Evaluation Contamination
Did the agent learn the test instead of demonstrating transferable capability?
Reference notes by Min · Reviewed October 15, 2026
Definition
Evaluation contamination occurs when material intended to measure performance is exposed during training, tuning or development in a way that weakens the independence of the measurement. Agent systems can also leak test answers through retrieval, memory or evaluator configuration.
An evaluation can look successful because the system encountered its answers before the test. For an agent, the source may be application memory rather than model training. Audit the complete workflow that supplies information during evaluation.
What it covers
- Training exposure
- Possible inclusion in model development.
- Application exposure
- Answers in retrieval or memory.
- Iteration
- Repeated tuning on final evaluation cases.
- Reporting
- Known overlap and remaining uncertainty.
Distinguish known leakage from uncertain prior exposure. Closed model training data may be unavailable, so absolute claims that a public benchmark is unseen can be unsupported. State that limitation and supplement public tasks with controlled workflow-specific evidence.
In practice
Practical design patterns by Min.
- Separate development cases from final decision sets.
- Inspect retrieval sources and remembered answers.
- Document access to evaluation material and changes over time.
Restrict final-set access and preserve a record of development use. Refresh cases thoughtfully while retaining comparable subsets where useful. Do not simply rename entities in a known test and assume the underlying solution is now independent.
Research notes
Source-grounded details and editorial applications. Follow each link for the original document.
Overlap can affect measured capability
Jiang and colleagues study pretraining contamination using controlled GPT-2 experiments and discuss limitations of n-gram overlap definitions. The results support careful assessment, not a single reliable detector for all models.
Source: Investigating Data Contamination for Pre-training Language Models (2024) ↗
Working example
A benchmark answer enters the knowledge base
A team stores evaluation explanations in the same corpus the agent searches.
- Remove final-test answers from the accessible corpus.
- Reassess results using independently held cases.
- Report which earlier measurements were affected.
Why the controls matter
The knowledge-base example makes the answer directly retrievable. Excellent performance then says little about solving a new case. However, retrieving approved guidance may be legitimate in the real task; the evaluation must specify which resources are intentionally allowed.
Common pitfalls
- Assuming exact-string checks detect all overlap.
- Publishing test answers and then claiming the set remained unseen.
- Treating suspected contamination as proof that every result is invalid.
Contamination is not the only reason benchmarks mislead. Unrepresentative cases and weak outcome checks can also inflate confidence. Investigate exposure without discarding the need to examine what the evaluation actually measures.
Questions for operational review
Use these questions to decide what evidence your team needs. They are editorial prompts, not a compliance checklist.
- Which components can access expected answers?
- How often was the final set used during tuning?
- What uncertainty about pretraining exposure is disclosed?
Articles linking here
References
Primary documents reviewed for this entry. Examples and implementation patterns are editorial synthesis.
- Investigating Data Contamination for Pre-training Language Models (2024) ↗Accessed October 15, 2026
- Gebru et al.: Datasheets for Datasets ↗Accessed October 15, 2026