Token-Level Inference-Time Alignment for Vision-Language Models

• URL: http://arxiv.org/abs/2510.21794v1
• Date: Mon, 20 Oct 2025 09:58:03 GMT
• Title: Token-Level Inference-Time Alignment for Vision-Language Models
• Authors: Kejia Chen, Jiawen Zhang, Jiacong Hu, Kewei Gao, Jian Lou, Zunlei Feng, Mingli Song,
• Abstract summary: Vision-Language Models (VLMs) have become essential backbones of modern multimodal intelligence.<n>We present TITA, a lightweight framework that freezes the base VLM and instead trains a reward model to approximate its distribution.<n>During inference, implicit preference signals are extracted as log-probability ratios between the reward model and the target VLM, yielding dense autoregressive feedback.
• Score: 58.41370989069588
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Vision-Language Models (VLMs) have become essential backbones of modern multimodal intelligence, yet their outputs remain prone to hallucination-plausible text misaligned with visual inputs. Existing alignment approaches often rely on expensive fine-tuning with annotated preference data or sequence-level inference strategies that provide only coarse, delayed feedback. To overcome these limitations, we present TITA (Token-level Inference-Time Alignment), a lightweight framework that freezes the base VLM and instead trains a reward model to approximate its distribution. During inference, implicit preference signals are extracted as log-probability ratios between the reward model and the target VLM, yielding dense autoregressive feedback. This formulation can be viewed as an inference-time variant of Direct Preference Optimization (DPO), providing token-level corrective signals without retraining the backbone. Extensive evaluations on LLaVA-1.5-7B and 13B show consistent gains across 12 benchmarks, with improvements of 8.6% on MMVet and 6.7% on POPE, indicating stronger general understanding and reduced hallucinations. Additional experiments on Qwen2.5-VL-7B and DeepSeek-VL2-27.5B show comparable gains, especially in hallucination reduction and VQA accuracy, while incurring negligible inference overhead.

Related papers

• More Than the Final Answer: Improving Visual Extraction and Logical Consistency in Vision-Language Models [74.10138874771852]
  We propose PeRL-VL (Perception and Reasoning Learning for Vision-Language Models), a decoupled framework that separately improves visual perception and textual reasoning on top of RLVR.<n>For perception, PeRL-VL introduces a VLM-based description reward that scores the model's self-generated image descriptions for faithfulness and sufficiency.<n>For reasoning, PeRL-VL adds a text-only Reasoning SFT stage on logic-rich chain-of-thought data, enhancing coherence and logical consistency independently of vision.
  arXiv  Detail & Related papers  (2025-12-13T23:06:18Z)
• Enhancing the Outcome Reward-based RL Training of MLLMs with Self-Consistency Sampling [90.87033586963828]
  Outcome-reward reinforcement learning (RL) is a common and increasingly significant way to refine the step-by-step reasoning of multimodal large language models (MLLMs)<n>We propose Self-Consistency Sampling (SCS) to correct this issue.<n>Based on Qwen2.5-VL-7B-Instruct, SCS improves accuracy by up to 7.7 percentage points on six multimodal benchmarks with negligible extra computation.
  arXiv  Detail & Related papers  (2025-11-13T18:59:57Z)
• Adaptive Residual-Update Steering for Low-Overhead Hallucination Mitigation in Large Vision Language Models [13.32858759983739]
  Large Vision-Language Models (LVLMs) often suffer from object hallucination, generating text inconsistent with visual inputs.<n>Existing inference-time interventions to mitigate this issue present a challenging trade-off.<n>We present Residual-Update Directed DEcoding Regulation (RUDDER), a framework that steers LVLMs towards visually-grounded generation.
  arXiv  Detail & Related papers  (2025-11-13T13:29:38Z)
• Weights-Rotated Preference Optimization for Large Language Models [30.25242193651982]
  We propose a novel Weights-Rotated Preference Optimization (RoPO) algorithm, which implicitly constrains the output layer logits with the KL divergence inherited from DPO.<n>Our RoPO achieves up to a 3.27-point improvement on AlpacaEval 2, and surpasses the best baseline by 6.2 to 7.5 points on MT-Bench with merely 0.015% of the trainable parameters.
  arXiv  Detail & Related papers  (2025-08-25T03:57:17Z)
• 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)
• VL-GenRM: Enhancing Vision-Language Verification via Vision Experts and Iterative Training [23.391643634478587]
  Vision-Language Reward Model (VL-RM) is key to aligning VL models by providing structured feedback.<n> bootstrapping dilemma arises as high-quality training data depends on already strong VL models.<n>We propose an iterative training framework leveraging vision experts, Chain-of-Thought rationales, and Margin-based Rejection Sampling.
  arXiv  Detail & Related papers  (2025-06-16T18:10:51Z)
• SoTA with Less: MCTS-Guided Sample Selection for Data-Efficient Visual Reasoning Self-Improvement [100.85923086072204]
  We introduce ThinkLite-VL, a family of visual reasoning models that achieve state-of-the-art (SoTA) performance using an order of magnitude fewer training samples.<n>We use Monte Carlo Tree Search (MCTS) to measure sample difficulty via the number of reasoning iterations a vision-language model (VLM) requires to solve each instance.<n>ThinkLite-VL-7B and ThinkLite-VL-72B significantly outperform their respective base models across eight visual reasoning benchmarks.
  arXiv  Detail & Related papers  (2025-04-10T17:49:05Z)
• VLFeedback: A Large-Scale AI Feedback Dataset for Large Vision-Language Models Alignment [55.7956150385255]
  We investigate the efficacy of AI feedback to scale supervision for aligning vision-language models. We introduce VLFeedback, the first large-scale vision-language feedback dataset. We train Silkie, an LVLM fine-tuned via direct preference optimization on VLFeedback.
  arXiv  Detail & Related papers  (2024-10-12T07:56:47Z)
• Self-Supervised Visual Preference Alignment [21.552415796397206]
  This paper makes the first attempt towards unsupervised preference alignment in Vision-Language Models (VLMs). We generate chosen and rejected responses with regard to the original and augmented image pairs, and conduct preference alignment with direct preference optimization. It is based on a core idea: properly designed augmentation to the image input will induce VLM to generate false but hard negative responses, which helps the model to learn from and produce more robust and powerful answers.
  arXiv  Detail & Related papers  (2024-04-16T12:19:54Z)
• Silkie: Preference Distillation for Large Visual Language Models [56.10697821410489]
  This paper explores preference distillation for large vision language models (LVLMs) We first build a vision-language feedback dataset utilizing AI annotation. We adopt GPT-4V to assess the generated outputs regarding helpfulness, visual faithfulness, and ethical considerations. The resulting model Silkie, achieves 6.9% and 9.5% relative improvement on the MME benchmark regarding the perception and cognition capabilities.
  arXiv  Detail & Related papers  (2023-12-17T09:44:27Z)

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.