NoLan: Mitigating Object Hallucinations in Large Vision-Language Models via Dynamic Suppression of Language Priors

Lingfeng Ren; Weihao Yu; Runpeng Yu; Xinchao Wang

cs.CVFeb 25, 2026

NoLan: Mitigating Object Hallucinations in Large Vision-Language Models via Dynamic Suppression of Language Priors

Authors: Lingfeng Ren, Weihao Yu, Runpeng Yu, Xinchao Wang

Organizations: National University of Singapore, Singapore · Peking University Shenzhen Graduate School, China

Abstract

Object hallucination is a critical issue in Large Vision-Language Models (LVLMs), where outputs include objects that do not appear in the input image. A natural question arises from this phenomenon: Which component of the LVLM pipeline primarily contributes to object hallucinations? The vision encoder to perceive visual information, or the language decoder to generate text responses? In this work, we strive to answer this question through designing a systematic experiment to analyze the roles of the vision encoder and the language decoder in hallucination generation. Our observations reveal that object hallucinations are predominantly associated with the strong priors from the language decoder. Based on this finding, we propose a simple and training-free framework, No-Language-Hallucination Decoding, NoLan, which refines the output distribution by dynamically suppressing language priors, modulated based on the output distribution difference between multimodal and text-only inputs. Experimental results demonstrate that NoLan effectively reduces object hallucinations across various LVLMs on different tasks. For instance, NoLan achieves substantial improvements on POPE, enhancing the accuracy of LLaVA-1.5 7B and Qwen-VL 7B by up to 6.45 and 7.21, respectively. The code is publicly available at: https://github.com/lingfengren/NoLan.

Figures & tables

Figure 1: No-Language-Hallucination Decoding (NoLan). Given an LVLM, an image $v$ , and a language question $x$ , NoLan mitigates hallucinations in responses by comparing outputs generated from multimodal and unimodal (text-only) inputs. Step 2 can also be simplified by setting $\alpha$ to a fixed value of $1$ . In this example, the hallucinated object “ whale ” is suppressed by reducing the influence of language priors during token generation, while the ground truth object “ bear ” is effectively enhanced.

Figure 2: Experimental pipeline to test whether LLaVA’s vision encoder can detect the presence of an object in an image.

Samples on COCO of POPE- Random
where LLaVA experiences hallucinations
Metric	Accuracy	Precision	Recall	F1 Score
Score	83.01	83.71	98.33	90.43

Table 1: The Vision encoder can robustly detect object presence in samples. On the MSCOCO dataset of POPE- random ( 40 ) , for samples where LLaVA-1.5 experiences hallucinations, its vision encoder can indeed predict object presence with high accuracy.

Panel of Figure 3 — Figure 3: An illustration of model prediction misdirected by the language priors. Given an image depicting six dwarfs in front of Snow White, LLaVA-1.5-13b provides the same token “ seven ” regardless of whether the image is provided as input or not.

Dataset	$D_{\mathrm{KL}}(p_{m}\\|p_{u})$	$D_{\mathrm{KL}}(p_{u}\\|p_{m})$	$D_{\mathrm{JS}}(p_{m},p_{u})$
POPE no-hallucination	1.20	0.58	0.28
POPE hallucination	0.46	0.28	0.11

Table 2: Token probability distribution difference between multimodal and unimodal inputs. We split MSCOCO with POPE- random ( 40 ) into two subsets according to whether the answers from LLaVA-1.5-7B ( 50 ) contain hallucinations or not. Here,

p_{m}

and

p_{u}

represent the token probability distributions conditioned on multimodal and unimodal (text-only) inputs, respectively. The lower KL Divergence and JS Divergence values in the hallucination subset indicate greater similarity between the two distributions, suggesting that language priors heavily influence the outputs.

Dataset	Model	Decoding	Random		Popular		Adversarial
Dataset	Model	Decoding	Accuracy $\uparrow$	F1 Score $\uparrow$	Accuracy $\uparrow$	F1 Score $\uparrow$	Accuracy $\uparrow$	F1 Score $\uparrow$
GQA	LLaVA1.5	Regular	$83.73_{(\pm 0.27)}$	$82.95_{(\pm 0.28)}$	$78.17_{(\pm 0.17)}$	$78.37_{(\pm 0.18)}$	$75.08_{(\pm 0.33)}$	$76.06_{(\pm 0.24)}$
		VCD	${86.65}_{(\pm 0.45)}$	${86.99}_{(\pm 0.41)}$	${80.73}_{(\pm 0.47)}$	${82.24}_{(\pm 0.35)}$	${76.09}_{(\pm 0.43)}$	${78.78}_{(\pm 0.36)}$
		NoLan-Base (Ours)	${88.35}_{(\pm 0.16)}$	${87.68}_{(\pm 0.17)}$	${84.13}_{(\pm 0.30)}$	${83.94}_{(\pm 0.26)}$	${80.65}_{(\pm 0.19)}$	${81.08}_{(\pm 0.21)}$
		NoLan-Plus (Ours)	$\textbf{88.53}_{(\pm 0.10)}$	$\textbf{87.84}_{(\pm 0.12)}$	$\textbf{84.62}_{(\pm 0.33)}$	$\textbf{84.35}_{(\pm 0.21)}$	$\textbf{81.23}_{(\pm 0.17)}$	$\textbf{81.56}_{(\pm 0.19)}$
	Qwen-VL	Regular	$80.97_{(\pm 0.32)}$	$79.01_{(\pm 0.40)}$	$75.99_{(\pm 0.33)}$	$74.84_{(\pm 0.34)}$	$75.46_{(\pm 0.63)}$	$74.33_{(\pm 0.71)}$

Table 3: Results on POPE ( 40 ) . Regular decoding denotes direct sampling, VCD ( 33 ) indicates sampling from visual contrastive distribution, while methods prefixed with NoLan refers to sampling from our proposed contrastive distribution

p_{\mathrm{nolan}}

. The best performances within each setting are bolded. The mean and the standard deviation over 5 runs of POPE.

