Large language models learn to detect polite but hostile language better
Honeyed Words, Hidden Daggers: Benchmarking and Training LLMs to See Through Mock Politeness
Computation and LanguageArtificial Intelligence
Summary
People sometimes say things that sound polite but are actually meant to be hurtful or hostile. The authors built a special test called ZhenhuanBench to measure how well language AI models can spot this kind of sneaky speech. They tested many models and found ways to train smaller AIs to get better at recognizing this hidden hostility. Their methods showed improvements on several different tests about language intent and sarcasm. This helps AI better understand the real meaning behind polite words that have a hidden negative intent.
What this means in practice
- •For customer support teams: Identify and flag customer messages that are polite but contain hidden dissatisfaction or hostility to improve response strategies.
- •For online content moderators: Detect subtle hostile speech masked as politeness to better maintain healthy online community interactions.
Authors
Caixin Kang, Sitong Gong, Tianyu Yan, Kaipeng Zhang, Yoichi Sato, Yifei Huang
Abstract
The Chinese idiom "kou mi fu jian" ("honeyed words, hidden daggers") names speech that is polite in form but hostile in function. We present ZhenhuanBench, a 1,592-item benchmark built from expert commentary on Empresses in the Palace, covering speaker intent, emotion-language mismatch, and per-listener uptake. Text-only evaluation covers 34 LLMs. GPT-6 Astra leads the leaderboard at 73.48% overall, while its listener-reaction score is 58.37%. The annotations also supply supervised traces and string-verifiable training rewards. With 338 gold traces, an 8B model outperforms teacher distillation on identical items. Subsequent RL saturates, and chance correction reduces the estimated recoverable performance gap by a factor of 7.5. Our CPR-GRPO loop (Compress-Probe-Replenish) uses this estimate to weight gold supervision, reaching 67.0 on public multiple-choice tasks and 74.5 under the training-matched hidden protocol. Without target-task training, CPR outperforms Base, SFT, and GRPO across six Chinese and English pragmatic and sarcasm evaluations. On three external datasets, training with CPR yields higher mean Macro-F1 than SFT and GRPO.