Shortcuts or Memorisation? Revisiting Data Corruption in Natural Language Understanding with Large Language Models
Organizations: University of Helsinki · Institute of Language and Speech Processing Athena Research Centre, Athens
Abstract
Pre-trained encoders fine-tuned on natural language inference (NLI) and other language understanding tasks keep much of their accuracy after the input is perturbed, e.g. by removing all nouns or shuffling the words. Does the same hold for large language models (LLMs), and if so, is it because they have seen the benchmarks during pre-training? In this paper we revisit these perturbation experiments with eight LLMs alongside fine-tuned encoders on GLUE and the three rounds of ANLI. Regardless of model size, every system that answers directly retains 36-72% of its above-chance performance without nouns on MNLI compared to only 14-29% on ANLI. Only a model that reasons before answering departs from this pattern. We also check contamination indicators (including exact -gram counts in the public pre-training data of two of the models) and find no link between exposure and robustness. Our strongest evidence comes from 368 new, human-validated items that none of the models can have seen. All eight LLMs keep more of their performance without nouns on items written in MNLI's style than on items written without its cues. When the perturbation instead removes the function words carrying such cues, this advantage reverses. In light of these results, we argue that robustness to perturbation on familiar NLI benchmarks reflects dataset shortcuts rather than memorisation, and tells us more about the benchmark than about the model.
Figures & tables
| MNLI | RTE | ANLI (R1–R3 average) | |||||||
| System | Clean | noun | verb | Clean | noun | verb | Clean | noun | verb |
| RoBERTa (fine-tuned) ‡ | 87.8 | 72 | 81 | 70.8 | 19 | 85 | 56.1 | 24 | 66 |
| Qwen3.8-27B | 88.9 | 63 | 74 | 91.3 | 40 | 73 | 75.2 | 29 | 60 |
| OLMo 3 32B | 89.3 | 70 | 79 | 90.6 | 30 | 76 | 61.6 | 19 | 49 |
| OLMo 3 7B | 88.8 | 68 | 76 | 87.0 | 40 | 66 | 51.4 | 14 | 50 |
| Muse Glimmer 30B | 67.2 | 47 | 66 | 80.9 | 29 | 76 | 55.6 | 25 | 56 |
| Clean | no nouns | nouns+verbs | Easy | ||||
|---|---|---|---|---|---|---|---|
| Model | MNLI | free | MNLI | free | MNLI | free | hard |
| Qwen3.8-27B | 100.0 | 97.3 | 72 | 47 | 49 | 71 | +16.0 ∗ |
| OLMo 3 32B | 99.4 | 98.0 | 75 | 57 | 50 | 62 | +10.9 |
| OLMo 3 7B | 98.2 | 94.6 | 73 | 54 | 59 | 74 | +17.8 ∗ |
| Muse Glimmer 30B | 98.8 | 98.0 | 54 | 45 | 50 | 68 | +18.5 ∗ |
| GPT-4o | 98.2 | 94.6 | 60 | 51 | 56 | 69 | +17.1 ∗ |
Appendix figures & tables8 assets
Supplementary material from the paper’s appendix.
Appendix
| Model | Task | Clean | Corrupt-Test | Corrupt-Train | Corrupt-Train-and-Test |
|---|---|---|---|---|---|
| RoBERTa-base | MNLI | 87.9 / 87.8 | 72.8 / 72.8 | 85.6 / 85.8 | 77.5 / 77.7 |
| RoBERTa-base | RTE | 70.0 / 70.8 | 53.8 / 54.0 | 47.3 / 62.9 | 47.3 / 56.7 |
| RoBERTa-base | MRPC | 88.7 / 88.1 | 82.3 / 79.6 | 86.3 / 82.0 | 80.2 / 82.3 |
| RoBERTa-base | CoLA | 64.0 / 58.8 | 17.8 / 15.3 | 39.7 / 41.3 | 34.3 / 34.9 |
| RoBERTa-base | STS-B | 90.1 / 89.9 | 56.1 / 55.8 | 81.7 / 81.8 | 63.5 / 63.3 |
