Learning Dynamics of VLM Finetuning

• URL: http://arxiv.org/abs/2510.11978v1
• Date: Mon, 13 Oct 2025 22:22:49 GMT
• Title: Learning Dynamics of VLM Finetuning
• Authors: Jusheng Zhang, Kaitong Cai, Jing Yang, Keze Wang,
• Abstract summary: Preference-based finetuning of vision-language models (VLMs) is brittle.<n>We introduce textbfCooling-Weighted DPO (CW-DPO), a two-stage recipe that explicitly models and exploits the training trajectory.<n>CW-DPO yields textbfmore stable optimization, textbfbetter calibration, and textbfhigher pairwise win-rates than SFT-only and vanilla DPO.
• Score: 12.966077380225856
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: 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 $\Delta\!\log p$ 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.

Related papers

• Rethinking Preference Alignment for Diffusion Models with Classifier-Free Guidance [8.038055165320195]
  We propose a simple method that improves alignment without retraining the base model.<n>To further enhance generalization, we decouple preference learning into two modules trained on positive and negative data.<n>We evaluate on Stable Diffusion 1.5 and Stable Diffusion XL with Pick-a-Pic v2 and HPDv3, showing consistent quantitative and qualitative gains.
  arXiv  Detail & Related papers  (2026-02-21T11:18:52Z)
• Improving LLM-based Recommendation with Self-Hard Negatives from Intermediate Layers [80.55429742713623]
  ILRec is a novel preference fine-tuning framework for LLM-based recommender systems.<n>We introduce a lightweight collaborative filtering model to assign token-level rewards for negative signals.<n>Experiments on three datasets demonstrate ILRec's effectiveness in enhancing the performance of LLM-based recommender systems.
  arXiv  Detail & Related papers  (2026-02-19T14:37:43Z)
• Correct and Weight: A Simple Yet Effective Loss for Implicit Feedback Recommendation [36.820719132176315]
  This paper introduces a novel and principled loss function, named Corrected and Weighted (CW) loss.<n>CW loss systematically corrects for the impact of false negatives within the training objective.<n> experiments conducted on four large-scale, sparse benchmark datasets demonstrate the superiority of our proposed loss.
  arXiv  Detail & Related papers  (2026-01-07T15:20:27Z)
• MaP: A Unified Framework for Reliable Evaluation of Pre-training Dynamics [72.00014675808228]
  Instability in Large Language Models evaluation process obscures true learning dynamics.<n>We introduce textbfMaP, a framework that integrates underlineMerging underlineand the underlinePass@k metric.<n>Experiments show that MaP yields significantly smoother performance curves, reduces inter-run variance, and ensures more consistent rankings.
  arXiv  Detail & Related papers  (2025-10-10T11:40:27Z)
• Don't Waste Mistakes: Leveraging Negative RL-Groups via Confidence Reweighting [24.822152032771736]
  Reinforcement learning with verifiable rewards (RLVR) has become a standard recipe for improving large language models.<n>We show that negative groups can be leveraged without extra supervision.
  arXiv  Detail & Related papers  (2025-10-09T18:01:44Z)
• Margin Adaptive DPO: Leveraging Reward Model for Granular Control in Preference Optimization [0.0]
  Margin-Adaptive Direct Preference Optimization provides a stable, data-preserving, and instance-level solution.<n>We provide a comprehensive theoretical analysis, proving that MADPO has a well-behaved optimization landscape.<n>It achieves performance gains of up to +33.3% on High Quality data and +10.5% on Low Quality data over the next-best method.
  arXiv  Detail & Related papers  (2025-10-06T20:09:37Z)
• From Noisy Traces to Stable Gradients: Bias-Variance Optimized Preference Optimization for Aligning Large Reasoning Models [90.45197506653341]
  Large reasoning models generate intermediate reasoning traces before producing final answers.<n> aligning LRMs with human preferences, a crucial prerequisite for model deployment, remains underexplored.<n>A common workaround optimized a single sampled trajectory, which introduces substantial gradient variance from trace sampling.
  arXiv  Detail & Related papers  (2025-10-06T17:58:01Z)
• Reinforcement Learning with Verifiable yet Noisy Rewards under Imperfect Verifiers [90.50039419576807]
  Reinforcement Learning with Verifiable Rewards (RLVR) trains policies against automated verifiers to avoid costly human labeling.<n>To reduce vulnerability to verifier hacking, many RLVR systems collapse rewards to binary $0,1$ during training.<n>This choice carries a cost: it introduces textitfalse negatives (rejecting correct answers, FNs) and textitfalse positives (accepting incorrect ones, FPs)
