Research map / Probabilistic and scientific modeling / Molecules and proteins
Model–Behavior Alignment under Flexible Evaluation: When the Best-Fitting Model Isn’t the Right One
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.
Model-recovery simulations with noise-calibrated synthetic THINGS odd-one-out data show that linear-probing predictive accuracy identifies the data-generating vision model less than 80% of the time, even with millions of triplets.
Problem
Whether predictive accuracy under flexible (linear-probing) evaluation of deep vision networks against human behavioral judgments identifies genuinely aligned representations, i.e., whether model comparison is identifiable.
Approach
Applies model-recovery simulations to representational alignment in a noise-calibrated behavioral setting: each of 20 ANNs is linearly aligned to THINGS odd-one-out, temperature-calibrated to the human noise ceiling, used to generate synthetic triplet responses, and all candidates are refit to check whether the generator wins on cross-validated accuracy. It further links misidentification to align
Key results
- Model recovery accuracy improved with training-set size but plateaued below 80% for 20 models even at 4.2 million training triplets.
- Expanded 30-model set plateaued near 70% recovery; objective-level recovery reached only 73.7% at 4.2M triplets.
- Regression showed misidentification linked to alignment-induced representational shifts and higher post-alignment effective dimensionality (beta = -0.455 for ED of generator).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Increase training data from thousands of triplets to 4.2 million | Full-matrix linear probing model recovery · 20-model set | recovery accuracy: below 20% -> below 80% | Helped |
| Use top 500 principal components for every model, fixing the transformation to 500 × 500 (ablation) | Full-matrix model recovery · Model recovery | recovery accuracy | No clear effect |
| Add 10 candidate models to the model set (ablation) | Full-matrix model recovery · Model recovery, 30-model set | recovery accuracy: 80% -> near 70% | Hurt |
| Recover objective category rather than individual model identity | Linear-probing model comparison · Objective-group recovery, 4.2 million training triplets | recovery accuracy: 73.7% | Helped |
| Recover architecture category rather than individual model identity | Linear-probing model comparison · Architecture-group recovery, 4.2 million training triplets | recovery accuracy: 70.3% | Helped |
| Increase evaluation flexibility from zero-shot to diagonal, rectangular, and full transformations | ANN predictions of human THINGS judgments · THINGS odd-one-out | accuracy: 0.5370 -> 0.5558 -> 0.5937 -> 0.5927 | Helped |
| Use scalar-matrix shrinkage rather than Frobenius-norm regularization (ablation) | ANN alignment · THINGS odd-one-out | accuracy | No clear effect |
| Increase transformation flexibility | Model comparison under linear probing · THINGS odd-one-out model recovery | recovery accuracy | Hurt |
| Increase candidate-model alignment-induced geometric shift | Model recovery · Simulated THINGS judgments | accuracy difference between generator and candidate: β = 0.495 | Helped |
| Increase data-generating model's alignment-induced geometric shift | Model recovery · Simulated THINGS judgments | accuracy difference between generator and candidate: β = -0.2510 | Hurt |
| Increase post-alignment effective dimensionality of the data-generating model | Model recovery · Simulated THINGS judgments | accuracy difference between generator and candidate: β = -0.455 | Hurt |
Surprising findings
- Controlling transformation dimensionality by using top-500 PCs did not improve model recovery.
- Under the most flexible evaluation, several models reached near noise-ceiling accuracy with no single model significantly better, so accuracy alone cannot distinguish them.
Limitations
- Only the THINGS odd-one-out behavioral task was simulated; neural data may show a different trade-off.
- Recovery levels are specific to the 20-model candidate set, and the true biological model is absent from the closed set.
Experimental setup
Datasets: THINGS odd-one-out
Benchmarks: THINGS odd-one-out
Metrics: predictive accuracy model recovery accuracy mean rank confusion matrix
Baselines: zero-shot evaluation diagonal linear probing rectangular linear probing VICE CKA
Builds on: linear probing of ANN representations (Muttenthaler et al.) SPoSe VICE model recovery simulations
Closest papers by meaning
- Bridging Critical Gaps in Convergent Learning: How Representational Alignment Evolves Across Layers, Training, and Distribution Shifts (NeurIPS 2025 · similarity 0.88)
- Scaling Laws for Task-Optimized Models of the Primate Visual Ventral Stream (ICML 2025 · similarity 0.88)
- Quantifying Uncertainty in Error Consistency: Towards Reliable Behavioral Comparison of Classifiers (NeurIPS 2025 · similarity 0.88)
- Evaluating alignment between humans and neural network representations in image-based learning tasks (NeurIPS 2024 · similarity 0.87)
- Not Too Generative, Not Too Discriminative: The Human Alignment Sweet Spot (NeurIPS 2026 · similarity 0.87)
- Evaluating Multiview Object Consistency in Humans and Image Models (NeurIPS 2024 · similarity 0.87)
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.