- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 1
- Soundness
- 2
Read review
Paper summary
The paper asks when a learned model that tracks state over a stream of events can be trusted. It separates what the data determines, what training learns and whether step-by-step execution stays correct. Its main new result shows that final answers alone can, in principle identify operations like “add” and “multiply” when both orders are observed. It also shows with known structure checking 75 transitions guarantees correctness on every sequence. On a synthetic task, per-key memory and intermediate-state labels improve learning and rounding learned transitions keeps accuracy at 1.0 up to length 16 while probabilistic execution collapses.
Critical strengths
1.Under the stated assumptions, the paper shows that primitive operations can be recovered using only final-output labels. Seeing compositions in one order leaves an internal ambiguity, while seeing both orders removes it for PatchWorld arithmetic (Thm. 7, Cor. 9). The stochastic extension gives a checkable condition for recovering the true operations through argmax projection (Thm. 8).
- Soft and projected rollouts use the same learned scores, so their difference in length-16 joint accuracy isolates the execution rule. The matched-information baselines receive the same intermediate labels and the paper reports non-significant differences openly, including for held-out composition. These controls strengthen the empirical evidence within the synthetic setting studied.
Critical weaknesses
-
The main theory is not demonstrated by terminal-only training. Theorems 7–8 and Corollary 9 explain when final-output labels can identify the true operations. However, no terminal-only experiment meets Theorem 8’s recovery condition. With the reset prior, terminal-only TOF recovers about 55 of 75 rows on average, models passing the full audit use intermediate labels. The paper correctly separates identifiability from successful learning, but whether its training procedure can reach the sufficient condition remains unclear. Reporting the minimum terminal target probability would directly address this gap.
-
The results depend on strong supplied information. Successful TOF models receive known semantic states, correct key routing, intermediate labels, and exact reset rules. The reset prior fixes 20 rows absent from training. The certificate checks the learned table against this reference, it does not establish that the supplied assumptions hold. Using identical scores for soft and projected rollouts isolates the execution rule, but Table 1b does not fully separate supervision, reset-prior and projection effects. Transition margins and a temperature-sharpened soft-rollout control would help show how robust the length-generalization advantage.
Questions for authors
-
For each seed in Table 1b, what are the audit count and minimum correct-terminal probability across all states and arithmetic pairs in both orders? Does any seed satisfy Theorem 8? If so, does its projected table pass the length tests? If not, how close is the best seed to the threshold?
-
For the Table 10 checkpoints, what are the smallest correct-transition and largest incorrect-transition probabilities? Do these explain the sharp accuracy drop from length three to four? How were checkpoints selected and soft distributions propagated?
-
Does temperature sharpening of the same kernels improve soft-rollout accuracy at lengths four and sixteen while preserving their argmax choices? This would help distinguish transition-confidence effects from the benefit of projection.