- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper proposes Abstract Token Curriculum (ATC), which trains continuous latent thoughts using answer supervision without explicitly supervising intermediate reasoning steps. ATC jointly increases task difficulty and the number of latent thoughts, with additional backtracking and truncated backpropagation for longer reasoning chains. Experiments on parity, graph reachability, and multi-digit addition demonstrate its effectiveness, alongside a theoretical analysis of attention behavior.
Critical strengths
-
ATC achieves 99.9% accuracy on graph reachability using only answer supervision, while removing either the curriculum or continuous thoughts reduces accuracy to approximately 50%. These results provide evidence that combining curriculum learning with latent computation is effective in this setting.
-
On graph reachability, truncated backpropagation alone achieves 51.0% accuracy, while adding curriculum backtracking restores it to 97.4%. This also reduces peak memory from 5.78 GiB to 1.55 GiB and training step time from 198.6 ms to 124.7 ms.
Critical weaknesses
-
In the eight-digit addition experiment, ATC progressively increases the number of input digits, while Coconut receives the full eight-digit problem throughout training and predicts intermediate partial sums. Therefore, the reported 98.7% versus 0% comparison involves substantially different curriculum designs, not just different supervision strategies. Similarly, on graph reachability, ATC advances based on answer accuracy while Coconut uses intermediate-node prediction accuracy. Without comparisons under more closely matched curricula and computational budgets, it is difficult to determine how much of the improvement comes from the proposed training method.
-
Theorem 1 assumes that the preceding thought already perfectly encodes the correct prefix parity and analyzes only a single gradient update on the key parameters of a one-layer attention model. It does not establish how such representations emerge from answer-only supervision under the actual training procedure. Moreover, the empirical attention analysis measures approximately 50% attention on the last thought, which is also the readout position itself. This does not directly validate the claimed attention mechanism.
-
The main text ends immediately after the attention analysis, without a discussion or conclusion, while the conclusion is placed at the end of the appendix. Important limitations of the theoretical assumptions and baseline comparisons are not adequately discussed in the main text. There are also reporting and presentation inconsistencies: Section 4 states that the experiments use 20 seeds, while Appendix I.3 reports 10 seeds for addition; some figure captions are minimal, and several cross-references in the appendix are incorrect. These issues make the paper's technical contributions harder to assess.
Questions for authors
-
Can the authors compare ATC and Coconut under the same input-difficulty curriculum, with matched training budgets? In particular, how much of the addition performance gap is caused by ATC's progressively shortened input problems rather than the difference in intermediate supervision?
-
How does the correlation-one assumption in Theorem 1 relate to the representations actually learned by ATC? Can the authors directly verify that intermediate thoughts encode the expected partial computations and clarify how the attention analysis distinguishes dependence on the previous thought from self-attention at the readout position?
-
The experiments use task-specific difficulty schedules based on prefix parity, graph distance, and the number of input digits. How much task-specific knowledge is required to construct these curricula, and what evidence supports applying ATC to tasks without such explicitly available decompositions?
-
Can the authors clarify the inconsistent number of random seeds reported for the addition experiments and provide a consistent description of the evaluation settings?
6 comments
Author response-1 (thread-2)
(Q2): How does the correlation-one assumption in Theorem 1 relate to the representations actually learned by ATC? Can the authors directly verify that intermediate thoughts encode the expected partial computations and clarify how the attention analysis distinguishes dependence on the previous thought from self-attention at the readout position?
(Answer): We perform several intervention experiments (Section-4 and Figure-5) to highlight this assumption. In graph experiments, we show that just corrupting the intermediate last stage thought can degrade the answer, while preserving it, does not change the answer. So this assumption is highly consistent with our experiments (Figure-5).
In addition to the evidence already in the paper, we ran further probing experiments on the thoughts themselves (https://anonymous.4open.science/r/iclr_27_rebuttal-94D6/ATC_interpretable_thoughts.pdf). For graph tasks, ATC’s thoughts encode the depth-by-depth progression of breadth-first search. https://anonymous.4open.science/r/iclr_27_rebuttal-94D6/token_node.pdf . For addition, an MLP probe accurately decodes the true carry bit at each digit position from the hidden state of the corresponding ATC thought https://anonymous.4open.science/r/iclr_27_rebuttal-94D6/addtion_probing.png
Moreover, we consistently observe that the attention concentration shown in the theorem,which is reflected in the graph experiments. For ATC, 63% of the attention is on the last thought, and for Coconut it is 36%, in both cases above the earlier thoughts and the rest of the prompt (https://anonymous.4open.science/r/iclr_27_rebuttal-94D6/attention_conc.png). Together, the interventions and the attention show that the model relies on the last thought.
