Do LLM-judge biases compose? A label-free, scale-aware test (SSIT) for position×length interaction in LLM-as-a-judge pipelines, with proofs, tie-coding and curvature diagnostics, 7-judge/6-family data, and the TMLR manuscript.
reproducible-research bias-correction difference-in-differences ssit length-bias llm-evaluation tmlr llm-as-judge randomization-inference position-bias verbosity-bias
-
Updated
Oct 20, 2026 - Python