Evaluating Sycophancy in Chinese Large Language Models on Factual Questions Derived from Online Search Queries
Authors: Geng Liu, Feng Li, Mengxiao Zhu, Francesco Pierri
Organizations: Department of Electronics, Information and Bioengineering, Politecnico di Milano, Milan, Italy · University of Science and Technology of China, Hefei, China
As large language models increasingly mediate information access, factually accurate and independent answers are critical. However, these models can exhibit sycophancy by aligning their responses with users' stated beliefs even when those beliefs are incorrect, potentially presenting misinformation as independently verified and reinforcing users' confidence in false claims. Prior work leaves unresolved whether introducing user beliefs causes correct responses to become incorrect or uncertain, or causes uncertain responses to become belief-aligned incorrect answers. It also remains unclear whether anti-sycophancy interventions preserve or restore factual accuracy or merely shift responses toward uncertainty. We analyze factual sycophancy in Chinese-language information seeking using yes/no fact-checking questions. Our analysis covers 364,941 responses from three frontier Chinese-based LLMs (DeepSeek, Qwen, and Doubao) to 12,165 factual questions derived from real-world Chinese search queries. We evaluate the models with and without reasoning across baseline, belief-conditioned, and anti-sycophancy prompting, tracing matched shifts among correct, incorrect, and uncertain responses. Under incorrect user beliefs, we distinguish belief-aligned errors from losses of factual confidence, in which initially correct answers become uncertain. Patterns vary across models and reasoning settings: reasoning is not a consistent safeguard, and anti-sycophancy instructions can reduce incorrect agreement while increasing uncertainty. In Chinese-language factual question answering, avoiding agreement with false beliefs is therefore not equivalent to preserving factual accuracy, highlighting the value of transition-level evaluation. Such behavior may undermine the reliability of LLM-mediated information access by reinforcing misinformation or weakening users' confidence in factually correct answers.
Figures & tables
Figure 1: Responses to each factual yes/no question are generated under baseline, belief-only, and anti-sycophancy conditions and classified as correct ( C ), incorrect ( I ), or uncertain ( U ). Stage 1 compares matched baseline and belief-only responses. Stage 2 compares matched belief-only and anti-sycophancy responses while holding the stated user belief fixed.
Figure 2: Changes in the proportion of correct responses under incorrect user beliefs. Panel A reports the change from baseline to belief-only prompting, and Panel B reports the change from belief-only to anti-sycophancy prompting. Uncertain responses are included in the denominator when computing the accuracy. Error bars indicate 95% question-level bootstrap confidence intervals.
Figure 3: Response transitions from baseline to belief-only prompting under incorrect user beliefs. Panels A and C show transitions to belief-aligned incorrect answers ( C → I and U → I ), whereas Panel B shows the shift from correctness to uncertainty ( C → U ). Panels A and B share the same y-axis scale to facilitate comparison between transitions originating from correct baseline responses. Colors indicate whether reasoning is disabled or enabled. Error bars indicate 95% question-level bootstrap confidence intervals.
Figure 4: Direct Stage 2 response transitions after adding the anti-sycophancy instruction under incorrect user beliefs. The factual question and stated user belief are held fixed between the belief-only and anti-sycophancy conditions. Panels A and C show transitions to belief-aligned incorrect answers ( C → I and U → I ), whereas Panel B shows a shift from a correct response to uncertainty ( C → U ). Rates are conditional on the response state in the belief-only condition: correct for Panels A and B and uncertain for Panel C. Error bars indicate 95% question-level bootstrap confidence intervals.
Figure 5: Changes in the response pathways after anti-sycophancy prompting under incorrect user beliefs. For each pathway, rates are calculated among the same responses grouped by their baseline state. Values show the anti-sycophancy rate minus the belief-only rate. Panels A and C report changes in the frequency of belief-aligned incorrect answers, whereas Panel B reports changes in the frequency of baseline-correct responses becoming uncertain. Negative values indicate that the outcome became less frequent after anti-sycophancy prompting, and positive values indicate that it became more frequent. Error bars indicate 95% question-level bootstrap confidence intervals for the transition-rate differences.
Figure 6: Logistic regression estimates for preserving initially correct responses after user beliefs are introduced. The outcome is whether an initially correct baseline response remains correct in the belief-only condition. Points show log-odds coefficients and horizontal bars show 95% cluster-robust confidence intervals based on standard errors clustered at the factual question level. Topic coefficients are estimated relative to Health.
Appendix figures & tables29 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 7: Baseline prompt. Chinese template used in data collection (left) and faithful English translation (right).
Figure 8: Belief-only prompt: Yes-belief version. Yes indicates the polarity of the injected user belief, not its correctness.
Figure 9: Belief-only prompt: No-belief version. No indicates the polarity of the injected user belief, not its correctness.
Figure 10: Anti-sycophancy prompt: Yes-belief version. The green sentence is the polarity-specific anti-sycophancy instruction.
Figure 11: Anti-sycophancy prompt: No-belief version. The green sentence is the polarity-specific anti-sycophancy instruction.
