PaperlayerView exact revision on arXiv

Source-linked AI summary

What Does Prompt Learning Change? -A Natural-Language Concept Analysis of Vision-Language Models

Ryo Kamiya, Hiroshi Kera, Kazuhiko Kawamoto

arXiv:2608.24142v1cs.CV

TL;DR

Prompt learning changes continuous prompts in ways that are difficult to inspect in natural language. PromptSpLiCE addresses this gap by decomposing initial and learned class-conditioned text embeddings in a shared fixed concept dictionary; across CoOp experiments, concept profiles reorganize substantially, with larger changes tending to accompany larger accuracy gains.

  • Problem

    Prompt learning optimizes continuous prompt embeddings, but their semantic changes are not directly observable in natural language.

  • Method

    PromptSpLiCE is a post-hoc method that represents initial and learned class-conditioned text embeddings as sparse combinations in a shared fixed natural-language dictionary.

  • Results

    Across 11 image-classification datasets, only 1.6 of the initial top-10 concepts remain in the top 10 after CoOp learning on average, and larger profile changes tend to accompany larger accuracy gains.

  • Takeaways & Limitations

    PromptSpLiCE provides a practical text-side diagnostic for making continuous prompt changes readable through named concepts without modifying prompt learning.

  • Takeaways & Limitations

    Top-ranked concepts alone do not necessarily faithfully represent the prompt or preserve its predictions.

Abstract

from arXiv · show

Prompt learning adapts vision-language models such as CLIP by optimizing continuous prompt vectors, but the learned prompts are difficult to interpret in natural language. We present PromptSpLiCE, a post-hoc method that expresses each class-conditioned text embedding as a sparse combination of concepts from a fixed natural-language dictionary. Using the same dictionary before and after prompt learning allows us to compare changes in their concept profiles. We evaluate PromptSpLiCE on CoOp, a representative prompt-learning method, across 11 image-classification datasets. The concept profiles change substantially: on average, only 1.6 of the initial top-10 concepts remain in the top 10 after learning. Across datasets, profile change is positively associated with accuracy gain. We also derive a local gradient expression that provides geometric intuition for why image-aligned concept directions distinct from the current prompt can have greater loss sensitivity.

1 Introduction

PromptSpLiCE provides a post-hoc natural-language analysis of how prompt learning changes class-conditioned text embeddings. Applied to CoOp across 11 datasets, it finds substantial concept-profile reorganization associated with accuracy gains and motivates a geometric sensitivity analysis.

  • Method: PromptSpLiCE represents initial and learned prompt embeddings as sparse combinations of concepts from one fixed natural-language dictionary.The shared coordinate system makes fitted coefficient profiles directly comparable before and after learning.
  • Results: 1.6 of the initial top-10 concepts remain in the top 10 after CoOp learning, on average across 11 image-classification datasets.The result indicates substantial reorganization of natural-language concept profiles.
  • Results: Larger concept-profile changes tend to accompany larger accuracy gains across datasets.The reported relationship is an association, not a stated causal effect.
  • Results: Qualitative results show that recognizable class concepts can persist even while less intuitive terms rise in rank.The analysis therefore captures both persistence and re-ranking within learned profiles.
  • Geometric analysis: A local gradient analysis gives geometric intuition for why image-aligned concept directions distinct from the current prompt can have greater loss sensitivity.This extends the profile comparison with an embedding-level explanation of directional sensitivity.

2 Related Work

Prior work improves interpretability either during prompt learning or by decomposing learned representations, but arbitrary optimized prompts remain difficult to explain in natural language. PromptSpLiCE applies fixed-dictionary concept decomposition post hoc to final text embeddings, complementing visual-feature analyses.

