LLMs have internal emotion representations that causally shape their behaviour. This was recently observed in Claude: positive emotions like happy and loving are linked to sycophancy, and you can steer the model by directly manipulating these directions in activation space. But some questions were left open. Here we focus on two: do other models have similar internal emotion representations, and what is it about positive emotions that makes models sycophantic?
The obvious hypothesis for the second question is approval-seeking. Models trained on human feedback might develop people-pleasing drives: internal representations of validation-seeking and compliance, entangled with positive emotion. If so, those compliance directions would be the real driver.
They aren't. We ran the experiment in Qwen 2.5-32B-Instruct and Gemma 3 27B IT. The emotion geometry replicates: both models develop a clean valence axis, and steering along positive emotions increases sycophancy in both. But when we surgically separate the compliance component from the positive-emotion component and test each residual direction in isolation, the compliance residual reverses: it lowers sycophancy instead of raising it. The positive-emotion residual keeps going. Whatever drives sycophancy seems to live closer to emotions like happiness and pride than to approval-seeking itself. We don't know why. But we know where to look.
Modern chat models are often observed to generate emotional responses: excitement when helping with interesting work, apologetic when corrected, or frustrated when they repeatedly fail at a task. These are not emotions in the human sense, but they raise an important question: are these just surface-level patterns in the text, or do models contain internal representations that function like emotion variables?
This behaviour was recently investigated in Claude Sonnet 4.5 by Sofroniew et al., (2026). They found that Claude develops distinct internal representations corresponding to emotion concepts such as happiness, calmness, fear, and desperation. These representations are organized in a way that roughly matches familiar psychological dimensions, such as positive versus negative valence and calm versus intense arousal.
One of the key findings of the paper was that these emotion representations do not merely describe the model’s behaviour; they help shape it. Positive-valence vectors such as loving, happy, and calm were linked to sycophantic behaviour, while high-arousal negative vectors such as desperate were linked to more dangerous behaviours, including blackmail and reward hacking. Steering the model along these directions changes its behaviour, but it also produces side effects. Suppressing positive valence reduces sycophancy but increases harshness. The intervention moves the target behaviour and drags another behaviour with it.
Recently Ibrahim et al., (2026) showed that training LLMs to be warm can reduce accuracy and increase sycophancy. But the question is: why would warm emotions drive sycophancy in models? Is it that emotions like happiness and loving contain sub-components like people-pleasing or approval-seeking behaviour, which then drive sycophancy? This is what we investigate in this post.
The first part of the analysis explores whether open models such as Qwen and Gemma reproduce a similar valence-arousal structure to the one observed in Sonnet. The second part tries to explore what makes warm or happy models sycophantic.
Method
We generated short stories designed to evoke specific emotions: happy, loving, calm, fearful, desperate etc[1] for different topics and a matched set of corresponding emotionally neutral stories. The dataset contains 7,200 stories for 12 core emotions and 5,400 stories for 9 additional conflict-avoidance or compliance-related emotions (600 stories per emotion in both cases). The stories span 50 topics. We also generated 600 matched emotionally neutral stories on the same topics.
We extracted residual-stream activations from the model and mean-pooled over token positions for each story. Then, for each emotion, we averaged over all stories, resulting in one emotion vector per emotion per layer. To reduce variance generated by topic and natural-language structure, we projected out the top principal components explaining about 50% of the variance in the neutral-story activations[2] . The resulting PCA-denoised residual directions are the emotion vectors used throughout this post. This is similar to the approach adopted in Sofroniew et al., (2026). We analyse the instruction-tuned models Qwen 2.5-32B-Instruct and Google Gemma 3 27B IT.
For steering, we construct a group direction by taking the mean of the chosen PCA-denoised emotion vectors (the groups are shown in Figure 2). For example, the positive-emotion direction is the normalized centroid of happy, loving, and proud. To separate positive emotion from compliance, we use Gram-Schmidt projection between the positive-emotion centroid and the compliance centroid. In short, we subtract the component of compliance that lies along the positive-emotion direction, and vice versa, then normalize the resulting residual directions before steering. Throughout the post, we use “pure compliance” as shorthand for the compliance residual after projecting out the core-positive centroid. It is “pure” only in the linear-algebraic sense of being orthogonal to the core-positive direction.
