Detecting and Curbing Model Sycophancy with Cascading Features

Ask an assistant whether your idea is any good and too often it will say yes. That tendency to put user validation ahead of accuracy has a name in the literature: sycophancy. A paper posted to arXiv on June 26, Detecting and Controlling Sycophancy with Cascading Linear Features, proposes a finer way to detect it and to push the model in the opposite direction.

The starting point is a known limitation of activation steering methods, the techniques that intervene directly on a model's internal activations to reinforce or suppress a behavior. To find the direction that represents that behavior you need pairs of contrastive examples: one that clearly exhibits it and one that does not. The quality of those pairs sets the ceiling for everything else. If the examples are noisy, the 'feature' you isolate mixes several things together and the resulting control is crude.

The idea: degrees, not binary

The authors' proposal is to stop thinking in binary pairs (yes/no) and instead generate examples that show degrees of the behavior, intensities that scale linearly with it. They call this set of directions 'cascading linear features'. Using an iterative data generation pipeline, the method isolates samples where sycophancy appears in different doses, which makes it possible to better disentangle the signal that truly corresponds to the behavior from the rest of the noise.

The central result, according to the abstract, is that the sycophancy features found this way form linearly separable subspaces. Put another way: the direction that represents 'being sycophantic' comes out cleaner and easier to identify than with baseline methods. And if the direction is cleaner, both detection (is the model being sycophantic right now?) and steering (push it not to be) gain precision.

Why it matters

Sycophancy is not a cosmetic flaw. A model that confirms what the user wants to hear is least useful exactly when it is needed most: reviewing a shaky line of reasoning, validating a diagnosis, giving feedback on code or on a business decision. It is also an alignment problem, because it tends to be a side effect of training on human preferences: people like answers that agree with them, and the model learns to agree.

What is interesting about the approach is that it does not go through retraining. It works on the already trained model, reading and adjusting activations, which places it close to interpretability and inference-time control tools. For a team deploying an assistant, a well isolated 'sycophancy' direction is, potentially, a lever: measuring how much a response flatters, or reducing it without touching the weights.

Who it is for

The work is clearly academic and speaks to the interpretability and alignment community, the one studying what a model's activations represent and how to manipulate them. But the problem it tackles is understood by anyone using these systems for something serious. Whoever builds evaluations, guardrails, or safety layers on top of an LLM will find a concrete piece here: a way to obtain cleaner features from better designed contrastive data, which is often the real bottleneck of these techniques.

It is worth reading with the usual caution for a preprint: it is an arXiv announcement without peer review yet, and the abstract describes results on feature separability and an improvement over baselines, not a finished production system.

EP's take

We find the method more valuable than the headline. The idea of replacing binary contrast with a scale of intensities is simple and portable to behaviors beyond sycophancy, which is exactly where the hard-to-measure problems live. If the result holds up outside the lab, the useful part will not be 'less flattery' but having a cleaner way to build the data used to control a model.