Evaluating LLM Sycophancy: Which Models Hold Their Ground?
zero0_one1 · reddit · 2026-08-06
This GitHub project compares the sycophancy of major LLMs, testing whether models maintain their judgment or simply agree with the current narrator.
- Mechanism: Uses first-person framing to measure shifts in the model's stance. Positive values indicate a tendency to agree with the narrator.
- Contradiction Rate: Counts how often a model contradicts itself across opposite narrators (agreeing to or rejecting both); lower is better.
- Findings: Models differ sharply in their willingness to decide who is right, with some being highly susceptible to narrative framing.
Related event: Latest LLMs Show Reduced Sycophancy, Studies Find(2 posts)→
More from Models
- Muse Spark 1.2 Launches with Muse Code Coding Agent — alexandr_wang · 2026-08-06
- Muse Spark Team Announces Major Improvements in Coding Capabilities — alexandr_wang · 2026-08-06
- Prime Agent Coding Harness Tops ARC-AGI-3 with 95.5% Beating Human Experts — xeophon · 2026-08-06
- Rumor: Ilya's SSI Model to Be Smaller Yet More Capable Than Fable — AIandDesign · 2026-08-06
- Report: Ilya's SSI Model to be a Small Reasoning Engine, Outperforming Fable — iruletheworldmo · 2026-08-06
- Meta Offers Near-Free AI Models For Training Data, Disrupting Google's Strategy — scaling01 · 2026-08-06