An exploratory study demonstrates that large language models exhibit a functional asymmetry between language production and perception, despite using the same next-token prediction mechanism for both. By measuring token probabilities directly rather than relying on metalinguistic prompting, researchers found that prompt framing alone induces distinct probability distributions in decoder-only architectures.
- Using the base Llama-3.1-8B model, the study generated poems under production prompts and re-scored them under perception-oriented prompts.
- Production-perception distances consistently exceeded production-production distances, with an overall average ratio of approximately 1.8.
- The effect replicated across five open-weight models: Llama-3.1-8B, EuroLLM-9B, gemma-2-9b-it, Mistral-7B-Instruct-v0.3, and Qwen2.5-7B-Instruct.
- Temporal analysis showed the perception prompt's influence is strongest at the beginning of the sequence, with divergence decaying as generated context accumulates.
These findings suggest that communicative framing significantly impacts how LLMs process information, distinguishing between generating text and evaluating it.