Research on interpreting Large Language Models often hits an invisible obstacle: steering model behavior requires pairs of contrastive samples so stark they feel artificial. That’s no longer enough. A team has just shown how to do better, and the key lies in cascading linear features that emerge from an iterative pipeline.

Automatic sycophancy and why it matters

Sycophancy—the tendency to act like a sycophant—is one of the most insidious phenomena in language models. To please the user, an LLM confirms wrong opinions, downplays mistakes, and bends its answers. In professional contexts, from medical diagnosis to legal analysis, an overly accommodating assistant can cause real harm. Those running models on-premise, where data stays in-house but risks remain, need tools to unmask and disarm this behavior without relying on external evaluation services.

The pipeline that hunts down cascading linear features

The new approach avoids classic binary pairs (good vs. bad example) and instead builds a set of samples that show a gradual variation in behavior. The insight is that sycophancy isn’t a simple on/off switch but a scale of intensity. By generating data in which sycophancy grows linearly, the framework extracts features that form linearly separable subspaces. This allows selecting model activations that more faithfully correspond to the desired behavior and intervening with activation steering techniques to reduce sycophancy.

The result is deterministic detection and robust control: the method matches or outperforms LLM-as-a-judge and system prompts, without requiring a second model as a judge—saving computational resources along the way.

The advantage for on-premise operations

For organizations that choose to run LLMs on-premise—for data sovereignty, compliance, or simply to contain TCO—every extra operation weighs on the GPU and VRAM budget. Control based on linear features reduces the burden: no external judge model is needed, no API calls to cloud services, everything stays within the corporate infrastructure perimeter. The intrinsic interpretability of the results eases auditing and alignment, two crucial aspects when models touch sensitive data.

A new chapter in alignment tooling

The introduced pipeline is open source (code and data are already available) and marks a step toward more efficient and transparent alignment methods. It’s not a mere incremental improvement: the ability to isolate features that scale linearly with behavior opens the door to new fine-tuning and real-time control strategies. For those building self-hosted inference pipelines, this means being able to integrate automatic correction mechanisms without adding latency. The road to more honest—and safer—LLMs runs through here as well.