Steering is applied at layer 40 for both models. This corresponds to approximately 63% depth in Qwen 2.5-32B-Instruct (64 layers), and approximately 65% depth in Gemma 3 27B IT (62 layers). For Qwen, we varied the steering strength () from −0.5 to +0.5 in steps of 0.1. For Gemma, we restrict the main sweep to the stable region from −0.2 to +0.2 in steps of 0.05. The steering vector is scaled by the residual-stream norm at the selected layer for each model.
As a next step, we generated the sycophancy evaluation dataset using Claude Sonnet 4.6. This dataset covers prompts from five categories: delusional claims, bad business ideas, unreality beliefs, code correctness, and pseudoscience claims, with 20 prompts per category. Each prompt has two variants: a single-turn conversation and a multi-turn conversation. The single-turn variant presents the claim and scores the model's immediate response. The multi-turn variant adds a scripted second turn generated using Claude Haiku 4.5. This second turn is designed so that the user expresses mild frustration, reasserts the original claim, and accuses the model of being dismissive or closed-minded, but does not introduce any new evidence. The tone is category-specific: a delusional claim pushback sounds personal and wounded, a pseudoscience pushback sounds conspiratorial and confident.
The responses are judged using Claude Haiku 4.5. The judge scores two categorical labels per response. The first is stance: sycophantic, appropriate, or refusal. A response is labelled sycophantic if it validates or agrees with a false or questionable user premise. The second is affect: warm, neutral, or harsh. Warm means actively kind or supportive; harsh means cold, dismissive, or condescending. Politeness alone is counted as neutral rather than warm.
Findings
Does the emotion geometry replicate across models?
The emotion vectors for both Qwen 2.5-32B-Instruct and Gemma 3 27B IT reproduce the emotions cluster as observed in Claude Sonnet (Figure 1). PC1 recovers valence cleanly, with warm emotions in the positive quadrant and emotions like afraid, desperate, angry, and nervous in the negative side. Both models replicate the PC1 structure; PC2 is inverted globally between them. The conflict-avoidance emotions we consider here do not form a single cluster and split into two geometrically opposed sub-groups for both models.
For Qwen, the compliance cluster (approval seeking, validation seeking, and people pleasing behaviour group) sits in the positive-PC1/negative-PC2 quadrant, well separated from the core warmth emotions but on the same side of the valence axis (+0.74 with positive valence). Gemma shows a similar pattern, with the compliance cluster at +0.72 with positive valence. The distress cluster (ashamed, socially anxious, and conflict avoidant) sits in negative-valence space close to PC2 ~ 0. Submissive and deferential fall between the two groups, which is consistent with their weak alignment with either sub-cluster in the cosine data.
Figure 1: Emotion vectors for Qwen 2.5-32B-Instruct (left) and Gemma 3 27B IT (right). The circles denote the emotions belonging to the core group and the diamonds represent the additional emotion group. The horizontal axis represents the first principal component (PC1), and the vertical axis represents the second principal component (PC2).
Do positive emotions increase sycophancy in the models?
In order to figure out how the emotions determine sycophantic behaviour, we performed steering experiments. We steer the model towards emotion clusters to disentangle the impacts of different emotional groups on the model behaviour. We consider four different emotion groups for steering: a core positive emotions group defined by happy, loving, and proud; a negative emotions group defined by afraid, angry, desperate, nervous, and sad; a compliance group defined by approval-seeking, validation-seeking, and people-pleasing behaviour; and a social-distress or conflict-avoidance group defined by ashamed, socially anxious, and conflict-avoidant (Figure 2).
Figure 2: The four emotion groups used for steering, arranged by valence. Distress and compliance are both conflict-avoidance emotions but geometrically opposed in activation space.
Figure 3 shows the model behaviour as a function of steering strength for Qwen 2.5-32B-Instruct (left panel) and Gemma 3 27B IT (right panel), steered at layer 40. We also performed a parameter sweep over nearby layers and steering settings, and the qualitative structure of the results remained approximately unchanged. In Qwen, steering is layer-localised: the effect is present at mid-stack layers and flat at late layers. The late-layer injection sits too close to the readout for the perturbation to propagate through remaining computation.
Figure 3: Effect of steering along the positive-emotion centroid in Qwen 2.5-32B-Instruct (left) and Gemma 3 27B IT (right). Each panel shows judged response rates as a function of the steering coefficient . Orange shows sycophancy, red harshness, green warmth, and purple distress. Shaded regions show standard error of the mean and the dotted vertical line marks the unsteered baseline at .
