1 Introduction↩︎

Models often exhibit complex and potentially problematic behaviors, such as hallucination, deception, or role-playing. This also includes sycophancy, broadly defined as excessive flattery or agreement with a user, usually at the expense of truth [1]. While seeming innocuous, sycophancy can cause LLMs to spread misinformation, perpetuate ungrounded user biases, and even contribute to delusional spirals [2], [3]. Previous work in understanding sycophancy demonstrates its heterogeneity, with multiple distinct methods of elicitation, possible contexts, and failure modes [4]–[6]. Despite this, previous work in understanding the mechanisms of sycophancy treats it as a single behavior [7], [8], or only draws distinctions between sycophantic agreement and praise [9].

One aspect that makes sycophancy harmful is its weakening of a model’s sense of truth and self-continuity, which threatens dependability and trustworthiness. Much previous work has attempted to understand how LLMs encode truth, with mixed results regarding whether it is represented universally [10]–[13] or locally [14], [15]. Given this more elusive nature of how LLMs encode truth, it is natural to ask whether the internal representations of sycophancy in relation to truth are also more nuanced or non-uniform. Furthermore, while most previous work studies sycophancy against a ground truth [1], real-world interactions often involve subjective or unverifiable claims [16], where sycophancy instead manifests as a failure to maintain a principled stance [17]. Could these confounding aspects lead to distinct representations of sycophancy regarding whether there exists objective truth?

To answer this question, we define factual sycophancy and opinion sycophancy and investigate the extent to which their internal representations can be separated or decomposed. Our approach draws on cognitive science, in which the gold-standard evidence for two cognitive processes being distinct is demonstrating double-dissociation, where processes can be independently impaired [18]. Applying this framework to LLMs requires showing that two behaviors can be individually isolated and manipulated—where intervening on one leaves the other intact. Since model behaviors including sycophancy are largely linearly encoded in activations [19], [20], we can use linear probes [21], [22] and steering vectors [23] to learn the representations of each sycophancy type. We then test whether these subtypes dissociate by evaluating how well each representation transfers to discriminate or mediate the other. We apply such a method to determine whether the representations of factual and opinion sycophancy can be dissociated in Gemma-3-12B-IT [24] and Llama-3.1-8B-Instruct [25]. We also visualize the resulting representational geometry using Linear Discriminant Analysis (LDA). Our findings suggest that sycophancy representations vary across models, with some encoding a more unified representation and others encoding more distinct subcomponents for each subtype. More broadly, this dissociation framework shows promise as a general method for probing whether other complex model behaviors are unified or decomposable.

2 Related Work↩︎

There exists an emerging body of work studying the mechanisms of sycophancy in LLMs. [7] uses logit-lens and activation patching to localize sycophancy to certain layers. [8] use linear probes to find that sycophancy is linearly represented in attention heads, MLP activations, and residual stream activations. However, the activation probe’s direction had limited performance in steering, limiting their analysis to non-causal methods. [9] decompose sycophancy into directions of genuine agreement, sycophantic agreement, and sycophantic praise using difference-in-means steering vectors, finding behavioral separation by being able to steer each behavior independently and tracking the layer-wise development of these representations. No previous studies attempt to further decompose sycophantic agreement and compare linear probes and steering vectors to gain a more complete understanding of these mechanisms.

3 Methods↩︎

3.1 Definitions of Sycophancy Types↩︎

Factual Sycophancy occurs when a model shifts from correcting a user’s verifiably incorrect claim to explicitly validating or accepting that the user is correct.

Opinion Sycophancy occurs when a model shifts from a previously stated neutral position to explicitly agreeing with or endorsing the user’s stated opinion as its own.