Figure 12: Aggregate changes in incorrect response rates under incorrect user beliefs. Panel A reports the change from baseline to belief-only prompting, and Panel B reports the change from belief-only to anti- sycophancy prompting. Values are percentage-point changes in incorrect response rates.
Figure 13: Aggregate changes in uncertain response rates under incorrect user beliefs. Panel A reports the change from baseline to belief-only prompting, and Panel B reports the change from belief-only to anti-sycophancy prompting. Values are percentage-point changes in uncertain response rates.
Figure 14: Aggregate changes in accuracy under correct user beliefs. Panel A reports the change from baseline to belief-only prompting, and Panel B reports the change from belief-only to anti- sycophancy prompting. Values are percentage-point changes in accuracy, with uncertain responses retained in the denominator.
Figure 15: Aggregate changes in incorrect response rates under correct user beliefs. Panel A reports the change from baseline to belief-only prompting, and Panel B reports the change from belief-only to anti- sycophancy prompting. Values are percentage-point changes in incorrect response rates.
Figure 16: Aggregate changes in uncertain response rates under correct user beliefs. Panel A reports the change from baseline to belief-only prompting, and Panel B reports the change from belief-only to anti- sycophancy prompting. Values are percentage-point changes in uncertain response rates.
Figure 17: Ground-truth polarity: Stage 1 accuracy changes under incorrect user beliefs. Values show percentage-point changes from baseline to belief-only prompting, computed within each Ground-truth subset.
Figure 18: Ground-truth polarity: Stage 1 incorrect response-rate changes under incorrect user beliefs. Values show percentage-point changes from baseline to belief-only prompting, computed within each Ground-truth subset.
Figure 19: Ground-truth polarity: Stage 1 uncertain response-rate changes under incorrect user beliefs. Values show percentage-point changes from baseline to belief-only prompting, computed within each Ground-truth subset.
Figure 20: Ground-truth polarity: Stage 2 accuracy changes under incorrect user beliefs. Values show percentage-point changes from belief-only to anti-sycophancy prompting, computed within each Ground-truth subset.
Figure 21: Ground-truth polarity: Stage 2 incorrect response-rate changes under incorrect user beliefs. Values show percentage-point changes from belief-only to anti-sycophancy prompting, computed within each Ground-truth subset.
Figure 22: Ground-truth polarity: Stage 2 uncertain response-rate changes under incorrect user beliefs. Values show percentage-point changes from belief-only to anti-sycophancy prompting, computed within each Ground-truth subset.
Figure 23: Ground-truth polarity: Stage 1 accuracy changes under correct user beliefs. Values show percentage-point changes from baseline to belief-only prompting, computed within each Ground-truth subset.
Figure 24: Ground-truth polarity: Stage 1 incorrect response-rate changes under correct user beliefs. Values show percentage-point changes from baseline to belief-only prompting, computed within each Ground-truth subset.
Figure 25: Ground-truth polarity: Stage 1 uncertain response-rate changes under correct user beliefs. Values show percentage-point changes from baseline to belief-only prompting, computed within each Ground-truth subset.
Figure 26: Ground-truth polarity: Stage 2 accuracy changes under correct user beliefs. Values show percentage-point changes from belief-only to anti-sycophancy prompting, computed within each Ground-truth subset.
Figure 27: Ground-truth polarity: Stage 2 incorrect response-rate changes under correct user beliefs. Values show percentage-point changes from belief-only to anti-sycophancy prompting, computed within each Ground-truth subset.
Figure 29: Complete response-state transition matrices from baseline to belief-only prompting under incorrect user beliefs. The top and bottom rows show results with LLM reasoning disabled and enabled, respectively. Source rows represent response states in the baseline condition, and target columns represent response states in the belief-only condition. Each cell reports the transition rate from the corresponding source state to the corresponding target state. C, U, and I denote correct, uncertain, and incorrect responses, respectively.
Figure 30: Complete response-state transition matrices from baseline to belief-only prompting under correct user beliefs. The top and bottom rows show results with LLM reasoning disabled and enabled, respectively. Source rows represent response states in the baseline condition, and target columns represent response states in the belief-only condition. Each cell reports the transition rate from the corresponding source state to the corresponding target state. C, U, and I denote correct, uncertain, and incorrect responses, respectively.
Figure 31: Complete response-state transition matrices from belief-only to anti-sycophancy prompting under incorrect user beliefs. The top and bottom rows show results with LLM reasoning disabled and enabled, respectively. Source rows represent response states in the belief-only condition, and target columns represent response states in the anti-sycophancy condition. Each cell reports the transition rate from the corresponding source state to the corresponding target state. C, U, and I denote correct, uncertain, and incorrect responses, respectively.
Figure 32: Complete response-state transition matrices from belief-only to anti-sycophancy prompting under correct user beliefs. The top and bottom rows show results with LLM reasoning disabled and enabled, respectively. Source rows represent response states in the belief-only condition, and target columns represent response states in the anti-sycophancy condition. Each cell reports the transition rate from the corresponding source state to the corresponding target state. C, U, and I denote correct, uncertain, and incorrect responses, respectively.
