Step-TP: A Grounded, Step-Level Dataset with Chain-of-Thought Reasoning for LLM-Guided Tensor Program Optimization

Researchers introduced Step-TP, a post-training dataset for LLM-guided tensor program optimization.

It breaks optimization into atomic, verifiable steps over intermediate states using a token-efficient IR and explicit state transitions.

The dataset adds grounded chain-of-thought supervision and strategy filtering to better support multi-step optimization.

The goal is to make automated code and performance tuning more reliable, interpretable, and easier to verify.