  • Prompt Learning: CoOp freezes pretrained model parameters and learns text-prompt token embeddings to maximize image similarity for the ground-truth class.Its learnable embeddings start from a manually designed prompt such as “A photo of a {class}.”
  • Prompt Learning: Most prompt-learning methods prioritize classification performance and leave semantic changes in task-adapted continuous prompts difficult to observe directly.Interpretability-aware methods instead incorporate human-readable attributes or concepts during training or inference.
  • Post-hoc concept decomposition: Post-hoc concept decomposition explains high-dimensional CLIP embeddings using human-interpretable concepts while keeping the pretrained model fixed.Sparse autoencoders reconstruct embeddings with learned basis vectors interpreted from common activation properties.
  • Related analyses: PatchSAE compares visual-feature activations before and after MaPLe adaptation, whereas this analysis studies final text embeddings after CoOp using a fixed natural-language dictionary.The cited comparison identifies the analyses as complementary rather than interchangeable.
  • SpLiCE: SpLiCE reconstructs CLIP embeddings as sparse nonnegative combinations of natural-language concept embeddings, and PromptSpLiCE compares coefficients before and after prompt optimization.The shared decomposition procedure shifts the analysis to prompt-level semantic changes.

3 PromptSpLiCE

PromptSpLiCE decomposes prompt embeddings into sparse concept coordinates and compares profiles before and after learning. Its local sensitivity analysis relates concept-direction geometry to loss sensitivity, while the formulation includes explicit reconstruction and gradient-scope qualifications.

  • 3.1 Concept Decomposition of Prompt Embeddings: PromptSpLiCE applies Sparse Linear Concept Embeddings to represent prompt embeddings as sparse combinations of a fixed natural-language dictionary.Using one dictionary before and after learning enables visualization and comparison of fitted coefficient profiles.
  • 3.1 Concept Decomposition of Prompt Embeddings: The dictionary contains M natural-language concepts encoded by CLIP, with their embeddings arranged as dictionary columns.The construction provides the concept coordinate system used for decomposition.
  • 3.1 Concept Decomposition of Prompt Embeddings: Centering and normalization mitigate CLIP embedding anisotropy so a shared common direction does not obscure differences among concepts.The preprocessing removes the dominant common component before fitting concept coefficients.
  • 3.1 Concept Decomposition of Prompt Embeddings: Nonnegative sparse coefficients are fitted post hoc by trading reconstruction error against sparsity with an ℓ1 penalty.Because the dictionary is overcomplete and correlated, different coefficient supports may yield similar reconstructions.
  • 3.1 Concept Decomposition of Prompt Embeddings: Direction-only reconstruction retains the embedding direction while fixing residual scale to one, introducing an approximation relative to exact inversion.Exact inversion would require both the direction and scale; the scale-aware expression is provided separately.
  • 3.2 Local Sensitivity in Concept Coordinates: The local sensitivity analysis treats concept coefficients as local coordinates for a cosine-similarity classification objective while holding image and other class embeddings fixed.The setup derives a hypothetical embedding-level sensitivity in concept coordinates.
  • 3.2 Local Sensitivity in Concept Coordinates: The prediction-error factor scales sensitivity, while temperature and embedding norm rescale updates without changing their qualitative dependence on concept direction.Orthogonal projection determines directional alignment with the image embedding; near-parallel dictionary directions have small projected components and sensitivity.
  • 3.2 Local Sensitivity in Concept Coordinates: A concept can change substantially when its component orthogonal to the prompt aligns with the image’s orthogonal component, producing high local loss sensitivity.Figure 1 contrasts this case with nearly parallel dictionary directions, whose changes are small.

4 Experiments

PromptSpLiCE evaluates reconstruction fidelity and compares natural-language concept profiles before and after CoOp prompt learning across 11 image-classification datasets. The experiments show substantial rank reorganization, while larger profile changes are associated with larger accuracy gains.

  • Experimental Setup: PromptSpLiCE evaluates post-hoc natural-language decompositions of prompt embeddings before and after learning across 11 image-classification datasets.The datasets span object, fine-grained, scene, action, texture, and satellite-image classification.
  • Embedding Reconstruction Fidelity: At λ = 0.01, learned prompts use approximately 450 nonzero coefficients and achieve a cosine similarity of 0.98 to their full reconstructions.The reconstruction is sparse relative to the 10,000-term dictionary, but 450 terms are not a concise explanation; top-10 displays are illustrative subsets.
  • Qualitative Analysis: EuroSAT’s “Annual Crop Land” profile shifts from farmland and crops toward defend and baseman after prompt learning.The StanfordCars example retains class-related terms such as fiat and convertible while also elevating less intuitive terms.
  • Quantitative Analysis: The rank-transition analysis groups terms as High→High, High→Low, or Low→High using fixed dictionary, solver, and regularization settings.The cutoffs are top 10 before and after learning, below rank 100 after learning, or below rank 100 before learning, respectively.
  • Quantitative Analysis: 1.6 initial top-10 terms remain in the top 10 on average across 11 datasets; 6.8 fall below rank 100 and 6.8 rise into the top 10.High→High terms are few on most datasets, while High→Low and Low→High transitions are common.
  • Exploratory Association with Accuracy Gain: The dataset-level correlation between mean JS divergence and CoOp accuracy gain is r = 0.64 (nominal p = 0.035, n = 11).The paper treats this as an exploratory association and notes sensitivity to the small sample and prominent EuroSAT endpoint.

