Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
The paper introduces CAROL, a probabilistic framework that reduces hallucinations in LLMs by optimizing for semantic consistency between generated responses and trusted contexts. CAROL formulates hallucination mitigation as a Markov chain accept-reject process over a lattice of textual sequences, guided by a string-submodular objective function. Experiments on QA and multi-agent reasoning tasks demonstrate that CAROL significantly reduces hallucinations and improves reliability compared to existing methods, with competitive efficiency.
LLMs can iteratively refine their outputs to reduce hallucinations by optimizing for semantic consistency with trusted contexts, achieving better reliability than likelihood-based or retrieval-augmented methods.
We introduce CAROL (Chain-based Adaptive Reconfiguration Over Lattices), a probabilistic framework for test-time hallucination reduction in large language models. Rather than relying on token-level uncertainty, CAROL defines a semantic uncertainty measure based on the consistency between generated responses and a trusted context, inducing a string-submodular objective over a lattice of textual sequences. This formulation enables hallucination mitigation to be cast as a Markov chain accept-reject process with provable convergence and near-optimality guarantees, allowing the model to iteratively refine outputs toward semantic consistency. By operating at the level of meaning, CAROL unifies hallucination detection and mitigation within a single framework. Empirical results on question answering and multi-agent reasoning benchmarks show that CAROL significantly reduces hallucinations and improves reliability and interpretability compared to likelihood-based and retrieval-augmented baselines, while maintaining competitive computational efficiency.