Research map / Label-efficient and robust learning / Recognition and long-tail
Prospector Heads: Generalized Feature Attribution for Large Models & Data
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces prospector heads, lightweight modules equipped on frozen encoders that quantize token embeddings into concepts and learn class-specific concept monogram/skip-bigram kernels to perform feature attribution without explaining end-to-end classifiers.
Problem
Feature attribution (localizing class-relevant regions of inputs) in scientific and biomedical data is imprecise and computationally expensive when done by explaining end-to-end classifiers, and it is data-inefficient in small-sample, high-dimensional settings.
Approach
Attach a parameter-efficient head to any frozen encoder: k-means quantizes token embeddings into K concepts (data sprites), then a K2conv kernel scores concept monograms and skip-bigrams within a receptive field, fit by either logistic regression or a fold-change with significance masking, trained directly to localize class-specific tokens under coarse datum-level labels.
Key results
- Prospectors improve mean test AUPRC on text retrieval to 0.711 from 0.626 for the top supervised baseline (8.5-point gain).
- On Camelyon16 prospectors gain 26.3 AUPRC points and on MetalPDB 22.0 AUPRC points over the top modality-specific baselines.
- Prospector-equipped domain foundation models (PLIP, COLLAPSE) give the strongest localization, and fold-change variant outperforms linear classifier in almost all settings.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Equip MiniLM with a prospector head instead of using the MLP head | MiniLM · WikiSection genetics test | localization AUPRC: 0.626 -> 0.711 | Helped |
| Equip MiniLM with a prospector head instead of DeBERTa NLI entailment | MiniLM · WikiSection genetics test | localization AUPRC: 0.584 -> 0.711 | Helped |
| Use prospector heads instead of the top modality-specific baseline | Prospector-equipped models · Camelyon16 | localization AUPRC: +26.3 points AUPRC | Helped |
| Use prospector heads instead of the top modality-specific baseline | Prospector-equipped models · Camelyon16 | localization AP: +8.8 points AP | Helped |
| Use prospector heads instead of the top modality-specific baseline | Prospector-equipped models · MetalPDB | localization AUPRC: +22.0 points AUPRC | Helped |
| Use prospector heads instead of the top modality-specific baseline | Prospector-equipped models · MetalPDB | localization AP: +8.8 points AP | Helped |
| Use the fold-change kernel-fitting variant instead of the linear-classifier variant (ablation) | Prospector heads · Almost all evaluated settings | localization performance | Helped |
| Equip the pathology model with a domain-specialized encoder such as PLIP rather than a less specialized encoder (ablation) | Prospector heads · Camelyon16 | localization performance | Helped |
| Use COLLAPSE embeddings rather than ESM2 sequence embeddings for protein structures (ablation) | Prospector heads · MetalPDB | localization performance | Helped |
Surprising findings
- The parameter-free fold-change variant outperformed the trained linear classifier in almost all evaluated settings.
- The amino-acid one-hot encoder (AA) was an exception to encoder generalization, performing well with prospectors.
Limitations
- No sensitivity analysis of design choices such as token resolution, connectivity, clustering method, and embedding sample size.
- Only open-source encoders were tested.
- Bimodal performance on MetalPDB with distribution shift between train and test concepts.
Experimental setup
Datasets: WikiSection Camelyon16 MetalPDB
Benchmarks: WikiSection (genetics section retrieval) Camelyon16 MetalPDB zinc binding sites
Metrics: AUPRC average precision precision dice coefficient Matthews correlation coefficient
Baselines: MLP probe one-class SVM DeBERTa NLI entailment concatenated mean attention concatenated prediction probability GAT Attention GAT Shapley values (SHAP) GNNExplainer
Builds on: concept bottleneck models TF-IDF elastic net logistic regression
Closest papers by meaning
- GEFA: A General Feature Attribution Framework Using Proxy Gradient Estimation (ICML 2025 · similarity 0.85)
- Provably Better Explanations with Optimized Aggregation of Feature Attributions (ICML 2024 · similarity 0.85)
- Regional Explanations: Bridging Local and Global Variable Importance (NeurIPS 2025 · similarity 0.85)
- Fast Data Attribution for Text-to-Image Models (NeurIPS 2025 · similarity 0.84)
- OrdShap: Feature Position Importance for Sequential Black-Box Models (NeurIPS 2025 · similarity 0.84)
- Less is More: Fewer Interpretable Region via Submodular Subset Selection (ICLR 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.