HypoVeil: A Hypothesis-Driven Pragmatic Inference-Time Control Framework for Privacy–Utility-Aware LLM-Agent Dialogue

ICLR 2026 Conference Submission20511 Authors

Keywords: Test Time Inference, Hypothesis-Driven, Pragmatic decision, Rational Speech Acts

TL;DR: We introduce HypoVeil, an inference-time method that couples a hypothesis-driven belief store with a pragmatic RSA planner to optimize utility and minimize privacy cost. HypoVeil yields a significant improvement over our baselines on V-Bench.

Abstract: Large language model (LLM) agents are increasingly used as personal assistants with privileged data access, raising privacy concerns not just from training, but also from information disclosed during conversations at inference time. The key tradeoff is providing enough information to accomplish tasks while minimizing unintended disclosure; yet, prior evaluations show LLMs still struggle to consistently respect contextual privacy norms. We introduce HYPOVEIL, an inference time privacy method that combines a hypothesis-driven mental model with pragmatic decision-making. The agent maintains a dimension-aware belief store composed of concise natural language hypotheses about the counterpart’s knowledge, goals, and likely interpretations, then couples it with a Rational Speech Act (RSA) module that selects utterances by maximizing task utility minus privacy cost under the current hypothesis. To showcase the effectiveness of our method, we create and test on V-BENCH, a benchmark where two agents must interact in multi-turn privacy scenarios, structured as Party B strategically probing for information and Party A needing to collaborate without violating contextual privacy norms. Across GPT-4o, Llama-3.1-8B, and Gemma-3-27B, our method (Mental Model w/ RSA) significantly improves the privacy–utility trade-off, increasing the trade-off score by 5.2\% on average, reducing privacy risk by 6.4\%, and increasing helpfulness by 2.8\% over the baseline. These findings indicate that a hypothesis-driven mental model combined with pragmatic reasoning at inference time provides a practical path to privacy-preserving and context-aware LLM agents.

Primary Area: alignment, fairness, safety, privacy, and societal considerations

Submission Number: 20511