CanaryBench: Stress Testing Privacy Leakage in Cluster-Level Conversation Summaries

Aggregate analytics over conversational data are increasingly used for safety monitoring, governance, and product analysis in large language model systems. A common practice is to embed conversations, cluster them, and publish short textual summaries describing each cluster. While raw conversations may never be exposed, these derived summaries can still pose privacy risks if they contain personally identifying information (PII) or uniquely traceable strings copied from individual conversations. We introduce CanaryBench, a simple and reproducible stress test for privacy leakage in cluster-level conversation summaries. CanaryBench generates synthetic conversations with planted secret strings ("canaries") that simulate sensitive identifiers. Because canaries are known a priori, any appearance of these strings in published summaries constitutes a measurable leak. Using TF-IDF embeddings and k-means clustering on 3,000 synthetic conversations (24 topics) with a canary injection rate of 0.60, we evaluate an intentionally extractive example snippet summarizer that models quote-like reporting. In this configuration, we observe canary leakage in 50 of 52 canary-containing clusters (cluster-level leakage rate 0.961538), along with nonzero regex-based PII indicator counts. A minimal defense combining a minimum cluster-size publication threshold (k-min = 25) and regex-based redaction eliminates measured canary leakage and PII indicator hits in the reported run while maintaining a similar cluster-coherence proxy. We position this work as a societal impacts contribution centered on privacy risk measurement for published analytics artifacts rather than raw user data.

Does Inference Scaling Improve Reasoning Faithfulness? A Multi-Model Analysis of Self-Consistency Tradeoffs

Self-consistency has emerged as a popular technique for improving large language model accuracy on reasoning tasks. The approach is straightforward: generate multiple reasoning paths and select the most common answer through majority voting. While this reliably boosts accuracy, it remains unclear whether these gains reflect genuine improvements in reasoning quality. We investigate a fundamental question that has not been studied before: does inference scaling improve reasoning faithfulness? We conduct a comprehensive empirical study across four frontier models (GPT-5.2, Claude Opus 4.5, Gemini-3-flash-preview, and DeepSeek-v3.2) on 100 GSM8K mathematical reasoning problems. Our analysis employs bootstrap confidence intervals, McNemar's tests for paired comparisons, and Cohen's d effect sizes to quantify the effects rigorously. The results reveal striking differences across models that challenge common assumptions about self-consistency. GPT-5.2 shows the expected pattern: accuracy improves from 78% to 90% at N=5, with faithfulness remaining relatively stable (0.540 to 0.510). Claude Opus 4.5 tells a completely different story. Its accuracy actually drops from 78% to 74.3% while faithfulness jumps dramatically from 0.270 to 0.891 at N=5. DeepSeek-v3.2, already at 98% accuracy, shows ceiling effects with modest faithfulness gains (0.440 to 0.541). Gemini-3-flash improves from 81% to 86% accuracy with a slight faithfulness decrease (0.260 to 0.212). Problem difficulty analysis reveals that GPT-5.2 solves 82% of hard problems while breaking only 13% of easy ones. Claude, in contrast, breaks 23% of easy problems, explaining its accuracy decrease. These findings matter for practitioners: self-consistency is not universally beneficial, and teams should test their specific models before deployment. We release our code and provide practical recommendations for navigating these tradeoffs.