AI for Code Review

ai-for-code-review uses AI to automate and enhance the evaluation of software code.

AI for Smart Contract Audits

AI can perform security audits for smart contracts, addressing critical security challenges in decentralized web services Decoupled Smart Contract Audits.
These audits can include severity evaluations and actionable remediation recommendations Decoupled Smart Contract Audits.
Frameworks can decouple comprehensive audit tasks into vulnerability detection, explanation, severity classification, and remediation recommendation Decoupled Smart Contract Audits.
Lightweight open-source-models (0.6B-4B parameters) can be optimized for efficiency in these tasks Decoupled Smart Contract Audits.
Techniques like Rank-Stabilized Low-Rank Adapters (rsLoRA), knowledge distillation, and Chain-of-Verification (CoVe) aggregation can maintain high accuracy with smaller models Decoupled Smart Contract Audits.
This decoupled approach can outperform larger models (7B to 34B parameters) in vulnerability detection and generative explanation tasks Decoupled Smart Contract Audits.
Research has identified a "severity centrality bias" in LLM-assisted auditing Decoupled Smart Contract Audits.

SAIL is a framework that uses LLMs to synthesize globally sound abstract transformers for abstract interpreters SAIL: Sound Abstract Interpreters with LLMs.
This automates the construction of sound abstract interpreters, which is a bottleneck in abstract interpretation for safely approximating program behaviors SAIL: Sound Abstract Interpreters with LLMs.
The synthesis task is formalized as a constrained optimization problem SAIL: Sound Abstract Interpreters with LLMs.
SAIL employs a mathematically grounded cost function to measure the degree of unsoundness of generated candidate transformers, while enforcing syntactic and semantic validity SAIL: Sound Abstract Interpreters with LLMs.
It combines model generation, syntactic and semantic validation, and cost-function-based refinement SAIL: Sound Abstract Interpreters with LLMs.
SAIL can synthesize non-trivial sound abstract transformers across diverse abstract domains, focusing on neural network verification SAIL: Sound Abstract Interpreters with LLMs.
Evaluation results indicate SAIL matches the performance of manually designed transformers SAIL: Sound Abstract Interpreters with LLMs.
It can also synthesize sound and high-precision transformers for complex non-linear operators that do not exist in the literature SAIL: Sound Abstract Interpreters with LLMs.