- Rating
- 3(+1)
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 3
Read review
Paper summary
The paper investigates when a transformer can solve a task without relying on the content of its chain of thought. It introduces a task-specific accuracy ceiling for shallow computation and shows, within its theoretical framework, that exceeding this ceiling requires sequential computation in the chain, regardless of how the model was trained. Experiments on synthetic tasks and mathematical benchmarks show that removing tokens or shuffle tokens in the chain reduces accuracy.
Critical strengths
-
A precise framework for when CoT is necessary. The task’s “ceiling” provides a quantitative criterion for the accuracy achievable through shallow computation.
-
Results that do not depend on a particular training method. Within its stated computational assumptions, the framework applies to any trained transformer. The necessity, depth, and locality results connect the importance of chain content, the computation needed to generate it, and the extraction of the final answer.
Critical weaknesses
The empirical interpretation is stronger than the evidence supports. The theory adresses what shallow computation can achieve, but the experiments measure what particular trained models achieve after their chain of thoughts are modified. These quantities need not coincide: a model may rely on its chain even when the task can be solved without one. Therefore, the accuracy drops on realistic benchmarks are consistent with the framework but do not establish that the task’s computational ceiling causes those drops. This distinction should be reflected more clearly in the main discussion, particularly when describing the experiments as matching theoretical predictions. Besides, a dedicated related work section is missing.
Questions for authors
None.
3 comments
Response to Reviewer 78QL (2/2): Changes in the revised paper, and related work
2. How the revised paper describes the experiments.
The revised paper reflects the distinction in the main text. At the start of Section 6 it states what the theory bounds and what each experiment measures, and Table 5 lists the statement that each experiment tests. It says that the measured accuracy matches the predicted value on the word problems, where the ceiling is known, and it reports the new experiment there. It describes the drops on MATH-500 and AIME as consistent with the theory, which is consistent with what the reviewer suggests.
The benchmark experiments are in the paper for what they establish without the value of the ceiling. The word problems are the controlled test, and the benchmarks show that the same behavior appears in the models that people use, which are pretrained reasoning LLMs with 1.5B to 32B parameters, including 19 checkpoints that we trained with GRPO. Tokens without content do not replace the chain, since filler, random tokens, and a template give the same accuracy as an empty chain within 0.11 (Figure 4c). The answer head reads the chain and does not continue the reasoning, since accuracy after truncation rises in step with how often the kept part already states the answer (Figure 4b). Training does not change this, since the checkpoints trained with a correct or a random reward behave alike (Figure 4d). The benchmarks also give a measurement. The accuracy that one pass reaches, 0.13 to 0.33 on MATH-500, is a lower bound on the ceiling.
3. Related work.
The revised paper has a dedicated related work section in the main text (Section 7). It places the paper against four lines of work. The first is the expressive power of transformers with and without a chain (Merrill and Sabharwal, 2023, 2024; Li et al., 2024; Liu et al., 2023; Amiri et al., 2025). These results are worst-case statements about what a transformer can compute, and we add the average-case ceiling and bounds for a trained model whose chain is modified. The second is intervention studies of faithfulness (Lanham et al., 2023; Turpin et al., 2023; Chen et al., 2025; Bentham et al., 2024; Chen et al., 2026; Ballon et al., 2026). These studies measure what happens when a chain is erased, truncated, or shuffled, and our theorems say how such measurements can be read. The third is methods that shorten or replace the chain. These are System One models (Almeida, 2026), filler and pause tokens (Pfau et al., 2024; Goyal et al., 2024), latent reasoning (Deng et al., 2024; Hao et al., 2025; Geiping et al., 2025), diffusion decoding (Nie et al., 2025), and chain compression (Xu et al., 2025). The ceiling bounds each of them when its number of steps is fixed (Appendix A). The fourth is the average-case complexity of group products (Barrington, 1989; Miles and Viola, 2013), which gives the ceilings of the tasks in our experiments.
The revised paper also states the computational model and the main definitions before the first claims that use them, and it uses the word problem of A₅ as a running example, which should reduce the effort that the first sections required. We hope that these answers resolve the reviewer's concerns, and we are glad to clarify any remaining point.
Additional References:
Bentham, Stringham, and Marasović. Chain-of-Thought Unfaithfulness as Disguised Accuracy. TMLR, 2024. Chen, Liao, Shiu, and Lee. Rethinking Dense Sequential Chains. arXiv 2605.07307, 2026.
