Learning Dynamics of VLM Finetuning
This addresses the problem of unstable optimization in VLM alignment for researchers and practitioners, offering an incremental improvement over existing finetuning techniques.
The paper tackles the brittleness of preference-based finetuning in vision-language models by introducing Cooling-Weighted DPO, a two-stage method that stabilizes training and improves performance, achieving higher pairwise win-rates and faster convergence compared to baseline methods.
Preference-based finetuning of vision--language models (VLMs) is brittle: trivially wrong negatives inject uninformative gradients that destabilize training. We recast alignment as \textbf{learning-dynamics--aware optimization} and introduce \textbf{Cooling-Weighted DPO (CW-DPO)}, a two-stage recipe that explicitly models and exploits the training trajectory. \textbf{Stage 1} performs supervised finetuning with \textbf{gentle negatives}: \textbf{low-weight smoothed supervision} that regularizes the base policy and curbs overconfidence without explicit penalties. \textbf{Stage 2} applies a DPO objective in which the \textbf{negative term is scaled by a cooling weight} computed from the model's \textbf{average token log-probability} on each negative, suppressing uninformative gradients from easy or off-distribution samples while preserving signal from hard negatives. In practice, we emphasize \textbf{on-policy negatives} and allow \textbf{mixed negatives} by blending a controllable fraction of dataset negatives to maintain contrast freshness. Throughout, we instrument training with $Δ\!\log p$ probes on positives and negatives as first-class signals for early stopping, curriculum design, and failure diagnosis. Across diverse VLM tasks, CW-DPO yields \textbf{more stable optimization}, \textbf{better calibration}, and \textbf{higher pairwise win-rates} than SFT-only and vanilla DPO, while \textbf{converging in fewer steps}. Ablations isolate the \textbf{cooling-weight mechanism} as the primary driver of these gains and show complementary benefits from mixing on-policy and dataset negatives. Taken together, our results show that \textbf{smoothing learning dynamics before cooling preferences} is a simple, general principle for robust VLM alignment.