| BERT-base | MNLI | 83.7 / 83.9 | 69.8 / 70.2 | 80.7 / 81.3 | 73.4 / 73.3 |
| R1 | R2 | R3 | |
|---|---|---|---|
| Clean | 73.8 / 73.5 | 48.9 / 49.3 | 44.4 / 45.5 |
| Corrupt-Test NOUN | 43.7 / 46.5 | 36.0 / 34.7 | 32.4 / 33.8 |
| Corrupt-Test VERB | 54.6 / 57.2 | 44.7 / 42.3 | 39.3 / 39.0 |
| Model | Examples | MNLI | RTE | ANLI R1 | ANLI R2 | ANLI R3 | QNLI |
|---|---|---|---|---|---|---|---|
| OLMo 3 7B | 0-shot | 88.3 (68) | 89.2 (40) | 63.0 (25) | 50.1 (3) | 47.5 (13) | 82.0 (56) |
| 5-shot, clean | 88.4 (67) | 88.5 (35) | 66.2 (28) | 50.9 (18) | 50.2 (18) | 87.0 (56) | |
| 5-shot, drop-NOUN | 87.5 (67) | 88.5 (33) | 63.6 (25) | 49.4 (12) | 49.6 (20) | 86.5 (57) | |
| OLMo 3 32B | 0-shot | 90.8 (70) | 91.3 (30) | 74.0 (23) | 61.1 (19) | 56.8 (17) | 88.5 (44) |
| 5-shot, clean | 89.0 (72) | 92.1 (33) | 75.0 (31) | 64.4 (19) | 60.8 (17) | 91.0 (56) | |
| 5-shot, drop-NOUN | 89.2 (72) | 91.7 (34) | 75.1 (31) | 62.8 (22) | 60.8 (18) | 90.8 (59) |
| Task | Model | Acc | Cen | AUC | Top answer |
|---|---|---|---|---|---|
| MNLI | Qwen3.8-27B | -25.9 | -20.4 | -13.4 | 35 57 |
| OLMo 3 32B | -19.5 | -17.5 | -10.1 | 38 44 | |
| OLMo 3 7B | -19.8 | -17.4 | -10.4 | 35 47 | |
| Muse Glimmer 30B | -31.9 | -25.7 | -21.3 | 63 96 | |
| ANLI R1 | Qwen3.8-27B | -32.3 | -32.8 | -27.2 | 39 59 |
| OLMo 3 32B | -31.1 | -31.5 | -29.8 | 47 52 |
| Perturbation | Qwen3.8 | OLMo 32B | OLMo 7B | Glimmer | GPT-4o | DeepSeek |
|---|---|---|---|---|---|---|
| drop-NOUN | 63/29 | 70/19 | 68/14 | 47/25 | 56/26 | 49/25 |
| drop-VERB | 74/60 | 79/49 | 76/50 | 66/56 | 71/64 | 64/58 |
| drop-ADJ | 78/70 | 82/72 | 80/64 | 77/70 | 79/75 | 78/69 |
| drop-ADV | 74/82 | 77/83 | 74/84 | 73/82 | –/– | –/– |
| drop-DET | 92/93 | 92/91 | 93/93 | 91/93 | 91/95 | 92/95 |
| drop-PRON | 97/99 | 97/97 | 97/92 | 94/97 | 99/97 | 96/99 |
| Model | Task | Clean | No nouns | No verbs | Shuffled | Hyp.-only |
|---|---|---|---|---|---|---|
| RoBERTa-base | CoLA | 58.8 | 15.3 | 4.4 | 1.7 | – |
| MRPC | 88.1 | 79.6 | 84.2 | 75.2 | 31.6 | |
| RTE | 70.8 | 54.0 | 67.6 | 57.2 | 48.0 | |
| STS-B | 89.9 | 55.8 | 85.3 | 85.2 | 5.1 | |
| BERT-base | CoLA | 56.4 | 16.2 | 4.4 | 1.1 | – |
| MRPC | 81.9 | 77.8 | 80.1 | 73.5 | 31.6 |
| TS-guessing | Min-K%++ | infini-gram | ||||
|---|---|---|---|---|---|---|
| Task | copy | no copy | Qwen | OLMo | Prem. | Hyp. |
| MNLI | 56.3 | 29.0 | -1.89 | -1.88 | 21 | 18/448 |
| RTE | 66.4 | 32.4 | -1.78 | -1.49 | 24 | 3/25 |
| ANLI R1 | 67.5 | 34.0 | -1.79 | -1.07 | 72 | 10/273 |
| ANLI R2 | 69.2 | 31.3 | -1.84 | -1.03 | 71 | 5/218 |
| ANLI R3 | 59.2 | 26.6 | -2.18 | -1.49 | 34 | 3/228 |
| No evidence | No nouns | Nouns+verbs | ||||
|---|---|---|---|---|---|---|
| Model | MNLI | BioASQ | MNLI | BioASQ | MNLI | BioASQ |
| Qwen3.8-27B | 26 | 83 | 72 | 97 | 44 | 95 |
| OLMo 3 32B | 51 | 77 | 79 | 94 | 47 | 89 |
| OLMo 3 7B | 41 | 64 | 78 | 88 | 43 | 83 |
| GPT-4o | 27 | 89 | 60 | 98 | 38 | 92 |
| DeepSeek-V4-Pro | 29 | 88 | 55 | 98 | 40 | 96 |
Language Model Robustness
Language model robustness studies how reliably language models preserve correct, safe behavior under input changes, distribution shifts, and adversarial pressure.
See all 514 papers