Pushback in a Haystack: Language Models Respond to Factual Errors Ambiguously and Ineffectively

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.

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.

Humans naturally form and express beliefs in daily communication, e.g., "I think the answer is 3" or "I suppose that's right." Such beliefs inevitably intertwine with fact and knowledge, making the ability to handle them in tandem desirable for large language models (LLMs), as they are increasingly deployed in user-facing settings. Prior work showed that even capable LLMs exhibit a systemic weakness in acknowledging user beliefs grounded in incorrect information. We extend this evaluation to 10 LLMs across 18 epistemic expressions and find that the size and direction of this weakness depend on the verb used to express the belief, with the accuracy gap between factual and false information ranging from +50% on "I vaguely remember" to -14% on "I seriously doubt". We further show that the phenomenon stems from what we call task confusion: models default to fact-checking the underlying claim, overriding the user's stated belief. We provide evidence where chains of thought that explicitly fact-check show lower accuracy on false information than those that do not, and a single instruction can reverse the failure across verb families. Mechanistically, models attend more to false beliefs they fail to confirm, but suppressing this attention at decoding time recovers accuracy only partially and only in some models, calling for future work on intervention methods. Our findings clarify prior results and show how fact-checking, a generally desirable behavior, can interfere with belief tracking in LLMs.