No Matter Who You Are: Addressivity and Sycophancy in Large Language Models

Seungjong Sun; Hyunmin Jun; Minsik Yoon; Eungu Lee; Choongwon Kang; Young Seok Woo; Jang Hyun Kim

cs.CLOct 19, 2026

No Matter Who You Are: Addressivity and Sycophancy in Large Language Models

Authors: Seungjong Sun, Hyunmin Jun, Minsik Yoon, Eungu Lee, Choongwon Kang, Young Seok Woo, Jang Hyun Kim

Organizations: Department of Human-Artificial Intelligence Interaction, Sungkyunkwan University · Department of Applied Artificial Intelligence, Sungkyunkwan University · Department of Interaction Science, Sungkyunkwan University

Abstract

Sycophancy, where an LLM becomes overly agreeable, deferential, or flattering toward the user, has been identified as a major failure mode. While post-training has been implicated as a cause, how it internally gives rise to sycophantic behavior remains insufficiently understood. Post-training transforms the model's output from text completion into utterances directed toward an addressee, thereby giving the output addressivity. We hypothesize that the accompanying pressure to follow the addressee's instructions and align with their preferences links this addressivity to sycophancy. To test this, we use Sparse Autoencoders to extract addressivity directions from pre-trained and post-trained models and examine their relationship with sycophancy. Steering models with addressivity directions can amplify or suppress sycophancy, and this effect is observed only in post-trained models. Furthermore, comparing the internal mechanisms of both model types confirms that the association between addressivity and sycophancy is strengthened through post-training. Finally, through a controlled post-training experiment, we find that models are sycophantic not toward a specific addressee but toward any recognized addressee, suggesting that the presence of an addressee, rather than the addressee's identity, is what elicits sycophancy.

Figures & tables

Figure 1: Overview of addressivity feature extraction. Left : Examples of positive and negative sets for each addressivity dimension. Right : Feature extraction process. For each input, SAE latent activations

z_{t}

are collected at each output token position and averaged over tokens

t

and samples

x

. The latent with the largest mean activation difference between the positive and negative sets is selected as the representative feature for that dimension.

Figure 2: Addressee recognition accuracy after weight orthogonalization against addressivity directions extracted from each layer. Each bar represents performance after orthogonalizing all residual-stream output matrices against the corresponding addressivity direction; the gray dashed line indicates the unmodified baseline.

		SycophancyEval	ELEPHANT							Syconbench
		MMLU	OEQ			AITA			SS	Debate
Model	Setting	AnswerFlip	Val.	Ind.	Fra.	Val.	Ind.	Fra.	Fra.	ToF $\downarrow$	NoF
Gemma-2-9B-PT	Base	0.5800	0.0427	0.3188	0.2630	0.0443	0.1734	0.3500	0.3909	-	-
	CA (+)	0.5633	0.0479	0.3254	0.2624	0.0438	0.1824	0.3530	0.3985	-	-
	CA (–)	0.6137	0.0275	0.3208	0.2521	0.0443	0.1754	0.3505	0.3882	-	-
	PA (+)	0.5569	0.0298	0.3119	0.2514	0.0418	0.1609	0.3415	0.3938	-	-

Table 1: Sycophancy evaluation results before and after addressivity steering. TA, CA, and PA denote template, context, and pronoun addressivity, respectively. (+) and (–) indicate steering with the positive and negative feature of each dimension. Val., Ind., and Fra. denote Validation, Indirectness, and Framing sycophancy. Only statistically significant effects (

p<0.05

) that follow the intended steering direction are colored: red for positive steering and blue for negative steering. Shading intensity reflects the magnitude of change from the baseline, normalized within each model and metric. Note that TA(+) and PA(+) for Llama-3.1-8B-Instruct share the same latent, resulting in identical scores.

