acceptodds
Under review as a conference paper at ICLR 2027

Growing Prototypes on Demand: Scalable Prototype Learning for Causal Language Models

Abstract

Enabling causal language models to produce interpretable evidence remains challenging. Prototype-based reasoning offers “this looks/reads like that" interpretability, but compressing diverse contextual training evidence into reusable prototypes remains challenging at scale. We propose a non-parametric approach that frames prototype learning as clustering in a frozen model's hidden space using a Dirichlet Process Gaussian Mixture Model (DP-GMM), which allows the number of components to grow only when needed. To scale to millions of hidden states, we introduce a tree-structured DP-GMM that performs coarse-to-fine mixture fitting, with a separate prototype set and prediction head fitted for each task and training distribution on a frozen backbone. The expectations of leaf means are projected to training hidden states to obtain example-grounded prototypes, whose normalised distance similarities drive a learned prediction head. This design enables scalable DP-GMM fitting with node-local cost independent of global prototype capacity for fixed local sample size, hidden dimension, truncation, and fitting settings. Experiments across language modelling, captioning, and VQA fit approximately 2.3M-7.4M hidden states into 5.8k-7.6k prototypes, with small-to-moderate performance drops. Paired interventions show that retaining a small top-matched set largely preserves performance, while removing it severely degrades performance, which supports the computational faithfulness.

Primary area:
interpretability and explainable AI
Keywords:
Language Model, Prototype Learning, Interpretability

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_PHj3OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a scalable approach to example-grounded prototype-based reasoning for language and vision-language models. Hidden states from a frozen backbone are clustered with a tree-structured DP-GMM, which recursively fits smaller DP-GMMs. This lets the number of prototypes grow only as needed and keeps memory tractable for large datasets. A new prediction head, trained on a softmax over negative Euclidean distances to the resulting prototypes, replaces the original head. Experiments on language modeling, image captioning, and VQA show that the prototype head provides example-based interpretability with only a small-to-moderate drop in performance.

Critical strengths

Scalable prototype learning. The paper extends prototype-based reasoning to language and vision-language modeling domains. The method uses hierarchical non-parametric clustering to adapt to large dataset size and prototype count. Since only one prediction head is trained, it is cheap to apply, and the comparison to the original backbone isolates the cost of the interpretability bottleneck. The appendix also includes a memory-cost analysis, although the baseline assumes a naive implementation (see Weaknesses).

Broad evaluation with a modest performance cost. The method is evaluated on three tasks spanning language and image modalities, each with two backbones. The prototype head retains most of the original performance.

Critical weaknesses

Insufficient positioning and limited novelty. The tree-structured DP-GMM is motivated mainly by scalability, but the paper does not position itself relative to other work with similar design choices. There is an extensive literature both on scaling DP mixtures to millions of points [1–3] and on hierarchical clustering methods [4–5], which also split clusters recursively and use a local stopping criterion. None of these works are mentioned, and the paper does not explain why the tree-structured DP-GMM was chosen or what it offers over these alternatives. Since the clustering method is closely related to existing approaches, the main contribution seems to be applying scalable clustering to prototype learning, which on its own is incremental. As an aside, the memory comparison in Appendix C is pessimistic, since it materializes the full responsibility matrix, which existing methods can potentially avoid.

Experiments do not convincingly support the claims. Standard baselines are missing, such as k-means or hierarchical k-means, a flat DP-GMM, and the existing scalable nonparametric methods discussed above. Without these comparisons, it is unclear whether the tree-structured DP-GMM is necessary for scalable prototype learning in the first place.

The interpretability evidence is also weak. Since hidden states that predict the same token tend to lie close together, the matched prototype is typically just a training example with the same next token or answer (as seen in every example in Figure 1). It is unclear whether these explanations convey more than the prediction itself, and explanation quality is never evaluated quantitatively.

The comprehensiveness and sufficiency results in Table 5 are difficult to interpret without additional controls. The hidden space is high-dimensional, and a softmax over negative Euclidean distances without a temperature parameter is likely (close to) one-hot. The results may be due to the concentration after the softmax.

[1] - B. Kulis, M. Jordan. Revisiting k-means: New Algorithms via Bayesian Nonparametrics. ICML, 2012.

[2] - M. Hughes, E. Sudderth. Memoized Online Variational Inference for Dirichlet Process Mixture Models. NeurIPS, 2013.

[3] - M. Hughes, E. Sudderth. Fast Learning of Clusters and Topics via Sparse Posteriors. arXiv, 2016

[4] - D. Pelleg, A. Moore. X-means: Extending K-means with Efficient Estimation of the Number of Clusters. ICML, 2000.

[5] - G. Hamerly, C. Elkan. Learning the k in k-means. NeurIPS, 2003.

Questions for authors

See weaknesses. In particular, the authors should (1) clarify how the tree-structured DP-GMM differs from existing scalable nonparametric or hierarchical clustering methods and what makes it particularly scalable or well suited to prototype-based reasoning, and (2) include additional baselines and a more rigorous interpretability analysis where appropriate. Additionally,

  1. How were the hyperparameters of each DP-GMM chosen, and are they shared across tree depths? Since deeper nodes contain fewer points, the prior may dominate and cause early stopping.

  2. Can the authors add runtime comparisons in addition to the memory analysis?

Reviewer_ssJzOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The method collects hidden states from a frozen backbone (LLaMA 3.2 1B/3B, BLIP) and fits a tree of small truncated DP-GMMs. Each node has K = 10 components. A node splits into its active components until only one component is active or depth 5 is reached. Leaf means are snapped to the nearest training hidden state to form prototypes. The original output head is replaced by softmax(−Euclidean distance to prototypes) followed by a linear head over the vocabulary.

