BALTO: Balanced Token-Level Policy Optimization for Hallucination Mitigation

A new framework called BALTO aims to reduce hallucinations in large language models by optimizing token-level policy, according to a paper submitted June 14, 2026. The approach addresses a granularity mismatch that has limited the effectiveness of response-level reinforcement learning rewards. [1] Hallucinations — generated content not grounded in provided evidence — remain a significant barrier to deploying large language models in knowledge-intensive tasks. While reinforcement learning has shown promise for mitigation, response-level faithfulness rewards often penalize supported content when localized hallucinations occur. [1] The BALTO framework, short for Balanced Token-level Policy Optimization, extracts checkable factual claims from model outputs and verifies them against a reference context. It then projects claim-level judgments down to token-level labels, introducing a balanced credit assignment mechanism that redistributes probability mass from unsupported content toward faithful content rather than suppressing the entire response. [1] The paper includes a theoretical analysis of response-level reward limitations and demonstrates BALTO's advantages in training stability and optimization efficiency. [1] Experiments were conducted across three benchmarks: ConFiQA, RAGTruth, and FinLLM-Eval. Across all six model–benchmark settings, BALTO achieved the highest faithfulness scores and consistently outperformed existing post-training baselines in Q-Score, indicating a stronger trade-off between faithfulness and informativeness. [1] The work was submitted to arXiv on June 14, 2026. [1]