Source-linked AI summary
How Can We Accelerate Progress Towards Human-like Linguistic Generalization?
Tal Linzen
TL;DR
The paper asks whether benchmark-driven NLP progress is advancing human-like language learning, given differences in data exposure and generalization. It critiques PAID evaluation and advocates a parallel ecosystem rewarding models that generalize human-like from limited data. Its conclusion is that human-like inductive biases could better align model behavior with human expectations and reduce reliance on superficial heuristics.
Problem
The paper examines whether NLP evaluation measures progress toward human-like sample-efficient language learning and consistent generalization.
Method
The paper describes PAID’s pretraining, fine-tuning, and same-distribution evaluation practices, then analyzes their incentives and failure modes.
Results
PAID favors models that scale to unlimited data and capture arbitrary fine-tuning correlations, while controlled examples reveal generalization behavior unlike humans.
Takeaways & Limitations
A parallel evaluation ecosystem should reward human-like generalization from limited data and reduce the allure of superficial heuristics.
Takeaways & Limitations
Human–transformer comparisons can be unfair because models receive hundreds of thousands of task-output examples before testing.
Abstract
from arXiv · showhide
This position paper describes and critiques the Pretraining-Agnostic Identically Distributed (PAID) evaluation paradigm, which has become a central tool for measuring progress in natural language understanding. This paradigm consists of three stages: (1) pre-training of a word prediction model on a corpus of arbitrary size; (2) fine-tuning (transfer learning) on a training set representing a classification task; (3) evaluation on a test set drawn from the same distribution as that training set. This paradigm favors simple, low-bias architectures, which, first, can be scaled to process vast amounts of data, and second, can capture the fine-grained statistical properties of a particular data set, regardless of whether those properties are likely to generalize to examples of the task outside the data set. This contrasts with humans, who learn language from several orders of magnitude less data than the systems favored by this evaluation paradigm, and generalize to new tasks in a consistent way. We advocate for supplementing or replacing PAID with paradigms that reward architectures that generalize as quickly and robustly as humans.
1 Introduction
The paper distinguishes broad application progress from progress toward human-like language learning and generalization. It argues that PAID evaluation may advance in a direction orthogonal to that goal.
- The paper asks whether machines can acquire language as efficiently as humans and generalize to new structures, contexts, and tasks.
- PAID has become a major progress-measurement paradigm but may encourage progress orthogonal to human-like generalization.The paper identifies PAID as Pretraining-Agnostic Identically Distributed evaluation.
2 Human-like Generalization
Humans learn language from limited exposure, generalize consistently to rare or unseen structures, and quickly learn new language tasks. Current deep learning systems do not reliably match these abilities.
- Humans generalize linguistic knowledge consistently to structures that are infrequent or absent from corpora.
- Humans quickly learn to do new things with language, including new tasks.
- Current deep learning systems behave unpredictably under distribution shift and require extensive instruction for each new task.
- Human rapid, consistent generalization likely relies on inductive biases from innate building blocks and diverse learning experience.
- Human-like generalizing systems would support both NLP applications and scientific study of language acquisition and processing.
3 Pretraining-Agnostic Evaluation
PAID combines large-scale pretraining, task-specific fine-tuning, and same-distribution testing, rewarding models that exploit extensive data and benchmark correlations. The paper argues this can obscure progress toward human-like sample efficiency.
- GLUE and SuperGLUE aggregate multiple supervised classification tasks into leaderboard scores, including a single average headline score.
- The standard recipe pretrains a transformer-style network on vast denoising data, then fine-tunes it on smaller supervised task datasets.
- 88.9% average accuracy on SuperGLUE for T5 was slightly below untrained humans’ 89.8% and over 20 percentage points above BERT’s earlier score.The paper attributes the jump mainly to scale differences between similar transformer systems and objectives.
- Pretraining corpus sizes increased from BERT’s 3.3 billion words to T5’s approximately 130 billion words, but SuperGLUE rankings omit this resource difference.The corpus estimates use reported gigabytes converted using average English word length and spacing assumptions.
- A single leaderboard conflates architecture, corpus size, and computational resources rather than comparing architectures alone.
- Because humans receive fewer than 100 million words before adulthood, PAID disadvantages models pretrained on human-scale data relative to systems such as T5.
- PAID may favor massive transformers whose performance depends on data quantities orders of magnitude beyond human exposure, unlike harder-to-scale symbolic architectures that perform well on smaller datasets.
4 Identically Distributed Training Set and Test Set
PAID commonly trains and tests on data generated by the same process, which can reward dataset-specific correlations rather than intended human-like generalizations. Evidence from NLI and syntax shows high benchmark accuracy can coexist with systematic errors on controlled examples.
- Identically distributed evaluation usually randomly splits one dataset into training and test portions generated by the same process.
- Same-distribution evaluation can obscure mismatches between human-intended generalizations and the patterns systems actually extract.
- NLI models exploit annotator correlations such as negation words, even though negation does not generally determine contradiction.
- BERT fine-tuned on MultiNLI achieves high in-distribution accuracy but shows little sensitivity to word order on constructed syntactic examples.
- Random splits allow models to reproduce training correlations and achieve high test accuracy while making mistakes humans rarely make on controlled cases.
- BERT’s syntactic behavior varies with fine-tuning initialization and example order, but augmentation with word-order examples can make it more human-consistent.
- Comparisons with human annotators are potentially unfair because models see hundreds of thousands of task-output examples before the test set.
- BERT may ignore syntax in inference in ways humans would likely find difficult to replicate.
5 The Generalization Leaderboard
The paper proposes a parallel Generalization Leaderboard to complement PAID by rewarding sample-efficient, human-like generalization across controlled phenomena, languages, and learning conditions.
- Standard, moderately sized pretraining corpora: Standardize moderately sized pretraining corpora and evaluate learning curves to measure sample efficiency more directly.The paper suggests comparing models using standardized corpora, including child-directed speech from CHILDES.
- Independent evaluation in multiple languages: Evaluate models independently across multiple languages with matched corpora and benchmarks rather than relying primarily on transfer from English.A composite score could summarize average performance across languages while preserving separate training and testing for each language.
- Normative evaluation: Use normative evaluation with expert-created controlled datasets designed to resist heuristics that ignore linguistic principles.The paper also supports continued exploration of adversarial filtering to limit crowdworker shortcuts.
- Test-only benchmarks: Keep benchmark models separate from evaluation data and favor systems that perform transfer tasks with minimal instruction.The proposed test-only approach aims to reduce the chance that fine-tuning exposes models to unintended benchmark correlations.
- Efficiency: Efficiency can be incorporated directly into modified benchmarks through factors such as parameter count, inference speed, and training compute.These implementation-level properties are distinct from the paper’s input–output generalization concerns.
- Breakdown by task and phenomenon: Provide detailed breakdowns by task and linguistic phenomenon so strong performance on specific phenomena is not hidden by average scores.Alternative headline metrics could reward large gains in one category even when offset by small losses elsewhere.
6 Conclusion
The paper argues that PAID favors models trained on unlimited data and tuned to arbitrary dataset statistics, and advocates a parallel evaluation ecosystem for human-like generalization. Such evaluations should reward models that learn language structure and new tasks from limited data while reducing reliance on superficial heuristics and adversarially sensitive behavior.
- PAID selects models that scale easily with unlimited data and capture arbitrary statistical patterns in fine-tuning datasets.
- A parallel evaluation ecosystem should reward human-like generalization from limited data and align model behavior more closely with human expectations.
- Human-like inductive biases may reduce reliance on superficial heuristics and adversarial examples that exploit human-insignificant input changes.