  arXiv  Detail & Related papers  (2025-10-01T13:56:44Z)
• Winning the Pruning Gamble: A Unified Approach to Joint Sample and Token Pruning for Efficient Supervised Fine-Tuning [71.30276778807068]
  We propose a unified framework that strategically coordinates sample pruning and token pruning.<n>Q-Tuning achieves a +38% average improvement over the full-data SFT baseline using only 12.5% of the original training data.
  arXiv  Detail & Related papers  (2025-09-28T13:27:38Z)
• GrAInS: Gradient-based Attribution for Inference-Time Steering of LLMs and VLMs [56.93583799109029]
  GrAInS is an inference-time steering approach that operates across both language-only and vision-language models and tasks.<n>During inference, GrAInS hidden activations at transformer layers guided by token-level attribution signals, and normalizes activations to preserve representational scale.<n>It consistently outperforms both fine-tuning and existing steering baselines.
  arXiv  Detail & Related papers  (2025-07-24T02:34:13Z)
• SPPD: Self-training with Process Preference Learning Using Dynamic Value Margin [16.346540681903804]
  We propose textbfSelf-training framework integrating textbfProcess textbfPreference learning using textbfDynamic value margin (SPPD)<n>Experiments on 7B-scale models demonstrate superior performance across in-domain and out-domain mathematical benchmarks.
  arXiv  Detail & Related papers  (2025-02-19T08:11:26Z)
• Minusformer: Improving Time Series Forecasting by Progressively Learning Residuals [14.741951369068877]
  We find that ubiquitous time series (TS) forecasting models are prone to severe overfitting. We introduce a dual-stream and subtraction mechanism, which is a deep Boosting ensemble learning method. The proposed method outperform existing state-of-the-art methods, yielding an average performance improvement of 11.9% across various datasets.
  arXiv  Detail & Related papers  (2024-02-04T03:54:31Z)
• Prototypical Contrastive Learning through Alignment and Uniformity for Recommendation [6.790779112538357]
  We present underlinePrototypical contrastive learning through underlineAlignment and underlineUniformity for recommendation. Specifically, we first propose prototypes as a latent space to ensure consistency across different augmentations from the origin graph. The absence of explicit negatives means that directly optimizing the consistency loss between instance and prototype could easily result in dimensional collapse issues.
  arXiv  Detail & Related papers  (2024-02-03T08:19:26Z)
• Coupling Global Context and Local Contents for Weakly-Supervised Semantic Segmentation [54.419401869108846]
  We propose a single-stage WeaklySupervised Semantic (WSSS) model with only the image-level class label supervisions. A flexible context aggregation module is proposed to capture the global object context in different granular spaces. A semantically consistent feature fusion module is proposed in a bottom-up parameter-learnable fashion to aggregate the fine-grained local contents.
  arXiv  Detail & Related papers  (2023-04-18T15:29:23Z)
• Feature-Level Debiased Natural Language Understanding [86.8751772146264]
  Existing natural language understanding (NLU) models often rely on dataset biases to achieve high performance on specific datasets. We propose debiasing contrastive learning (DCT) to mitigate biased latent features and neglect the dynamic nature of bias. DCT outperforms state-of-the-art baselines on out-of-distribution datasets while maintaining in-distribution performance.
  arXiv  Detail & Related papers  (2022-12-11T06:16:14Z)
• Debiased Contrastive Learning of Unsupervised Sentence Representations [88.58117410398759]
  Contrastive learning is effective in improving pre-trained language models (PLM) to derive high-quality sentence representations. Previous works mostly adopt in-batch negatives or sample from training data at random. We present a new framework textbfDCLR to alleviate the influence of these improper negatives.
  arXiv  Detail & Related papers  (2022-05-02T05:07:43Z)
• Sequential Recommendation via Stochastic Self-Attention [68.52192964559829]
  Transformer-based approaches embed items as vectors and use dot-product self-attention to measure the relationship between items. We propose a novel textbfSTOchastic textbfSelf-textbfAttention(STOSA) to overcome these issues. We devise a novel Wasserstein Self-Attention module to characterize item-item position-wise relationships in sequences.
  arXiv  Detail & Related papers  (2022-01-16T12:38:45Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.