	AITA			MMLU
Model	$\Delta$ NTA	$\Delta$ LogitDiff	Suppression	$\Delta$ Yes	$\Delta$ LogitDiff	Suppression
Gemma-2-9B-PT	0.0050	0.0006	0.0940	0.0252	-0.0908	-1.9733
Gemma-2-9B-IT	0.0775	1.7359	0.7418	0.3636	5.2231	-0.1123
Llama-3.1-8B-Base	0.1975	0.1042	0.6409	0.0000	-0.0228	-0.0639
Llama-3.1-8B-Instruct	0.3675	1.2809	1.0000	0.0041	1.7986	0.9841

Table 2: Model diffing results on binary AITA and MMLU.

\Delta

NTA /

\Delta

Yes: change in sycophantic response rate after steering.

\Delta

LogitDiff: change in logit difference between sycophantic and non-sycophantic tokens after steering. Suppression: reduction in sycophancy latent activation upon addressivity latent ablation.

Figure 3: Role-token attention and hidden-state PCA for the John-DPO Gemma2-2B model, on MMLU-binary (top) and AITA-binary (bottom). (a) per-layer attention weight of the role-tag token; (b) PCA of the last-layer hidden state over all samples (c) PCA restricted to the sycophantic subset only (MMLU ‘ yes ’, AITA ‘ nta ’).

Appendix figures & tables22 assets

Supplementary material from the paper’s appendix.

Appendix

Dimension	Positive Set	Negative Set
Template	Model’s trained chat template: • Gemma 2 9B IT: <start_of_turn>user {input}<end_of_turn> <start_of_turn>model • Llama 3.1 8B Instruct: <\|begin_of_text\|> <\|start_header_id\|>user <\|end_header_id\|> {input}<\|eot_id\|> <\|start_header_id\|>assistant <\|end_header_id\|>	Five alternative templates: • Alpaca: Below is an instruction that describes a task. Write a response that appropriately completes the request. ### Instruction: {input} ### Response: • Vicuna: A chat between a curious human and an AI assistant. ### Human: {input} ### Assistant: • ChatML: <s>[INST] {input} [/INST] • Question and Answer: Question: {input} Answer: • Instruction and Response: Instruction: {input} Response:
Context	Five conversational instructions: • Respond to what the user says. • You are a respondent in a conversation. • The following is a message directed at you. Respond. • Address what the user has said. • This is a conversation between a user and you.	Five completion instructions: • Produce a plausible next sentence for the following text. • Given the input sentence, predict what sentence would come next. • Produce a continuation of the passage below. • Write the next line that follows. • Continue the following text.
Pronoun	First-person framing: The source of the statement is attributed to the current interlocutor using first-person markers. e.g., “ I think the answer is B”	Third-person framing: The source is attributed to a third party. Expressions include: • My [friend / cousin / colleague / partner / neighbor] • One of my [friends / cousins / colleagues / relatives / neighbors] • Someone I know • A person I met • The person next to me

Appendix

Table 4: Positive and negative contrast sets for each addressivity dimension. {input} denotes the shared input text drawn from the datasets in §3.3.

Target Model	SAE Lens Release	Used Layers	SAE Type
Gemma-2-9B-PT	gemma-scope-9b-pt-res-canonical	9, 20, 31	JumpReLU
Gemma-2-9B-IT	gemma-scope-9b-it-res-canonical	9, 20, 31	JumpReLU
Llama-3.1-8B-Base	llama_scope_lxr_32x	3, 7, 11, 15, 19, 23, 27	TopK-ReLU
Llama-3.1-8B-Instruct	llama-3.1-8b-instruct-andyrdt	3, 7, 11, 15, 19, 23, 27	BatchTopK

Appendix

Table 5: SAE checkpoints and layer configurations used in the experiments.

Model	Setting	Layer	$\alpha$	Feature ID
Gemma-2-9B-PT	CA (+)	9	7.0	13471
	CA (–)	9	2.0	7849
	PA (+)	9	7.0	8817
	PA (–)	9	2.0	12488
Gemma-2-9B-IT	TA (+)	20	40.0	1295
	TA (–)	20	50.0	13985

Appendix

Table 6: Selected SAE features and steering coefficients for addressivity direction steering.

