Knowledge Tree
exposure bias
Facts (21)
Sources
Hallucination Causes: Why Language Models Fabricate Facts mbrenndoerfer.com Mar 15, 2026 21 facts
claimExposure bias causes hallucinations because teacher forcing creates a training-inference mismatch where the model is never trained to handle its own errors, causing early mistakes in generation to cascade across subsequent tokens.
claimExposure bias in large language models creates compounding errors where a small factual inaccuracy or semantic drift at a specific position changes the conditioning context for subsequent positions, leading the model to generate statistically likely continuations based on erroneous premises.
procedureExposure bias accumulation occurs when a model generates an incorrect token, which then becomes part of the conditioning context for all subsequent tokens, causing the model to amplify details consistent with the error and suppress details associated with the correct information.
claimExposure bias is not unique to large language models; it arises in any sequence-to-sequence system trained with teacher forcing, including neural machine translation systems from the pre-transformer era.
claimExposure bias in large language models is the discrepancy between the distribution of conditioning contexts seen during training, which uses ground-truth tokens via teacher forcing, and the distribution seen during inference, which uses model-generated tokens.
claimBenchmarks for large language models that test only high-frequency factual questions fail to reveal tail entity hallucination, and benchmarks that test only short responses fail to reveal exposure bias accumulation.
claimThe mathematical consequence of exposure bias is that large language model probability estimates are well-calibrated for contexts matching the training distribution but poorly calibrated for out-of-distribution contexts that contain errors.
claimExposure bias is a cause of hallucination in large language models that arises from a mismatch between training efficiency and inference realism.
claimExposure bias in large language models does not require the model to lack the correct answer; rather, hallucinations arise because an error changes the input distribution, activating incorrect associations despite the model potentially possessing reliable knowledge.
claimRobust approaches to mitigating large language model hallucinations target multiple causes simultaneously, including using retrieval augmentation for knowledge gaps, better data curation for training data issues, scheduled sampling variants for exposure bias, and calibration training for generation pressure.
claimUncertainty calibration through Reinforcement Learning from Human Feedback (RLHF) addresses the surface expression of completion pressure in large language models but does not change the underlying lack of a world model or the exposure bias structure.
procedureScheduled sampling is an approach proposed to reduce exposure bias, which involves randomly replacing ground-truth conditioning tokens with model-generated ones during training at increasing rates.
claimThe max_new_tokens parameter controls sequence length in large language models, and longer generations face higher cumulative exposure bias divergence, which increases hallucination risk as the sequence grows.
claimInstruction tuning and reinforcement learning from human feedback (RLHF) improve a large language model's ability to express uncertainty and abstain from answering when knowledge is insufficient, but they do not retroactively fill knowledge gaps or undo exposure bias present in the base model.
claimRetrieval augmentation, which adds external knowledge at inference time, addresses knowledge gaps in large language models but leaves exposure bias and completion pressure untouched.
claimExposure bias in large language models causes errors to be self-reinforcing, where each subsequent token conditions on an initial incorrect context rather than the ground truth.
claimTraining data issues, exposure bias, knowledge gaps, and generation pressure are recognized phenomena contributing to large language model hallucinations, but quantifying their individual contributions to specific hallucinations is difficult.
claimThe four major categories of root causes for large language model hallucinations are training data issues, exposure bias during learning, structural knowledge gaps, and generation pressure at inference time.
claimLarge language model hallucinations are driven by the interaction of four causes: training data issues (noisy web data), knowledge gaps (questions about tail entities), completion pressure (generating confident-sounding responses), and exposure bias (early errors compounding in long-form answers).
claimLarge language models exhibit a 3% floor of irreducible hallucination even at high training frequencies, which is caused by exposure bias, completion pressure, and conflicting signals in training data.
claimFinetuning large language models modifies the model's response style regarding expressed confidence, but the underlying knowledge gaps and exposure bias patterns remain encoded in the base model from pretraining.