Measure Less, Know More: Self-Supervised Test-Time Feature Acquisition

Eeshaan Jain, Linus Bleistein, Bart Deplancke, Charlotte BunnearXiv 2026 · arXiv

In the era of foundation models (FMs), we’ve grown accustomed to the "more is better" philosophy. Feed a model more modalities—images, text, sensor data, genomic markers—and its representations become more robust. But in the real world, data isn't free. In a clinical setting, every additional lab test costs money and time. In satellite imaging, downloading every available spectral band is a bandwidth nightmare.

The challenge is: Which features should we acquire when we are on a budget, especially if we don't know exactly what the downstream task will be?

A recent paper from researchers at EPFL introduces ECHO-$k$, a self-supervised framework that teaches models to "shop" for the most informative data points iteratively.

TL;DR

ECHO-$k$ is a task-agnostic reinforcement learning (RL) approach for sequential feature acquisition. Instead of needing labeled data to decide what to measure next, it uses a foundation model’s own internal representations as a "gold standard" to reconstruct, allowing for cost-effective deployment across diverse fields like proteomics and geospatial sensing.

What Problem Does This Paper Address?

Most existing "active feature acquisition" methods are task-driven. They require a specific target (like a classification label) to calculate which feature reduces uncertainty the most. This creates two major bottlenecks:

  1. Label Dependency: You need a labeled dataset to train the acquisition policy.
  2. Task Rigidity: A policy trained to pick features for "Task A" might be useless for "Task B."

As we move toward a world of "frozen" foundation model backbones that serve dozens of downstream applications, we need an acquisition strategy that is task-agnostic. We need a way to pick the most "representative" features of an instance without knowing exactly how that instance will be used later.

Methods

The core "aha!" moment of the paper is treating the full-view latent representation of a foundation model as the target.

Figure 1: ECHO-k learns an RL policy to iteratively se- lects views to approximate the full latent representation learned by any pretrained model, without any knowledge of downstream tasks. Once learned, the policy can be deployed for budgeted feature acquisition.

The ECHO-$k$ Workflow:

  1. The Proxy Target: During training, the researchers take fully observed samples and pass them through a pretrained FM to get a "perfect" embedding ($z^*$).
  2. Reinforcement Learning: They train an RL agent (using Proximal Policy Optimization, or PPO) to select features one by one.
  3. The Reward: The agent receives a reward based on how much the new feature moves the current partial embedding closer to the "perfect" $z^*$.
  4. Long-Horizon Planning: Unlike "greedy" methods that just pick the best next step, the RL agent learns to account for correlations between features, planning its acquisition path over a fixed budget.

The authors also provide a theoretical backbone for this, proving that in a linear setting, recovering the latent representation effectively bounds the error for any Lipschitz-continuous downstream predictor.

Datasets

The authors stress-tested ECHO-$k$ across a remarkably diverse set of benchmarks:

  • hPancreas: Single-cell RNA-seq data (200 genes) using the TABULA foundation model.
  • NSCLC Patient Cohort: Spatial proteomics (40 protein markers) using the VirTues foundation model.
  • MMEarth-Bench: A global geospatial dataset with 12 modalities (radar, optical, meteorological, etc.).
  • MiniBooNE: A particle physics dataset (33 variables) for neutrino identification.
  • Vision: Downsampled MNIST and Fashion-MNIST (196 "pixel-features").

Key Insights and Ideas

  • Representations are Universal: The internal state of a foundation model is a rich summary of cross-modal information. By aiming to reconstruct this state, ECHO-$k$ naturally captures the features most relevant to a wide array of downstream tasks.
  • The "Reverse-U" Regret: The theory and experiments show a distinct "reverse-U" shape in the regret curve. The gap between an "oracle" (who sees everything) and a test-time policy is largest in the middle of the acquisition process, where uncertainty about unobserved features is highest.
Figure 2: Expected true loss gap curve be- tween the oracle and the test-time optimal, random, and prior-based (E[Ej]) policies.
  • Emergent Clinical Logic: In proteomics, without any medical supervision, ECHO-$k$ learned to prioritize markers like panCK (tumor identity) and Ki-67 (proliferation) early in the sequence—markers that human pathologists also prioritize.

Key Results

ECHO-$k$ consistently outperformed 12 different baselines, including standard unsupervised feature selection methods like Variance and Laplacian Score.

Figure 3: a. Results of ECHO-k on acquiring geospatial modalities to predict soil pH regression on MMEarth-Bench (d = 12) b. ECHO-k can operate on single-cell RNA-seq foundation models to sequentially acquire informative genes that support broad set of downstream tasks. We further present comparison of ECHO-k to baselines, evaluated on a cell type annotation task on hPancreas (with d = 200).
  • Geospatial: On the MMEarth-Bench, ECHO-$k$ reached strong predictive performance for soil pH after acquiring only 20-30% of the available modalities.
  • Task Transfer: A single policy trained on the VirTues backbone performed excellently for both cancer subtyping and cancer grading, proving its task-agnostic nature.
  • Robustness: The method remains effective even when the training "reference" data is itself incomplete (up to 50% missingness), suggesting it can be trained on the messy, real-world datasets it is meant to solve.

Limitations and Future Work

While powerful, ECHO-$k$ is not a silver bullet:

  • Backbone Bias: The policy is only as good as the foundation model it follows. If the FM ignores a specific feature, ECHO-$k$ will never learn to acquire it.
  • Distribution Shift: The paper doesn't deeply explore what happens if the test-time data comes from a significantly different distribution than the training data used to build the reference embeddings.
  • Future Directions: The authors suggest extending this to "adaptive stopping"—where the model itself decides it has seen enough data to make a confident prediction and stops spending its budget.

By shifting the focus from "labels" to "representations," ECHO-$k$ provides a principled path toward making our most powerful AI models resource-aware and ready for the constraints of the real world.

Summary updated

Abstract

Recent progress in multimodal, high-dimensional learning has enabled foundation models to process heterogeneous, large-scale data. However, at test time, acquiring all features or modalities can be prohibitively costly and often redundant. Sequentially selecting informative modalities is therefore critical, yet challenging when the downstream task or prediction target is unknown. To this end, we introduce ECHO-$k$, a task-agnostic and self-supervised learning principle for modality acquisition: we use a deep model's internal pretrained representations (e.g., from a foundation model) as proxy targets that summarize cross-modal information. We provide theoretical guarantees in a stylized linear setting that motivate a reinforcement learning (RL) policy for sequential modality selection. Across task-agnostic and label-free acquisition baselines, ECHO-$k$ consistently improves budgeted downstream performance across diverse foundation-model backends. Our method provides a principled route to cost-aware test-time deployment, with implications for any multimodal system where measurements are expensive or time-constrained, and downstream tasks unknown a priori.
BibTeX
@misc{jain2026measure,
  author = {Eeshaan Jain and Linus Bleistein and Bart Deplancke and Charlotte Bunne},
  title = {{Measure Less, Know More: Self-Supervised Test-Time Feature Acquisition}},
  year = {2026},
  eprint = {2610.03454},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2610.03454},
}
Figure