NVIDIA Spatial Intelligence Lab (SIL) NVIDIA Research

Mitigating Multimodal Hallucinations via Gradient-based Self-Reflection

Shan Wang13†,  Maying Shen2,  Nadine Chang2,  Chuong Nguyen3,  Hongdong Li1,  Jose M. Alvarez2
1Australian National University
2NVIDIA
3CSIRO
†Work completed during an internship at NVIDIA.
CVPR 2026
history_edu arXiv description Paper format_quote BibTeX

What are hallucinations? Outputs that are not grounded in the provided visual inputs are typically caused by:

Text–Visual Bias: Overreliance on text inputs (prompts or past outputs) over visual information, which becomes more severe in longer generations.

Co-occurrence Bias: Models hallucinate objects because they often appear together in the training data.

Teaser showing example of hallucination and our proposed method.

We propose a training-free inference-based method that identifies model biases using gradient signals. In this example, the base MLLM hallucinates “beer” even though the glasses are empty. The prediction is overly influenced by the prompt (gray bar), previously generated tokens (white bar), and visual tokens associated with previously mentioned objects (green and magenta bars), while relying less on other visual evidence (blue bars). Our method increases the influence of the blue part and reduces the influence of text inputs thereby mitigating text–visual hallucination. Similarly, by suppressing visual tokens associated with previously mentioned objects, it also reduces co-occurrence hallucination.

Abstract

Multimodal large language models (MLLMs) achieve strong performance across diverse tasks but remain prone to hallucinations, where outputs are not grounded in visual inputs. This issue can be attributed to two main biases: text–visual bias, the overreliance on prompts and prior outputs, and co-occurrence bias, spurious correlations between frequently paired objects. We propose Gradient-based Influence-Aware Constrained Decoding (GACD), an inference-based method that addresses both biases without auxiliary models and is readily applicable to existing models without finetuning. The core of our approach is bias estimation, which uses first-order Taylor gradients to understand the contribution of individual tokens—visual features and text tokens—to the current output. Based on this analysis, GACD mitigates hallucinations through two components: (1) suppressing spurious visual features correlated with the output objects, and (2) rebalancing cross-modal contributions by strengthening visual features relative to text. Experiments across multiple benchmarks demonstrate that GACD effectively reduces hallucinations and improves the visual grounding of MLLM outputs.

Method Overview

Token Influence

We propose using a first-order Taylor expansion to estimate token influence, defined as the sum of the Manhattan norms of gradients from different contexts (visual inputs, prompts, and past outputs), formulated as:

$$ \begin{aligned} \mathbf{z}^{\star}_m \;\approx\; \underbrace{\sum^S_{s=1} \mathbf{g}^v_{ms}}_{\text{visual influence}} {t}^v_s \;+\; \underbrace{ \underbrace{\sum^N_{n=1} \mathbf{g}^p_{mn}}_{\text{prompt influence}} {t}^p_n \;+\; \underbrace{\sum^{m-1}_{i=1} \mathbf{g}^{y}_{mi}}_{\text{past output influence}} }_{\text{text influence}} y_i \;+\; \text{Const}, \end{aligned} $$

where:

  • \(\mathbf{z}^{\star}_m\): output logits
  • \(t^{v}_{s}\): input visual tokens.
  • \(t^{p}_{n}\): input language prompts.
  • \(y_{i}\): past outputs.
  • \(\mathbf{g}^{v}_{ms}\): token-wise Jacobians w.r.t to input visual tokens.
  • \(\mathbf{g}^{p}_{mn}\): token-wise Jacobians w.r.t to input prompt tokens.
  • \(\mathbf{g}^{y}_{mi}\): token-wise Jacobians w.r.t to past output tokens.
  • visual influence: \(\sum_{s=1}^{S}\|\mathbf{g}^{v}_{ms} \|_1\).
  • prompt influence: \(\sum_{n=1}^{N}\|\mathbf{g}^{p}_{mn} \|_1\).
  • past output influence: \(\sum_{i=1}^{m-1}\|\mathbf{g}^{y}_{mi} \|_1\).
  • text influence: prompt influence + past output influence

GACD

GACD pipeline

GACD consists of two core components:

  • Object-aware Visual Token Grouping: visual tokens are partitioned into previously generated object-related to and unrelated tu.
  • Influence-Weighted Decoding: To suppress biases from the prompt, prior outputs, and previously mentioned objects.
    • We first isolate their influence into a negative guidance logit, zom.
    • We then obtain the final prediction, ẑm, by subtracting this biased guidance from the original prediction z*m.
    • During this process, we dynamically scale the weight of this subtraction to explicitly force the influence of unrelated visual tokens (tu) to match the influence of the dominant text tokens.

Results

Quantitative Results

Strong and Consistent Improvements: GACD improves performance across all evaluated MLLMs and achieves the best AMBER Score compared with existing SOTA methods.

Method LLaVA v1.5 InstructBLIP mPLUG-Owl2 LLaVA v1.6 Qwen2-VL InternVL2
Base 83.5 86.5 84.0 88.5 90.1 90.0
RLAIFv 89.0 87.1 88.6 88.6 90.4 90.1
VCD 83.8 86.7 86.6 88.0 90.0 89.6
M3ID 84.7 86.9 86.8 88.4 90.2 89.6
AVISC 85.5 87.2 87.5 88.3 90.1 89.4
Ours 90.2 88.1 89.6 89.2 91.1 90.3
POPE ratio comparison

Adaptive Visual Enhancement: GACD strengthens visual influence only when the model relies too little on visual information (below 50%). This avoids unnecessary changes when visual grounding is already strong, such as in InternVL2.

AMBER score improvements

Category Gains: GACD consistently improves F1 scores across categories, with the largest gains on direct visual errors.

Text-Visual Hallucination Qualitative Results

Text-Visual hallucination.

(Left) Without GACD: The model relies too little on visual information, with visual influence ratio averaging only about 20%. Hallucinations often appear when visual influence is low.

(Right) With GACD: GACD increases overall visual influence and prevents it from dropping as generation continues, leading to fewer hallucinations.

Analysis Insight: Our token influence helps analyze MLLMs. It naturally shows that punctuation and suffix tokens depend less on visual input, and also reveals how visual reliance decreases as the output becomes longer.

Co-occurrence Hallucination Qualitative Results

Co-occurrence hallucination.

(Left) Within the visual tokens, the influence distributions for the prediction Chair and the hallucinated Table are highly correlated; notably, the visual token contributing most to Table also contributes most to Chair.

(Right) Shared most influential tokens are common in hallucinated Table predictions, while GACD reduces this issue and lowers hallucination rates.

BibTeX

@inproceedings{wang2026mitigating,
  title={Mitigating Multimodal Hallucinations via Gradient-based Self-Reflection},
  author={Wang, Shan and Shen, Maying and Chang, Nadine and Nguyen, Chuong and Li, Hongdong and Alvarez, Jose M},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year={2026}
}