Research map / Continual learning / Online and replay-based
$\phi$-DPO: Fairness Direct Preference Optimization Approach to Continual Learning in Large Multimodal Models
Thanh-Dat Truong, Huu-Thien Tran, Jackson Cothren, Bhiksha Raj, Khoa Luu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use DPO preference loss instead of knowledge distillation (ablation) | ϕ-DPO (LLaVA-7B) · MLLM-CL Domain | MFN: 69.21 -> 71.74 | Improved |
| use Fair DPO (γ=2) instead of vanilla DPO (ablation) | ϕ-DPO (LLaVA-7B) · MLLM-CL Domain | MFN: 71.74 -> 74.00 | Improved |
| use Fair DPO instead of vanilla DPO (ablation) | ϕ-DPO (LLaVA-7B) · MLLM-CL Domain | BWT: -1.33 -> -0.37 | Improved |
| lower β from 0.10 to 0.01 (ablation) | ϕ-DPO · MLLM-CL Domain | MFT / BWT: MFT 74.29 -> 75.43; BWT -0.37 -> -2.66 | Mixed |
| raise β from 0.10 to 0.50 (ablation) | ϕ-DPO · MLLM-CL Domain | MFN: 74.00 -> 71.85 | Hurt |
| increase γ from 0 to 2.0 (ablation) | ϕ-DPO · MLLM-CL Domain | MFN: 71.74 -> 74.00 | Improved |
| raise γ from 2.0 to 5.0 (ablation) | ϕ-DPO · MLLM-CL Domain | MFN: 74.00 -> 72.18 | Hurt |
| use ϕ-DPO instead of DPO on LLaVA-13B (ablation) | ϕ-DPO · MLLM-CL Domain | MFN: 74.93 -> 75.81 | Improved |
| use ϕ-DPO instead of DPO on InternVL-7B (ablation) | ϕ-DPO · MLLM-CL Domain | MFN: 72.37 -> 74.23 | Improved |
| scale backbone from LLaVA-7B to LLaVA-13B (ablation) | ϕ-DPO · MLLM-CL Domain | BWT: MFN 74.00 -> 75.81; BWT -0.37 -> -0.59 | Mixed |
| ϕ-DPO vs best prior method MR-LoRA | ϕ-DPO · MLLM-CL Domain | MFN: 69.63 -> 74.00 | Improved |
| ϕ-DPO vs HiDe on CoIN | ϕ-DPO · CoIN | MFN: 62.82 -> 68.86 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- SEFE* — SEFE: Superficial and Essential Forgetting Eliminator for Multimodal Continual Instruction Tuning (ICML 2025) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Doubly Perturbed Task Free Continual Learning (AAAI 2024 · similarity 0.86)
- Overcoming Catastrophic Forgetting in Visual Continual Learning with Reinforcement Fine-Tuning (NeurIPS 2026 · similarity 0.86)
- Pareto Continual Learning: Preference-Conditioned Learning and Adaption for Dynamic Stability-Plasticity Trade-off (AAAI 2025 · similarity 0.85)
- D-Fusion: Direct Preference Optimization for Aligning Diffusion Models with Visually Consistent Samples (ICML 2025 · similarity 0.85)
- Earlier Tokens Contribute More: Learning Direct Preference Optimization From Temporal Decay Perspective (ICLR 2025 · similarity 0.85)
- MaPPO: Maximum a Posteriori Preference Optimization with Prior Knowledge (NeurIPS 2026 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
