Novelty:
Existing work already covers many of the ideas in this paper. For example, Macfarlane & Bonnet "Searching Latent Program Spaces" do gradient search during TTT on small task-like embeddings. There are many other related works such as Zintgraf et al. "Fast Context Adaptation via Meta-Learning" also updates only small context parameters via meta-learning, and Autodecoders such as DeepSDF by JJ Park et al. freeze the model during inference and only fine-tune a small task vector. Also, in other applications such as prompt tuning in general it is similar to freezing the model and fine-tuning only one task vector. The related work does not position the paper well in the context of this existing literature. It should be clearer what was already there before and what the novelty is, which is probably mainly the task-specific evaluation on Arc.
Experiments and Drawn Conclusions:
While the numbers themselves look correct, the interpretation of them needs to be more careful. For instance, Table 1 shows that freezing the backbone and only fine-tuning the task embeddings makes those embeddings more aligned with the training distribution (which is to some extent expected). However, the used baseline is Full-TTT and this one trains both, the task embeddings and the backbone. Thus, the embeddings only carry meaning for its also fine-tuned backbone. Therefore, the comparison to the original training embeddings is expected to fail and yield the wrong measure, as an updated backbone likely shifts the distribution of the task embeddings. The authors mention this issue briefly and also report results where the probe is fitted on duplicated training tasks of the same Full-TTT protocol instead of the original training embeddings obtained after initial training. This comparison is more meaningful and should be shown in the main table (Table 1) and should be discussed. Focusing primarily on the first experiments as the main result is misleading, and statements like "loses any semantically meaningful information" are overclaimed in a general sense.
Also, it remains unclear how general these findings are since only a single model is used. Also, the claim that good ARC performance conflates rule induction and rule execution is only measured indirectly on ARC (Section 3.2 says it could be memorization) while the appendix provides much stronger evidence based on synthetic data.
Alternative Explanations:
Some of the shown interpretations might be correct, but there are also alternative solutions that are not ruled out. Specifically, in Section 4.4, combined moves are compared with single-step movements (unit move). The comparison is done with cosine similarity, which only considers the direction from the origin, not the magnitude of these moves. To provide some intuition: the moves "up 1", "up 2", and "up 3" could be arranged on a perfect line with same distance between them but the line is not passing through the origin. That would mean that they represent distance well but cosine similarity could still be low. In such a case, R3-U2 is expected to be more similar to R3-U0 than to R1-U0 as it was observed in the paper. Thus, it cannot be concluded that the model cannot handle scaling.
Another example is the compositionality in Appendix F: The network is trained to move and change seperately and during inference it should compose these two actions. The network fails on this compositionality task, even though the embedding is close to the two relevant tasks. The authors interpret this phenomen that the correct task embedding was found but the model failed to execute it. A much simpler alternative explanation is that this is just a side effect of the loss and the embedding drifts kind of towards the middle of the two relevant embeddings without any understanding of it. A simple analogy just for understanding: We could have a color yellow and color blue learned during training and if both are done together, it drifts somewhere in the middle of these two embeddings representing green. But this is just the closest position in embedding space between the two, it does not mean we understood the concept of colors nor of green. These alternative explanations could simple be ruled out with control tasks (e.g. use two training tasks where the target overlaps pixel-wise but where it is not a composition of the rules -> if the embeddings are still between the two, they reflect an overlap not rule induction).
Presentation & General comments:
- The structure is not optimal and methods, motivation, and results are blurred. E.g., the problem statement is defined inside related work, Section 3 refers to preliminary results as motivation but the details of it remain unclear
- At some places, equations could make things a bit clearer, e.g., for the objective and the metrics
- For concept transfer: Since it is up to 3 concepts, chance is up to 3/16=19% and not 1/16=6.3%
- Moves should also have chance level reported. Why did you not do it for this method but all the others? It is very hard to interpret what an RMSE of 2.x means.
- Table 1: It could be interesting to have a very simple baseline without TTT such as using a linear probe with the average input-output distance to identify the rule -> such a simple method could already work on the synthetic dataset and make the 100% probe accuracy less informative
- The results seem to rest on a single model run once with a single seed. Since the gain is relatively small (about 3 tasks at pass@1), we need confidence intervals.
- It is not clear whether the compute is matched between the runs or if the two-stage run is just trained longer
- Description of training details, hyperparameters, etc. are missing and should be in the appendix
Answers to weaknesses and questions
We thank the reviewer for the overall positive and supporting review. We reply to weaknesses and questions hereafter.
Weakness:
"We first evaluate how well test-time embeddings recover known rules in Section 4.1. We then examine the training tasks retrieved for ARC-1 test tasks to show the embeddings capture meaningful similarities in Section 4.2. Next, we compare performance across datasets in Section 4.3. Finally, in Section 4.4, we study the geometry of the embedding space to see where task generalization succeeds."
We also revised the opening of Section 4.1 to explain the retrieval and probing evaluations before introducing the datasets and their splits:
"Datasets and experimental details. We use two types of evaluation, according to Section 3.3. On ARC-1 and ConceptARC, we test whether an embedding fitted to a duplicate of a training task retrieves that task's original embedding. On ConceptARC, LUCD, and Moves, we use linear probes to measure whether the embeddings encode concept labels, rule labels, or displacement coordinates. The dataset descriptions below specify how we construct the training and evaluation sets."
Questions: