- Rating
- 3(+1)
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper aims at understanding that in continuous chain of thought, how the current continuous token and the previous KV cache help reasoning respectively. First, by careful intervention on a already trained model, this paper shows that the previous KV cache history is still important for correct generation in continuous chain of thought reasoning. Next, this paper shows a expressiveness result: on graph reachability task, there exists a Transformer that use both continuous token and previous KV cache to find the correct answer. Empirical result on this task show the some consistent features between the constructed model and real trained model.
Critical strengths
- This paper analyzes an interesting problem in continuous chain of thought. By careful controlled experiments, this paper aims at showing the role of continuous token itself and previous KV cache in continuous CoT.
Critical weaknesses
My soundness concern is mainly about the theoretical results comparing to [1] as both are considering the same task.
- The theoretical result is not as strong as that in [1]. In Theorem 1 of [1], the number of required continuous tokens is , where is the diameter of the graph. However, in the Theorem 1 of this paper, the number of required continuous tokens scales as where is the number of the nodes. It is counter-intuitive that after using KV cache in construction, the required tokens increases.
[1] Reasoning by Superposition: A Theoretical Perspective on Chain of Continuous Thought. https://arxiv.org/abs/2505.12514
Questions for authors
Please refer to the weaknesses part.
- Could the authors explain if continuous tokens is necessary in this construction, or can we improve it to continuous tokens, where is the diameter of the graph?
3 comments
Thanks for the response. I would increase my score to 3 to support acceptance.
We would like to thank you again for your insightful feedback, which indeed helped us strengthen the paper's theoretical results. We also appreciate your support for acceptance, and we would be happy to address any further questions or discuss any remaining concerns.
Response to Reviewer bG6b
We sincerely thank you for your thoughtful feedback. Our responses follow.
Response. Thank you for raising this important point. We agree that the token bound in Theorem 1 can be sharpened to depend on the graph diameter. The original N−1 bound is a worst-case sufficient budget over graphs with N nodes, rather than a requirement for every graph.
Specifically, each continuous step performs one level of BFS expansion: the feedback represents the current frontier Fc, while the current and cached frontier records collectively recover the visited set (Rc=⋃j=0cFj). The next frontier is therefore Fc+1=N(Fc)∖Rc. After C continuous steps, answer readout succeeds whenever the reachable candidate belongs to RC, equivalently whenever C≥dist(r,c⋆). Thus, D continuous tokens are enough to obtain a correct answer, where D is the maximum shortest-path distance between any reachable pair of nodes.
Accordingly, using cached history does not increase the required number of continuous tokens in our construction: it achieves the same diameter-dependent token bound as [1]. The distinction is how information is stored: [1] maintains the cumulative reachable set in each continuous thought, whereas our construction represents the current frontier in feedback and retains earlier frontiers in the KV cache.
For your convenience, we display our revised Theorem 1 below:
Let zc be the input at position ic=3M+6+c, and let iA=3M+C+7 be the answer position.
Theorem (frontier-history). We consider integer budgets C satisfying C≥dist(r,c⋆) and 3M+C+7≤Pmax. For any fixed N, Mmax, and Pmax, there exists a two-block Transformer with the architecture above and parameters depending only on these quantities. (In particular, D continuous tokens suffice for correct answer readout, and any larger budget also yields the correct answer, subject to the same context bound.)
For every legal prompt and every budget C satisfying the conditions above, initializing z0=er and running C additional feedback steps before ⟨A⟩ gives:
i. Frontier updates. For 0≤c≤C, hic(2)=ϕ(Fc+1). Feeding each output into the next step gives zc=ϕ(Fc) for 1≤c≤C.
ii. Visited-set recovery. For 0≤c≤C, Block 1 retains zc in frontier coordinates; a uniform causal head in Block 2 returns bc∈RN with supp(bc)=Rc, where supp denotes the nonzero node coordinates.
iii. Answer readout. The answer output is hiA(2)=ec⋆, giving the reachable candidate a unit logit margin over every other node.
We have revised the paper to make this guarantee explicit: