arXiv:2509.26476v2 Announce Type: replace-cross Abstract: We study code-to-metric regression: predicting numeric outcomes of code executions, a challenging task due to the open-ended nature of programming languages. While prior methods have resorted to heavy and domain-specific feature engineering, we show that a single unified Regression Language Model (RLM) using a frozen LLM encoder can simultaneously predict directly from text, (i) the memory footprint of code across multiple high-level languages such as Python and C++, (ii) the latency of Triton GPU kernels, and (iii) the accuracy and spe
The proliferation of Large Language Models (LLMs) allows for novel applications in performance prediction and code optimization, building on existing capabilities in code generation and analysis.
This development allows for more accurate and unified prediction of code resource consumption and performance, critical for efficient software development, infrastructure planning, and cost management in the age of AI.
Instead of domain-specific feature engineering, a single LLM-based system can now predict diverse code metrics across multiple programming languages and hardware, streamlining development and research processes.
- · Software developers
- · Cloud providers
- · AI/ML engineers
- · Hardware designers
- · Manual performance tuning experts
- · Domain-specific code analysis tool vendors
Reduced compute costs and improved efficiency in software deployment and execution.
Faster innovation cycles in areas reliant on code optimization, such as AI model development and high-performance computing.
Potential for autonomous code self-optimization and self-repair systems, further abstracting human intervention in software engineering.
This signal links to a primary source. Continuum Brief monitors and indexes it as part of the live intelligence stream — we do not republish source content.
Read at arXiv cs.LG