Technical Summary: When Do LLMs Reason? A Dynamical Systems View via Entropy Phase Transitions

Problem Statement

Chain-of-thought (CoT) reasoning has become the default strategy for enhancing Large Language Model (LLM) capabilities. However, empirical evidence reveals a paradox: CoT often yields marginal or even negative performance gains on factual and open-ended tasks while significantly increasing token consumption and latency. Current approaches to adaptive reasoning typically rely on static, offline profiling of task difficulty or heavy token-level interventions, lacking a reliable, real-time signal to determine when explicit reasoning is beneficial for a specific model-query pair. The core challenge is to identify a mechanism that can dynamically detect the emergence of a "reasoning state" during generation to enable selective, efficient inference.

Methodology: EDRM

The authors propose EDRM (Entropy Dynamics-based Reasoning Manifold), a lightweight, training-free routing framework that views reasoning not as a static capability but as a dynamic decoding state emerging from the interaction between a model and a query.

1. Theoretical Framework: Entropy Phase Transitions

The paper posits that successful reasoning corresponds to a "phase-transition-like" shift in decoding dynamics:

• Convergent Regime: Tasks benefiting from CoT exhibit a consistent reduction in token-level entropy (uncertainty) during early decoding steps, shifting from a high-entropy exploratory state to a low-entropy structured state.
• Divergent/Unstable Regime: Tasks where CoT is ineffective display unstable, oscillatory, or increasing entropy trajectories, indicating that the model is drifting or over-thinking without converging.

2. Entropy Dynamics Descriptors

To quantify these dynamics, EDRM analyzes the first $N=64$ decoding steps under a "Standard" probing setting (minimal prompting, no forced CoT). It extracts three complementary statistics from the entropy trajectory $\{H_i\}_{i=1}^N$:

• Cumulative Entropy ($S_H$): The sum of token-level entropies, representing the total uncertainty budget. High $S_H$ indicates an "uncertainty-overload" regime where CoT is unlikely to converge.
• Spearman Correlation ($V_{sp}$): The rank correlation between step index and entropy. A negative value indicates a monotonic decrease in uncertainty (convergent reasoning), while positive values suggest divergence.
• Von Neumann Ratio ($avnr$): The ratio of mean square successive differences to variance, measuring the volatility or smoothness of the trajectory. Low $avnr$ indicates stable dynamics; high $avnr$ suggests oscillatory behavior.

These three descriptors form a compact, interpretable Reasoning Manifold $(S_H, V_{sp}, avnr)$.

3. Routing Mechanism

EDRM uses these descriptors to adaptively select one of three decoding strategies: Direct (no reasoning), Standard (minimal prompting), or CoT (explicit step-by-step reasoning).

• Heuristic Router: A training-free policy routes to CoT if $V_{sp} < -k \cdot avnr$ (stable convergence), to Direct if $V_{sp} > k \cdot avnr$ or if $S_H$ exceeds a threshold (instability or overload), and to Standard otherwise.
• Learned Router: An optional lightweight MLP can be trained on a small calibration set to predict the optimal strategy based on the manifold coordinates, allowing for finer-grained adaptation.

Key Contributions

1. Dynamic View of Reasoning: The paper conceptualizes LLM reasoning as an emergent decoding state rather than a fixed property, identifying a phase-transition-like shift in entropy dynamics as the signal for when explicit reasoning becomes beneficial.
2. EDRM Framework: Introduction of a training-free, instance-adaptive routing framework that embeds early-stage entropy dynamics into a compact manifold for real-time strategy selection.
3. Efficiency and Accuracy: Demonstration that entropy-driven control can significantly reduce token consumption while maintaining or improving accuracy across diverse models and benchmarks, without architectural modifications.

Experimental Results

The framework was evaluated on 15 benchmarks (spanning mathematical, commonsense, scientific, and formal logic reasoning) across 4 LLMs of varying scales (3B to 8B) and architectures (including base and reasoning-enhanced models).

• Dataset-Level Performance: EDRM achieves 41–55% token reduction compared to always-on CoT while improving accuracy by up to 1.05 percentage points on base models. For instance, on Llama-3.2-3B, it reduced tokens from 251.7 to 113.2 while increasing accuracy from 60.84% to 61.89%.
• Instance-Level Performance: Fine-grained instance-level routing further improves accuracy by up to 4.7% over static CoT baselines while maintaining 27–45% token savings.
• Robustness: EDRM effectively suppresses "over-reasoning" in reasoning-enhanced models (e.g., Qwen3-4B), reducing token costs by ~34–38% while maintaining comparable accuracy.
• Generalization: The routing signals generalize across different model families and task types, successfully distinguishing between tasks requiring deep reasoning (e.g., GSM8K, FOLIO) and those where direct recall is superior (e.g., ARC-E, CSQA).

Significance and Claims

The paper claims that reasoning should be invoked selectively rather than by default. By interpreting reasoning as a controllable decoding state detectable via early entropy dynamics, EDRM provides a principled, low-overhead mechanism for adaptive inference.

The authors emphasize that their approach:

• Requires no training for the heuristic variant and only minimal calibration for the learned variant.
• Operates without architectural modifications to the underlying LLM.
• Offers a theoretical insight into the nature of reasoning as a dynamic transition, moving beyond static task-model categorizations.
• Demonstrates that entropy-driven decoding control is a viable path toward efficient, adaptive LLM inference, particularly for resource-constrained and latency-sensitive applications.

The paper concludes that while EDRM currently requires a short probing stage and is validated on open-source text models in the 3B–8B range, it establishes a foundational step toward adaptive reasoning control in autonomous agents and production systems.