Presentation
Late Breaking Results: Astfusion -- Two-Stage Structural Enhancement Learning for Robust HLS Code Generation
DescriptionRecent studies show that abstract syntax trees (ASTs) can improve the syntactic validity of large language model (LLM)-generated high-level synthesis (HLS) code. However, structure-enhanced methods often degrade at inference due to reliance on explicit structural signals during training. We show that robustness depends on how structure is introduced, rather than simply whether it is used. We propose ASTFusion, a learn--then--utilize framework that decouples structural acquisition from structural conditioning. ASTFusion internalizes hierarchical and dependency-aware structure through reconstruction from incomplete implementations, and exploits partial structural cues as optional guidance during generation. On HLSEval, ASTFusion improves Func@1 from 40.00% to 72.34% over standard HLS fine-tuning and from 57.69% to 72.34% over prior AST-enhanced fine-tuning, while reaching 98.71% Synth@10. These results show that progressive structural enhancement is effective for robust HLS code generation under realistic inference constraints.
Event Type
Late Breaking Results
TimeMonday, July 276:34pm - 6:37pm PDT
LocationExhibit Hall