Do Generalisation Results Generalise?

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

TL;DR

Just because a Large Language Model (LLM) generalizes well to one out-of-distribution (OOD) dataset doesn't mean it will perform well on another. By "regressing out" in-domain performance to calculate partial OOD correlations, this paper reveals that generalization is highly inconsistent and depends heavily on the specific model architecture and the nature of the data shift.

What Problem Does This Paper Address?

When we train an LLM on a specific task (like Natural Language Inference), we often test it on "out-of-distribution" (OOD) data—data that looks different from the training set—to see if the model has truly "learned" the task or is just relying on shortcuts and spurious patterns.

The current standard in research is to use a single OOD benchmark. If the model does well, researchers often claim the model is "robust." But is generalization a single, unified capability? If a model handles one type of distribution shift (e.g., adversarial examples), does that mean it can handle another (e.g., a shift in subject matter)? This paper investigates whether "generalization results themselves generalize," or if robustness is a fragmented, idiosyncratic property.

Methods

The authors finetuned two families of models—OPT (ranging from 2.7B to 30B parameters) and OLMo2 (7B to 32B)—on Natural Language Inference (NLI) tasks using a few-shot setup (32, 64, and 128 examples).

To get a clear picture of generalization, they introduced a metric called partial OOD correlations. The logic is as follows:

  1. The Confounder: As a model gets better at its training task (in-domain), it naturally tends to get better at OOD tasks. This "trival" correlation masks whether the model is actually learning a specific type of robustness.
  2. The Solution: The researchers used Generalized Additive Models (GAMs) to "regress out" the in-domain performance. They calculated a "residual"—essentially measuring how much better or worse a model performed on an OOD set relative to what was expected given its in-domain score.
  3. The Correlation: They then measured how these residuals correlated across different OOD test sets. If the correlation is high, it means generalization transfers; if it's low or negative, generalization is inconsistent.

Datasets

The study focused on Natural Language Inference (NLI), using a total of 8 datasets:

  • Training/In-Domain: MNLI and SNLI.
  • OOD Evaluation: The authors evaluated models on whichever of the two above wasn't used for training, plus WNLI, SciTail, RTE, HANS, ANLI, and PAWS.
  • Note on Preprocessing: To ensure consistency across different annotation styles, the authors removed "neutral" labels, focusing on a binary classification (entailment vs. non-entailment).

Key Insights and Ideas

  • Generalization is not a monolith: There is no "universal" generalization factor. A model might show a strong positive correlation between two OOD sets (like ANLI and RTE) in one training run, but a negative correlation in another model family.
  • The "Robustness" Illusion: Naively looking at accuracy tables might suggest that larger models are more robust. However, once you control for in-domain performance, that "robustness" often vanishes. Larger models aren't necessarily better at generalizing; they are just better at the core task.
  • Model Stability Matters: The researchers found that OPT models exhibit "oscillating" OOD performance—their ability to generalize swings wildly during a single training run. In contrast, OLMo2 models were much more stable, suggesting that pre-training recipes significantly impact how "predictable" a model's generalization will be.
  • Negative Correlations: In some cases, becoming "more robust" to one type of shift actually made the model less robust to another. This suggests that different OOD datasets might require the model to rely on conflicting features.

Key Results

  • Inconsistent Trends: The authors observed no overarching trend in generalization results. The existence of a positive or negative correlation between any two OOD test sets depended almost entirely on the specific model being analyzed.
  • Performance Gaps: Even the best models showed massive variability. For example, an OPT 30B model achieved 86.0% accuracy on its in-domain task (MNLI) but plummeted to 49.7% on PAWS (an OOD set involving word scrambling).
  • Spider Web Visualizations: The paper uses "correlation graphs" (Figure 1 and 6) to show the relationships between OOD sets. These graphs look like tangled webs with no consistent pattern, reinforcing the idea that generalization is a complex, non-transferable phenomenon.

Limitations and Future Work

  • Task Specificity: The study is limited to NLI. It remains to be seen if these findings hold for other domains like coding, mathematics, or common-sense reasoning.
  • Model Scale: While 30B-32B parameters are significant, the authors acknowledge they couldn't test "frontier" models (100B+ parameters) due to compute constraints.
  • Data Contamination: Although the authors performed preliminary checks, it is notoriously difficult to guarantee that OOD test sets were not partially leaked into the massive pre-training corpora of these models.
  • Future Direction: The authors suggest that future research should move away from single-testset OOD evaluations and instead adopt a "multi-OOD" approach to provide a more honest assessment of model reliability.

Summary updated

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.
BibTeX
@inproceedings{boglioni2026do,
  author = {Matteo Boglioni and Andrea Sgobbi and Gabriel Tavernini and Francesco Rita and Marius Mosbach and Tiago Pimentel},
  title = {{Do Generalisation Results Generalise?}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2026},
  doi = {10.18653/v1/2026.findings-acl.1284},
}

Loading related content...

Figure