Editorial illustration for Study formalises LLM reasoning redundancy as truncatable steps in correct traces
Study formalises LLM reasoning redundancy as truncatable...
How many reasoning steps does an LLM actually need to solve a problem? A new study offers a crisp, empirical answer: far fewer than it typically takes. The researchers formalise this redundancy as a measurable quantity, the fraction of trailing steps you can chop off a correct reasoning trace, forcing the model to stop thinking and emit a final answer, without spoiling the result.
Across four frontier reasoning models and two math benchmarks, that fraction lands between 61% and 93%. In six of eight model-benchmark conditions, the median critical prefix is just a single segmented step. Even on the hardest Level-5 problems from MATH-500, redundancy remains substantial, ranging from 46% to 85%.
The finding is robust across judge families and problem difficulty. The implication is stark: these models are doing a lot of unnecessary thinking. The question isn’t whether they can reason, it’s how much of that reasoning is actually required.
We formalise reasoning redundancy directly in terms of the reasoning model itself: the redundancy of a correct trace is the largest fraction of its trailing segmented steps that can be truncated while $\pi$, forced to terminate thinking and emit a final answer, still produces the correct answer. A large-scale quantification across four frontier reasoning models and two mathematical benchmarks shows that step-level redundancy is consistently high -- between 61% and 93% across the 8 (model, benchmark) conditions we study, with the median critical prefix equal to a single segmented step in six of the eight conditions -- that the finding is robust to the choice of judge family, and that although $\rho$ decreases with problem difficulty on MATH-500, all four models remain substantially redundant ($\rho \in [46\%, 85\%]$) even on the hardest Level-5 problems.
The numbers are striking: four frontier models, two benchmarks, and in six out of eight conditions, the median critical prefix is a single segmented step. That means any extra thinking beyond that first step is, mathematically, redundant , a luxury, not a necessity. The formalisation is clean, the result is stark, and the implication is uncomfortable for anyone who assumed that more reasoning means better reasoning.
Yes, harder problems demand more thinking. Even then, redundancy hovers between 46% and 85%. The model is still talking to itself long after it has already arrived.
This is not a failure of intelligence , it is a failure of efficiency. The model does not know when to stop. What makes this finding powerful is its grounding in the model’s own logic.
No external oracle, no hand-crafted heuristic. The trace itself reveals where the thinking becomes surplus. And the fact that a single early step so often contains the critical reasoning suggests that much of what follows is not refinement, but recitation , a kind of cognitive loop, a rhetorical flourish in the machine’s internal monologue.
The path forward is clear. Stop trying to make models think harder. Start teaching them to think shorter.
Related Reading
LLMs & Generative AI ChatGPT's 'Nerdy' tweak rewards goblin metaphors in answers, study finds
LLMs & Generative AI Google tests visual 'magazine-style' UI for Gemini 3 Pro users
LLMs & Generative AI AI Engineers Face Rising Costs, Need New Strategies for Efficiency
From Archives Direct and Surrogate Verification Encode Transformer Circuits into SMT Solvers
From Archives AWS Agent Toolkit Shows Invocation, Success, UserError, SystemError Stats
Common Questions Answered
What does the study mean by 'truncatable steps' in LLM reasoning traces?
Truncatable steps refer to the trailing reasoning steps that can be removed from a correct reasoning trace without affecting the final answer. The researchers formalized this concept as a measurable quantity, showing that LLMs can be forced to stop thinking early and still produce correct results, indicating that many of their reasoning steps are mathematically redundant.
What percentage of reasoning steps were found to be redundant across the frontier models tested?
The study found that between 61% and 93% of trailing steps in correct reasoning traces could be truncated without spoiling the result across four frontier reasoning models and two math benchmarks. This striking range demonstrates that LLMs typically use far more reasoning steps than they actually need to solve problems correctly.
What is the significance of the 'median critical prefix' being a single segmented step?
In six out of eight conditions tested, the median critical prefix was just a single segmented step, meaning that any extra thinking beyond that first step is mathematically redundant. This finding suggests that for many problems, LLMs only truly need one reasoning step to arrive at the correct answer, challenging the assumption that more reasoning inherently leads to better reasoning.
How does problem difficulty affect the amount of redundant reasoning in LLMs?
While harder problems do demand more thinking than simpler ones, the study found that redundancy still hovers at substantial levels even for more complex problems. This indicates that even when tackling difficult tasks, LLMs maintain significant amounts of unnecessary reasoning steps beyond what is mathematically required to reach the correct solution.
Further Reading
- Less is Not Worse: Effective Reasoning Without Complete Reasoning Traces — OpenReview / ICLR 2026 submission
- Less is Not Worse: Effective Reasoning Without Complete Reasoning Chains — NeurIPS 2025
- Guiding LLM reasoning trajectories with Prefix consensus — arXiv
- Thought Anchors: Which LLM Reasoning Steps Matter? — NASA ADS / arXiv
- Evaluating Step-by-step Reasoning Traces: A Survey — ACL Anthology
