MGDA-Decoupled: Geometry-Aware Multi-Objective Optimisation for DPO-based LLM Alignment
This addresses procedural unfairness in LLM alignment for AI developers by providing a more equitable trade-off method, though it is incremental as it builds on existing DPO frameworks.
The paper tackles the problem of aligning large language models to multiple conflicting objectives like helpfulness and truthfulness by introducing MGDA-Decoupled, a geometry-aware multi-objective optimization algorithm that operates within the DPO paradigm, achieving the highest win rates against golden responses on the UltraFeedback dataset.
Aligning large language models (LLMs) to desirable human values requires balancing multiple, potentially conflicting objectives such as helpfulness, truthfulness, and harmlessness, which presents a multi-objective optimisation challenge. Most alignment pipelines rely on a fixed scalarisation of these objectives, which can introduce procedural unfairness by systematically under-weighting harder-to-optimise or minority objectives. To promote more equitable trade-offs, we introduce MGDA-Decoupled, a geometry-based multi-objective optimisation algorithm that finds a shared descent direction while explicitly accounting for each objective's convergence dynamics. In contrast to prior methods that depend on reinforcement learning (e.g., GAPO) or explicit reward models (e.g., MODPO), our approach operates entirely within the lightweight Direct Preference Optimisation (DPO) paradigm. Experiments on the UltraFeedback dataset show that geometry-aware methods -- and MGDA-Decoupled in particular -- achieve the highest win rates against golden responses, both overall and per objective.