(Q3): The experiments use task-specific difficulty schedules based on prefix parity, graph distance, and the number of input digits. How much task-specific knowledge is required to construct these curricula, and what evidence supports applying ATC to tasks without such explicitly available decompositions?
(Answer:) We believe that coming up with a curriculum for a task is substantially easier than designing intermediate supervision or identifying what kind of intermediate state is useful for the model, which ATC avoids. For example providing the BFS tree as the intermediate supervision might be difficult if someone is not familiar with BFS algorithm for solving graph connectivity while distance form source to node is a simple heuristic metric we use for defining the curriculum in this case, which encourages the model to solve the problem by saving the BFS tree levels in intermediate states.
However, one can be fully unaware of BFS algorithm, and the point of ATC is the model grows the useful relevant intermediate representations itself in the hidden thoughts without explicit supervision. Similarly in addition, one can be unaware of the addition algorithm and simply define the curriculum to be the number of digits of the input, which we use for ATC.
So the point of ATC is exactly that the model designer can come up with simple heuristic difficulty measures to define the curriculum, and leave the formation of useful intermediate representations to the model itself. We see that this paradigm is highly successful and importantly avoids the need of direct intermediate supervision for every specific tasks.
(Q4): Can the authors clarify the inconsistent number of random seeds reported for the addition experiments and provide a consistent description of the evaluation settings?
(Answer): Please see response to W3.
If these clarifications and new results address your concerns, we would appreciate it if you would reconsider your score.
Thank you for the detailed clarifications and additional probing experiments, which have addressed some of my concerns. However, my main concern about the fairness of the comparison with Coconut remains, as the different curriculum designs make it difficult to isolate the source of the performance gains. I would be happy to reconsider my score if this concern can be further addressed.
Author response-1 (thread-1)
We thank the reviewer for their time to review and the detailed response. Please see our response below
(W1): In the eight-digit addition experiment, ATC progressively increases the number of input digits, while Coconut receives the full eight-digit problem throughout training and predicts intermediate partial sums.
(Answer): The difference in prompts is exactly the point of the comparison. We highlight that using curriculum with input prompt being the intermediate difficult question is the reason for ATC’ success. The original Coconut method uses the full prompt to an intermediate stage question and changing this would make the baseline a “Coconut” method. We believe the rest of the evaluation is fair in comparison.
(W2): Theorem 1 assumes that the preceding thought already perfectly encodes the correct prefix parity and analyzes only a single gradient update on the key parameters of a one-layer attention model. It does not establish how such representations emerge from answer-only supervision under the actual training procedure. ..
(Answer): Proving how the whole representation emerges is out of scope of the theorem. Our theorem proves that once the last thought encodes the parity of the earlier bits, a single gradient step on the attention keys can direct attention to that thought.
We find sufficient empirical evidence to believe that this step is the key to learning chain of thought. For example, in the graph experiments, the last intermediate thought encodes the answer information and corrupting that thought drastically drops the answer accuracy.
(W3): The main text ends immediately after the attention analysis, without a discussion or conclusion, while the conclusion is placed at the end of the appendix. Important limitations of the theoretical assumptions and baseline comparisons are not adequately discussed in the main text. There are also reporting and presentation inconsistencies: Section 4 states that the experiments use 20 seeds, while Appendix I.3 reports 10 seeds for addition; some figure captions are minimal, and several cross-references in the appendix are incorrect. These issues make the paper's technical contributions harder to assess
(Answer): Thank you for this suggestion and we have included a conclusion in the manuscript since the page limit is now increased to 10. We will highlight the scope of the theorem more clearly in our answer. The seeds number was a typo and we fixed it to 10 and revise the figure captions as well. To clarify, graph experiments used 20 seeds and the addition experiments used 10 seeds. We will make the following changes to give a consistent description of the evaluation settings:
(Q1): Can the authors compare ATC and Coconut under the same input-difficulty curriculum, with matched training budgets? In particular, how much of the addition performance gap is caused by ATC's progressively shortened input problems rather than the difference in intermediate supervision?
(Answer:) The addition experiments already match the training budget in epochs. Both methods use the same 400-epoch, model architecture, optimizer settings, and batch size (Appendix I.3).
The difference in prompts is exactly the point of the comparison. ATC presents shorter addition problems at early stages and gradually increases their length. Our Coconut-style baseline keeps the full eight-digit problem in the prompt at every stage and changes the partial sum it is trained to predict. Giving Coconut the same prompts as ATC would change its curriculum and yield a modified baseline. The reported accuracy gap compares the two curricula as designed.
(continued)