Research map / Label-efficient and robust learning / OOD and anomaly detection
Predicting the Performance of Foundation Models via Agreement-on-the-Line
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.
Shows that agreement-on-the-line (AGL) for OOD accuracy estimation can be induced in ensembles of lightly finetuned foundation models by randomizing only the linear head initialization, and that ensembles across different base FMs also exhibit AGL on language QA and classification, enabling label-free OOD estimation with ALine.
Problem
Label-free estimation of out-of-distribution performance for foundation models, since lightly finetuned pretrained models may lack the ensemble diversity needed for agreement-on-the-line to hold.
Approach
Diversity sources (random linear head init, data ordering, data subsetting) are compared for inducing AGL in finetuned FM ensembles; random head initialization alone reliably produces AGL, and ensembles of different pretrained FMs on language tasks also show AGL, so ALine-S/D can be applied to predict OOD performance.
Key results
- Random linear head initialization yields the lowest ALine-D MAPE on CIFAR10C (14.64%) versus data ordering (37.01%) and data subsetting (35.85%) for linear-probed CLIP.
- ALine-S achieves 6.34% MAPE on SQuAD-Shifts Amazon and 2.10% on SNLI, with the next best confidence baseline reaching up to ~20% higher error on SQuAD.
- Ensembles of finetuned models from different base LMs (Llama, GPT, OPT) on SQuAD show both ACL and AGL, and AGL-based estimates are accurate without any OOD labels.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use random linear-head initialization rather than data ordering or data subsetting to diversify the ensemble (ablation) | CLIP linear probes · CIFAR10C | ALine-D MAPE (%): 14.64 vs 37.01; 14.64 vs 35.85 | Helped |
| Use random linear-head initialization rather than data ordering or data subsetting to diversify the ensemble (ablation) | GPT2-Medium/OPT-125M full finetuning · SQuAD-Shifts Amazon | ALine-D MAPE (%): 6.34 vs 10.30; 6.34 vs 16.21 | Helped |
| Use random head initialization rather than data ordering or data subsetting (ablation) | CLIP full finetuning · CIFAR10C | ALine MAPE (%): 29.37 vs 69.98; 29.37 vs 33.57 | Helped |
| Use random head initialization rather than data ordering or data subsetting (ablation) | OPT-125M full finetuning · SQuAD-Shifts Amazon | ALine-D MAPE (%): 6.54 vs 11.37; 6.54 vs 11.15 | Helped |
| Use random head initialization rather than data ordering or data subsetting (ablation) | BERT full finetuning · SQuAD-Shifts Reddit | ALine-D MAPE (%): 8.62 vs 9.16; 8.62 vs 18.64 | Helped |
| Use ALine-S rather than the strongest listed baseline | OOD performance estimation with diverse foundation-model ensembles · Most evaluated high-agreement benchmarks | MAPE (%) | Mixed |
| Use ALine-S rather than DOC-Feat | OOD performance estimation · ImageNetC | MAPE (%): 14.17 vs 13.42 | Hurt |
| Use ALine-D rather than ATC | OOD performance estimation · ImageNetV2 MatchFreq | MAPE (%): 8.44 vs 3.02 | Hurt |
| Use ALine-D rather than ATC | OOD performance estimation · SQuAD-Shifts New Wiki | MAPE (%): 6.33 vs 5.18 | Hurt |
Surprising findings
- Only random linear head initialization reliably induces agreement-on-the-line; data ordering and data subsetting do not in lightly finetuned FMs.
- Linear probes on CLIP features can exhibit AGL, contrary to prior work that linear models on raw pixels do not.
Limitations
- AGL is unreliable when agreement linear fit has low R^2 (e.g., iWildCam, Camelyon17, some CIFAR10C/ImageNetC corruptions), so the method is filtered by R^2 <= 0.95.
- No theoretical guarantees are provided for why random head initialization is key.
- Long-form generation tasks are not studied.
Experimental setup
Datasets: CIFAR10 CIFAR100 ImageNet FMoW-WILDS iWildCam-WILDS Camelyon17-WILDS OfficeHome MNLI SQuAD
Benchmarks: CIFAR10C CIFAR10.1 CIFAR100C ImageNetC ImageNetV2 FMoW-WILDS iWildCam-WILDS Camelyon17-WILDS OfficeHome MNLI-Mismatched SNLI SQuAD-Shifts
Metrics: MAPE accuracy F1 exact match R^2 of agreement linear fit
Baselines: Naive Agreement ATC AC DOC-Feat ProjNorm
Builds on: agreement-on-the-line accuracy-on-the-line ALine-S ALine-D average thresholded confidence
Closest papers by meaning
- Specialized Foundation Models Struggle to Beat Supervised Baselines (ICLR 2025 · similarity 0.85)
- Understanding Overadaptation in Supervised Fine-Tuning: The Role of Ensemble Methods (ICML 2025 · similarity 0.85)
- Towards Few-Shot Adaptation of Foundation Models via Multitask Finetuning (ICLR 2024 · similarity 0.85)
- Fine-Grained Benchmark Generation for Comprehensive Evaluation of Foundation Models (NeurIPS 2026 · similarity 0.84)
- Rethinking Weight Decay for Robust Fine-Tuning of Foundation Models (NeurIPS 2024 · similarity 0.84)
- Prompting is a Double-Edged Sword: Improving Worst-Group Robustness of Foundation Models (ICML 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.