Browse Teach Curate

Dynamic Scaling of Unit Tests for Code Reward Modeling

Research contribution: scaling the number of LLM-generated unit tests improves the pass/fail reward signal used to rerank and select correct code solutions (with larger gains on harder problems), delivered via CodeRM-8B, a fine-tuned lightweight unit-test generator, plus a difficulty-aware dynamic mechanism that allocates more unit tests to harder problems.

Data Text

Dynamic Scaling of Unit Tests for Code Reward Modeling Zeyao Ma1,3†, Xiaokang Zhang1,3, Jing

This NLP/ML research paper (cs.CL) improves how correct solutions are selected when a large language model generates code. A common pipeline samples many candidate programs and reranks them with LLM-…

Dynamic Scaling of Unit Tests for Code Reward Modeling

Dynamic Scaling of Unit Tests for Code Reward Modeling Zeyao Ma1,3*†, Xiaokang Zhang1,3*, Jing

Dynamic Scaling of Unit Tests for Code Reward Modeling Zeyao Ma1,3†, Xiaokang Zhang1,3, Jing