Hao Zhang

Boming Xia, Qinghua Lu, Liming Zhu et al.

Large Language Models (LLMs) have enabled the emergence of LLM agents, systems capable of pursuing under-specified goals and adapting after deployment. Evaluating such agents is challenging because their behavior is open ended, probabilistic, and shaped by system-level interactions over time. Traditional evaluation methods, built around fixed benchmarks and static test suites, fail to capture emergent behaviors or support continuous adaptation across the lifecycle. To ground a more systematic approach, we conduct a multivocal literature review (MLR) synthesizing academic and industrial evaluation practices. The findings directly inform two empirically derived artifacts: a process model and a reference architecture that embed evaluation as a continuous, governing function rather than a terminal checkpoint. Together they constitute the evaluation-driven development and operations (EDDOps) approach, which unifies offline (development-time) and online (runtime) evaluation within a closed feedback loop. By making evaluation evidence drive both runtime adaptation and governed redevelopment, EDDOps supports safer, more traceable evolution of LLM agents aligned with changing objectives, user needs, and governance constraints.

Xiaoxue Ma, Shangru Wu, Ernest Pobee et al.

Atomicity is a correctness criterion to reason about isolated code regions in a multithreaded program when they are executed concurrently. However, dynamic instances of these code regions, called transactions, may fail to behave atomically, resulting in transactional atomicity violations. Existing dynamic online atomicity checkers incur either false positives or false negatives in detecting transactions experiencing transactional atomicity violations. This paper proposes RegionTrack. RegionTrack tracks cross-thread dependences at the event, dynamic subregion, and transaction levels. It maintains both dynamic subregions within selected transactions and transactional happens-before relations through its novel timestamp propagation approach. We prove that RegionTrack is sound and complete in detecting both transactional atomicity violations and non-serializable traces. To the best of our knowledge, it is the first online technique that precisely captures the transitively closed set of happens-before relations over all conflicting events with respect to every running transaction for the above two kinds of issues. We have evaluated RegionTrack on 19 subjects of the DaCapo and the Java Grande Forum benchmarks. The empirical results confirm that RegionTrack precisely detected all those transactions which experienced transactional atomicity violations and identified all non-serializable traces. The overall results also show that RegionTrack incurred 1.10x and 1.08x lower memory and runtime overheads than Velodrome and 2.10x and 1.21x lower than Aerodrome, respectively. Moreover, it incurred 2.89x lower memory overhead than DoubleChecker. On average, Velodrome detected about 55% fewer violations than RegionTrack, which in turn reported about 3%-70% fewer violations than DoubleChecker.