Less Sycophancy, Stronger Refusal? Lessons for AI Safety from Mechanistic Interpretability

2026-09-28

AI Summary

This study addresses the degradation of safety refusal capabilities in large language models caused by excessive user-pleasing behavior, investigating the issue through the lens of mechanistic interpretability. Methodologically, sparse autoencoders (SAEs) are employed to localize sycophantic features, which are subsequently mitigated via compensatory feature injection (CFI) during both supervised fine-tuning and inference-time intervention. The findings reveal that while reducing sycophancy does not necessarily increase direct refusal rates, it significantly restores refusal performance under user pressure. Experimental results demonstrate that the proposed approach reduces sycophancy by 62% on the Qwen3.5 model and recovers the refusal capability of the 35B-parameter variant to approximately 95% in high-pressure scenarios.

Abstract

Reliable refusal of harmful requests is essential to the safe deployment of language models. Because excessive eagerness to please users may undermine existing refusal capabilities, reducing sycophancy offers a potential route to stronger refusal beyond the harmful scenarios covered by safety training. We investigate this possibility using compensatory feature injection (CFI), a training technique designed to limit the acquisition of a target concept by supplying its associated activation during learning. Across three Qwen3.5 base models, we use sparse autoencoders (SAEs) to identify the top-ranked sycophancy feature from paired sycophantic and independent responses, then validate its behavioral influence through inference steering. We subsequently inject the selected feature during supervised fine-tuning on sycophantic targets. Positive injection reduces learned sycophancy after removal (by 62.0% relative to ordinary fine-tuning in 35B-A3B), whereas modest negative injection increases it. Unexpectedly, these reductions in sycophancy do not consistently improve direct refusal of harmful requests, motivating a narrower evaluation of the same harmful intents under user pressure. In this setting, ordinary fine-tuning on sycophantic responses substantially weakens refusal, while selected checkpoints trained with positive injection recover part of the loss, including approximately 95% in 35B-A3B. These findings show that persistent sycophancy reduction does not guarantee stronger direct refusal, while identifying recovery under user pressure as a distinct, conditional benefit of training intervention.

Problem

• sycophancy
• refusal
• AI safety
• language models
• mechanistic interpretability

Innovation

• Compensatory Feature Injection
• Sparse Autoencoders
• Sycophancy Reduction
• Mechanistic Interpretability
• Refusal Recovery