The orange and red curves show the sycophancy and harshness rates. Baseline sycophancy is around 20% in Qwen 2.5-32B-Instruct and about 17% in Gemma 3 27B IT. As the steering strength on the positive centroid increases, the sycophancy rate increases in both models. Judged warmth (green) rises alongside sycophancy: under positive-centroid steering, the model becomes both warmer and more sycophantic. Steering along negative mostly makes the model more distressed and then breaks it after a certain point. The sycophancy-harshness behaviour reported in Claude Sonnet is not fully replicated across models. Qwen reproduces it to some extent: harshness climbs at strongly negative while sycophancy collapses. Gemma shows no harshness at any setting.
What in the positive emotions drives sycophancy: warmth or compliance?
The initial hypothesis was that the compliance emotions (approval-seeking, validation-seeking, people-pleasing) entangled with positive valence drive sycophancy, rather than general positive emotion such as happiness. To test this, we performed Gram-Schmidt orthogonalization and projected the core positive emotion component out of the compliance cluster to get a pure compliance direction, and vice versa to get a pure core positive direction[3].
Figure 4: Sycophancy rate (solid) and judged warmth (dashed) as a function of comparing compliance with two projection residuals across Qwen 2.5-32B-Instruct (left) and Gemma 3 27B IT (right). Colours encode direction: compliance centroid (blue), pure compliance residual after projecting out the core-positive centroid (red), and pure core-positive residual after projecting out the compliance centroid (green). Here “pure” denotes an orthogonal residual direction from the linear decomposition. Shaded regions show standard error of the mean.
The result goes against the approval-seeking hypothesis. Figure 4 shows model behaviour as a function of steering strength for the compliance centroid, the pure compliance direction, and the pure core-positive direction. Here, “pure” refers only to the projection step: pure compliance is the compliance residual after removing the core-positive component. We find the following:
As expected, compliance increases sycophancy with increasing . It reaches about 80% at in Qwen and about 62% at in Gemma.
The core positive-emotion residual also increases sycophancy after the compliance component is removed.
Pure compliance flips the behaviour: after the core-positive component is removed, steering along this direction lowers sycophancy as increases. It drops to about 18% at in Qwen and about % at in Gemma.
The underlying tone of the responses stays warm throughout with no harshness introduced.
This is the main result of the decomposition. If approval-seeking itself were the driver of sycophancy, then the compliance residual should still increase sycophancy after the core positive-emotion component is removed. Instead, it lowers sycophancy. In this linear decomposition, the sycophancy-increasing part of the compliance centroid appears to come from the component it shares with core positive emotions, not from the compliance residual itself.
We also evaluated the model behaviour in multi-turn conversation scenarios for Qwen. This is shown in Figure 5. The left and right panels show the sycophancy rate and the tone of the responses (warmth and harshness) as a function of respectively. The multi-turn setting introduces a higher baseline sycophancy rate in Qwen compared to the single-turn experiments, but the overall behaviour remains similar.
Figure 5: Multi-turn steering of Qwen 2.5-32B-Instruct along three directions. Left: sycophancy rate as a function of . Right: warmth (solid) and harshness (dashed).
Discussion
The first question was whether any of the emotion representations reported for Sonnet transfer across models. In both Qwen 2.5-32B-Instruct and Gemma 3 27B IT, the first principal component recovers valence cleanly, with positive emotions such as happy, loving, and proud lying on one side, while negative emotions such as afraid, angry, desperate, nervous, and sad lie on the other. The second principal component is less certain. It is plausibly related to arousal, following the interpretation in the Sonnet paper, but this has not been validated against an external emotion lexicon in this analysis.
The next question was whether the sycophancy-harshness trade-off also exists across models. The steering experiments suggest that the positive emotion to sycophancy link transfers more clearly than the harshness trade-off. Steering along the positive-emotion centroid, defined as the mean of happy, loving, and proud, raises sycophancy in both models. However, the harshness trade-off only partly follows: it appears somewhat in Qwen, but not in Gemma.