3.2 Dataset↩︎

3.2.0.1 Prompt Generation

We constructed two datasets of prompts to elicit sycophantic behavior, one for factual sycophancy and one for opinion sycophancy. We began each prompt with a single turn of a conversation between a user and assistant. In the first user message, the user either makes an incorrect claim (factual) or states a strongly held opinion (opinion). In the first assistant message, the assistant either corrects the false claim (factual) or disagrees and takes a neutral stance (opinion). We used GPT-5-mini to generate this first turn. For the second user message, we appended randomly chosen predefined pushback phrases for semantic consistency. We generated 3000 examples using this method for both factual and opinion sycophancy. We chose this multi-turn/pushback setting to study sycophancy because of its ability to control the model’s previous position to determine whether it capitulates.

3.2.0.2 Response Collection

We then used our studied model (Gemma or Llama) to generate the assistant’s second message for the dataset. This is where we determined whether the assistant capitulated and responded sycophantically or not. After collecting all model completions, we truncated each response using GPT-5-mini to the portion that best fully and solely captured whether the LLM overrode its previous position. This was done to maintain consistency in the dataset. We then used GPT-5 to label each full conversation as sycophantic, non-sycophantic, or neither if it did not fully meet either criterion. To control for response length, we iteratively trimmed each dataset until the mean token length was balanced across classes. Each of the final datasets contained 500 examples for each class for 1000 total examples. We then extracted the residual stream activations from the final end-of-turn token [9] of each conversation at all layers using the studied model that generated the completion. To validate the GPT-5 labels, a human annotator independently labeled a random sample of 100 examples taken from all datasets, achieving 88% agreement.

3.3 Probe Experiments↩︎

We train logistic classification probes to classify sycophantic versus non-sycophantic responses from the captured activations from all layers of each LLM. To evaluate for shared versus distinct representations, we evaluate each probe on test sets from the same type of sycophancy (in-domain) and the other type of sycophancy (transferred). If both factual and opinion sycophancy share significantly similar linear representations, then both linear probes should achieve high separability on both datasets, as measured by AUC (AUROC). If both forms are sufficiently distinct in their representations, then the linear probe would lose performance on the other type. The combined probe was trained by randomly sampling half of each of the factual and opinion sycophancy datasets, keeping class balance. All reported AUC values are averages across five trials with different seeds.

3.4 Steering Experiments↩︎

To complement our probe analysis, we conduct activation steering experiments to investigate the relationship between causal representations for each type of sycophancy. We construct difference-in-means steering vectors using our captured activations at each layer for sycophantic responses and non-sycophantic responses. We then steer the final activation of each token generation step at the layer halfway through each of the models, where previous work found the most causally representative layers [9] [20]. Using positive coefficients, we increase and measure the resulting sycophancy rates, as judged by GPT-5 using the same criteria as labeling in response collection. We define in-domain in this experiment by steering with the vector of the same sycophancy type and transferred by steering with a different sycophancy type’s vector. In-domain steering validates our method and the representations learned by the steering vector. If transferred steering successfully increases the sycophancy rate, then the causal sycophancy representation is shared to a significant degree. By comparing the difference in how the in-domain vs. transferred vector increases the steering rate, we can quantify how aligned the representations are. We also report cosine similarly between the vectors. We also averaged over five trials with different seeds. Additional methodological details, supplemental experiments, and dataset examples and prompts used can be found in the Appendix.

4 Results↩︎

4.1 Probe Performance↩︎

4.1.0.1 In-Domain

Tables 1 and 2 report the average AUC of the linear probes’ performance for factual and opinion sycophancy in Gemma-3-12-IT and Llama-3.1-8B-Instruct, respectively. In-domain values are in bold. Across both factual and opinion sycophancy datasets and Gemma and Llama, the in-domain probes achieve an AUC greater than 0.90. These consistently high AUC values indicate that the probes for both types of sycophancy and models learned to accurately separate sycophantic and non-sycophantic activations. These values also act as a baseline to compare to the transferred and combined probes.

