- Rating
- 1
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper studies how operations learned from source examples can be combined to solve new tasks with little target data. Circuit-TR assumes the correct composition is known, while Circuit-AD learns it from target examples. The theory separates errors in learned operations from errors in choosing their composition. Neural-AD provides a practical search method. Experiments test synthetic computations, GCD, electricity-data aggregates, and supplied compositions of operations extracted from language models.
Critical strengths
-
The paper separates errors in learned operations from errors in choosing how to combine them. It also shows why operations that produce the same answer without noise can behave differently with noise.
-
Testing every possible circuit in small settings provides a strong reference for the search method. Comparing learning with and without intermediate steps helps explain when learning fails.
Critical weaknesses
-
Very Limited novelty over earlier methods. Earlier work like Modular meta-learning and Neural Relational Inference with Fast Modular Meta-learning already learns reusable modules and searches for ways to combine them, including learned search proposals. The theory brings these ideas together, but the neural guarantee assumes good search results without showing that the method achieves them.
-
The comparisons leave the method’s benefit unclear. The paper lacks a comparison with earlier search methods using the same learned operations and computing budget. Several electricity baselines perform worse than uniform guessing under log loss. The method also struggles with a nine-position target despite an available accurate solution.
Questions for authors
1.How does Neural-AD compare with earlier modular search methods using the same frozen operations, objective, and computing budget?
2.In the small enumerable settings, what are the actual optimization residual and Jensen gap in Theorem C.5 for the mixture predictor?
3.Why do several electricity baselines perform worse than uniform prediction under log loss? Does improving their calibration or optimization change the conclusions?
We appreciate Reviewer i6rm’s suggestions and feedback. We invite them to view the updated manuscript, as we address their concerns below.
Reviewer i6rm: “Very Limited novelty over earlier methods.” The Reviewer points to modular meta-learning and its use of learned search proposals.
We made this connection explicitly in the submission.
Our question is what makes two sequence tasks related enough that pretraining on one can make the other learnable from few examples. Tasks with different answers can share the mechanisms that compute individual steps. We show when next-value prediction identifies those operations and how many new examples suffice to approach the predictive risk of an oracle given their arrangement. The transformer gives us one way to test that claim, and it does not refute possibility of other algorithms, generic or specialized alike, to demonstrate this capacity. A central message of the paper is precisely in this regards: adaptability from task A to task B is a quality of the data-generation mechanisms underlying the task A and B, not the quality of a particular learning algorithm. We show, as a proof-of-concept, it is plausible that optimizing a next-token prediction objective on task A allows a structure-agnostic model, such as the one introduced in this paper, to adapt faster to a domain B, as a mere side-effect of the training objective. even when the true intermediate execution trace that hints at the true composition is hidden to the learner.
Reviewer i6rm: “How does Neural-AD compare with earlier modular search methods using the same frozen operations, objective, and computing budget?”
Before responding, we would like to emphasize, once again, that the main message of the paper is not about a particular architectural choice and training procedure. We believe it would be clear from the exposition, but your feedback indicated that we need to make the distinction more clear to avoid misconceptions like this. With the same frozen gates, starting circuits and trace scores, modular annealing and our learned proposal selected identical circuits from 16 traces on each of two tasks. Whole-trace accuracy was 96.8% on the first task and 96.9% on the second. Annealing used at most as many nominal gate evaluations (revised Appendix K.5). With 1,024 pretraining pairs per task and 32 new examples, BounceGrad and Fast Modular learned libraries that gave 79.3% and 61.5% trace accuracy, ahead of Neural-AD’s 42.7%.
Reviewer i6rm: “[...] what are the actual optimization residual and Jensen gap in Theorem C.5 for the mixture predictor?”
We did not measure either quantity for the submitted result. Submitted Table 6 reports the fitting gap of one decoded circuit. The revised version returns a single circuit, for which the Jensen gap is zero by construction. With full traces, both searches attained the exact empirical minimum in all 120 unary fits (revised Appendix K.13).
Reviewer i6rm: “The method also struggles with a nine-position target despite an available accurate solution.”
Correct. Through 500 answers the penalized objective favored the one-gate circuit. At 1,000 and 2,000 answers it favored the supplied circuit, and search missed it (submitted lines 2725–2730). In a separate revised experiment on three-input range, changing the annealing schedule at the same proposal budget raised 16-answer accuracy from 59.0% to 99.9% (Appendix K.4).
Reviewer i6rm: “Why do several electricity baselines perform worse than uniform prediction under log loss? Does improving their calibration or optimization change the conclusions?”
The learner must transfer arithmetic from synthetic traces to real electricity readings. On largest adjacent change, a fresh baseline trained on 100 examples assigned the true answer probability below 0.00011 on 81% of held-out windows. Uniform assigns it 0.01. Averaging each of three validation-selected predictors with uniform brought their losses to 4.39–4.50 nats, below uniform’s 4.61. On all 20,000 test windows, a smoothed answer histogram scored 4.34 nats. The circuit learner scored 0.067 (submitted Table 10). Better calibration reduces the margin. The learned operations still transfer.