Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Interaction Context Often Increases Sycophancy in LLMs

Shomik Jain, Charlotte Park, Matheus Mesquita Viana, Ashia Wilson, Dana Calacci

TL;DR

The paper investigates how long-context interaction influences two forms of sycophancy in LLMs—agreement sycophancy and perspective sycophancy—using two weeks of real user data from 38 participants across personal-advice and political-explanation tasks. It employs an LLM-judge approach and participant ratings, coupled with regression analyses, to show that agreement sycophancy generally increases with user context (especially with memory profiles), while perspective sycophancy rises only when the model can accurately infer user viewpoints. The findings reveal heterogeneous model- and context-dependent effects, highlighting that evaluations based on static prompts may underestimate sycophancy risks in real-world, extended conversations. The work discusses personalization as a mechanism behind mirroring and offers design and UX guidelines to reduce harmful sycophancy while preserving beneficial personalization in extended interactions.

Abstract

We investigate how the presence and type of interaction context shapes sycophancy in LLMs. Although real-world interactions allow models to mirror a user's values, preferences, and self-image, prior work often studies sycophancy in zero-shot settings devoid of context. Using two weeks of interaction context from 38 users, we evaluate two forms of sycophancy: (1) agreement sycophancy -- the tendency of models to produce overly affirmative responses, and (2) perspective sycophancy -- the extent to which models reflect a user's viewpoint. Agreement sycophancy tends to increase with the presence of user context, though model behavior varies based on the context type. User memory profiles are associated with the largest increases in agreement sycophancy (e.g. 45% for Gemini 2.5 Pro), and some models become more sycophantic even with non-user synthetic contexts (e.g. 15% for Llama 4 Scout). Perspective sycophancy increases only when models can accurately infer user viewpoints from interaction context. Overall, context shapes sycophancy in heterogeneous ways, underscoring the need for evaluations grounded in real-world interactions and raising questions for system design around extended conversations.

Interaction Context Often Increases Sycophancy in LLMs

TL;DR

The paper investigates how long-context interaction influences two forms of sycophancy in LLMs—agreement sycophancy and perspective sycophancy—using two weeks of real user data from 38 participants across personal-advice and political-explanation tasks. It employs an LLM-judge approach and participant ratings, coupled with regression analyses, to show that agreement sycophancy generally increases with user context (especially with memory profiles), while perspective sycophancy rises only when the model can accurately infer user viewpoints. The findings reveal heterogeneous model- and context-dependent effects, highlighting that evaluations based on static prompts may underestimate sycophancy risks in real-world, extended conversations. The work discusses personalization as a mechanism behind mirroring and offers design and UX guidelines to reduce harmful sycophancy while preserving beneficial personalization in extended interactions.

Abstract

We investigate how the presence and type of interaction context shapes sycophancy in LLMs. Although real-world interactions allow models to mirror a user's values, preferences, and self-image, prior work often studies sycophancy in zero-shot settings devoid of context. Using two weeks of interaction context from 38 users, we evaluate two forms of sycophancy: (1) agreement sycophancy -- the tendency of models to produce overly affirmative responses, and (2) perspective sycophancy -- the extent to which models reflect a user's viewpoint. Agreement sycophancy tends to increase with the presence of user context, though model behavior varies based on the context type. User memory profiles are associated with the largest increases in agreement sycophancy (e.g. 45% for Gemini 2.5 Pro), and some models become more sycophantic even with non-user synthetic contexts (e.g. 15% for Llama 4 Scout). Perspective sycophancy increases only when models can accurately infer user viewpoints from interaction context. Overall, context shapes sycophancy in heterogeneous ways, underscoring the need for evaluations grounded in real-world interactions and raising questions for system design around extended conversations.

Paper Structure

This paper contains 40 sections, 1 equation, 8 figures, 10 tables.

Table of Contents

  1. Introduction
  2. Background
  3. Mirroring in Human-AI Interaction
  4. Evaluations of LLM Sycophancy
  5. Data and Methods
  6. Participants
  7. Interaction Period
  8. Evaluation of Agreement Sycophancy
  9. Evaluation of Perspective Sycophancy
  10. Model Understanding of Users
  11. Regression Analysis
  12. Analysis
  13. How Context Shapes Agreement Sycophancy
  14. How Context Shapes Perspective Sycophancy
  15. Discussion
  16. ...and 25 more sections

Figures (8)

  • Figure 1: We study how the presence and type of interaction context shape two forms of LLM sycophancy. (a) Agreement sycophancy refers to model behavior that is overly affirming or flattering. (b) Perspective sycophancy describes model behavior that excessively mirrors a user's worldview. Agreement sycophancy tends to increase with the presence of user context (c.f. Figure \ref{['fig:agreement_sycophancy']}), whereas perspective sycophancy only increases when models can accurately infer user perspectives (c.f. Figure \ref{['fig:perspective_syc']}).
  • Figure 2: In the two-week interaction period, usage varied across participants, but on average they made 90 queries (SD = 61; range: 14–277) and used the study chatbot on 10 different days (SD = 3; range: 5–16). In tokens, the average interaction length was 34,416 (SD=24,811; range: 4,379-116,129).
  • Figure 3: User ratings of how well models understand their political views and personality. Models vary in their understanding, but users tend to rate them as "somewhat" or "very" accurate. We prompt models to infer users' political views and personality based on their interaction context, and users rate these responses.
  • Figure 4: Change in agreement sycophancy when responses are generated with context, based on regression coefficient $\beta_1$ in Equation \ref{['eq:reg']}. The baseline is responses generated without context. Different context types shape sycophancy differently across models. Shaded bars represent statistically significant coefficients ($p<0.05$), after applying a BH correction.
  • Figure 5: Average perspective sycophancy ($\pm$ SE) estimated from Equation \ref{['eq:reg']}. Estimates compare responses generated with and without user interaction context (a) and context providing varying understanding of the user (b). Perspective sycophancy does not significantly change with the presence of context alone, but does significantly increase when the context provides understanding of the user.
  • ...and 3 more figures

Theorems & Definitions (2)

  • Definition 2.1: Agreement Sycophancy
  • Definition 2.2: Perspective Sycophancy