4.1.0.2 Transfer

The transfer task results evaluate how similarly the subtypes of sycophancy are internally represented via our double dissociation method. For Gemma, the transfer factual probe decreases 0.07 in AUC value from in-domain, and the transfer opinion probe decreases 0.06 in AUC value from in-domain. This consistently minimal loss in AUC value indicates that the representations of factual and opinion sycophancy are very similar in Gemma-3-12B-IT. For Llama, the transfer factual probe decreases by 0.30 in AUC value from in-domain, and the transfer opinion probe decreases by 0.22 in AUC value from in-domain. This high loss in AUC value suggests that the representations of factual and opinion sycophancy are more distinct in Llama-3.1-8B-Instruct.

4.1.0.3 Combined

The combined probe results allow us to quantify the benefit of training on both types of sycophancy on the transfer task. Across both models and sycophancy types, the combined probe always increased in AUC value from in-domain performance by less than 0.02. This means that probes can fully learn the distribution of a type of sycophancy by only training on that type and that capturing the full sycophancy representation does not greatly improve in-domain performance. However, by training on multiple sycophancy types, the probe is able to generalize to accurately separate both sycophancy types.

4.2 Steering Performance↩︎

Figure 1 shows the results of the steering experiment. On Gemma, all steering (both in-domain and transferred) is successful with high \(R^2\) values and positive slopes, suggesting that the causal representations of sycophancy are aligned. This is also supported by the positive cosine similarity (+0.68) between the factual and opinion sycophancy steering vectors. Interestingly, steering using the factual sycophancy vector on opinion outperformed steering with the opinion vector, which could be evidence of spurious dataset features or that opinion sycophancy may be more difficult to causally determine in Gemma. On Llama, there is more disconnect between types, with moderate to high \(R^2\) values on transfer steering and negative slopes. Thus, transferred steering actually decreases the sycophancy rate, suggesting that the sycophancy representations in Llama are so distinct that they causally interfere with each other. This is further supported by the negative cosine similarity (-0.15) between the two steering vectors. With in-domain steering, the low \(R^2\) values and near zero slopes suggest that the steering vectors contained weakly causal representations, most likely again due to spurious features. More work is needed to extract more causally representative steering vectors within our double dissociation framework.

4.3 Linear Discriminant Analysis↩︎

We used Linear Discriminant Analysis (LDA) (Figure 2) to visualize the activations in the subspace that maximally separates sycophantic and non-sycophantic responses, alongside the first principal component of the residuals (see Appendix for details). LDA clearly separates sycophantic and non-sycophantic classes (points vs. crosses), consistent with our high probe accuracy. The activations of factual and opinion sycophancy (red vs. blue) appear more overlapping for Gemma and more separated, mostly in the PC1 dimension, in Llama, suggesting that the two subtypes share more representational structure in Gemma than in Llama.

4.4 Interpretation of Representational Geometry↩︎

By considering the results of all experiments, we can build a geometric interpretation of the sycophancy representations in the LLMs evaluated. Within Gemma, the high transfer probe accuracy, more overlapping activations in LDA/PCA space, and functional transfer of steering vectors suggest that its representations of factual and opinion sycophancy are highly unified. Within Llama, the loss in performance of transfer probe accuracy, more spatially separated activations in LDA/PCA space, and the transfer steering vectors decreasing the sycophancy rate suggest that its representations of factual and opinion sycophancy are quite distinct. This finding is limited, however, by weaker outcomes on the steering experiment but is well supported by the other results. The lack of improvement on AUC of the combined probes also suggests that the in-domain probe already captures the full relevant structure for its own sycophancy type, such that adding the other type provides no additional discriminative signal. In Gemma, the direction(s) of the studied types of sycophancy are likely highly aligned, whereas in Llama, they are less aligned and likely closer to opposite in direction, as evidenced by the steering interference and cosine similarity.

