Effort as Ceiling, Not Dial: Reasoning Budget Does Not Modulate Cognitive Cost Alignment Between Humans and Large Reasoning Models

2026-05-19 · Source: cs.CL updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Social Sciences & Behavioral Studies · Depth: Expert, extended

Summary

A study investigating Large Reasoning Models (LRMs) found that inference-time reasoning effort does not modulate the alignment between LRM token usage and human cognitive costs. Researchers tested GPT-OSS-20B and GPT-OSS-120B across three effort levels (low, medium, high) and six reasoning tasks, including arithmetic, formal logic, and relational reasoning. The alignment between LRM chain-of-thought (CoT) trace length and human reaction times remained invariant, with Bayes Factors leaning towards the null hypothesis and mean alignment being numerically near-identical across conditions. A manipulation check revealed that the "reasoning_effort" parameter acts as an upper budget for token generation rather than a real-time allocation dial. Arithmetic complexity contrasts further showed that token allocation tracks fine-grained, format-dependent human difficulty patterns, with model scale improving this match. These findings suggest that cognitive cost alignment is a training-time achievement, robust to inference-time perturbations.

Key takeaway

For AI Scientists and Machine Learning Engineers developing or deploying LRMs, understand that the alignment of model reasoning costs with human cognition is primarily determined during training, not by inference-time effort settings. Your efforts to achieve human-like cognitive cost scaling should focus on reinforcement learning with verifiable rewards (RLVR) training objectives, as post-training adjustments to reasoning effort parameters will likely only set a token budget rather than dynamically reconfigure the model's problem-solving policy.

Key insights

Human-LMM cognitive cost alignment is a stable, training-time achievement, robust to inference-time effort changes.

Principles

• Alignment is a structural property, not a calibration-sensitive artifact.
• Model scale improves fidelity of tracking human difficulty patterns.

Method

Evaluated GPT-OSS-20B and GPT-OSS-120B under three reasoning effort conditions across six tasks. Measured within-task and cross-task Pearson correlations between log-transformed token counts and human RTs, using Bayesian paired-samples t-tests for effort invariance.

In practice

• Focus on training objectives for human-aligned cost allocation.
• Inference-time effort parameters act as budgets, not dynamic controls.

Topics

• Large Reasoning Models
• Cognitive Cost Alignment
• Chain-of-Thought
• Effort Invariance
• Arithmetic Cognition

Best for: AI Scientist, Research Scientist, Machine Learning Engineer

Related on AIssential

• Humans Disengage, Reasoning Models Persist: Separating Difficulty Registration from Deliberation Allocation — LRMs spend more tokens on incorrect answers due to uncertainty, while humans disengage from problems they expect to fail.
• Reinforcement Learning with Semantic Rewards Enables Low-Resource Language Expansion without Alignment Tax — Semantic-space alignment with reinforcement learning mitigates "alignment tax" in low-resource language expansion for LLMs.
• Humans Disengage, Reasoning Models Persist: Separating Difficulty Registration from Deliberation Allocation — LRMs increase deliberation on failures due to uncertainty, while humans decrease it due to abandonment.
• Inference Scaling (Test-Time Compute): Why Reasoning Models Raise Your Compute Bill — Inference scaling dynamically allocates compute for reasoning, but requires strategic management to avoid excessive costs and performance issues.
• My research: a computational cognitive neuroscience perspective on alignment — The first takeover-capable AI will likely be augmented LLMs, requiring alignment strategies that account for emergent human-like cognitive capacities.
• The $250K Inverse Scaling Prize and Human-AI Alignment — Inverse scaling reveals a core misalignment between LLM training objectives and desired human values.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CL updates on arXiv.org.