Lingjun Liu

Random fuzzing rarely reaches the code paths where real compiler bugs live — but thousands of past bug reports show exactly which paths those are. IssueMut mines 1,760 historical GCC and LLVM bug reports into structured test-case pairs, then runs a multi-agent LLM pipeline — generation, validation, and iterative-revision agents — that synthesizes 587 fuzzing mutators from real past bugs. It found 65 new bugs, 60 of which GCC and LLVM maintainers confirmed or fixed.

Deep-learning compilers can silently diverge across execution modes and hardware — the same model may produce different results in eager vs. compiled mode, or on CPU vs. GPU. This pipeline fuzzes model graphs by generating API sequences, then differentially tests each graph across those settings to flag divergent results, parallelized via SLURM so many campaigns run concurrently on a cluster. It found 76 bugs across PyTorch, JAX, TensorFlow, and XLA, with 45 confirmed or fixed.

An automated test generator for a nuclear reactor protection system. It models Function Block Diagram programs and their mutations as SMT constraints, then solves them with Yices to derive test sequences. It detected 100% of real industrial faults, cut manual test-suite creation from days to minutes, and ran 20x faster than leading structural-coverage tools.

Research on self-adaptive systems normally relies on virtual simulations — there's a lack of physical testbeds to validate self-adaptive approaches in the real world. This open physical exemplar rebuilds the problem on LEGO hardware, so other researchers can reproduce and extend platooning experiments cheaply. It won the SEAMS 2021 Best Artifact Award.