Ballon, Verbeken, Ginis, and Algaba. Probing the Trajectories of Reasoning Traces in Large Language Models. NeurIPS 2026 (arXiv 2601.23163).
Dear Reviewer,
As the discussion period comes to a close, we wanted to make sure we have addressed all of your concerns. In our response we separated what the theory bounds from what each experiment measures, and the revised paper now describes the results on MATH-500 and AIME as consistent with the theory. It also has a dedicated related work section in the main paper. If any points remain unclear, or you would like us to expand on anything, please let us know. We would be glad to respond before the deadline.
Thank you for your time and for the feedback that has helped us improve the paper.
Best regards,
The Authors
Response to Reviewer 78QL (1/2): What the theory bounds and what the experiments measure
We thank the reviewer for the careful reading, and we are glad that the reviewer found the framework precise and its independence from the training method useful.
1. What the theory bounds and what the experiments measure.
We thank the reviewer for stating this distinction clearly, and we agree with it. The ceiling is what the best shallow computation achieves on a task. An erasing experiment measures what one trained model achieves without its chain. Theorem 3.1 says that the second quantity is at most the first. A model can rely on its chain where the task does not require it, and the models in the paper did. They were trained only with the chain, and their answer head copies the last token of the chain, so they fall to the baseline even on groups whose ceiling is 1.
We therefore ran a new experiment that removes this dependence wherever it is only a habit, so that the two quantities can be compared directly. In half of the training examples we replace the chain with filler, in the same format as the erasure at test time, and we train the model to give the correct answer there. These models see the erased chain throughout training, so training pushes them to answer without the chain wherever one pass can compute the answer. We trained them from scratch with supervision, with four layers, on four groups, at five lengths from n = 8 to 128, and with three seeds. Each model trains for twice as many steps as the models in the paper. Before the main runs we fixed in writing the outcome that would contradict Theorem 3.1, which is an accuracy above 0.03 without the chain on A₅ at n = 64 or 128. It did not occur. The table reports the two longest inputs, n = 64 and 128. Ranges are over seeds.
The ceiling is the best accuracy that a shallow circuit of fixed depth and size reaches as the input grows. The counting value is the accuracy of a circuit that computes the commutative part of the product by counting and guesses the rest (Proposition D.2 and Table 3 of the revised paper). The accuracy without the chain is the better of the two predictions that a model makes in one pass, after the filler and at the last input token. The table covers the 22 of 24 models that learned the task with the chain, two of them after a longer run.
We evaluate each model on 10,000 fresh instances, so the standard error at 1/60 is 0.0013. The values slightly above 0.0167 in the last column are within this error. The largest, 0.0197, is 2.4 standard errors above 1/60 and inside the band of three standard errors that we fixed before the runs. The 12 one-pass predictions of the six A₅ models average 0.0163.
Where the ceiling is below 1, the theorem forces the drop, and the new models show it although they were trained against it. A₅ is maximally serial, and its ceiling is the guessing baseline 1/60. Every seed still falls from 1.000 to between 0.016 and 0.020. S₅ is average-case serial, since one pass can compute the sign of the product and nothing more, and no model exceeds its ceiling of 2/120 beyond sampling error. Where the ceiling is 1, the theorem forces no drop. Z₆₀ is commutative, one pass computes its product by counting, and the new models lose nothing. The drop on this group in the paper therefore reflected the trained model and not the task, as the reviewer describes. A₄ × Z₅ is solvable and not commutative. Counting gives the commutative part of its product, which narrows the answer to 4 candidates, and the models reach this value of 1/4.
Two controls show that the new training does teach the models to answer without the chain. On short inputs of A₅ one pass can still compute the product, and the accuracy without the chain is 1.000 at n = 8 in every seed. The length up to which one pass suffices also grows with the depth. With two layers every seed is at the baseline from n = 16, and with four layers from n = 64. With eight layers two of three seeds still answer without the chain at n = 128, with accuracy 0.995 and 0.999. This is what the theory says, since the ceiling of 1/60 is the limit for long inputs at a fixed depth.
On A₅ the link between the ceiling and the drop is therefore not an inference from one trained model. The theorem says that every model must fall to 1/60 on long inputs, and models that are trained to avoid the drop still fall to it.