INSIGHT: Automatic Generation of Explanations for Efficient Identification of Hardware Bugs and Underspecifications
Vincent Quentin Ulitzsch, Alessandro Bertani, Peter W. Deutsch, David Langus Rodriguez, Kelly Xu, Aarti Gupta, Sharad Malik, Mengjia Yan
Abstract
Hardware verification is a critical task in processor development. Functional correctness bugs that escape verification can manifest as serious security vulnerabilities, allowing attackers to violate control and dataflow integrity. A common verification strategy to eliminate functional correctness bugs during design time is to compare the architectural behavior of a device under test, i.e., the hardware implementation, against a golden reference implementation, using model checking or fuzzing to expose mismatches. In practice, a single bug often manifests in many mismatch-triggering programs. Moreover, due to ISA underspecification, not every mismatch corresponds to a bug. Without additional guidance, model checkers repeatedly rediscover programs that trigger the same bug or ISA underspecification, while fuzzers generate large numbers of triggering programs that stem from the same underlying cause, causing substantial manual overhead. This paper presents INSIGHT, a framework that automatically synthesizes generalized, machine-readable explanations of mismatches between a processor implementation and its reference model. INSIGHT (i) guides model checkers towards exposing distinct bugs and underspecifications and (ii) clusters large program corpora by shared explanations, enabling automatic deduplication. INSIGHT translates explanation synthesis into a combinatorial optimization problem and solves it efficiently using Integer Linear Programming solvers. We implement and evaluate INSIGHT on an in-order core and the out-of-order BOOM RISC-V core. We show that with INSIGHT, model checkers can discover more distinct bugs in the same timeframe. INSIGHT can also reduce a large corpus of fuzzing testcases into a small number of clusters.
Ask Lune
Know where this paper stands in minutes, not an afternoon.
Lune can read the top-tier papers that cite it and the ones closest to it, and tells you what it claims, what came after and where it falls short, each point cited to the paper it comes from. No skimming, no AI slop.
Related papers
- GoldenFuzz: Generative Golden Reference Hardware FuzzingLichao Wu, Mohamadreza Rostami, Huimin Li, Nikhilesh Singh et al.NDSS 2026 · 3 citations
- Encarsia: Evaluating CPU Fuzzers via Automatic Bug InjectionMatej Bölcskei, Flavien Solt, Katharina Ceesay-Seitz, Kaveh RazaviUSENIX Security 2025
- Interplay of Efficient Model Checking and Secure Processor Design: A Case Study on Secure SpeculationTingzhen Dong, Qinhan Tan, Kunpeng Wang, Thomas Bourgeat et al.S&P 2026
- Lost and Found in Speculation: Hybrid Speculative Vulnerability DetectionMohamadreza Rostami, Shaza Zeitouni, Rahul Kande, Chen Chen et al.DAC 2024 · 6 citations
- HARTBREAKER: Deterministic Fuzzing of Multi-Hart RISC-V CPUs with Non-Deterministic ProgramsQuentin Bordier, Tobias Kovats, Flavien Solt, Kaveh RazaviISCA 2026
