Survey traces binary type inference from heuristics to Transformers

Researchers from multiple institutions published a survey on type inference from stripped binaries, covering techniques from early heuristics to modern deep learning. Stripped binaries lack symbol tables and debugging info, making reverse engineering difficult. The paper tracks the shift from 'duck typing' and constraint-solving (e.g., BITY, BinSub) to context-aware neural models (EKLAVYA, CATI) and finally to Transformer and GNN-based architectures (SeeType, TYGR).

Core challenges include semantics loss from compiler optimizations and difficulty recovering structural types (arrays, structs). The survey proposes future directions in neuro-symbolic inference, combining neural pattern recognition with symbolic reasoning for more robust type recovery. This work is foundational for advancing decompilers, vulnerability analysis, and malware understanding.