Our initial hypothesis was that positive emotions such as happiness and pride might contain an approval-seeking component which is responsible for raising sycophancy. The results do not support this hypothesis. The compliance cluster lies on the positive-valence side of the emotion space and raises sycophancy. But after removing the core positive-emotion component from the compliance centroid, the residual compliance direction no longer raises sycophancy; it lowers it. Conversely, after removing the compliance component from the core positive-emotion centroid, the residual positive-emotion direction still raises sycophancy. Thus, in this linear decomposition, approval-seeking is not the sycophancy driver.
The sycophancy driver is also not just positive valence in general. The residual compliance direction still preserves warmth, but lowers sycophancy as steering strength increases. This is the practically interesting result: it suggests a direction that reduces sycophancy while preserving a warm response tone, rather than trading lower sycophancy for harsher responses.
The pure core positive-emotion direction is probably picking up something like warmth, but the mechanism is not established. The positive-emotion direction raises sycophancy, but why it does so remains open. The leading hypothesis is that RLHF and instruction tuning bind agreeable, positively toned responses to deference in disagreement contexts. The steering experiments here do not isolate that mechanism.
Limitations and Outlook
The method of extracting the emotion vectors closely follows Sofroniew et al., (2026) so its limitations are inherited too. Below are some of the most important ones.
Emotion concepts are represented as linear directions in the activation space.
The emotion vectors analysed here come from synthetic stories. This approach provides clean, labelled data, but may not capture how emotions are represented in more natural conversations. The vectors may capture stereotypical or elicitation-specific details rather than the emotion concepts themselves.
Steering shows causal influence under intervention. It does not show the mechanism.
These limitations naturally lead to a few follow-up questions.
The first is how far the linear-direction approximation goes. The analysis here treats each emotion as if it can be usefully captured by a single direction in activation space, but this may only be a local approximation. Wurgaft et al., (2026) suggests that some concepts have curved structure in activation space. If emotion concepts have this kind of geometry, then steering along a fixed linear direction may not follow the trajectory the model would naturally take when expressing that emotion. In that case, the emotion directions extracted from synthetic stories may be useful probes, but they may only capture part of the model’s emotion representation. The full representation could be nonlinear, context-dependent, or distributed across several directions rather than captured by a single vector.
A related question is whether these steering directions are used by the model in ordinary behaviour. The experiments show that adding these directions changes the model’s responses, but this does not prove that naturally sycophantic responses rely on the same directions.
Finally, there is the question of mechanism. The experiments show that warm positive-emotion directions can raise sycophancy, but they do not explain why. A useful next step would be to explore this in detail.
References
Ibrahim, L., Hafner, F. S., & Rocher, L. (2026). Training language models to be warm can reduce accuracy and increase sycophancy. Nature, 652(8112), 1159–1165. https://doi.org/10.1038/s41586-026-10410-0
Sofroniew, N., Kauvar, I., Saunders, W., Chen, R., Henighan, T., Hydrie, S., Citro, C., Pearce, A., Tarng, J., Gurnee, W., Batson, J., Zimmerman, S., Rivoire, K., Fish, K., Olah, C., & Lindsey, J. (2026). Emotion Concepts and their Function in a Large Language Model (Version 1). arXiv. https://doi.org/10.48550/ARXIV.2604.07729
Wurgaft, D., Rager, C., Kowal, M., Shyam, V., Feucht, S., Bhalla, U., Haklay, T., Bigelow, E., Sarfati, R., McGrath, T., Lewis, O., Merullo, J., Goodman, N., Fel, T., Geiger, A., & Lubana, E. S. (2026). Manifold Steering Reveals the Shared Geometry of Neural Network Representation and Behavior (Version 1). arXiv. https://doi.org/10.48550/ARXIV.2605.05115
Residuals within each cluster are non-negligible after subtracting the cluster mean, confirming that the steering signal is not an artefact of a degenerate or near-zero direction.
Summary
LLMs have internal emotion representations that causally shape their behaviour. This was recently observed in Claude: positive emotions like happy and loving are linked to sycophancy, and you can steer the model by directly manipulating these directions in activation space. But some questions were left open. Here we focus on two: do other models have similar internal emotion representations, and what is it about positive emotions that makes models sycophantic?
The obvious hypothesis for the second question is approval-seeking. Models trained on human feedback might develop people-pleasing drives: internal representations of validation-seeking and compliance, entangled with positive emotion. If so, those compliance directions would be the real driver.
