Driftbench

DriftBench is a benchmark for measuring whether large language models maintain fidelity to structured research briefs across multiple turns of iterative pressure. The dataset, created by 'driftbench', was last updated on April 25, 2026. It reveals a recall-adherence dissociation where models can restate constraints they simultaneously violate.

Use Cases

• Benchmarking LLM fidelity to hard constraints based on structured research briefs mentioned in the description
• Analyzing recall-adherence dissociation in model behavior across iterative dialogue turns
• Evaluating trajectory drift in multi-turn LLM-assisted scientific ideation workflows

Strengths

• Focuses on a specific and emerging evaluation problem: trajectory drift in multi-turn LLM interactions
• Benchmark design targets a measurable phenomenon: the dissociation between constraint recall and adherence

Limitations

• Column-level documentation is absent; field semantics must be inferred after download
• Row count is unknown, which may limit suitability assessment
• Description metadata is limited; actual data quality requires manual inspection after download

Provenance

Source

huggingface

Collection Method

Likely constructed for benchmark evaluation purposes.

Time Range

null

Freshness

Last updated 2026-04-25 18:29:26; freshness should be verified

Geography

null