When Are Predictions Enough? An Evaluation Protocol for Frozen Expert Composition

Alessandro Pereira, Keith J Ransom, Lewis MitchellarXiv 2026 · arXiv

In the era of foundation models and specialized "off-the-shelf" classifiers, it is common practice to combine multiple frozen models to solve a task. But how should we combine them? Should we just average their final predictions, or is it worth the extra engineering effort to tap into their internal features?

This paper, presented at NeurIPS 2026, introduces a rigorous evaluation protocol to answer these questions, using deepfake detection as a primary testbed.

TL;DR

Combining the internal features of frozen expert models can significantly outperform combining their final predictions, but these gains are highly sensitive to architectural diversity and evaluation design. The authors provide a new protocol to ensure that "feature-space" advantages aren't just artifacts of weak baselines or data leakage.

What Problem Does This Paper Address?

When we use a "frozen" expert (a model whose weights we don't change), we typically only look at its final output—for example, a single probability score in a binary deepfake detector. This "prediction-space" (PS) approach is simple and cheap. However, the model's internal "feature-space" (FS) contains much richer information that the final prediction might discard.

The challenge is that comparing PS and FS is often unfair. If a researcher compares a complex neural network trained on features against a simple average of predictions, the "gain" might just come from the complexity of the combiner, not the features themselves. Furthermore, hidden data overlap (train/test leakage) in common benchmarks can lead to over-optimistic results.

Methods

The authors propose a four-part evaluation protocol to make the comparison fair:

  1. PS Baseline Ladder: Instead of comparing against a single baseline, they use a "ladder" of increasingly strong prediction combiners, including probability averaging, logistic/neural stacking, and tuned XGBoost models.
  2. Feature-Space Anchor & Capacity Check: They compare a simple linear model on raw features (FS1) against a high-capacity MLP (FS2). This helps determine if gains come from richer information or just more parameters.
  3. Axis-Isolated Pools: They construct expert pools that vary only one thing at a time—either the model architecture (holding training data constant) or the training data (holding architecture constant).
  4. Grouped Evaluation (LOGO): They use a "Leave-One-Generator-Out" (LOGO) protocol. By holding out an entire deepfake generator during training, they measure how well the combination generalizes to truly unseen domains.
Figure 1: Prediction-space (top) vs feature-space (bottom) assembly. PS combines few-class outputs yi ∈RK, or functions h(y) of them, with a baseline ladder (§3.2); FS combines penultimate represen- tations zi ∈RDi, either directly (FS1) or via a learned projection and MLP (FS2).

Datasets

  • DF40: A deepfake detection benchmark featuring 40 different generation methods. The authors used 21 frozen detectors built on various backbones (CLIP, Xception, SRM, etc.).
  • SpeechFake & PartialSpoof: Used for exploratory checks to see if the findings hold true for synthetic speech detection.

Key Insights and Ideas

  • Information > Capacity: Surprisingly, the simple linear model (10K parameters) performed just as well as the 3.5M parameter MLP. This suggests that the bottleneck is the information available in the interface, not the complexity of the model processing it.
  • Diversity is Essential: Feature-space combination showed a clear advantage (+0.081 AUROC) when experts had diverse architectures. However, when experts shared the same architecture but were trained on different data, the advantage disappeared.
  • The "Weak Baseline" Trap: The authors found that using weak prediction baselines can inflate the apparent advantage of features by as much as 0.172 AUROC.
  • Transfer Matters: Features help most when they retain information that the final prediction omits and that information generalizes to new, unseen generators.

Key Results

The primary findings are summarized in the table below (referenced as Table 1 in the paper):

Table 1: FS1 versus the PS ceiling, measured by macro AUROC across the eight held-out generators. Under LOGO, the PS ceiling is selected retrospectively on each fold’s test data, favoring PS. ∆= FS1 −PS ceiling; wins count generators with ∆>0.
  • Architecture-Diverse Pool (arch6): Under the strict LOGO protocol, the linear feature combiner (FS1) outperformed the best prediction-space baseline by 0.081 AUROC, winning on 7 out of 8 generators.
  • Same-Architecture Pool (clip4): The ordering reversed; prediction-space combination actually performed better by 0.074 AUROC.
  • Standard Split Inflation: On the standard (overlap-affected) split, the gain for arch6 jumped to 0.105, illustrating how data leakage can exaggerate performance.

Limitations and Future Work

  • Task Scope: The study focused primarily on binary classification. The authors note that many-class or structured-output tasks (like object detection) remain untested.
  • Sample Size: The evaluation relied on only eight held-out generators, which limits the statistical precision of the findings.
  • Operational Costs: While features provide more info, they are much "heavier" to store and transfer than simple predictions. The paper provides a "cost ledger" but notes that end-to-end deployment costs will vary.
  • Future Directions: The authors suggest applying this protocol to large-scale foundation models and exploring how feature access affects calibration and fairness.

Summary updated

Abstract

Practitioners often combine independently trained models while keeping their weights frozen. These experts are commonly combined through their final predictions, although their internal features may retain useful information that those predictions omit. We ask when combining features improves on combining predictions, and how to make that comparison fair. We introduce a reusable evaluation protocol, apply it to 21 frozen deepfake detectors from DF40 (Yan et al., 2024), and audit how evaluation choices change the conclusions. The protocol uses strong prediction baselines, compares simple and larger feature combiners, varies architecture and training data separately, and reports results by held-out generator with train/test overlap controls. Under leave-one-generator-out evaluation, logistic regression on the features of six architecture-diverse experts improves macro AUROC by 0.081 over the best tested prediction combiner in each fold, with gains on seven of eight generators. The larger network shows no clear improvement over the linear model (approximately 3.5M versus 10K parameters), suggesting that access to richer information matters more than additional combiner capacity on this pool. Features may help when they retain information that the predictions omit and that transfers to new generators. Predictions may be enough when the strongest prediction baseline already performs well, while neither approach reliably compensates for poor expert transfer. Weakening prediction baselines on the overlap-affected standard split inflates the apparent feature advantage by up to 0.172 AUROC. Overlap-aware evaluation also changes the estimated gain, though the contrast does not isolate overlap. Results by generator and expert pool show where the advantage holds, weakens, or reverses. Without these controls, the same benchmark can support different conclusions.
BibTeX
@misc{pereira2026when,
  author = {Alessandro Pereira and Keith J Ransom and Lewis Mitchell},
  title = {{When Are Predictions Enough? An Evaluation Protocol for Frozen Expert Composition}},
  year = {2026},
  eprint = {2611.13356},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.13356},
}
Figure