Language models shift answers based on their prior responses
Written Once, Read Differently How Language Models Own Answers Steer Their Next Ones
Computation and LanguageArtificial Intelligence
Summary
When language models answer yes or no questions, their first answers can influence how they respond to later questions about the same topic. The authors found that these models keep a kind of 'commitment direction' in their processing that affects future answers. This effect varies with the model’s size, training, and instructions, sometimes even reversing depending on how the model is prompted. Understanding this helps explain how AI models internally link related questions even when their reading of these links changes.
What this means in practice
- •For ai developers: Improve AI response consistency by accounting for how models internally commit to previous answers when designing interaction flows.
- •For natural language application teams: Adjust prompts and tuning strategies knowing that answer commitments may read differently across model sizes and instruction sets.
Authors
Marcus Armstrong
Abstract
When a language model answers a yes/no question, that answer changes how it answers a different question about the same entity. We study this \emph{self-commitment} in 22 open-weight models from four families (0.5B--72B parameters), fixing the model's first answer and measuring the shift in its second across 3,600 question pairs about 120 entities. Each answer writes a \emph{commitment direction} into the residual stream. The similarity of two questions' directions predicts their coupling on held-out entities ($ρ= 0.43$--$0.85$), and ablating a single direction shared across topics leaves at most 17\% of the effect in every model. How this direction is read is far less stable: the reading circuit spreads across more attention heads as models grow, instruction tuning makes more questions read it in the opposite sense, and in one model changing only the system prompt reverses the circuit's sign. Ablating the direction while sparing the attention sink removes most of the coupling at under 1.3\% perplexity cost in 18 of 19 instruction-tuned models, though it shifts standalone answers in two. Without the sink spared, the direction's alignment with the leading principal component of ordinary text predicts which models pay a high cost. Written once, the commitment is read differently across scale, training and prompt.