NumAnchor grounds numeric operands for assembly language models
Original title: Beyond Tokens: NumAnchor for Address-Grounded Numeric Operand Representation in Assembly Language Models
Numeric-like operands were 40.9% to 45.7% of occurrences; retrieval Recall@1 rose by up to 35.0% without changing the host model.
Researchers present NumAnchor, an input adapter that represents recoverable numeric-like operands in assembly as typed program entities. In two corpora those operands were 45.7% and 40.9% of operand occurrences. Descriptors for code, data, and stack are fused into embeddings without lengthening the host sequence or changing training. Frozen probes reached 98.7% hierarchical-location accuracy. Across assembly language model families, retrieval improved by up to 35.0% Recall@1 and 31.7% MRR.
- Jump and call target alignment AUROC scores were 0.932 and 0.907.
- Execution-aware type recovery improved by up to 5.2 F1 points.
- Ablations found complementary gains from code, data, and stack descriptors.
From the original article
1 min read at arxiv.orgAssembly language models (ALMs) are becoming important for automated binary analysis, yet their token-level interfaces often obscure numeric-like operands. In two corpora, such operands account for 45. 7% and 40.
By Kun Liu · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.