LRMs' Human-Like Reasoning is Fixed, Not Adjustable by Budget

A new study from arXiv (2026) investigates whether the alignment between Large Reasoning Models (LRMs) and human cognitive costs can be tuned by adjusting inference-time reasoning budgets. Across models GPT-OSS-20B and GPT-OSS-120B, three effort levels, and six reasoning tasks, the researchers measured how well chain-of-thought token lengths matched human reaction times. The results were striking: within-task and cross-task alignment remained numerically near-identical across all conditions, with Bayes Factors leaning decisively toward the null hypothesis. A manipulation check revealed that the effort parameter actually sets an upper generation ceiling, not a real-time allocation dial—meaning the model's policy for distributing reasoning tokens is crystallized during training and cannot be adjusted on the fly. However, scaling model size did improve the granularity of alignment on arithmetic complexity tasks, suggesting better learned difficulty patterns at larger scales.

These findings have significant implications for how we think about interpretability and control in AI systems. The paper supports a "compiled" (training-time) account of LRM problem-solving rather than an "online" (inference-time) one. For practitioners, this means adjusting reasoning budgets (e.g., increasing max tokens in chain-of-thought) won't make models emulate human cognitive patterns more closely—it merely adds verbose but shallow output. The alignment observed between LRM token allocation and human difficulty is a robust, built-in property of the training process. This challenges intuitive efforts to improve model alignment by simply giving models more compute at inference. Instead, any improvements to human-like reasoning must come from changes in training data, architecture, or objectives.