LLM Performance on Adversarial Physics Traps

LLM Physics Law-Breaker Benchmark Results evaluate how 21 large language models perform against 34 adversarial physics-based reasoning traps. The dataset contains benchmark scores assessing model robustness to logical inconsistencies and physical fallacies. The original author and creation date are unknown.

Use Cases

• Compare performance scores across 21 different LLM architectures on physics reasoning tasks.
• Analyze failure patterns across 34 distinct adversarial physics trap categories to identify model weaknesses.
• Benchmark new LLM versions against established model scores from this evaluation suite.
• Rank models based on aggregate scores from the physics law-breaker test suite.

Strengths

• Benchmark covers 21 distinct models for comparative analysis.
• Evaluation includes 34 different adversarial physics scenarios.

Limitations

• Specific sample size, row count, and detailed scoring columns are unknown.
• Lacks metadata on model versions, evaluation parameters, and raw response data.

Provenance

Collection Method

Adversarial benchmark testing of pre-trained LLMs.