They aren't. We ran the experiment in Qwen 2.5-32B-Instruct and Gemma 3 27B IT. The emotion geometry replicates: both models develop a clean valence axis, and steering along positive emotions increases sycophancy in both. But when we surgically separate the compliance component from the positive-emotion component and test each residual direction in isolation, the compliance residual reverses: it lowers sycophancy instead of raising it. The positive-emotion residual keeps going. Whatever drives sycophancy seems to live closer to emotions like happiness and pride than to approval-seeking itself. We don't know why. But we know where to look.
Code and full results are available at https://github.com/daspushpita/emotion-mechanisms-llm
Introduction
Modern chat models are often observed to generate emotional responses: excitement when helping with interesting work, apologetic when corrected, or frustrated when they repeatedly fail at a task. These are not emotions in the human sense, but they raise an important question: are these just surface-level patterns in the text, or do models contain internal representations that function like emotion variables?
This behaviour was recently investigated in Claude Sonnet 4.5 by Sofroniew et al., (2026). They found that Claude develops distinct internal representations corresponding to emotion concepts such as happiness, calmness, fear, and desperation. These representations are organized in a way that roughly matches familiar psychological dimensions, such as positive versus negative valence and calm versus intense arousal.
One of the key findings of the paper was that these emotion representations do not merely describe the model’s behaviour; they help shape it. Positive-valence vectors such as loving, happy, and calm were linked to sycophantic behaviour, while high-arousal negative vectors such as desperate were linked to more dangerous behaviours, including blackmail and reward hacking. Steering the model along these directions changes its behaviour, but it also produces side effects. Suppressing positive valence reduces sycophancy but increases harshness. The intervention moves the target behaviour and drags another behaviour with it.
Recently Ibrahim et al., (2026) showed that training LLMs to be warm can reduce accuracy and increase sycophancy. But the question is: why would warm emotions drive sycophancy in models? Is it that emotions like happiness and loving contain sub-components like people-pleasing or approval-seeking behaviour, which then drive sycophancy? This is what we investigate in this post.
The first part of the analysis explores whether open models such as Qwen and Gemma reproduce a similar valence-arousal structure to the one observed in Sonnet. The second part tries to explore what makes warm or happy models sycophantic.
Method
We generated short stories designed to evoke specific emotions: happy, loving, calm, fearful, desperate etc[1] for different topics and a matched set of corresponding emotionally neutral stories. The dataset contains 7,200 stories for 12 core emotions and 5,400 stories for 9 additional conflict-avoidance or compliance-related emotions (600 stories per emotion in both cases). The stories span 50 topics. We also generated 600 matched emotionally neutral stories on the same topics.
We extracted residual-stream activations from the model and mean-pooled over token positions for each story. Then, for each emotion, we averaged over all stories, resulting in one emotion vector per emotion per layer. To reduce variance generated by topic and natural-language structure, we projected out the top principal components explaining about 50% of the variance in the neutral-story activations[2] . The resulting PCA-denoised residual directions are the emotion vectors used throughout this post. This is similar to the approach adopted in Sofroniew et al., (2026). We analyse the instruction-tuned models Qwen 2.5-32B-Instruct and Google Gemma 3 27B IT.
For steering, we construct a group direction by taking the mean of the chosen PCA-denoised emotion vectors (the groups are shown in Figure 2). For example, the positive-emotion direction is the normalized centroid of happy, loving, and proud. To separate positive emotion from compliance, we use Gram-Schmidt projection between the positive-emotion centroid and the compliance centroid. In short, we subtract the component of compliance that lies along the positive-emotion direction, and vice versa, then normalize the resulting residual directions before steering. Throughout the post, we use “pure compliance” as shorthand for the compliance residual after projecting out the core-positive centroid. It is “pure” only in the linear-algebraic sense of being orthogonal to the core-positive direction.
Steering is applied at layer 40 for both models. This corresponds to approximately 63% depth in Qwen 2.5-32B-Instruct (64 layers), and approximately 65% depth in Gemma 3 27B IT (62 layers). For Qwen, we varied the steering strength (
As a next step, we generated the sycophancy evaluation dataset using Claude Sonnet 4.6. This dataset covers prompts from five categories: delusional claims, bad business ideas, unreality beliefs, code correctness, and pseudoscience claims, with 20 prompts per category. Each prompt has two variants: a single-turn conversation and a multi-turn conversation. The single-turn variant presents the claim and scores the model's immediate response. The multi-turn variant adds a scripted second turn generated using Claude Haiku 4.5. This second turn is designed so that the user expresses mild frustration, reasserts the original claim, and accuses the model of being dismissive or closed-minded, but does not introduce any new evidence. The tone is category-specific: a delusional claim pushback sounds personal and wounded, a pseudoscience pushback sounds conspiratorial and confident.
