DuPO: Enabling Reliable LLM Self-Verification via Dual Preference Optimization
This addresses the need for scalable and annotation-free optimization methods for LLMs, offering a general paradigm that could benefit various AI applications, though it builds on existing dual learning concepts with incremental innovations.
The paper tackles the problem of enabling reliable self-verification in large language models (LLMs) without costly annotations by introducing DuPO, a dual preference optimization framework that uses self-supervised rewards from reconstruction tasks, resulting in gains such as a 2.13 COMET improvement in translation quality and a 6.4-point accuracy boost in mathematical reasoning.
We present DuPO, a dual learning-based preference optimization framework that generates annotation-free feedback via a generalized duality. DuPO addresses two key limitations: Reinforcement Learning with Verifiable Rewards (RLVR)'s reliance on costly labels and applicability restricted to verifiable tasks, and traditional dual learning's restriction to strictly dual task pairs (e.g., translation and back-translation). Specifically, DuPO decomposes a primal task's input into known and unknown components, then constructs its dual task to reconstruct the unknown part using the primal output and known information (e.g., reversing math solutions to recover hidden variables), broadening applicability to non-invertible tasks. The quality of this reconstruction serves as a self-supervised reward to optimize the primal task, synergizing with LLMs' ability to instantiate both tasks via a single model. Empirically, DuPO achieves substantial gains across diverse tasks: it enhances the average translation quality by 2.13 COMET over 756 directions, boosts the mathematical reasoning accuracy by an average of 6.4 points on three challenge benchmarks, and enhances performance by 9.3 points as an inference-time reranker (trading computation for accuracy). These results position DuPO as a scalable, general, and annotation-free paradigm for LLM optimization.