Toward a Gricean Retreat: Probing LLMs for Knowledge Boundaries and Referent Specificity
A study on arXiv (cs.AI) finds that LLMs fabricate plausible details for entities outside their knowledge boundary instead of retreating to safer, more general claims. Using a T-REx-based benchmark, the authors show that model activations encode whether a referent is inside the knowledge boundary and anticipate referent specificity, but these signals are not reconciled in generation. Models prefer specific referents even for unknown entities, even when correct generic alternatives are offered.
The paper 'Toward a Gricean Retreat: Probing LLMs for Knowledge Boundaries and Referent Specificity' investigates why LLMs hallucinate specific details about unknown entities. It frames the issue through Gricean cooperation: a cooperative speaker should retreat up the specificity hierarchy when uncertain. The authors probe LLMs and find that activations contain signals about knowledge-boundary membership and referent specificity, but generation does not use these signals. Models overwhelmingly choose specific referents over generic ones, even when the entity is unknown and correct generic alternatives are provided. The authors propose 'Gricean alignment' as a future direction to couple knowledge-boundary awareness with referent specificity.
The study demonstrates that LLM internal activations encode two distinct signals: (1) whether a referent falls inside the model's knowledge boundary, and (2) the specificity of the referent about to be generated. However, these signals are not integrated into the generation policy. This suggests a disconnect between the model's implicit knowledge-state representation and its decoding strategy. The T-REx-based benchmark varies entity familiarity and referent specificity, enabling controlled probing. The finding that models prefer specific referents even when correct generic alternatives are offered indicates a bias toward informativeness over truthfulness, violating Gricean maxims.
This research highlights a fundamental limitation in current LLM behavior: the inability to gracefully acknowledge ignorance. For industry applications, this leads to confident but false outputs, undermining trust in high-stakes domains. The proposed 'Gricean alignment' could become a new training or steering objective, potentially reducing hallucinations without sacrificing usefulness. Companies building LLM-based products may need to incorporate knowledge-boundary detection and dynamic specificity adjustment to improve reliability.
Reducing hallucinations through Gricean retreat could significantly enhance the reliability of LLMs in enterprise and consumer applications, reducing the need for external fact-checking and increasing user trust. This could lower operational risk and open up new use cases in regulated industries where accuracy is critical. The research also suggests a potential differentiator for model providers who can demonstrate better knowledge-boundary awareness.
Next observable signals include follow-up papers on Gricean alignment methods, such as fine-tuning or reinforcement learning from human feedback (RLHF) that rewards retreat behavior. We may see benchmarks that measure calibration of specificity to knowledge confidence. If successful, this line of work could lead to models that naturally say 'I don't know' or provide appropriately hedged answers, improving safety and user trust. Commercial adoption may follow if such methods are integrated into major LLM training pipelines.