Christian Medeiros Adriano

Julius Porbeck, Christian Medeiros Adriano, Holger Giese

Large language model (LLM)-based debugging systems can generate failure explanations, but these explanations may be incomplete or incorrect. Misleading explanations are harmful for downstream tasks (e.g., bug triage, bug fixing). We investigate how explanation quality is affected by various LLM context configurations. Existing work predominantly treats LLM-generated failure explanations as an ad hoc by-product of debugging or repair workflows, using generic prompting over undifferentiated artifacts such as code, tests, and error messages rather than targeting explanations as a first-class output with dedicated quality assessment. Consequently, existing approaches provide limited support for assessing whether these explanations capture the underlying fault-error-failure mechanism and for actionable next steps, and most techniques instead prioritize task success (e.g., patch correctness or review quality) over the explicit causal explanation quality. We systematically vary the debugging information to study how distinct context compositions affect the quality of LLM-generated failure explanations. Across 93 context configurations on real bugs and three economically viable models (gpt-5-mini, DeepSeek-V3.2, and Grok-4.1-fast), we evaluate explanations with six criteria and validate the LLM-as-a-judge scores against human ratings in a user study. Our results indicate that explanation quality is causally affected by context composition. Evidence-rich, failure-specific artifacts improve causal and action-oriented quality, whereas overly large contexts tend to yield vague explanations. Higher explanation-score quartiles are associated with higher downstream repair pass rates and, for some models, with fixes that are closer to the reference minimal fixes. In contrast, low-score quartiles can even underperform the no-explanation baseline. Reproduction package is publicly available.

Christian Medeiros Adriano, Andre van der Hoek

Microtask crowdsourcing is the practice of breaking down an overarching task to be performed into numerous, small, and quick microtasks that are distributed to an unknown, large set of workers. Microtask crowdsourcing has shown potential in other disciplines, but with only a handful of approaches explored to date in software engineering, its potential in our field remains unclear. In this paper, we explore how microtask crowdsourcing might serve as a means of fault localization. We particularly take a first step in assessing whether a crowd of workers can correctly locate known faults in a few lines of code (code fragments) taken from different open source projects. Through Mechanical Turk, we collected the answers of hundreds of workers to a pre-determined set of template questions applied to the code fragments, with a replication factor of twenty answers per question. Our findings show that a crowd can correctly distinguish questions that cover lines of code that contain a fault from those that do not. We also show that various filters can be applied to identify the most effective subcrowds. Our findings also presented serious limitations in terms of the proportion of lines of code selected for inspection and the cost to collect answers. We describe the design of our experiment, discuss the results, and provide an extensive analysis of different filters and their effects in terms of speed, cost, and effectiveness. We conclude with a discussion of limitations and possible future experiments toward more full-fledged fault localization on a large scale involving more complex faults.

Zainab Rehan, Christian Medeiros Adriano, Sona Ghahremani et al.

Rule-based systems remain central in safety-critical domains but often struggle with scalability, brittleness, and goal misspecification. These limitations can lead to reward hacking and failures in formal verification, as AI systems tend to optimize for narrow objectives. In previous research, we developed a neuro-symbolic causal framework that integrates first-order logic abduction trees, structural causal models, and deep reinforcement learning within a MAPE-K loop to provide explainable adaptations under distribution shifts. In this paper, we extend that framework by introducing a meta-level layer designed to mitigate goal misspecification and support scalable rule maintenance. This layer consists of a Goal/Rule Synthesizer and a Rule Verification Engine, which iteratively refine a formal rule theory from high-level natural-language goals and principles provided by human experts. The synthesis pipeline employs large language models (LLMs) to: (1) decompose goals into candidate causes, (2) consolidate semantics to remove redundancies, (3) translate them into candidate first-order rules, and (4) compose necessary and sufficient causal sets. The verification pipeline then performs (1) syntax and schema validation, (2) logical consistency analysis, and (3) safety and invariant checks before integrating verified rules into the knowledge base. We evaluated our approach with a proof-of-concept implementation in two autonomous driving scenarios. Results indicate that, given human-specified goals and principles, the pipeline can successfully derive minimal necessary and sufficient rule sets and formalize them as logical constraints. These findings suggest that the pipeline supports incremental, modular, and traceable rule synthesis grounded in established legal and safety principles.

Kathrin Korte, Christian Medeiros Adriano, Sona Ghahremani et al.

[Context] Multi-agent reinforcement learning (MARL) has achieved notable success in environments where agents must learn coordinated behaviors. However, transferring knowledge across agents remains challenging in non-stationary environments with changing goals. [Problem] Traditional knowledge transfer methods in MARL struggle to generalize, and agents often require costly retraining to adapt. [Approach] This paper introduces a causal knowledge transfer framework that enables RL agents to learn and share compact causal representations of paths within a non-stationary environment. As the environment changes (new obstacles), agents' collisions require adaptive recovery strategies. We model each collision as a causal intervention instantiated as a sequence of recovery actions (a macro) whose effect corresponds to a causal knowledge of how to circumvent the obstacle while increasing the chances of achieving the agent's goal (maximizing cumulative reward). This recovery action macro is transferred online from a second agent and is applied in a zero-shot fashion, i.e., without retraining, just by querying a lookup model with local context information (collisions). [Results] Our findings reveal two key insights: (1) agents with heterogeneous goals were able to bridge about half of the gap between random exploration and a fully retrained policy when adapting to new environments, and (2) the impact of causal knowledge transfer depends on the interplay between environment complexity and agents' heterogeneous goals.

Leonhard Hennicke, Christian Medeiros Adriano, Holger Giese et al.

Generative foundation models like Stable Diffusion comprise a diverse spectrum of knowledge in computer vision with the potential for transfer learning, e.g., via generating data to train student models for downstream tasks. This could circumvent the necessity of collecting labeled real-world data, thereby presenting a form of data-free knowledge distillation. However, the resultant student models show a significant drop in accuracy compared to models trained on real data. We investigate possible causes for this drop and focus on the role of the different layers of the student model. By training these layers using either real or synthetic data, we reveal that the drop mainly stems from the model's final layers. Further, we briefly investigate other factors, such as differences in data-normalization between synthetic and real, the impact of data augmentations, texture vs.\ shape learning, and assuming oracle prompts. While we find that some of those factors can have an impact, they are not sufficient to close the gap towards real data. Building upon our insights that mainly later layers are responsible for the drop, we investigate the data-efficiency of fine-tuning a synthetically trained model with real data applied to only those last layers. Our results suggest an improved trade-off between the amount of real training data used and the model's accuracy. Our findings contribute to the understanding of the gap between synthetic and real data and indicate solutions to mitigate the scarcity of labeled real data.