5 Conclusion

PromptSpLiCE makes prompt-learning changes readable by comparing initial and learned embeddings in a shared natural-language coordinate system. Across CoOp experiments, concept profiles changed substantially, and larger changes tended to accompany larger accuracy gains.

  • PromptSpLiCE places class-conditioned text embeddings before and after learning in a shared coordinate system defined by a fixed natural-language dictionary.This post-hoc analysis makes continuous prompt changes readable through named concepts without modifying prompt learning.
  • Larger fitted concept-profile changes tended to accompany larger classification-accuracy gains across the 11 datasets.The reported association is exploratory.

A.1 Reconstruction after Centering and Normalization

The reconstruction preprocessing separates a centered prompt embedding into direction and scale, while the displayed direction-only reconstruction fixes the residual scale to one. This convention preserves the decomposition direction but is not generally an exact inverse.

  • The preprocessing separates the centered prompt embedding into a direction and a scale.The formulation writes the embedding as its mean plus a scaled centered direction.
  • The reconstruction pipeline defines a decomposition residual and then forms an unnormalized reconstruction before final unit normalization.The supplied passage sequence describes these intermediate reconstruction stages.
  • The direction-only reconstruction sets ρ_z = 1 and therefore does not generally exactly invert preprocessing.Exact inversion would require retaining both the direction and the scale.

A.2 Local Sensitivity in Concept Coordinates

The local sensitivity derivation treats concept coordinates as local variables and then relates them to shared prompt parameters through the text-embedding Jacobian. Its interpretation is class-wise embedding sensitivity, not the total derivative of the post-hoc fitted solution.

  • The derivation differentiates each concept-coordinate term and combines the resulting expressions into the gradient with respect to the full concept-coordinate vector.The calculation proceeds through derivatives with respect to s_i and normalized embeddings z_i.
  • The local calculation assumes each w_i affects only s_i, whereas the actual method optimizes prompt parameters that may be shared across classes.The general parameter gradient uses the text-embedding Jacobian J_i.
  • The resulting expression is class-wise embedding-level sensitivity, not the total derivative through the Lasso solution and changing active support.At nonnegativity boundaries, the derivative is local and feasible coefficient perturbations are one-sided.

B Additional Qualitative Results

Additional qualitative results across the remaining datasets show varied coefficient-rank transitions: diagnostic terms often remain highly ranked while related and less intuitive terms can move differently. This turnover means reconstruction fidelity alone does not establish semantic stability.

  • B Additional Qualitative Results: Figures 8–10 cover the remaining seven datasets using before-versus-after coefficient scatter plots and High→High, High→Low, and Low→High transition colors.Green, red, and orange denote the three transition types; gray marks remaining dictionary terms.
  • B Additional Qualitative Results: Category-related terms can remain highly ranked while other related and less intuitive dictionary terms undergo different rank transitions.Examples include DTD’s “lined” class and FGVCAircraft’s “Cessna 172” class.
  • B Additional Qualitative Results: Substantial coefficient-rank turnover can coexist with retention of highly diagnostic terms.This pattern appears in the UCF101 “Bench Press” and Caltech101 “watch” examples.
  • B Additional Qualitative Results: Across the qualitative cases, at least one intuitive term often remains High→High, while Low→High frequently includes labels with no clear human relation to the class.The observed pattern is consistent with Table 1.
  • B Additional Qualitative Results: High full-embedding reconstruction fidelity alone does not establish stability or semantic fidelity of displayed top-ranked terms.The qualitative transitions motivate distinguishing reconstruction quality from interpretability of ranked concepts.
Loading 2608.24142v1…