SAVOR: Self-Aware Visual Grounding via Confidence-Calibrated Reinforcement Learning for Multimodal Hallucination Mitigation

• URL: http://arxiv.org/abs/2609.16601v1
• Date: Tue, 15 Sep 2026 03:53:32 GMT
• Title: SAVOR: Self-Aware Visual Grounding via Confidence-Calibrated Reinforcement Learning for Multimodal Hallucination Mitigation
• Abstract summary: Multimodal large language models (MLLMs) produce fluent claims about objects, attributes, or relations that are not grounded in the image.<n>We introduce Savor, a training framework that augments the output schema with token and answer confidence.<n>We show that Savor reduces hallucination while preserving general capability on MME and MMBench, with lower Expected Error than DPO and decoding baselines.
• Score: 4.100479369058624
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Multimodal large language models (MLLMs) have made strong progress on visual question answering and image captioning, yet they still produce fluent claims about objects, attributes, or relations that are not grounded in the image. Many remedies either modify decoding at test time, which adds latency, or fine tune with preferences such as DPO variants, which teach which answer is preferred but not when the model's own answer is unreliable. We argue that calibrated self assessment is the missing signal. We introduce Savor, a training framework that (i) augments the output schema with token and answer confidence, (ii) optimises the policy with a Group Relative Policy Optimisation (GRPO) objective that penalises calibration error and poor abstention decisions, and (iii) uses the learned confidence at inference time to revisit visual evidence only when the model is uncertain. Experiments on POPE, HallusionBench, AMBER and MMHal-Bench across two recent backbones (InternVL3-8B and Qwen3-VL-8B) show that Savor reduces hallucination while preserving general capability on MME and MMBench, with lower Expected Calibration Error than DPO and decoding baselines.

Related papers

• Ask, Solve, Generate: Self-Evolving Unified Multimodal Understanding and Generation via Self-Consistency Rewards [52.42920996842378]
  Most unified large multimodal models (LMMs) that support both visual understanding and image generation rely on curated post-training supervision.<n>We propose a self-evolving training framework with three internal roles: a Proposer that generates visual questions, a solver that answers and evaluates them, and a Generator that synthesizes images.<n>For image generation, we design a multi-scale internal evaluation scheme that combines question-answer fidelity scoring with cycle-consistent captioning.
  arXiv  Detail & Related papers  (2026-06-25T17:59:57Z)
• Just how sure are you? Improving Verbalized Uncertainty Calibration in Medical VQA [0.16383644639245779]
  Multimodal large language models (MLLMs) applied to Medical Visual Question Answering (VQA) tend to produce overconfident outputs regardless of actual correctness.<n>This work proposes a training based framework that finetunes MLLMs to improve their calibration using a composite loss function.
  arXiv  Detail & Related papers  (2026-06-25T13:33:58Z)
• Grounded or Guessing? LVLM Confidence Estimation via Blind-Image Contrastive Ranking [7.845652284569666]
  We introduce BICR (Blind-Image Contrastive Ranking), a model-agnostic confidence estimation framework.<n>A lightweight probe is trained on the real-image hidden state and regularized by a ranking loss that penalizes higher confidence on the blacked-out view.<n> evaluated across five modern LVLMs and seven baselines on a benchmark covering visual question answering, object detection, medical imaging, and financial document understanding.
  arXiv  Detail & Related papers  (2026-05-11T17:35:10Z)
• Knowing When Not to Answer: Evaluating Abstention in Multimodal Reasoning Systems [52.83669998269706]
  Abstention has been studied in text-only settings but remains underexplored multimodally.<n>Current benchmarks either ignore unanswerability or rely on coarse methods that miss realistic failure modes.<n>We introduce MM-AQA, a benchmark that constructs unanswerable instances from answerable ones via transformations along two axes.
  arXiv  Detail & Related papers  (2026-04-16T09:23:22Z)
• Benchmarking Deflection and Hallucination in Large Vision-Language Models [25.176271096443482]
  Existing benchmarks overlook conflicts between visual and textual evidence.<n>We introduce VLM-DeflectionBench, a benchmark of 2,775 samples spanning diverse multimodal retrieval settings.<n>Our results highlight the need to evaluate not only what models know, but how they behave when they do not.
  arXiv  Detail & Related papers  (2026-04-13T20:22:22Z)
• Same Answer, Different Representations: Hidden instability in VLMs [65.36933543377346]
  We introduce a representation-aware and frequency-aware evaluation framework that measures internal embedding drift, spectral sensitivity, and structural smoothness.<n>We apply this framework to modern Vision Language Models (VLMs) across the SEEDBench, MMMU, and POPE datasets.
  arXiv  Detail & Related papers  (2026-02-06T12:24:26Z)
• Beyond Single Models: Mitigating Multimodal Hallucinations via Adaptive Token Ensemble Decoding [41.828387997311474]
  Large Vision-Language Models (LVLMs) have recently achieved impressive results in multimodal tasks such as image captioning and visual question answering.<n>They remain prone to object hallucination -- generating descriptions of nonexistent or misidentified objects.<n>We propose Adaptive Token Ensemble Decoding (ATED), a training-free, token-level ensemble framework that mitigates hallucination by aggregating predictions from multiple LVLMs during inference.
  arXiv  Detail & Related papers  (2025-10-21T06:11:24Z)
• Self-Consistency as a Free Lunch: Reducing Hallucinations in Vision-Language Models via Self-Reflection [71.8243083897721]
  Vision-language models often hallucinate details, generating non-existent objects or inaccurate attributes that compromise output reliability.<n>We present a novel framework that leverages the model's self-consistency between long responses and short answers to generate preference pairs for training.
  arXiv  Detail & Related papers  (2025-09-27T10:37:11Z)
• Measuring Epistemic Humility in Multimodal Large Language Models [17.490955813494693]
  We present HumbleBench, a new hallucination benchmark designed to evaluate MLLMs' ability to reject plausible but incorrect answers.<n>We leverage fine-grained scene graph annotations to extract ground-truth entities and relations, and prompt GPT-4-Turbo to generate multiple-choice questions.<n>HumbleBench fills a key gap in current evaluation suites, providing a more realistic measure of MLLM reliability in safety-critical settings.
  arXiv  Detail & Related papers  (2025-09-11T17:54:00Z)
• LACIE: Listener-Aware Finetuning for Confidence Calibration in Large Language Models [69.68379406317682]
  We introduce a listener-aware finetuning method (LACIE) to calibrate implicit and explicit confidence markers. We show that LACIE models the listener, considering not only whether an answer is right, but whether it will be accepted by a listener. We find that training with LACIE results in 47% fewer incorrect answers being accepted while maintaining the same level of acceptance for correct answers.
  arXiv  Detail & Related papers  (2024-05-31T17:16:38Z)
• Multi-Modal Hallucination Control by Visual Information Grounding [121.6983694815504]
  We show that Generative Vision-Language Models (VLMs) are prone to generate plausible-sounding textual answers that are not always grounded in the input image. We introduce Multi-Modal Mutual-Information Decoding (M3ID), a new sampling method for prompt amplification. M3ID amplifies the influence of the reference image over the language prior, hence favoring the generation of tokens with higher mutual information with the visual prompt.
  arXiv  Detail & Related papers  (2024-03-20T22:05:18Z)

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.