Can minimal reasoning steps match verbose reasoning accuracy?
This explores whether stripping reasoning down to its bare computational steps—cutting the verbose explanation—keeps accuracy intact, and what that tradeoff reveals about why chains-of-thought are long in the first place.
This explores whether minimal reasoning steps can match verbose reasoning accuracy. The short answer from the corpus is a fairly emphatic yes—and the more interesting finding is *why*. Chain of Draft achieves accuracy equal to standard chain-of-thought while using only 7.6% of the tokens Can minimal reasoning chains match full explanations?. The implication is that roughly 92% of a typical reasoning trace isn't doing computational work—it's style, documentation, and connective prose. The reasoning that matters is small; the verbosity is presentation.
Several notes converge on the same point from different angles. One finds that verbose and concise reasoning occupy distinct, linearly-separable regions of a model's activation space, so you can extract a single steering vector from 50 examples and cut chain length by 67% with no retraining and no accuracy loss Can we steer reasoning toward brevity without retraining?. Another shows that reasoning steps can be dynamically pruned at inference: by tracking which steps actually receive downstream attention, the verification and backtracking steps turn out to be largely ignored, so removing 75% of them preserves accuracy Can reasoning steps be dynamically pruned without losing accuracy?. Both suggest the 'fat' in a reasoning trace is identifiable and removable rather than load-bearing.
But the corpus complicates the picture in a useful way: shorter isn't always better, and longer isn't always worse—it depends on the task and the model. Optimal CoT length follows an inverted-U curve: accuracy peaks at an intermediate length, the sweet spot grows with task difficulty but shrinks as the model gets more capable Why does chain of thought accuracy eventually decline with length?. So a strong model on an easy problem genuinely should reason minimally, while a weaker model on a hard one needs the steps. Notably, RL training drifts toward shorter chains on its own as models improve—brevity emerges from the reward signal, it isn't imposed. There's even a sharper warning: reasoning accuracy degrades as inputs get longer, dropping from 92% to 68% with just 3,000 tokens of padding, far below the context limit Does reasoning ability actually degrade with longer inputs?. Verbosity can actively *hurt* by burying the signal.
If you want the deeper 'why minimal works,' the corpus points somewhere unsettling. Several notes argue that chain-of-thought isn't genuine step-by-step inference at all—it's the model reproducing familiar reasoning *forms* learned in training Does chain-of-thought reasoning reveal genuine inference or pattern matching?, degrading predictably the moment you shift the distribution Does chain-of-thought reasoning actually generalize beyond training data?. If the verbose trace is partly performance—imitating what reasoning looks like—then it makes sense that the performance can be cut without touching the answer. And when models genuinely fail, the cause is often not too few steps but execution bandwidth (they know the algorithm but can't run it at scale) Are reasoning model collapses really failures of reasoning? or premature abandonment of good paths Why do reasoning models abandon promising solution paths?—problems that more verbosity doesn't solve.
The thing you didn't know you wanted to know: the length of a reasoning chain is mostly decoupled from the reasoning itself. Brevity matched to capability is the actual target—and on current evidence, less is not a compromise but frequently the better setting.
Sources 9 notes
Chain of Draft achieves equivalent accuracy to standard chain-of-thought on arithmetic, symbolic, and commonsense tasks while using only 7.6% of tokens. The 92.4% of removed tokens served style and documentation, not computation.
Activation-Steered Compression extracts a single vector from 50 paired examples to reduce chain-of-thought length by 67% while maintaining accuracy and achieving 2.73x speedup. The method is training-free and generalizes across model sizes and domains.
The PI framework categorizes reasoning into six types and uses attention maps to identify that verification and backtracking steps receive minimal downstream attention. Selecting only high-attention steps preserves accuracy while cutting reasoning length substantially.
Task accuracy peaks at intermediate CoT length, with optimal length increasing alongside task difficulty but decreasing with model capability. RL training naturally gravitates toward shorter chains as models improve, revealing that simplicity emerges from reward signals rather than explicit training.
FLenQA shows reasoning accuracy drops from 92% to 68% at just 3000 tokens of padding, far below context window capacity. The degradation is task-agnostic, uncorrelated with language modeling performance, and persists even with chain-of-thought prompting.
CoT works by constraining models to reproduce familiar reasoning patterns from training, not by enabling novel symbolic reasoning. Performance degrades predictably under distribution shifts—the signature of imitation rather than capability emergence.
DataAlchemy experiments show CoT fails systematically under distributional shifts in task, length, and format. Models produce fluent but logically inconsistent reasoning — imitating reasoning form without valid underlying logic.
Models confined to text-only generation cannot execute multi-step procedures at scale, even when they know the underlying algorithm. Tool-enabled models solve problems beyond the supposed reasoning cliff, suggesting the bottleneck is procedural execution bandwidth.
Reasoning LLMs exhibit two reinforcing failures: wandering (invalid exploration) and underthinking (premature path-switching). Decoding-level interventions like thought-switching penalties improve accuracy without fine-tuning, suggesting viable solutions exist but are abandoned prematurely.