Predicting the Performance of Foundation Models via Agreement-on-the-Line

Motivation

• Foundation models are increasingly deployed under distribution shift, while their OOD degradation can vary substantially from finetuning-distribution performance.The motivation spans applications including clinical decision-making across hospitals and robot navigation through unseen terrains.Source: §1 · p. 1
• AGL is difficult to guarantee after light finetuning because models remain close to pretrained weights and may produce highly correlated OOD predictions.Matching ID–OOD trends require models to disagree more on OOD data than on ID data, which depends on sufficiently decorrelated errors.Source: §1 · p. 3

Contributions

• Random head initialization consistently induces agreement-on-the-line in finetuned foundation models across vision and language benchmarks.The study covers linear probing and full finetuning, whereas data ordering and data subsetting do not reliably provide sufficient OOD diversity.Source: §1 · p. 3
• Finetuned models from different base foundation models can also support AGL-based performance estimation, with language models showing both AGL and ACL across question-answering and text-classification tasks.Different base models may differ in pretraining datasets, architectures, and training regimens, yet the language-task ensembles retain the relevant linear trends.Source: §1 · p. 3
• AGL-based estimation predicts finetuned foundation-model OOD performance across tasks, modalities, and training procedures, outperforming baselines by up to 20% mean absolute percentage error.The evaluated settings include classification and question answering, vision and language, and linear probing and full finetuning.Source: §1 · p. 3

Setup

• The setup covers foundation models mapping inputs to discrete outputs, including GPT2, OPT, Llama2, BERT, and CLIP, evaluated with accuracy, exact match, and macro-F1.OOD estimation assumes labeled ID validation data and unlabeled samples from a related but different OOD distribution.Source: §2.1 · pp. 3–4
• Linear probing freezes the foundation-model backbone and trains a randomly initialized linear or span-prediction head, whereas full finetuning updates the backbone and head with a small learning rate.LoRA is treated as conceptually equivalent to full finetuning for the study’s empirical comparisons.Source: §2.1 · pp. 3–4

Related work

• Existing OOD estimators include confidence calibration, auxiliary-task approaches, and reference ensembles, while AGL uses generic finetuning and requires at least three models.Reference-ensemble methods have higher computational cost, whereas auxiliary-task approaches can overfit particular datasets or modalities.Source: §2.2 · p. 4

Accuracy and agreement on the line

• AGL estimates OOD performance by fitting the ID–OOD agreement relationship on unlabeled data and applying its slope and bias to ID validation performance.Agreement requires model predictions on unlabeled data, avoiding the OOD ground-truth labels needed to fit ID–OOD accuracy directly.Source: §2.3 · pp. 4–5
• AGL links strong linear ID–OOD accuracy and agreement trends, requiring their slopes and intercepts to match when the coupled phenomenon holds.The method uses a collection of models and can estimate each model’s OOD accuracy from ID accuracy, ID agreement, and OOD agreement.Source: §2.3 · p. 5

Experimental design

• The study varies random heads, data ordering, and independently sampled data subsets while changing training epochs to obtain ensembles spanning different in-distribution performance levels.The diversity sources are perturbed one at a time, and epoch-induced randomness does not change the observations.Source: §3 · p. 5

Evaluation

• The benchmarks cover corruptions, replication, style, geographical, temporal, and interlaboratory shifts, including iWildCam-WILDS and Camelyon17-WILDS as negative ACL conditions.Camelyon17-WILDS shows neither ACL nor AGL in the cited figure.Source: §3.1 · p. 6, Figure 2 · p. 6

Results

• Random head initialization is the only tested diversity source that reliably induces AGL across CLIP vision and language-model ensembles.Data ordering and data subsetting instead produce agreement trends closer to the diagonal y = x line.Source: §3.1 · p. 6, §3.2 · pp. 6–7

Vision results

• Across vision benchmarks, CLIP linear-probe ensembles show ACL and agreement correlations, but matching agreement and accuracy slopes and biases occur only with diverse initialization.The same pattern extends to fully finetuned CLIP models.Source: §3.1 · p. 6, Figure 2 · p. 6

Language results

• Across text classification and extractive question-answering, full-finetuned LLM ensembles show AGL with randomly initialized linear or span heads, but not with data ordering or subsetting.The evaluated metrics are accuracy for text classification and F1 for question-answering.Source: §3.2 · pp. 6–7

Implications

• These findings differ from [2] because lightly finetuned or linearly probed foundation models require a particular diversity source, unlike heavily trained networks where AGL is more robust to that choice.The comparison concerns models trained from scratch for tens to hundreds of epochs versus lightly adapted foundation models.Source: §3.3 · pp. 7–8

Model collection

• The evaluated language models span OPT, GPT-2, GPT-Neo, Llama2, Alpaca, and Vicuna, trained or adapted using different pretrained text corpora and finetuning procedures.OPT and GPT models are fully finetuned, while Llama, Alpaca, and Vicuna use LoRA finetuning.Source: §4 · p. 8

Accuracy trends

• Models pretrained on different text corpora produce finetuned language models lying on the same linear accuracy trend, unlike the differing robustness trends reported for vision foundation models.The authors suspect greater homogeneity among the language-model pretraining datasets in this study.Source: §4.1 · p. 8, §4 · p. 8

Results

• Ensembles combining Llama, GPT, and OPT models exhibit AGL for question-answering F1 and text-classification accuracy across the evaluated shifts.The shifts include SQuAD to SQuAD-Shifts and MNLI-Matched to MNLI-Mismatched and SNLI.Source: §4.1 · pp. 8–9