Figure 33: Logistic regression estimates for preserving correct responses after anti-sycophancy instructions. The outcome is whether a belief-only response that is initially correct remains correct under the anti-sycophancy condition. Points show log-odds coefficients and horizontal bars show 95% cluster-robust confidence intervals. Topic coefficients are estimated relative to Health.
Figure 34: State-persistence robustness models. Panels are ordered row-wise: Stage 1 \textscU→\textscU , Stage 1 \textscI→\textscI , Stage 2 \textscU→\textscU , and Stage 2 \textscI→\textscI . Points show log-odds coefficients and horizontal bars show 95% cluster-robust confidence intervals. Topic coefficients are estimated relative to Health.
Stage 1 Transition Rate (%)
Model
Reasoning
C → I
C → U
U → I
Qwen
Disabled
5.1±1.4
1.6±0.8
8.3±2.9
Enabled
6.0±2.7
0.8±1.6
39.9±6.8
DeepSeek
Disabled
6.6±3.6
8.2±4.2
19.2±5.1
Enabled
4.9±3.5
11.6±3.8
21.2±7.7
Doubao
Disabled
5.0±4.1
5.3±2.9
8.7±1.5
Appendix
Table 3: Stage 1 repeated-generation robustness check across 10 independent generation passes ( N=10 ). Values report the mean transition rate ± standard deviation across passes from baseline to belief-only prompting under incorrect user beliefs.
Stage 2 Transition Rate (%)
Model
Reasoning
C → I
C → U
U → I
Qwen
Disabled
7.5±1.5
3.6±1.2
12.0±0.0
Enabled
7.7±3.1
1.2±2.3
43.0±9.5
DeepSeek
Disabled
5.2±2.7
19.4±4.3
16.6±4.4
Enabled
5.2±2.7
15.9±3.2
10.5±6.7
Doubao
Disabled
4.4±3.5
26.4±4.7
0.0±0.0
Appendix
Table 4: Stage 2 repeated-generation robustness check across 10 independent generation passes ( N=10 ). Values report the mean transition rate ± standard deviation across passes from belief-only to anti-sycophancy prompting under incorrect user beliefs.
Sycophancy in Large Language Models, the tendency to align responses with user preferences over ground truth, has largely been studied using questions without a clear ground truth and with a single, plainly stated user opinion. We present a systematic study of which framings of a user's claim trigger sycophancy in factual question answering, where every answer is objectively verifiable without extra context. Evaluating 14 models on 42K questions, we express correct and incorrect user opinions through various claim framings over 8 dimensions and 21 types, and measure a broad set of opinion-driven behaviors. We find that all models are susceptible to user opinions regardless of size. Models are most influenced by how confidently an opinion is stated, whom it is attributed to, and whether it is presented as shared between the user and the model. In contrast to prior work, framings based on the user's expertise, authority over the model, or social pressure matter little in our factual setting. Larger models resist incorrect opinions better but are no better at recovering from their own errors. Our results call for sycophancy evaluations that probe how an opinion is framed, not merely whether one is present, and that measure multiple behaviors beyond a single flip rate.
When a language model agrees with a user's false belief, is it failing to detect the error, or noticing and agreeing anyway? We show the latter. Across twelve open-weight models from five labs, spanning small to frontier scale, the same small set of attention heads carries a "this statement is wrong" signal, whether the model is evaluating a claim on its own or being pressured to agree with a user. Silencing these heads flips sycophantic behavior sharply while leaving factual accuracy intact, so the circuit controls deference rather than knowledge. Edge-level path patching confirms that the same head-to-head connections drive sycophancy, factual lying, and instructed lying. Opinion-agreement, where no factual ground truth exists, reuses these head positions but writes into an orthogonal direction, ruling out a simple "truth-direction" reading of the substrate. Alignment training leaves this circuit in place: an RLHF refresh cuts sycophantic behavior roughly tenfold while the shared heads persist or grow, a pattern that replicates on an independent model family and under targeted anti-sycophancy DPO. When these models sycophant, they register that the user is wrong and agree anyway.
Large language models can answer a medical question correctly and still abandon that answer when a user pushes back. We study this failure as medical sycophancy and ask when models are most likely to give in. Across five open-weight models, 500 MedQuAD questions, and 1.2 million trials, we use a fully crossed design over four conversational factors: user role, user evidence, interaction structure, and grounding. Medical sycophancy is nearly three times more common when users challenge an answer the model has already given than when the false claim appears in the initial query. Models are also more susceptible to users presented as physicians or medical students. Most strikingly, fabricated evidence has opposite effects across interaction structures. It increases sycophancy in single-turn interactions but reduces it after the model has already answered. Grounding helps, but does not eliminate the behavior. Sycophancy varies more across medical questions than across models, making question selection an important part of benchmark design. Reasoning traces suggest that multi-turn failures are associated with models turning back toward their own prior answer, while fabricated evidence receives more scrutiny after an initial response. Together, the results show that medical sycophancy depends as much on how a model is challenged and evaluated as on which model is tested.
Kaike Ping, Buse Çarık, Caleb Wohn +3
Virginia Tech · Shanghai Tongren Hospital, Shanghai Jiao Tong University School of Medicine · Emory University