogeneous Models through Adaptive Routing and Local Repair

Abstract

Under the test-time compute scaling paradigm, improving the multi-step reasoning accuracy of lightweight large language models within a limited computational budget remains a critical challenge. Existing self-correction and multi-agent debate methods often suffer from rigid interaction schemes, coarse-grained correction, and homogeneous role configurations. Whether role-asymmetric constraints can enable a critic agent to overcome the reasoning biases of its generator has rarely been tested directly. This paper develops a collaborative reasoning framework combining adaptive routing with role-asymmetric local repair, treating this hypothesis as a falsifiable research question, and conducts systematic experiments on two backbone models, Qwen2.5-7B-Instruct and Qwen3-8B. Results show that the framework's disadvantage relative to simple baselines is capability-dependent: on the weaker model, it is statistically indistinguishable from standard chain-of-thought and other single-model baselines while significantly outperforming multi-agent debate; on the stronger model, it significantly underperforms most baselines. A sanity check confirms that the critic reliably identifies clearly structured arithmetic errors; however, its recall drops sharply on subtle errors requiring cross-step verification of quantitative relationships, and the net effect of repair mirrors this pattern, close to neutral on the weaker model and clearly negative on the stronger one. Ablation experiments show that the marginal effects of routing and role asymmetry reverse with model capability, turning negative on the stronger model even as removing either component improves its accuracy. These findings provide reproducible empirical evidence for systematic reasoning blind spots in self-criticism by homogeneous models, and indicate that role asymmetry implemented only through prompting cannot reliably overcome cognitive homogeneity.