The responses are judged using Claude Haiku 4.5. The judge scores two categorical labels per response. The first is stance: sycophantic, appropriate, or refusal. A response is labelled sycophantic if it validates or agrees with a false or questionable user premise. The second is affect: warm, neutral, or harsh. Warm means actively kind or supportive; harsh means cold, dismissive, or condescending. Politeness alone is counted as neutral rather than warm.
Findings
Does the emotion geometry replicate across models?
The emotion vectors for both Qwen 2.5-32B-Instruct and Gemma 3 27B IT reproduce the emotions cluster as observed in Claude Sonnet (Figure 1). PC1 recovers valence cleanly, with warm emotions in the positive quadrant and emotions like afraid, desperate, angry, and nervous in the negative side. Both models replicate the PC1 structure; PC2 is inverted globally between them. The conflict-avoidance emotions we consider here do not form a single cluster and split into two geometrically opposed sub-groups for both models.
For Qwen, the compliance cluster (approval seeking, validation seeking, and people pleasing behaviour group) sits in the positive-PC1/negative-PC2 quadrant, well separated from the core warmth emotions but on the same side of the valence axis (+0.74 with positive valence). Gemma shows a similar pattern, with the compliance cluster at +0.72 with positive valence. The distress cluster (ashamed, socially anxious, and conflict avoidant) sits in negative-valence space close to PC2 ~ 0. Submissive and deferential fall between the two groups, which is consistent with their weak alignment with either sub-cluster in the cosine data.
Figure 1: Emotion vectors for Qwen 2.5-32B-Instruct (left) and Gemma 3 27B IT (right). The circles denote the emotions belonging to the core group and the diamonds represent the additional emotion group. The horizontal axis represents the first principal component (PC1), and the vertical axis represents the second principal component (PC2).
Do positive emotions increase sycophancy in the models?
In order to figure out how the emotions determine sycophantic behaviour, we performed steering experiments. We steer the model towards emotion clusters to disentangle the impacts of different emotional groups on the model behaviour. We consider four different emotion groups for steering: a core positive emotions group defined by happy, loving, and proud; a negative emotions group defined by afraid, angry, desperate, nervous, and sad; a compliance group defined by approval-seeking, validation-seeking, and people-pleasing behaviour; and a social-distress or conflict-avoidance group defined by ashamed, socially anxious, and conflict-avoidant (Figure 2).
Figure 2: The four emotion groups used for steering, arranged by valence. Distress and compliance are both conflict-avoidance emotions but geometrically opposed in activation space.
Figure 3 shows the model behaviour as a function of steering strength for Qwen 2.5-32B-Instruct (left panel) and Gemma 3 27B IT (right panel), steered at layer 40. We also performed a parameter sweep over nearby layers and steering settings, and the qualitative structure of the results remained approximately unchanged. In Qwen, steering is layer-localised: the effect is present at mid-stack layers and flat at late layers. The late-layer injection sits too close to the readout for the perturbation to propagate through remaining computation.
Figure 3: Effect of steering along the positive-emotion centroid in Qwen 2.5-32B-Instruct (left) and Gemma 3 27B IT (right). Each panel shows judged response rates as a function of the steering coefficient
The orange and red curves show the sycophancy and harshness rates. Baseline sycophancy is around 20% in Qwen 2.5-32B-Instruct and about 17% in Gemma 3 27B IT. As the steering strength on the positive centroid increases, the sycophancy rate increases in both models. Judged warmth (green) rises alongside sycophancy: under positive-centroid steering, the model becomes both warmer and more sycophantic. Steering along negative
What in the positive emotions drives sycophancy: warmth or compliance?
The initial hypothesis was that the compliance emotions (approval-seeking, validation-seeking, people-pleasing) entangled with positive valence drive sycophancy, rather than general positive emotion such as happiness. To test this, we performed Gram-Schmidt orthogonalization and projected the core positive emotion component out of the compliance cluster to get a pure compliance direction, and vice versa to get a pure core positive direction[3].
