Research map / Large language models / Language, safety and interpretability
Task Contamination: Language Models May Not Be Few-Shot Anymore
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Shows that LLMs (GPT-3 series and open models) beat majority baselines more often on datasets released before their training data collection, and extracts task examples and membership-inference evidence suggesting task contamination in zero/few-shot evaluation.
Problem
Zero-shot and few-shot evaluations of LLMs may be inflated by task contamination, where task training examples appear in pre-training data, but this has not been systematically examined.
Approach
Uses chronology: compare LLM performance on datasets released before versus after each LLM's training-data collection date, controlling for difficulty, and complements this with training data inspection, task example extraction by prompting, and an exact-match membership inference check on Spider.
Key results
- Across models, datasets released before LLM training-data collection beat the majority baseline more often than post-collection datasets (statistically significant at 99% by Mann-Whitney U for zero and few-shot).
- For post-2021 tasks with no extracted task examples, only 1 of 51 model/dataset combinations (davinci-001 on MTSC-RW) significantly beat the majority baseline.
- On Spider, the number of exact-match generations correlates strongly with execution accuracy (R=0.88).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Evaluate on datasets released before rather than after the model's training-data collection date | LLMs across 16 classification datasets · Pre-collection vs post-collection datasets | percentage of model/dataset combinations above majority baseline | Helped |
| Evaluate on pre-collection rather than post-collection datasets | Individual LLMs · Pre-collection vs post-collection datasets | performance relative to majority baseline | Helped |
| Add task examples to instruction-tuning data | Alpaca and Vicuna relative to LLaMA · Tasks with more than one task example | accuracy | Helped |
| Use GPT-3.5-T rather than davinci on Spider | GPT-3 series · Spider | execution accuracy: R = 0.88 | Mixed |
| Evaluate classification tasks with no demonstrated possibility of task contamination | LLMs in zero- and few-shot settings · Post-collection datasets with no extracted task examples | statistically significant improvements over majority baseline: 1 out of 51, or $2\%$ | No clear effect |
| Advance from davinci-001 to davinci-002 | GPT-3 series · Pre-2021 datasets, zero-shot | performance | Helped |
| Advance from davinci-001 to davinci-002 | GPT-3 series · Post-2021 datasets, zero-shot | performance | Hurt |
Surprising findings
- Non-instruction-tuned open models rarely beat majority baselines on classification tasks, and GPT-3 series performance on pre-2021 datasets rose sharply over versions.
- Instruction tuning alone does not explain GPT-3 gains: davinci-001 to davinci-002 improved on pre-2021 datasets but declined on post-2021 datasets.
Limitations
- Chronological analysis has high recall but low precision; performance gains may stem from other factors.
- Training data inspection and task example extraction have low recall and cannot prove absence of contamination.
- Membership inference only applies to generation tasks, so it was run on a single task (Spider), and the effect on performance was not quantified.
Experimental setup
Datasets: RTE WNLI COPA SST-2 MRPC QNLI CB WiC BoolQ StrategyQA NewsMTSC-MT NewsMTSC-RW
Benchmarks: RTE WNLI COPA SST-2 MRPC QNLI CB WiC BoolQ StrategyQA NewsMTSC-MT NewsMTSC-RW
Metrics: accuracy majority-baseline comparison statistical significance (t-test, Mann-Whitney U, sign test) execution accuracy exact-match count
Baselines: majority baseline LLaMA
Builds on: GLUE SuperGLUE in-context learning
Closest papers by meaning
- Overestimation in LLM Evaluation: A Controlled Large-Scale Study on Data Contamination’s Impact on Machine Translation (ICML 2025 · similarity 0.91)
- Time Travel in LLMs: Tracing Data Contamination in Large Language Models (ICLR 2024 · similarity 0.91)
- How Much Can We Forget about Data Contamination? (ICML 2025 · similarity 0.91)
- Fixing It in Post: A Comparative Study of LLM Post-Training Data Quality and Model Performance (NeurIPS 2025 · similarity 0.91)
- Training on the Benchmark Is Not All You Need (AAAI 2025 · similarity 0.89)
- Rethinking LLM Unlearning Objectives: A Gradient Perspective and Go Beyond (ICLR 2025 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.