Research map / Large language models / Language, safety and interpretability
Training on the Test Task Confounds Evaluation and Emergence
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 differences between newer and older LLMs on MMLU and GSM8K largely stem from training on the test task, and proposes adjusting comparisons by fine-tuning all models on the same task-specific data before evaluation.
Problem
Benchmark comparisons of base LLMs are confounded by how much models were trained on task-relevant data (training on the test task), which inflates newer models' scores and creates spurious emergence and family superiority.
Approach
Define "training on the test task" as including task-relevant data in pretraining, and adjust for it by fine-tuning every model on the same sufficient task-specific dataset before evaluation; use this to show that model recency, model family differences, and emergence largely vanish or shift after adjustment.
Key results
- Before adjustment, models released after November 2023 outperform older ones at equal pretraining compute by over 7 points on MMLU and 19 points on GSM8K; after fine-tuning all models on task data, the difference is small and not significant.
- Fine-tuning older models on task data recreates the newer-vs-older performance gap (treatment effect similar in magnitude), and a further joint fine-tuning removes it.
- Adjusting shifts GSM8K rankings by 7.7 ranks on average (max 21), and the area of Pareto-frontier improvement since Nov 2023 shrinks sixfold; emergence point for MMLU moves from ~1e22 to ~6e20 FLOPs after 64,000 examples.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Fine-tune all models on the same sufficient task-specific data before evaluation | Older and newer base models · MMLU | accuracy difference by model recency: over 7 accuracy points to small and not statistically significant | No clear effect |
| Fine-tune all models on the same sufficient task-specific data before evaluation | Older and newer base models · GSM8K | accuracy difference by model recency: 19 accuracy points to small and not statistically significant | No clear effect |
| Fine-tune older models on task-relevant data | Older base models · MMLU and GSM8K | accuracy | Helped |
| Further fine-tune both control and treatment groups on the test task | Older control and already fine-tuned older models · MMLU and GSM8K | accuracy difference between groups | No clear effect |
| Reformulate ARC and HellaSwag as MMLU-style multiple-choice questions | All evaluated models · ARC Challenge and HellaSwag | accuracy difference by model recency | Helped |
| Use standard cloze evaluation instead of MMLU-style multiple-choice prompts | All evaluated models · ARC Challenge and HellaSwag | accuracy difference by model recency | No clear effect |
| Fine-tune all models on the MMLU auxiliary training set | All evaluated models · ARC Challenge and HellaSwag | accuracy difference by model recency | No clear effect |
| Fine-tune older models on task-relevant data | Older base models · MMLU | emergence point: around 10^{22} FLOPs to around 6\cdot 10^{20} FLOPs after training on 64,000 examples | Helped |
| Fine-tune older models on task-relevant data | Older base models · MMLU | log-linear fit R^2: 0.63 to 0.95 | Helped |
| Adjust benchmark comparisons for test-task training | Recent model families · Benchmark accuracy per pretraining compute | Pareto-frontier area of improvement: reduces by a sixfold | Hurt |
| Fine-tune older models on task-relevant data | Older models compared with newer models · MMLU | balanced accuracy for discriminating model recency: 64.6% to 52.2% | No clear effect |
| Fine-tune older models on task-relevant data | Older models compared with newer models · GSM8K | balanced accuracy for discriminating model recency: 73.9% to 52.5% | No clear effect |
Surprising findings
- Reformulating ARC and HellaSwag as MMLU-style multiple-choice questions reproduces the newer-vs-older gap seen on MMLU, while cloze evaluation shows none.
- Fine-tuning older models on task data produced a recency-like confound, and the mediated indirect effect through test task training (0.071 MMLU, 0.168 GSM8K) was significant while the direct effect was not.
Limitations
- Adjustment requires substantial task-relevant fine-tuning data and compute for each model, and may be infeasible for proprietary task data.
- Only minimal hyperparameter tuning was feasible; some models degraded after fine-tuning and needed different learning rates.
- Causal interpretation relies on assumptions such as test task training saturating and fine-tuning being similar to pretraining-stage training.
Experimental setup
Datasets: MMLU auxiliary training set MetaMathQA Orca-Math
Benchmarks: MMLU GSM8K ARC Challenge HellaSwag MMLU Pro GPQA BBH MuSR MATH Level 5
Metrics: accuracy Brier score regression coefficient theta (recency effect) rank shift balanced accuracy of discriminator
Baselines: Pythia Llama 2 Qwen 1.5
Builds on: LM Evaluation Harness HF Open LLM Leaderboard scaling laws DeepSpeed ZeRO
Closest papers by meaning
- Understanding Emergent Abilities of Language Models from the Loss Perspective (NeurIPS 2024 · similarity 0.91)
- EvaLearn: Quantifying the Learning Capability and Efficiency of LLMs via Sequential Problem Solving (NeurIPS 2025 · similarity 0.90)
- Predicting Emergent Abilities with Infinite Resolution Evaluation (ICLR 2024 · similarity 0.90)
- The Illusion of Diminishing Returns: Measuring Long Horizon Execution in LLMs (ICLR 2026 · similarity 0.90)
- Mind the Gap: Examining the Self-Improvement Capabilities of Large Language Models (ICLR 2025 · similarity 0.89)
- Raising the Bar: Investigating the Values of Large Language Models via Generative Evolving Testing (ICML 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.