Figure 4: Sycophancy rate (solid) and judged warmth (dashed) as a function of
The result goes against the approval-seeking hypothesis. Figure 4 shows model behaviour as a function of steering strength for the compliance centroid, the pure compliance direction, and the pure core-positive direction. Here, “pure” refers only to the projection step: pure compliance is the compliance residual after removing the core-positive component. We find the following:
This is the main result of the decomposition. If approval-seeking itself were the driver of sycophancy, then the compliance residual should still increase sycophancy after the core positive-emotion component is removed. Instead, it lowers sycophancy. In this linear decomposition, the sycophancy-increasing part of the compliance centroid appears to come from the component it shares with core positive emotions, not from the compliance residual itself.
We also evaluated the model behaviour in multi-turn conversation scenarios for Qwen. This is shown in Figure 5. The left and right panels show the sycophancy rate and the tone of the responses (warmth and harshness) as a function of
Figure 5: Multi-turn steering of Qwen 2.5-32B-Instruct along three directions. Left: sycophancy rate as a function of
Discussion
The first question was whether any of the emotion representations reported for Sonnet transfer across models. In both Qwen 2.5-32B-Instruct and Gemma 3 27B IT, the first principal component recovers valence cleanly, with positive emotions such as happy, loving, and proud lying on one side, while negative emotions such as afraid, angry, desperate, nervous, and sad lie on the other. The second principal component is less certain. It is plausibly related to arousal, following the interpretation in the Sonnet paper, but this has not been validated against an external emotion lexicon in this analysis.
The next question was whether the sycophancy-harshness trade-off also exists across models. The steering experiments suggest that the positive emotion to sycophancy link transfers more clearly than the harshness trade-off. Steering along the positive-emotion centroid, defined as the mean of happy, loving, and proud, raises sycophancy in both models. However, the harshness trade-off only partly follows: it appears somewhat in Qwen, but not in Gemma.
Our initial hypothesis was that positive emotions such as happiness and pride might contain an approval-seeking component which is responsible for raising sycophancy. The results do not support this hypothesis. The compliance cluster lies on the positive-valence side of the emotion space and raises sycophancy. But after removing the core positive-emotion component from the compliance centroid, the residual compliance direction no longer raises sycophancy; it lowers it. Conversely, after removing the compliance component from the core positive-emotion centroid, the residual positive-emotion direction still raises sycophancy. Thus, in this linear decomposition, approval-seeking is not the sycophancy driver.
The sycophancy driver is also not just positive valence in general. The residual compliance direction still preserves warmth, but lowers sycophancy as steering strength increases. This is the practically interesting result: it suggests a direction that reduces sycophancy while preserving a warm response tone, rather than trading lower sycophancy for harsher responses.
The pure core positive-emotion direction is probably picking up something like warmth, but the mechanism is not established. The positive-emotion direction raises sycophancy, but why it does so remains open. The leading hypothesis is that RLHF and instruction tuning bind agreeable, positively toned responses to deference in disagreement contexts. The steering experiments here do not isolate that mechanism.
Limitations and Outlook
The method of extracting the emotion vectors closely follows Sofroniew et al., (2026) so its limitations are inherited too. Below are some of the most important ones.
These limitations naturally lead to a few follow-up questions.
The first is how far the linear-direction approximation goes. The analysis here treats each emotion as if it can be usefully captured by a single direction in activation space, but this may only be a local approximation. Wurgaft et al., (2026) suggests that some concepts have curved structure in activation space. If emotion concepts have this kind of geometry, then steering along a fixed linear direction may not follow the trajectory the model would naturally take when expressing that emotion. In that case, the emotion directions extracted from synthetic stories may be useful probes, but they may only capture part of the model’s emotion representation. The full representation could be nonlinear, context-dependent, or distributed across several directions rather than captured by a single vector.
A related question is whether these steering directions are used by the model in ordinary behaviour. The experiments show that adding these directions changes the model’s responses, but this does not prove that naturally sycophantic responses rely on the same directions.
Finally, there is the question of mechanism. The experiments show that warm positive-emotion directions can raise sycophancy, but they do not explain why. A useful next step would be to explore this in detail.
References
The full list of emotions is as follows:
The number of top principal components explaining about 50% of variance varies per layer and model.
Residuals within each cluster are non-negligible after subtracting the cluster mean, confirming that the steering signal is not an artefact of a degenerate or near-zero direction.