Research map / Label-efficient and robust learning / Recognition and long-tail

Information Templates: A New Paradigm for Intelligent Active Feature Acquisition

NeurIPS 2026 · Label-efficient and robust learning › Recognition and long-tail · approach: Dense features and foundation priors

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.

Proposes TAFA, an RL-free active feature acquisition framework that learns a small library of feature templates via submodular-style genetic search and Gumbel-Softmax refinement, and distills it into interpretable step-wise decision trees.

Problem

Active feature acquisition requires sequentially choosing costly features at inference time; RL policies are hard to train and greedy/CMI or ACO-style methods either ignore joint informativeness, need distribution modeling, or are computationally prohibitive.

Approach

Restrict the AFA action space to a learned library of feature templates (jointly informative feature subsets) found by a submodular set-optimization objective, searched with mutation-guided greedy selection and refined by continuous relaxation with a Gumbel-Softmax actor; then distill into step-wise decision trees over templates for interpretability.

Key results

• TAFA consistently outperforms AACO, SEFA, DIME and static baselines on most datasets with higher accuracy at fewer acquisitions, and is substantially faster than AACO at inference.
• Distilled tafa-interp achieves higher reward than DDT and SYMPOL at comparable leaf budgets on interpretability comparisons.
• On CUBE-sigma, the distilled policy explains about 99.9% of instances with 5 distinct acquisition subsets.

Surprising findings

• Continuous-relaxation refinement gives 5-8% gains on high-dimensional datasets but the authors argue genetic search alone suffices for low-dimensional EngineDB.

Limitations

• Submodularity and greedy guarantees hold only under special monotone loss/cost choices; practical losses may violate them.
• Uses uniform acquisition costs in experiments.
• Continuous refinement benefits depend on dimensionality, and interpretable variant trades some accuracy.

Experimental setup

Datasets: CUBE-sigma=0.3 Big Five personality Gas sensor home activity Electrical grid stability MNIST Fashion MNIST EngineDB

Benchmarks: CUBE-sigma=0.3 Big Five Gas Grid MNIST Fashion MNIST EngineDB

Metrics: accuracy vs number of features acquired inference time reward vs number of leaves

Baselines: AACO SEFA DIME static DDT SYMPOL

Builds on: active feature acquisition (AFA) ACO / AACO DIME SEFA DAgger

Closest papers by meaning

• Stochastic Encodings for Active Feature Acquisition (ICML 2025 · similarity 0.89)
• Active feature acquisition via explainability-driven ranking (ICML 2025 · similarity 0.88)
• Active Learning with LLMs for Partially Observed and Cost-Aware Scenarios (NeurIPS 2024 · similarity 0.85)
• Efficient Adaptive Data Acquisition via Pretrained Belief Representations (NeurIPS 2026 · similarity 0.85)
• Generator Assisted Mixture of Experts for Feature Acquisition in Batch (AAAI 2024 · similarity 0.83)
• ALINE: Joint Amortization for Bayesian Inference and Active Data Acquisition (NeurIPS 2025 · similarity 0.83)