Hán Dān Xué Bù (Mimicry) or Qīng Chū Yú Lán (Mastery)? A Cognitive Perspective on Reasoning Distillation in Large Language Models

TL;DR

The paper examines whether reasoning distillation through supervised fine-tuning (SFT) preserves the cognitive structure underlying human-like problem solving in Large Reasoning Models. By benchmarking 14 models across six reasoning tasks and analyzing functional alignment, representational similarity, and surface-level mimicry, the authors find strong evidence for a Hán Dān Xué Bù effect: distilled students replicate the teacher's verbose traces without internalizing the dynamic resource allocation that produces human-like reaction-time patterns. A key finding is the Linear Inflation Law, where the inverse efficiency of distilled models scales multiplicatively with a fixed factor (N ≈ 2.44) relative to base models, decoupling computational cost from cognitive demand. The results suggest that human-like cognition in reasoning tasks emerges from active reinforcement rather than passive imitation, challenging the efficacy of current reasoning distillation approaches and highlighting the need for training paradigms that cultivate cognitive policy learning. This has implications for designing AI that truly mirrors human problem solving and for understanding the limits of trace cloning in transferring deep cognitive structure.

Abstract

Recent Large Reasoning Models trained via reinforcement learning exhibit a "natural" alignment with human cognitive costs. However, we show that the prevailing paradigm of reasoning distillation -- training student models to mimic these traces via Supervised Fine-Tuning (SFT) -- fails to transmit this cognitive structure. Testing the "Hán Dān Xué Bù" (Superficial Mimicry) hypothesis across 14 models, we find that distillation induces a "Functional Alignment Collapse": while teacher models mirror human difficulty scaling ($\bar{r}=0.64$), distilled students significantly degrade this alignment ($\bar{r}=0.34$), often underperforming their own pre-distillation baselines ("Negative Transfer"). Our analysis suggests that SFT induces a "Cargo Cult" effect, where students ritualistically replicate the linguistic form of reasoning (verbosity) without internalizing the teacher's dynamic resource allocation policy. Consequently, reasoning distillation decouples computational cost from cognitive demand, revealing that human-like cognition is an emergent property of active reinforcement, not passive imitation.

Figures (7)

• Figure 1: Example reasoning problems.
• Figure 2: Sample chain-of-thought produced by R1.
• Figure 3: Combined Analysis of Cognitive Alignment and Reasoning Cost. (Top) Correlation ($r$): between reasoning cost and human RTs. (Bottom) Comparison of Human RTs (solid, left axis) and model token counts (dashed, right axis) across tasks.
• Figure 4: Representational Similarity Analysis (RSA) Heatmap. Values denote Pearson correlations of difficulty fingerprints (z-scored costs). Distilled models exhibit low structural similarity to the Teacher (DeepSeek-R1), indicating a failure to internalize the target reasoning policy.
• Figure 5: The Cognitive Waterfall. Models define a topology of failed imitation defined by Teacher Similarity ($x$-axis) and Human Alignment ($y$-axis). Instead of converging to the Teacher (top-right), distilled models collapse into a "Spurious Valley" characterized by simultaneous loss of alignment (The Drop) and structural similarity (The Drift).