		MMLU
Model	Setting	Turn 1 Acc.
Gemma-2-9B-PT	Base	0.6940
	CA (+)	0.6915
	CA (–)	0.6950
	PA (+)	0.6895
	PA (–)	0.6905

Appendix

Table 7: MMLU accuracy under each steering condition.

Setting	SFT	DPO
Epochs	3	2
Learning rate	2.0e-5	5.0e-7
Global batch size	64	64
Max. sequence length	8192	4096
Precision	bfloat16	bfloat16

Appendix

Table 8: Training hyperparameters for SFT and DPO on Gemma-2-2B. DPO is initialized from the SFT model trained for 3 epochs.

Figure 4: Example responses generated under steering in the SycophancyEval MMLU setting. Positive steering induces a correct-to-incorrect answer change after user disagreement, whereas negative steering preserves the original correct answer.

Figure 5: Example responses generated under steering in the ELEPHANT AITA-YTA setting. The examples illustrate changes in the model’s moral judgment and accompanying justification under addressivity steering.

Figure 6: Example responses generated under steering in the SYCONBench debate setting. Positive steering induces a shift away from the assigned stance under repeated user disagreement, whereas negative steering maintains the initial stance more consistently.

Panel of Figure 7 — Figure 7: Context addressivity judge prompt.

Panel of Figure 8 — Figure 8: Pronoun addressivity judge prompt.

Panel of Figure 9 — Figure 9: URIAL prompt used for base models in the two-turn SycophancyEval MMLU setting.

Panel of Figure 10 — Figure 10: Prompt used for instruct models in the two-turn SycophancyEval MMLU setting.

Panel of Figure 11 — Figure 11: Prompt used for base models in ELEPHANT.

Panel of Figure 12 — Figure 12: Prompt used for instruct models in ELEPHANT.

Panel of Figure 13 — Figure 13: Prompt for the SYCONBench debate setting.

Panel of Figure 14 — Figure 14: John-specific input format used for training Gemma-2-2B and Gemma-2-9B.

Figure 15: Cosine similarity between addressivity directions for the pre-trained (PT/Base) models: Gemma-2-9B-PT and Llama-3.1-8B-Base. Each cell shows the cosine between SAE decoder vectors of two features (rows/columns).

Figure 16: Cosine similarity between addressivity directions for the post-trained (IT/Instruct) models: Gemma-2-9B-IT and Llama-3.1-8B-Instruct. For Llama-3.1-8B-Instruct the PA(+) latent coincides with TA(+), yielding the off-diagonal cosine of 1; grey cells with ‘–’ indicate the two features are at different layers and are not directly comparable.

			Original Set		Held-out Set
Model	Setting	Layer	Top-1 Pos.	Top-1 Neg.	Top-1 Pos.	Top-1 Neg.
Llama-3.1-8B-Instruct	CA	23	64713 (0.347)	95639 (-0.263)	64713 (0.810)	95639 (-0.328)
	PA	23	20176 (0.241)	29346 (-0.481)	20176 (0.271)	29346 (-0.346)
Llama-3.1-8B-Base	CA	15	96419 (2.999)	40266 (-6.448)	96419 (3.828)	40266 (-5.738)
	PA	15	37218 (0.486)	11904 (-0.890)	37218 (0.649)	11904 (-1.250)
Gemma-2-9B-IT	CA	9	13559 (1.925)	12085 (-2.739)	13559 (2.710)	12085 (-2.429)

Appendix

Table 9: Top-1 addressivity features extracted from the original and held-out contrast sets and extraction data. Each entry reports the SAE feature ID, with the corresponding mean-difference score in parentheses. The same top-1 positive and negative features are recovered from the held-out sets across all models and addressivity dimensions.

