When Do LLMs Reason? A Dynamical Systems View via Entropy Phase Transitions
The paper treats reasoning as a decoding state, using early entropy patterns to decide when chain-of-thought helps. It reports token savings and accuracy gains with a training-free router.
Intelligence analysis by GPT-5.4 Mini
The paper argues that LLM reasoning is not fixed by the task alone, but emerges during decoding. It uses early entropy behavior to route between inference strategies, aiming to avoid default chain-of-thought when it is costly and unhelpful.
A big language model can sometimes think step by step, but that takes more time and words. This paper tries to spot, very early, when that extra thinking will help and when it will just waste effort.
It uses a simple idea: watch how unsure the model is while it starts talking. If the model quickly becomes more settled, that may mean step-by-step thinking will work well. If it stays jumpy, the extra thinking may not help much.
The picture is like watching a kid solve a puzzle. Some puzzles need careful steps, and others are faster without them. The paper says a computer can learn when to switch modes instead of always using the slow path.
Analysis
What the paper claims
The authors frame reasoning as a dynamic state that appears during generation rather than a permanent trait of a model or a prompt. Their core observation is that early entropy behavior seems to separate tasks that benefit from chain-of-thought from tasks that do not.
Why entropy matters here
In their analysis, tasks that gain from chain-of-thought tend to show a steady drop in entropy early in decoding. Tasks with weak or negative chain-of-thought benefit more often show unstable or rising entropy. The paper interprets this as a phase-transition-like movement from a high-entropy exploratory phase to a lower-entropy structured reasoning phase.
EDRM
Using that signal, the authors propose EDRM, a training-free routing framework that maps entropy trajectories into a compact manifold representation. The goal is to choose inference strategies adaptively instead of always turning on reasoning. The method is described as usable in zero-shot settings and as capable of instance-level adaptation with a small calibration set.
Reported results
Across 15 benchmarks and 4 LLMs, the paper says EDRM beats static baselines. At the dataset level, it reports 41-55% token reduction while improving accuracy with as few as 50 calibration samples. At the instance level, it reports up to 4.7% higher accuracy while keeping 27-45% token savings.
Takeaway
The main message is not that chain-of-thought is bad, but that it should be used selectively. The paper’s contribution is a practical rule for deciding when to spend tokens on reasoning and when to skip it.
Key points
- The paper treats reasoning as a decoding state that emerges during generation.
- Early entropy patterns are used as a signal for whether chain-of-thought will help.
- The proposed EDRM method is training-free and adaptively routes inference strategies.
- The authors report token savings of 41-55% at the dataset level and up to 4.7% accuracy gains at the instance level.
- The main claim is that reasoning should be used selectively, not by default.
