acceptodds
Withdrawn from ICLR 2027

Learning to Adapt Without Gradients: Emergence of Mental Representations via Second-Order Learning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Mental representation, characterized by structured internal models that mirror external environments, is fundamental to advanced cognition but remains challenging to investigate empirically. Existing theory suggests that second-order learning mechanisms that adapt first-order learning, i.e., learning about the underlying task or domain, can promote the emergence of environment–cognition isomorphism. In this work, we empirically validate this hypothesis using a hierarchical framework with a Graph Convolutional Network (GCN) as a first-order learner and an MLP-based controller as a second-order learner. The controller dynamically adapts the GCN’s parameters in response to structurally novel environments, enabling fast adaptation without gradient-based optimization. Across synthetic maze tasks and real-world road network graphs, we show that second-order learning improves generalization under distributional shift. Notably, the approach achieves performance comparable to gradient-based meta-learning methods such as MAML, while requiring no test-time gradients or labeled data and performing adaptation in a single forward pass. We further show that effective adaptation is linked to the emergence of structured latent representations that are isomorphic to the environment; when this isomorphism is disrupted, performance degrades significantly. These results provide empirical evidence for the role of structured representations in second-order learning and establish a practical alternative to gradient-based meta-learning for graph-structured environments.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
meta-learning, graph neural networks, fast adaptation, out-of-distribution generalization, structured representations

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_Y34KOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary
  • summary/contribution: The paper proposes "second-order learning" (SOL) for adaptive reasoning in graph-structured environments: an MLP controller maps the outputs and parameters of a frozen Graph Convolutional Network (GCN) to a parameter update in one forward pass, with no labels or gradients at test time.
  • validation:
    • The task is to label shortest-path nodes in perturbed 10×10 grid mazes and in an OpenStreetMap road graph with closures.
    • Reported results: accuracy goes from 62.5% (unadapted GCN) to 80.2% and SOL is on par with MAML on an OOD split. SOL beats MAML on the road graph.
  • summary: The paper also argues that GCN embedding distances correlate with maze distances (r ≈ 0.90), and reads this as support for the hypothesis that second-order learning produces representations isomorphic to the environment.
Critical strengths
  1. The question (amortized, label-free adaptation as an alternative to MAML) is relevant for meta-learning, and the paper reports adaptation cost, several seeds, and code.
Critical weaknesses
  1. The accuracies do not beat a majority-class baseline.
    • On the road graph about 8% of nodes are positive (the Vanilla GCN has recall 1.0 and precision 0.082, Table 10). Predicting all negative gives about 92%, above SOL's 75.7%. SOL is also below MAML on recall, F1 and IoU.
    • Shortest-path labels are not unique (hundreds to thousands of equally short paths at low p), so node accuracy is not well defined.
  2. The isomorphism analysis cannot support the causal claim.
    • It is computed on the base GCN, which is frozen while the controller is trained (App. C.1, E.1), so second-order learning cannot have shaped it.
    • The node features include the (x, y) coordinates and the reference is coordinate distance, so a high correlation is expected. Replacing the features with noise (Exp. 5) also removes task information.
  3. On an empirical level, the contribution is the initiated methodology, since the parameter choices in the main content are manual, e.g., Empirical study 1 and 2. On a theoretical level, the paper refers to this alignment as an "isomorphism" between the environment and cognition; however, the correlation metrics themselves do not constitute a formal proof of graph isomorphism in the strict mathematical sense.
Questions for authors

Clarifications:

  1. Can you report F1, IoU, path validity and a majority-class baseline for all maze experiments?
  2. Does a GCN trained without a controller on the same perturbed mazes, or a deeper GNN, match SOL? A 2-layer GCN sees two hops, while the controller sees all nodes at once.
  3. Which nodes form the support set for MAML and Reptile at test time, and do they overlap with the evaluation nodes?
  4. What is the isomorphism score for the raw input features, a random GCN, and the adapted parameters θ′\theta'?
  5. Clear sections, unclear proof for the correctness of table 1.
  6. Figure 1 - Does SOL lead to the appearance of this type of tissue?

