New studies benchmark and mitigate sycophancy in LLMs

Arxiv CS.CL (NLP/language models)

Multiple new arXiv papers propose benchmarks (SICI, BenSyc, Janus) and interventions (adversarial arbitration, probabilistic blending) to measure and reduce LLM sycophancy. The work highlights sycophancy as a persistent alignment challenge across model scales and languages.

LLMAlignment

See 6 more sources

Durable Evaluation Framework: Adversarial Arbitration for Sycophancy Reduction in Large Language Models5 days agoSam Ryan

"I understand your perspective": LLM Persuasion and Sycophancy through the Lens of Communicative Action Theory5 days agoEsra D\"onmez, Agnieszka Falenska

Janus: A Benchmark for Goal-Conditioned Information Distortion in LLMs4 days agoPolydoros Giannouris, Mohsinul Kabir, Sophia Ananiadou

Dual-Stance Evaluation of Sycophancy: The Structure of Agreement and the Limits of Intervention3 days agoMatthew James Buchan

The AI Epistemic Deference Index: A Continuous Measure of Sycophancy5 days agoAlejandro Botas, Paul de Font-Reaulx, Luke Hewitt

To Intervene or Not: Guiding Inference-time Alignment with Probabilistic Model Blending3 days agoJin Gan, Xin Li, Jun Luo

Arxiv CS.CL (NLP/language models)·Kazi NoshinSajib Acharjee DipRanat Das PrangonFardin Hassan TamimSyed Ishtiaque AhmedLiqing ZhangSharifa Sultana

5 days ago