← Back to Publications

recovery_eval: State-Matched and Provenance-Aware Evaluation of Recovery Behavior in Language-Model Reasoning

Sham Satish Thakare

Abstract

Evaluating whether a reasoning language model can autonomously recover from intermediate verification errors requires isolating true recovery capacity from unconditional generative performance. In this work, we propose recovery_eval, a state-matched continuation evaluation protocol that controls for frozen structural covariates across reasoning rollouts. By benchmarking state-contrast continuations ($D_{\text{recovery}} = -0.1100$), we reveal significant discrepancies between nominal self-correction claims and true trajectory recovery under covariate controls.

Artifacts & Code

BibTeX

@inproceedings{thakare2026recoveryeval,
  author    = {Thakare, Sham Satish},
  title     = {recovery\_eval: State-Matched and Provenance-Aware Evaluation of Recovery Behavior in Language-Model Reasoning},
  booktitle = {11th IEEE Special Session on Machine Learning on Big Data (IEEE BigData)},
  year      = {2026},
  note      = {Under Review, Reference: BigD497},
  url       = {https://github.com/shamddd/recovery_eval}
}