Agreement trends

• Pairwise ID-versus-OOD agreement is strongly correlated, with slope and intercept closely matching accuracy, including pairs drawn from different base models.This occurs without special regularization for ensemble diversity and also holds for generative question-answering tasks.Source: §4.1 · p. 9

Baselines

• ALine-based estimation is evaluated against ATC [19], AC [24], DOC-Feat [21], and Naive Agreement [27].Confidence baselines are temperature-scaled using in-distribution validation data, with the lower estimated error retained.Source: §5 · p. 9

Metric scope

• For question-answering, AGL predicts accuracy, F1, and exact-match, whereas confidence baselines are compared using exact-match because they lack an easy analogous formulation for F1.Extractive question-answering is treated as a joint prediction of answer-span start and end indices.Source: §5 · p. 9

Results

• Across high-R2 shifts, ALine-S and ALine-D provide precise OOD performance estimates in finetuned foundation models and surpass other baselines by a large margin.The methods remain effective when the agreement line is substantially off y = x.Source: §5 · p. 9

Evaluation condition

• ALine-S and ALine-D are evaluated only on datasets whose agreement fit has R2 > 0.95, excluding specified low-correlation shifts from the main evaluation.The excluded cases include iWildCam-WILDS, Camelyon-WILDS, and selected CIFAR10C, CIFAR100C, and ImageNetC corruptions.Source: §5 · p. 9

• AGL is not guaranteed across all distribution shifts, so deployers should monitor additional performance metrics.Source: §6 · p. 10
• Zero-shot evaluation of different pretrained checkpoints does not reliably support AGL-based performance prediction.Source: §6 · p. 10
• The study covers classification and question-answering but leaves long-form generation for future work.Source: §6 · p. 10
• The paper provides no theoretical guarantees for its empirical findings.Source: §6 · p. 10

• AGL-based OOD prediction extends beyond from-scratch neural networks to foundation models across classification, question-answering, vision, and language tasks.The method applies to both linear probing and full finetuning.Source: §7 · p. 10
• Randomness in specific optimization choices, especially linear-head initialization, is crucial for observing AGL during foundation-model finetuning.The paper emphasizes that ensemble diversity must be tuned carefully because lightly finetuned models may otherwise lack the needed diversity.Source: §7 · p. 10
• Linear models can exhibit AGL in CLIP representation space, indicating that the phenomenon is not limited to neural networks.Source: §7 · p. 10
• All evaluated language models, including OPT, GPT2, GPT2-Neo, Alpaca, Llama, and Vicuna, lie on the same accuracy line.Source: §7 · p. 10

• Additional experiments examine the effect of diversity sources for extractive question answering.Source: A Appendix · p. 15

• ALine-S and ALine-D use agreement-based linear relationships to estimate OOD performance without directly fitting ID-to-OOD accuracy from labeled OOD data.The methods transfer a linear relation inferred from agreements to OOD performance estimation.Source: §A.1.1 · p. 16
• AGL states that ID-versus-OOD accuracy and agreement have matching slopes and biases when their corresponding linear correlations hold.Source: §A.1.1 · p. 16

ALine-S

• ALine-S estimates the accuracy slope and bias by fitting the agreement line, then uses them to transform ID performance into OOD estimates.Source: §A.1.1 · p. 16

ALine-D

• ALine-D treats the OOD accuracies as unknowns, constructs equations from model pairs, and solves the approximate system by least squares.Source: §A.1.1 · p. 16

• The study compares ALine with ATC, AC, and Doc-Feat, noting that these baselines perform better after in-distribution temperature scaling.Source: §A.1.2 · p. 17
• Confidence-based classification baselines optimize a separate temperature for each model using in-distribution validation cross-entropy.Source: §A.1.2 · p. 17
• For extractive question answering, start and end logits receive separate temperatures optimized with validation cross-entropy.Source: §A.1.2 · p. 17
• Exact match treats question answering as classification over start–end index pairs, enabling classification confidence methods with pair confidence σ(s(x))_i · σ(e(x))_j.Source: §A.1.2 · p. 17

A.1.3 Comparison to ProjNorm

• ALine-D estimates correlate more strongly with OOD performance than ProjNorm, which supplies only a correlated score rather than an OOD-performance estimate.Source: §A.1.3 · p. 17, Table 4 · p. 17

A.1.4 Failure Datasets with Low Linear Correlation

• Datasets with ID–OOD agreement correlation ≤0.95 are excluded from baseline comparisons because AGL-based estimates are not guaranteed reliable when agreement correlation is weak.Source: §A.1.4 · p. 18
• For excluded datasets, ATC is generally significantly better when AGL-based methods are unreliable.Source: §A.1.4 · p. 18

Diversity-source experiments

• For heavily trained ResNet18 models trained from scratch, AGL can appear irrespective of the diversity source, unlike lightly finetuned foundation-model ensembles.Source: §A.3.4 · p. 27
• Across CLIP, GPT2, OPT, BERT, and multiple distribution shifts, random head initialization yields the strongest AGL behavior and smallest or aligned ALine errors.Source: §A.3.1 · p. 21, §A.3.3 · p. 26, §A.4.1 · p. 30
• The same AGL pattern persists across full finetuning and parameter-efficient methods, while data ordering or subsetting generally does not reproduce it in the reported CLIP and GPT2 experiments.Source: §A.3.1 · p. 21, §A.3.3 · p. 26, §A.8 · p. 47