Evaluate Shared Residual Distillation on open-ended tasks

Investigate whether Shared Residual Distillation is effective on open-ended tasks beyond the mathematical-reasoning benchmarks evaluated in the paper.

Background

The empirical evaluation focuses on mathematical reasoning using AIME, HMMT, and MATH-500, with IFEval and MMLU-Pro used to assess retention of broader capabilities. The paper does not establish whether shared residual supervision transfers to open-ended tasks, where correctness may be difficult to verify and useful local guidance may have different characteristics.