Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Do Generalisation Results Generalise?

Matteo Boglioni, Andrea Sgobbi, Gabriel Tavernini, Francesco Rita, Marius Mosbach, Tiago Pimentel

TL;DR

This paper investigates whether out-of-distribution (OOD) generalisation results transfer across different distribution shifts in large language models by analyzing partial correlations across multiple OOD testsets while controlling for in-domain performance. Using OPT and OLMo2 in few-shot, pattern-based finetuning on 8 NLI datasets (with SNLI/MNLI as in-domain and seven OOD datasets), the authors compute residuals after regressing OOD performance on in-domain performance and examine cross-OOD correlations via GAM-based regressors. The key finding is that there is no universal trend: correlations between OOD generalisations are highly model- and dataset-dependent, sometimes positive, sometimes negative, and often unstable over training, underscoring the need for evaluating across multiple OOD testsets. The work highlights the instability and complexity of cross-OOD generalisation, cautioning against conclusions drawn from single-testset evaluations and advocating broader, more diverse benchmarking for deployment readiness.

Abstract

A large language model's (LLM's) out-of-distribution (OOD) generalisation ability is crucial to its deployment. Previous work assessing LLMs' generalisation performance, however, typically focuses on a single out-of-distribution dataset. This approach may fail to precisely evaluate the capabilities of the model, as the data shifts encountered once a model is deployed are much more diverse. In this work, we investigate whether OOD generalisation results generalise. More specifically, we evaluate a model's performance across multiple OOD testsets throughout a finetuning run; we then evaluate the partial correlation of performances across these testsets, regressing out in-domain performance. This allows us to assess how correlated are generalisation performances once in-domain performance is controlled for. Analysing OLMo2 and OPT, we observe no overarching trend in generalisation results: the existence of a positive or negative correlation between any two OOD testsets depends strongly on the specific choice of model analysed.

Do Generalisation Results Generalise?

TL;DR

This paper investigates whether out-of-distribution (OOD) generalisation results transfer across different distribution shifts in large language models by analyzing partial correlations across multiple OOD testsets while controlling for in-domain performance. Using OPT and OLMo2 in few-shot, pattern-based finetuning on 8 NLI datasets (with SNLI/MNLI as in-domain and seven OOD datasets), the authors compute residuals after regressing OOD performance on in-domain performance and examine cross-OOD correlations via GAM-based regressors. The key finding is that there is no universal trend: correlations between OOD generalisations are highly model- and dataset-dependent, sometimes positive, sometimes negative, and often unstable over training, underscoring the need for evaluating across multiple OOD testsets. The work highlights the instability and complexity of cross-OOD generalisation, cautioning against conclusions drawn from single-testset evaluations and advocating broader, more diverse benchmarking for deployment readiness.

Abstract

A large language model's (LLM's) out-of-distribution (OOD) generalisation ability is crucial to its deployment. Previous work assessing LLMs' generalisation performance, however, typically focuses on a single out-of-distribution dataset. This approach may fail to precisely evaluate the capabilities of the model, as the data shifts encountered once a model is deployed are much more diverse. In this work, we investigate whether OOD generalisation results generalise. More specifically, we evaluate a model's performance across multiple OOD testsets throughout a finetuning run; we then evaluate the partial correlation of performances across these testsets, regressing out in-domain performance. This allows us to assess how correlated are generalisation performances once in-domain performance is controlled for. Analysing OLMo2 and OPT, we observe no overarching trend in generalisation results: the existence of a positive or negative correlation between any two OOD testsets depends strongly on the specific choice of model analysed.

Paper Structure

This paper contains 25 sections, 1 equation, 14 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Out-of-distribution Generalisation in Language Models
  3. How to evaluate OOD generalisation?
  4. Do finetuned models generalise?
  5. Measuring Correlations between OOD Generalisations
  6. Experimental Setup
  7. Task.
  8. Models.
  9. Data.
  10. Results
  11. Finetuned models tend to generalise, but not everywhere.
  12. OPT's generalisation performance oscillates, but OLMo2's doesn't.
  13. Generalisation's generalisation is complicated.
  14. Conclusions
  15. Pattern-based finetuning details
  16. ...and 10 more sections

Figures (14)

  • Figure 1: OLMo2's partial OOD correlations on SNLI (top) and MNLI (bottom). No clear trends are observed. Edge thickness increases with absolute correlation value. Legend: 1a.MNLI, 1b.SNLI, 2.WNLI, 3.SciTail, 4.RTE, 5.HANS, 6.ANLI, and 7.PAWS.
  • Figure 2: Accuracy ($y$-axis) across training steps ($x$-axis) of OPT (top) and OLMo2 (bottom) for a single finetuning run on MNLI (left) and SNLI (right). Legend: MNLI, SNLI, WNLI, RTE, SciTail, ANLI, HANS and PAWS.
  • Figure 3: Partial correlations of OPT (top) and OLMo2 (bottom) across model sizes (ordered from left to right) trained on MNLI (left) and SNLI (right). All these correlations are obtained by fitting a GAM regressor over 3 independent training runs. See a larger version of this plot in \ref{['fig:corr_gam_4x4_128']} (in \ref{['app:detailed_results']}).
  • Figure 4: Partial correlations averaged across all OOD testset pairs for OPT and OLMo2 with different sizes.
  • Figure 5: Few-shots results throughout a finetuning run on either MNLI or SNLI. OPT OOD performances (first rows) frequently oscillate during training; OLMo2 OOD performances (second rows) are relatively stable across training. Legend: MNLI, SNLI, WNLI, RTE, SciTail, ANLI, HANS and PAWS
  • ...and 9 more figures