Verena Kaynig-Fittkau

Author directory

2026

Large language models are widely used for personal advice (relationship conflicts, financial decisions, mental-health support, etc.), but they often align their response with the user’s position rather than offering independent judgment, even when users need pushback rather than validation. Mitigating this, whether through training objectives, system-level guardrails, or evaluation design, requires knowing what triggers it. But what in the user’s message triggers this behavior is unknown. Natural advice-seeking messages vary simultaneously in semantic content (what is disclosed), linguistic style (how it is written), and narrative framing (whose side is told), but existing benchmarks cannot attribute model behavior to any single cue. Without this attribution, mitigations cannot be targeted, for example, suppressing warmth may harm users with genuine distress but might not fix the model’s tendency to endorse whichever side of a conflict speaks. In this work, we isolate which of these cues drive which aspect of sycophantic behavior across 787 advice-seeking scenarios and observe three patterns emerge: emotional warmth is content-driven; stance endorsement is framing-driven: endorsement drops by ∼50% when the same conflict is retold from the opposing perspective; and recommendation directness is style-driven. We also find that hedging reduces rather than increases model endorsement; models do not simply validate vulnerable-sounding users.

2022

We introduce DocTime - a novel temporal dependency graph (TDG) parser that takes as input a text document and produces a temporal dependency graph. It outperforms previous BERT-based solutions by a relative 4-8% on three datasets from modeling the problem as a graph network with path-prediction loss to incorporate longer range dependencies. This work also demonstrates how the TDG graph can be used to improve the downstream tasks of temporal questions answering and NLI by a relative 4-10% with a new framework that incorporates the temporal dependency graph into the self-attention layer of Transformer models (Time-transformer). Finally, we develop and evaluate on a new temporal dependency graph dataset for the domain of contractual documents, which has not been previously explored in this setting.