Intel / article

ExecuCritic: Calibrated Critic Shaping for Code Generation with Verifiable Rewards

Source: arxiv.org
Author: Junjie Cao, Yingjie He
Date: Published Sep 16, 2026

Context

A failed code test tells a training system that something went wrong, but a single pass-or-fail result offers little guidance about which behavior to improve. ExecuCritic trains a code generator and a critic on the same program executions. The critic predicts test outcomes and supplies short diagnostic feedback; the generator uses that signal only when the critic agrees with execution results for the current group. Across eight benchmarks and two model backbones, the authors report improvements over several comparison methods with fewer training steps and executions. The mechanism makes critic feedback conditional on demonstrated agreement with tests.

Key takeaways · AI-distilled

A key limitation of training code models with execution feedback is that a whole program is reduced to a single pass or fail signal, making it hard for reinforcement learning to know which part of the code to credit or blame.
ExecuCritic jointly trains a coder and a critic on the same execution rollouts, sampled attempts at running the generated code; the critic predicts pass or fail and gives brief diagnostic feedback that the coder uses only when the critic agrees with the actual test outcome.
Across eight code benchmarks and two open model backbones, the authors report ExecuCritic outperformed training without a critic, prompted reviewer systems, and scalar reward model baselines, while using fewer training steps and fewer sandbox executions.
The authors' ablation and reliability analyses suggest the performance gains come from better credit assignment rather than from running more sampling during training.

Terms in this piece · Glossary

sandbox — An isolated environment where AI-generated code or agent actions run without being able to touch anything real.
calibration — How well a model's confidence matches reality — a calibrated model saying "90% sure" is right about 90% of the time.

Read the source arxiv.org