Critical strengths
  • The problem is interesting: prototype layers have not scaled to open-vocabulary generation, and this is a simple, backbone-agnostic way to try.

  • The divide-and-conquer fit greatly reduces memory, and Appendix C quantifies the reduction clearly.

Critical weaknesses
  • The scalability claim isn’t tested at scale. WikiText-103 is subsampled to 2.5M hidden states, about the same as WikiText-2, so the “grows with data” story never sees a larger N. There are no runtime results, even though “node-local cost independent of global capacity” is a headline claim.

  • The “on demand” growth may be an artifact of the settings. With K = 10, α = 0.1 and depth capped at 5, the count lands at about 6–7k in every setting. Depths 4 and 5 give identical counts (6,433 and 6,796), meaning every depth-4 node had exactly one active component. That is surprising for 2,048-dimensional data with thousands of points per node, and it needs explaining. Is it full covariance, how is the NIW prior set, are clusters collapsing? It reads more like a property of the truncation, prior, and hard assignment than evidence the method “saturates where the data does.”

  • The interventions are close to tautological. A softmax over raw Euclidean distances in 2,048 dimensions with no temperature is nearly one-hot. Removing the top-k entries and renormalizing will wreck the prediction, and keeping them will preserve it, for any prototype set. Showing faithfulness needs controls: random or k-means prototypes with the same M under the same interventions, and the entropy of the activation distribution.

  • No baseline tests whether the DP-GMM matters. The only comparison is the original head. Missing comparisons:

  1. k-means with the same M,

  2. M randomly sampled training hidden states,

  3. kNN-LM and SNE, which Appendix D argues against but which are at least needed to judge how good the explanations are. Without these, the contribution of the tree DP-GMM itself is unshown.

  • Interpretability is claimed, not measured. The linear head can put arbitrary signed weights on about 6k activations. The figure examples look hand-picked, and there is no comparison with simple nearest-neighbor explanations of the original model.

  • Prior work on tree-structured nonparametric clustering is missing. Tree-structured stick-breaking (Adams et al. 2010), nested DPs, Bayesian hierarchical clustering and divisive mixture trees go uncited, so the novelty of the tree DP-GMM is overstated.

Questions for authors

N/A

Reviewer_T2RXOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a scalable prototype-based prediction framework for causal language and vision-language models. For each task, a tree-structured Dirichlet Process Gaussian Mixture Model clusters training hidden states from a frozen backbone, and cluster means are projected onto actual training hidden states to obtain example-grounded prototypes. A learned prediction head maps normalized prototype similarities to token probabilities. Experiments on language modeling, image captioning, and visual question answering show that thousands of prototypes can represent millions of training hidden states, with small-to-moderate performance degradation relative to the same backbones with their original prediction heads. Prototype removal and retention experiments assess dependence on top-matched prototype sets, while a tree-depth study examines prototype growth and capacity saturation.

Critical strengths
  1. The method provides a compact, example-grounded prediction mechanism. By replacing the original prediction head with a prototype-based bottleneck on a frozen backbone, the method makes training examples part of the actual computational pathway rather than merely providing post-hoc explanations. This is a meaningful direction toward more transparent language-model predictions.

  2. The experiments demonstrate feasibility across multiple tasks. The method constructs thousands of prototypes from millions of hidden states and preserves much of the original performance on language modeling, image captioning, and visual question answering. The prototype removal and retention experiments also provide meaningful evidence that a small set of highly matched prototypes is computationally important. These results support the practical promise of the approach.

Critical weaknesses
  1. The evidence for prediction-specific explanations remains incomplete. The removal and retention experiments support the importance of the top-matched prototype sets, but do not fully establish that the particular training examples displayed faithfully support individual predicted tokens. In Eq. 23, each prototype’s contribution depends on both its similarity and its token-specific learned weight; a highly similar prototype need not provide strong positive support for the predicted token. The paper does not sufficiently validate the correspondence between similarity-based explanation rankings and prediction-specific contributions. This gap matters because example-grounded interpretability is a central claimed benefit.

  2. The advantage of tree-structured DP-GMM over simpler prototype construction methods is not established. The tree-depth study demonstrates capacity saturation, and the appendix estimates a memory advantage over a monolithic DP-GMM. However, these results do not isolate the benefits of DP-GMM or adaptive growth relative to simpler alternatives such as K-means or hierarchical K-means. Changing tree depth also changes the prototype budget and prediction-head capacity. Matched-budget comparisons using the same hidden states, prototype projection, and head-training procedure are needed to determine whether the proposed complexity yields meaningful advantages in predictive performance, explanation quality, or computational cost.

Questions for authors
  1. Clarification and supporting evidence: How are the prototypes displayed as explanations selected? Does selection account for the token-specific prediction-head weights, or only similarity? Can the authors provide prediction-level evidence that the displayed examples positively support the predicted token, rather than only aggregate removal/retention results? This would materially affect my assessment of the interpretability claim.

  2. Evidence needed for the current contribution: Under comparable prototype budgets and otherwise matched settings, how does tree-structured DP-GMM compare with K-means or hierarchical K-means in predictive performance, explanation quality, fitting time, and peak memory? Such a comparison would materially affect my assessment of the proposed prototype-learning mechanism.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.