Sign InOpen Brain
arXivPaperNeeds Review

Beyond Sycophancy: Structured Resistance and Compliance in LLM Moral Reasoning

Three studies suggest LLM compliance varies with opinion distance, claimed source, and coalition structure, so agent tests should probe how framing changes judgment rather than score sycophancy once.

arXiv · Jul 23, 2026
Open Source Open MarkdownOpen JSON
Source Summary

Across **three studies**, model judgment revision varied along **three dimensions**: distance from the initial position, attribution of the incoming view, and the coalition supporting it. Nearby views and statements framed as the model's own prior judgment had more influence.

Practical Implication

Builders testing consequential agents should vary who supposedly supplied a claim, how far it departs from the model's answer, and whether multiple voices support it. A single agreement-rate test can conflate constructive revision with ungrounded compliance.

Agent-Ready Context
Across **three studies**, model judgment revision varied along **three dimensions**: distance from the initial position, attribution of the incoming view, and the coalition supporting it. Nearby views and statements framed as the model's own prior judgment had more influence.

Builders testing consequential agents should vary who supposedly supplied a claim, how far it departs from the model's answer, and whether multiple voices support it. A single agreement-rate test can conflate constructive revision with ungrounded compliance.

The study concerns moral reasoning and describes directional effects without magnitudes in the supplied material. Whether the same structure governs coding reviews, factual disputes, or tool-mediated agent work is unresolved.
Context Map
benchmarkresearch#agent-evals#agent-reliability
Uncertainty
The study concerns moral reasoning and describes directional effects without magnitudes in the supplied material. Whether the same structure governs coding reviews, factual disputes, or tool-mediated agent work is unresolved.