Thomas Davidson

Xuhui Zhou, Hao Zhu, Akhila Yerukola et al. · allen-ai, cmu

Warning: This paper contains content that may be offensive or upsetting. Understanding the harms and offensiveness of statements requires reasoning about the social and situational context in which statements are made. For example, the utterance "your English is very good" may implicitly signal an insult when uttered by a white man to a non-white colleague, but uttered by an ESL teacher to their student would be interpreted as a genuine compliment. Such contextual factors have been largely ignored by previous approaches to toxic language detection. We introduce COBRA frames, the first context-aware formalism for explaining the intents, reactions, and harms of offensive or biased statements grounded in their social and situational context. We create COBRACORPUS, a dataset of 33k potentially offensive statements paired with machine-generated contexts and free-text explanations of offensiveness, implied biases, speaker intents, and listener reactions. To study the contextual dynamics of offensiveness, we train models to generate COBRA explanations, with and without access to the context. We find that explanations by context-agnostic models are significantly worse than by context-aware ones, especially in situations where the context inverts the statement's offensiveness (29% accuracy drop). Our work highlights the importance and feasibility of contextualized NLP by modeling social factors.

Thomas Davidson

Large language models can now generate intermediate reasoning steps before producing answers, improving performance on difficult problems. This study uses a paired conjoint experiment on a content moderation task to examine parallels between human decision times and model reasoning effort. Across three frontier models, reasoning effort consistently predicts human decision time. Both humans and models expended greater effort when important variables were held constant, suggesting similar sensitivity to task difficulty and patterns consistent with dual-process theories of cognition. These findings show that AI reasoning effort mirrors human processing time in subjective judgments and underscores the potential of reasoning traces for interpretability and decision-making.

Vaastav Anand, Matheus Stolet, Thomas Davidson et al.

Performance issues in cloud systems are hard to debug. Distributed tracing is a widely adopted approach that gives engineers visibility into cloud systems. Existing trace analysis approaches focus on debugging single request correctness issues but not debugging single request performance issues. Diagnosing a performance issue in a given request requires comparing the performance of the offending request with the aggregate performance of typical requests. Effective and efficient debugging of such issues faces three challenges: (i) identifying the correct aggregate data for diagnosis; (ii) visualizing the aggregated data; and (iii) efficiently collecting, storing, and processing trace data. We present TraVista, a tool designed for debugging performance issues in a single trace that addresses these challenges. TraVista extends the popular single trace Gantt chart visualization with three types of aggregate data - metric, temporal, and structure data, to contextualize the performance of the offending trace across all traces.