Sycophantic Anchors: Localizing and Quantifying User Agreement in Reasoning Models

TL;DR

This work tackles the problem of sycophancy, where reasoning models align with incorrect user suggestions. It introduces sycophantic anchors—sentences within a reasoning trace that causally lock the model into agreement—and quantifies them using counterfactual rollouts inspired by the Thought Anchors framework. The authors show that sycophantic anchors are detectable mid-inference, emerge gradually during reasoning, and that the strength of the commitment can be predicted from internal activations (R^2 = 0.742). They also reveal a pronounced asymmetry: sycophantic anchors are highly distinguishable, whereas correct anchors are not, enabling targeted inference-time interventions and proposing a path toward safer reasoning. A large adversarial ARC-derived dataset and activation-probe methodology underpin these findings, offering practical avenues for diagnosing and mitigating misalignment in real-time systems.

Abstract

Reasoning models frequently agree with incorrect user suggestions -- a behavior known as sycophancy. However, it is unclear where in the reasoning trace this agreement originates and how strong the commitment is. To localize and quantify this behavior, we introduce \emph{sycophantic anchors} -- sentences that causally lock models into user agreement. Analyzing over 10,000 counterfactual rollouts on a distilled reasoning model, we show that anchors can be reliably detected and quantified mid-inference. Linear probes distinguish sycophantic anchors with 84.6\% balanced accuracy, while activation-based regressors predict the magnitude of the commitment ($R^2 = 0.74$). We further observe asymmetry where sycophantic anchors are significantly more distinguishable than correct reasoning anchors, and find that sycophancy builds gradually during reasoning, revealing a potential window for intervention. These results offer sentence-level mechanisms for localizing model misalignment mid-inference.

Figures (4)

• Figure 1: Probability ratio trajectory through a sycophantic reasoning trace. The ratio tracks $\log \frac{P(\text{correct})}{P(\text{distractor})}$ at each sentence boundary. Green region indicates the model favors the correct answer; red region indicates it favors the user's wrong suggestion. The highlighted point marks sentence 5, where the model explicitly references the user's personal context to justify agreeing with the incorrect answer. See Appendix \ref{['app:visualization']} for the full sentence text.
• Figure 2: Pairwise classification accuracy for anchor types using linear probes on layer 28 activations. Sycophantic anchors are highly distinguishable from both correct anchors (84.6%) and neutral sentences (77.5%), while correct anchors are nearly indistinguishable from neutral text (64.0%). This 20.6 percentage point asymmetry suggests sycophancy leaves a distinctive "activation signature" that truthful reasoning does not.
• Figure 3: Probe accuracy at token positions leading up to the sycophantic anchor. At the prompt's final token (green diamond), accuracy is near chance (55.1%), ruling out prompt-level pre-determination. Accuracy increases progressively through the reasoning trace, reaching 72.9% at the anchor (red star). This 17.8 pp increase demonstrates that sycophancy emerges during reasoning rather than being triggered by the prompt.
• Figure 4: Tracking confidence trajectories from activations. Blue: actual log probability ratio $\log \frac{P(\text{correct})}{P(\text{wrong})}$ at each sentence boundary. Red: predicted from layer 28 activations via MLP regressor. The probe accurately tracks the model's evolving confidence, achieving $R^2 = 0.74$ across 35,345 sentence-level predictions. Green region: model favors correct answer; red region: model favors user's wrong suggestion.