Suggestions (not affecting my score):

  • Relate the method to amortized meta-learning (LEO, CNAPs), label-free test-time adaptation (ARM, TENT) and neural algorithmic reasoning (Veličković et al., 2020).
Reviewer_qEHHOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a second-order / meta learning framework in which an MLP controller predicts parameter updates for a GCN, which computes the gradient / adaptation in a feedforward manner. Experiments on synthetic mazes, road-network graphs, and reward-value prediction tasks suggest that the controller can improve performance under distribution shift. The paper further argues that effective adaptation is associated with latent representations that preserve the structure of the environment.

Critical strengths
  • The framework is conceptually intuitive and offers an alternative to gradient-based test-time adaptation, without requiring test-time labels.
  • The analysis of latent representations provides a potentially interesting perspective on the relationship between adaptation and environmental structure.
Critical weaknesses
  • Insufficient evidence for the representation claims. It's unclear the training procedure actually allow the latent representation to be shaped by the second order learner (see question 1 below). Even assuming it does, current experimental results do not establish that second-order learning causes the representations to emerge. Since the model receives spatial coordinates as input, such alignment may arise from ordinary first-order training. Replacing spatial features with random noise also changes the information available to the model, therefore does not isolate the role of second-order learning.

  • Limited evidence for general applicability. The experiments primarily use small, synthetic maze tasks and a relatively shallow two-layer GCN. It remains unclear whether the controller’s benefits persist with stronger base models and more challenging graph-reasoning problems. The controller takes flattened node-level representations, outputs, and model parameters as input. Its input dimension therefore depends on the number of nodes, limiting its ability to generalize directly to graphs of varying sizes. The reported cross-size results evaluate the GCN’s representation, not the scalability of the complete adaptation framework.

Questions for authors
  1. The training procedure is not clearly explained in the main paper. Appendix A.5 seems to suggest that the training is staged: first train GCN on unperturbed maze, then freeze GCN and train the MLP on perturbed maze. Could the authors clarify if this the correct description of the training procedure? Does GCN ever get gradient from the training of second order learner? And if not, how can the second-order learner exert selection pressure on or shape the emergence of the first-order latent representations? Because gradients from the second-order controller never update the GCN, second-order learning cannot have caused the isomorphic representations in the first-order learner to emerge.

  2. What evidence isolates the contribution of second-order learning to the emergence of structured representations, beyond the effect of ordinary GCN training on spatially informative inputs? For example, if we train GCN without MLP, does the structured representation fade or disappear?

  3. Why is GCN limited to 2-layer network and does not iteratively propagate information across the graph until it has enough information to solve the problem? The first order learner's capacity seems to be curtailed to make space for the second order learner to be useful. Does SOL remain effective when the first-order learner is replaced with deeper or more capable graph neural networks? Can the authors show that its gains are not specific to the chosen GCN architecture?

  4. Could the authors report F1 score for all the experiments and clarify the percentage of the labels for the classification? It seems that the classification problem is quite imbalanced, therefore accuracy might not be a good performance metric. In table 2, SOL has the highest accuracy, however MAML has a lower BCE loss. Note that Table 4 has this similar pattern. If the labels are imbalanced, then MAML actually outperforms SOL.

  5. What is gained by using this particular MLP controller, conditioned on flattened GCN outputs and parameters, instead of an existing task-conditioned hypernetwork or learned update rule—and is that advantage robust when the underlying GNN and graph size change? Could the authors provide parameter-generation meta learning baselines which is closer to what this paper proposes?

Reviewer_3zo3OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This work tests if a GCN + MLP controller architecture can be trained to enable second order learning (learning to learn) without gradients or labels.

The results seems interesting, showing a 17% increase on a maze task, matching MAML.

Unfortunately the paper operated in a combination of a niche architecture x task x metric so it is very hard to understand what is going on.

The paper's presentation is really poor.

Critical strengths

The paper focuses on an important topic: continual learning and test time adaptation.

Critical weaknesses

CW1: Unfortunately, the presentation of the paper is very bad. While the paper grounds itself to a nice literature, it is very hard to understand how to position this paper's setting or finding within the literature. As CW2 says, if the task is niche the paper should do a better job at presenting its setup clearly, which it fails to do so.

CW2: Nicheness, the paper operates in a very niche setup across task, architecture, learning objective and metrics. Combined with CW1, it is very hard to judge the paper.

Questions for authors

Q1: Are there past papers operating on these same tasks?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.