research-algorithm-checker

To verify machine learning algorithm correctness, you can use a multi-agent review workflow that traces code flow from entry to outputs, validates edge cases, and checks numerical logic against your reference design. This process categorizes bugs and optimization findings for academic R&D.

Checking numerical correctness in loss and probability logic involves validating mathematical implementations for thresholds and gradient flow. A dedicated algorithm checker compares your code against papers or pseudocode to pinpoint deviations and identify computation mistakes.

Tracing tensor shape and dtype mismatches requires boundary and edge-case analysis that checks for indexing errors and null handling. An algorithm verification workflow systematically reviews these failure points to ensure your training pipeline runs correctly.

Algorithm code review does help find performance optimization opportunities by identifying wasted computation, memory inefficiency, and cache or parallelization opportunities. It ensures your research code is both correct and efficient before running experiment cycles.

To track and deduplicate bugs found during research code verification, use an ISSUES.md file with todo, doing, done, and discard statuses. This issue tracking method ensures discarded optimization ideas are not repeated during the algorithm review workflow.

You do need to provide a reference paper, pseudocode, or design document to check your training pipeline implementation. Design-to-implementation consistency validation requires comparing your code against these references to pinpoint exact deviations.