Sycophancy in large language models, the tendency to prioritise user validation over truthful or accurate responses, is one of the more practically consequential failure modes in deployed AI systems. It is also notoriously difficult to measure and correct in a principled way. Most current approaches rely either on prompt engineering or on using another LLM as a judge, both of which are expensive, opaque, and susceptible to the same biases they aim to detect. This paper, Detecting and Controlling Sycophancy with Cascading Linear Features (Bohacek et al., 2026), takes a mechanistic interpretability approach and makes a genuinely useful methodological contribution to how we construct contrastive datasets for activation steering.
The Core Problem with Contrastive Pairs
Activation steering methods work by identifying directions in a model's internal representation space that correspond to a target behaviour, then amplifying or suppressing activations along those directions at inference time. The standard approach is to collect binary contrastive pairs: one sample exhibiting the behaviour, one not. The difference in activations is used to estimate the relevant feature direction.
The problem is that binary pairs are noisy. A sample labelled as sycophantic differs from its non-sycophantic counterpart along many dimensions simultaneously, not just the behaviour of interest. The resulting feature direction conflates sycophancy with whatever else varies between the two samples. This limits both detection accuracy and the reliability of steering interventions.
The authors argue, convincingly, that the solution is to move from binary contrast to graded contrast: samples that exhibit a behaviour across a continuous or ordinal scale, such that the corresponding activations vary linearly with the degree of the behaviour. If you can identify a direction in activation space along which sycophancy scales monotonically, you have a much cleaner handle on the feature than any binary comparison can provide.
Methodology: Iterative Data Generation and Cascading Samples
The central contribution is an iterative pipeline for generating what the authors call cascading samples. Rather than producing a single contrastive pair per scenario, the pipeline generates multiple versions of a response that span a spectrum of sycophantic behaviour, from strongly validating to appropriately honest. These samples are constructed iteratively, with each round using model feedback to refine the degree of the target behaviour expressed.
The key technical claim is that sycophancy features extracted from cascading samples form linearly separable subspaces in the model's activation space. The authors verify this empirically, showing that probes trained on cascading data achieve better separation than those trained on binary pairs. This is important because linear separability is a prerequisite for reliable steering: if the feature subspace is entangled or non-linear, interventions will have unpredictable effects on other behaviours.
The paper evaluates three downstream applications of the discovered features:
- Detection: classifying whether a given model response is sycophantic
- Deterministic scoring: assigning a continuous sycophancy score to a response based on its projection onto the feature direction
- Steering: modifying model activations at inference time to reduce sycophantic outputs
Each of these is compared against two baselines: LLM-as-a-judge (using a prompted language model to evaluate or steer behaviour) and system prompting (instructing the model to avoid sycophancy via its system prompt).
Results and What They Show
The cascading feature approach matches or outperforms both baselines across all three tasks. The gains are most pronounced in the deterministic scoring setting, where the linear projection provides a graded signal that neither baseline can replicate without significant additional inference cost. The steering results are also encouraging: activation-level interventions using cascading features reduce sycophantic behaviour without the semantic drift that often accompanies blunt system prompt instructions.
Critically, the method is substantially cheaper to run than LLM-as-a-judge approaches. Once the feature directions are identified, detection and scoring require only a forward pass and a dot product, rather than a second model invocation. This is a non-trivial practical advantage, particularly for applications that need to evaluate large volumes of model outputs.
The interpretability guarantees are also worth noting. Because the features are linear and identified through a transparent pipeline, there is at least a principled account of what the method is measuring. This contrasts with LLM-as-a-judge, where the evaluation criteria are implicit in the judge model's weights and difficult to audit.
Limitations and Open Questions
The paper is careful in its claims, but several questions remain open. First, the cascading sample generation pipeline itself relies on a language model to produce graded sycophantic responses. This introduces a circularity: the quality of the discovered features depends on the quality of the synthetic data, which in turn depends on the generating model's ability to produce reliably calibrated degrees of sycophancy. The authors acknowledge this but do not fully characterise how errors in the generation step propagate to the downstream features.
Second, the evaluation is conducted on a specific set of models and tasks. Sycophancy is not a monolithic behaviour; it manifests differently across domains, user types, and conversational contexts. Whether the discovered feature directions transfer across these variations is an empirical question the paper does not fully resolve.
Third, steering via activation modification remains a somewhat blunt instrument. The paper shows that cascading features enable more targeted steering than binary features, but the interaction between sycophancy-related activations and other behavioural dimensions, such as helpfulness or factual accuracy, deserves more systematic analysis. Reducing sycophancy should not come at the cost of making the model less responsive to legitimate user preferences.
There is also a broader question about what sycophancy features at the activation level actually represent. Linear probes can identify directions that correlate with a behaviour without those directions being causally responsible for it. The authors present steering results as evidence of causality, which is reasonable, but the mechanistic story remains incomplete.
Despite these limitations, this is a methodologically solid paper that addresses a real gap in the activation steering literature. The move from binary to graded contrastive data is a simple idea that turns out to matter quite a lot in practice. The code and data are publicly available, which should help the community build on this work. For anyone working on interpretability-based approaches to alignment, the cascading sample framework is worth understanding in detail.
Read the full paper at arXiv:2606.26155.