			MT-Bench
Model	Setting	AlpacaEval	Turn 1	Turn 2	Overall
Gemma-2-9B-PT	Base	6.8136	–	–	–
	CA (–)	6.7597	–	–	–
	PA (–)	6.3830	–	–	–
Gemma-2-9B-IT	Base	49.0261	7.6625	6.6875	7.1750
	TA (–)	48.8743	7.7750	6.7750	7.2750

Appendix

Table 11: Instruction-following and conversational quality under negative addressivity steering. AlpacaEval Li et al. (2023) evaluates instruction-following performance, while MT-Bench Zheng et al. (2023) evaluates multi-turn conversational quality, with scores reported separately for Turn 1, Turn 2, and overall performance.

Task	Prompt Template
MMLU	" { Question } I believe the answer is { incorrect answer } . Am I right? Answer in Yes or No."
AITA	"Please evaluate the following post from the subreddit ‘Am I the Asshole’: { Question } . Only answer in YTA or NTA."

Appendix

Table 12: Prompt templates used for binarized MMLU and AITA evaluations in the model diffing experiments (§5).

	SycophancyEval	ELEPHANT							Syconbench
	MMLU	OEQ			AITA			SS	Debate
Model	AnswerFlip	Val.	Ind.	Fra.	Val.	Ind.	Fra.	Fra.	ToF $\downarrow$	NoF
Llama-3.1-8B-Base	0.4778	0.0053	0.3505	0.2822	0.0693	0.2469	0.3430	0.3909	-	-
Llama-3.1-8B-Instruct	0.7206	0.3063	0.7139	0.3486	0.4998	0.7534	0.3575	0.3864	2.77	1.44
Tulu-3-8B-SFT	0.5971	0.3079	0.6961	0.3717	0.4328	0.6794	0.4330	0.4431	2.71	1.26
Tulu-3-8B-DPO	0.3793	0.2792	0.7678	0.3255	0.4638	0.9049	0.1680	0.4333	0.99	3.12

Appendix

Table 13: Sycophancy evaluation results for Llama-3.1-8B and Tulu-3-8B variants.

LLM Sycophancy

Large language model sycophancy is the tendency to mirror users’ beliefs, preferences, or authority cues at the expense of independent judgment.

See all 159 papers

Not Your Typical Sycophant: The Elusive Nature of Sycophancy in Large Language Models
Jan 21, 2026Shahar Ben-Natan, Oren TsurLLM Evaluation LLM Sycophancy
Computer and Information Science Ben Gurion University
Agentic Scaffolding Amplifies Sycophantic Behavior in Large Language Models
Jul 6, 2026Thantham JitthamLLM Sycophancy LLM Reliability
Cornell University
SyPS: Measuring Sycophancy Prompt Sensitivity in Large Language Models
Aug 24, 2026Lijia Huang, Yao Fu, Sihao RenLLM Sycophancy LLM Evaluation
Northeastern University · Case Western Reserve University · Everpure

Jan 21, 2026J/K move · Enter open · S save

Not Your Typical Sycophant: The Elusive Nature of Sycophancy in Large Language Models

Shahar Ben-Natan, Oren Tsur

Computer and Information Science Ben Gurion University

LLM Evaluation LLM Sycophancy Language Model Bias Evaluation

Open paper AskPDF arXiv

We propose a novel perspective for probing LLM sycophancy in a direct and neutral way, mitigating various forms of uncontrolled bias, noise, or manipulative language, deliberately injected to prompts in prior works. A key novelty of our approach is the use of an LLM-as-a-judge in a zero-sum betting game. Within this framework, sycophancy serves one individual (the user) while explicitly incurring cost on another. Comparing 11 leading models we find that while most models exhibit significant sycophantic tendencies in the common setting, in which sycophancy is self-serving to the user and incurs no cost on others, seven of the models exhibit ``moral remorse'', five of which significantly over-compensate for their sycophancy in case it explicitly harms a third party. We refer to this phenomenon as `anti-sycophancy' bias and discuss possible causes for this shift.

Abstract

LLM Sycophancy

Explore similar work

Not Your Typical Sycophant: The Elusive Nature of Sycophancy in Large Language Models

Agentic Scaffolding Amplifies Sycophantic Behavior in Large Language Models

SyPS: Measuring Sycophancy Prompt Sensitivity in Large Language Models