From Loops to Oops: Fallback Behaviors of Language Models Under Uncertainty

TL;DR

This work reframes undesirable LLM behaviors—repetitions, degenerate text, and hallucinations—as fallback strategies that arise under epistemic uncertainty. By controlling uncertainty across multiple model families and training regimes, it demonstrates a consistent ordering where more capable models shift from repetitions to degenerate text to hallucinations, a pattern that also manifests within single generations as generation length grows. The study shows that decoding methods alleviate some forms of degeneration but tend to increase hallucinations, suggesting that current fixes merely trade one fallback for another. These findings have practical implications for deployment and scalable oversight of LLMs, underscoring the need for strategies that address epistemic gaps rather than relying on decoding tweaks alone.

Abstract

Large language models (LLMs) often exhibit undesirable behaviors, such as hallucinations and sequence repetitions. We propose to view these behaviors as fallbacks that models exhibit under epistemic uncertainty, and investigate the connection between them. We categorize fallback behaviors - sequence repetitions, degenerate text, and hallucinations - and extensively analyze them in models from the same family that differ by the amount of pretraining tokens, parameter count, or the inclusion of instruction-following training. Our experiments reveal a clear and consistent ordering of fallback behaviors, across all these axes: the more advanced an LLM is (i.e., trained on more tokens, has more parameters, or instruction-tuned), its fallback behavior shifts from sequence repetitions, to degenerate text, and then to hallucinations. Moreover, the same ordering is observed during the generation of a single sequence, even for the best-performing models; as uncertainty increases, models shift from generating hallucinations to producing degenerate text and finally sequence repetitions. Lastly, we demonstrate that while common decoding techniques, such as random sampling, alleviate unwanted behaviors like sequence repetitions, they increase harder-to-detect hallucinations.

Figures (48)

• Figure 1: When language models face uncertainty, they exhibit fallback behaviors, shifting from hallucinations to degenerate text generation (repeating previous facts in different phrasing) and finally verbatim repetitions.
• Figure 2: Larger models resort to more complex fallback behaviors.Pythia Models with larger parameter counts produce more correct facts (green) and hallucinations (orange) while less repeating facts (blue). The green line indicates the number of ground truth answers in Triv.
• Figure 3: Larger models hallucinate instead of abstaining. When completing a list of facts about fictitious entities (FQamp), larger Pythia models hallucinate more, while smaller models repeat facts. Models never abstain from giving incorrect facts.
• Figure 4: Models that train longer shift to complex fallbacks. As Pythia-6.9B checkpoints see more training tokens (in billions), they produce more hallucinations and fewer repetitions. The green line shows correct answers upper bound in Triv.
• Figure 5: Larger models hallucinate more when generating biographies of rare entities. Larger Pythia models produce more atomic facts and more hallucinations on Bio.