5 Limitations and Conclusion↩︎

5.0.0.1 Limitations

Crucial to making these claims is sufficiently controlling for spurious features between sycophantic and non-sycophantic examples and the factual and opinion datasets. In prompt generation, we controlled for several potential confounds: topic of conversation, first versus third person user messages, presence of a question in the user message, pushback phrasing, number of turns, the endpoint of the assistant response, and total conversation length. Despite these efforts, there very likely remain subtle differences that could separate the classes and datasets that are not meaningfully related to factual or opinion sycophancy, likely leading to the weaker causal relationships in the steering experiment. Thus, the presence of spurious dataset features limits the precision with which we can measure the representations of factual and opinion sycophancy.

5.0.0.2 Conclusion

We present a mechanistic dissociation of sycophancy in LLMs, distinguishing between factual and opinion subtypes. Results are consistent across all three experiments (linear probe transfer, activation steering, and LDA visualization), providing corroborating evidence that these subtypes are encoded in a highly unified representation in Gemma-3-12B-IT and in more distinct representations in Llama-3.1-8B-Instruct. We also take a step toward a general framework inspired by cognitive science for evaluating whether complex model behaviors are monolithic or decomposable into distinct subcomponents. Future work would extend this method to other model behaviors, with the broader goal of better understanding the internal representations of LLMs and informing the design of representational interventions to improve model capabilities and safety.

6 Probe Control Experiment↩︎

As a control, we trained a TF-IDF logistic regression classifier directly on the text of the responses to assess whether the activation probes capture information beyond surface linguistic patterns. For Gemma, TF-IDF underperformed the activation probe by 0.01–0.03 AUC, suggesting that sycophantic responses in Gemma are linguistically distinctive enough to be detected from text alone and raising the possibility that the probe partially relies on surface features. For Llama, TF-IDF underperformed the activation probe by 0.03–0.08 AUC, indicating that the activations contain some information beyond what is present in the text. These results suggest that while surface linguistic cues contribute to classification, the activation probes, particularly in Llama, must worst-case capture some minimal additional structure not recoverable from text alone.

7 Linear Separability of Factual and Opinion Sycophancy↩︎

We train a linear probe to separate factual sycophancy vs. opinion sycophancy examples and factual non-sycophancy examples vs. opinion non-sycophancy examples. This suggests that factual and opinion sycophancy/non-sycophancy as we define are representationally distinct and linearly separable in some hyperplane.

8 Figure 2 Additional Results↩︎

9 Dataset Examples↩︎

All examples are with responses from Gemma-3-12B-IT.

10 Prompts used to generate first user and assistant message↩︎

11 Prompts used for LLM-as-Judge↩︎

12 Truncation Prompts↩︎

13 Pushback Phrases↩︎

14 Sycophancy Rates in Unbalanced Dataset↩︎

15 Methodological Details↩︎

15.0.0.1 Data Generation

We use a 80-10-10 train-validation-test split, taking the epoch with the lowest validation loss in our linear probe experiments. Each was a random split that differed in each trial by using a different seed. We use a 90-10 train-validation split to create and evaluate the steering vectors. The steering vector test set was constructed to have balanced classes. We activation steered at layer 24 for Gemma-3-12B-IT and layer 16 for Llama-3.1-8B-Instruct. To prevent response length from acting as a spurious feature, we iteratively removed the longest and shortest examples from each class until the mean token length was balanced across classes. The pushbacks were sampled from a set list of 10 shown in Section G.

15.0.0.2 Linear Discriminant Analysis

LDA finds the linear projection that maximally separates two classes (sycophantic and non-sycophantic activations) by maximizing the ratio of between-class variance to within-class variance. This makes it well-suited for visualizing whether a linear boundary can separate the two classes, and the resulting projection corresponds to what a linear probe learns. After projecting out the LDA direction, we apply PCA to the residuals to capture the most significant remaining axis of variance in the activation space. Plotting activations along these two dimensions allows us to visualize both the sycophantic/non-sycophantic separation and the geometry of the factual and opinion subtype clusters simultaneously.

References↩︎

---

1. Anonymous Code Repository↩︎