A Composite Rubric-and-Evidence-Checkpoint-Chain Evaluation Framework for LLM Reasoning
Large language model reasoning evaluation heavily relies on manually designed rubrics, which often suffer from inconsistent scoring and weak interpretability as reasoning chains grow longer. This paper presents a composite evaluation framework that combines a structured multi-dimensional rubric with ordered evidence checkpoint chains. Existing paradigms each have a blind spot: rubric-only assessment tends to overlook flawed intermediate reasoning when the final answer is correct, while plain checklist evaluation is vulnerable to superficial keyword matching and cannot enforce sequential logical constraints. The proposed framework unifies two complementary scoring branches, namely a multi-dimensional rubric for holistic quality judgment and ordered evidence checkpoint chains that capture prerequisite relations between reasoning steps, and fuses their normalized scores by a single tunable weight. This work is a conceptual framework accompanied by qualitative illustrative cases rather than large-scale empirical experiments. It aims to provide an auditable evaluation paradigm for long-CoT reasoning models; while large-scale dataset curation and systematic quantitative validation remain reserved for future investigation.
Authors
- Liping Ding (ORCID: https://orcid.org/0009-0004-9929-456X)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-21
- DOI
- https://doi.org/10.5281/zenodo.24412811
- Primary Topic
- Topic Modeling
- Type
- preprint