EvalAwareBench: A Factor-Controlled Benchmark for Language Model Evaluation Awareness

A factor-controlled benchmark for studying evaluation awareness in language models, where eight psychology-grounded trigger factors can be independently manipulated. The dataset was created by researchers from ETH Zürich, the Max Planck Institute for Intelligent Systems, and other institutions, and was last updated on the platform in May 2026.

Use Cases

• Studying evaluation awareness in language models based on the described psychology-grounded trigger factors.
• Conducting controlled experiments on model behavior based on the factor-controlled benchmark design.
• Benchmarking language model robustness to specific psychological triggers as described in the dataset summary.

Strengths

• Designed with eight psychology-grounded trigger factors that can be independently manipulated, suggesting a structured experimental design.
• Created by a multi-institutional research team including ETH Zürich and the Max Planck Institute for Intelligent Systems.
• Last updated on the platform in May 2026, indicating recent maintenance.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment for large-scale training.
• The full description is hosted externally, requiring a click-through for complete documentation.

Provenance

Source

aisa-group on Hugging Face

Collection Method

Likely constructed as a controlled benchmark for research purposes, as described.

Freshness

Last updated 2026-05-20 14:01:40