Table 4: Results of accuracy on MSCOCO of POPE.

Model	Decoding	Object-level		Attribute-level		Total Scores $\uparrow$
Model	Decoding	Existence $\uparrow$	Count $\uparrow$	Position $\uparrow$	Color $\uparrow$	Total Scores $\uparrow$
LLaVA1.5	Regular	175.67	124.67	114.00	151.00	565.33
	VCD	184.66	138.33	128.67	153.00	604.66
	VDD	190.00	143.30	145.00	165.00	643.29
	NoLan-Base (Ours)	190.00	145.00	138.33	155.00	628.33
	NoLan-Plus (Ours)	190.00	151.67	143.33	175.00	660.00

Table 5: Results on the hallucination subset of MME ( 19 ) . Regular decoding denotes direct sampling, VCD ( 33 ) indicates sampling from visual contrastive distribution, and VDD ( 89 ) expresses visual debias decoding. In contrast, methods prefixed with NoLan refer to sampling from our proposed contrastive distribution

p_{\mathrm{nolan}}

. The best performances within each setting are bolded .

Figure 4: Illustration of hallucination mitigation by our proposed NoLan-Plus with two samples from LLaVA-Bench. Hallucinated objects from LVLM’s regular decoding are highlighted in red.

Appendix figures & tables19 assets

Supplementary material from the paper’s appendix.

Appendix

Decoding	POPE $\downarrow$	MME $\downarrow$	MM-Vet $\downarrow$	LLaVA-Bench $\downarrow$
Regular	0.6484	2.9863	1.5889	4.1875
Text-only	0.6948	3.7051	2.0020	3.5508
VCD	0.4646	0.8632	0.6854	2.1153
NoLan-Base	0.4692	0.9106	0.6040	0.8188
NoLan-Plus	0.3786	0.7439	0.4423	0.7931

Appendix

Table 6: Entropy-based uncertainty evaluation across four benchmarks: POPE, MME, MM-Vet, and LLaVA-Bench. Results indicate that NoLan achieves the lowest uncertainty in all settings.

Token Pos	0	1	2	3	4	5	6	7	8	9	10	11	12
KL Value	1.36	0.53	0.97	0.50	0.57	0.76	0.63	0.66	0.47	0.46	0.36	0.42	0.42

Appendix

Table 7: Token-wise KL values indicating hallucination.

		POPE				MME					MM-Vet
$\boldsymbol{\alpha}$	Model	Accuracy $\uparrow$	Precision	Recall	F1 Score $\uparrow$	MME-Hallu $\uparrow$	Existence $\uparrow$	Count $\uparrow$	Position $\uparrow$	Color $\uparrow$	total $\uparrow$
0.0	LLaVA-1.5-7B	83.29	92.13	72.80	81.33	565.33	175.67	124.67	114.00	151.00	31.1
1.0	LLaVA-1.5-7B	86.50	96.68	75.60	84.85	628.33	190.00	145.00	138.33	155.00	33.0
2.0	LLaVA-1.5-7B	86.27	96.66	75.13	84.55	620.00	190.00	143.33	138.33	148.33	32.5
3.0	LLaVA-1.5-7B	85.97	96.63	74.53	84.16	588.33	180.00	138.33	121.67	148.33	31.8
0.0	LLaVA-1.5-13B	84.35	93.22	74.04	82.60	616.67	185.00	136.67	131.67	163.33	36.1

Appendix

Table 8: Sensitivity to modulation rate

\alpha

. In NoLan-Base,

\alpha

is manually set to regulate the influence of the modulation distribution, defined in Equation 3 . When

\alpha=0

, NoLan-Base reverts to standard decoding.

		POPE				MME					MM-Vet $\uparrow$
$\boldsymbol{\beta}$	Model	Accuracy $\uparrow$	Precision	Recall	F1 Score $\uparrow$	MME-Hallu $\uparrow$	Existence $\uparrow$	Count $\uparrow$	Position $\uparrow$	Color $\uparrow$	total $\uparrow$
0.0	LLaVA-1.5-7B	83.29	92.13	72.80	81.33	565.33	175.67	124.67	114.00	151.00	31.1
0.2	LLaVA-1.5-7B	86.37	96.66	75.33	84.68	588.33	180.00	138.33	121.67	148.33	30.5
0.4	LLaVA-1.5-7B	86.33	96.58	75.33	84.64	626.67	190.00	155.00	128.33	153.33	32.8
0.6	LLaVA-1.5-7B	86.60	96.76	75.73	84.97	645.00	190.00	143.33	143.33	168.33	32.5
0.8	LLaVA-1.5-7B	87.00	97.27	76.13	85.42	660.00	190.00	151.67	143.33	175.00	33.3

Appendix

Table 9: Sensitivity to modulation rate

\beta

. In NoLan-Plus,

\beta

is manually set to regulate the influence of the modulation distribution, defined in Equation 8 . When

\beta=0

, NoLan-Plus reverts to standard decoding.

Decoding		MSCOCO of POPE- random
logits		Accuracy $\uparrow$	Precision	Recall	F1 Score $\uparrow$
LLaVA-1.5-7B
$l_{m}+\alpha(l_{m}-l_{u})$	NoLan-Plus	87.00	97.27	76.13	85.42
$l_{m}+\alpha(l_{m}-l_{u})$	NoLan-Base	86.50	96.68	75.60	84.85
$l_{m}$	multimodal	83.29	92.13	72.80	81.33
$l_{u}$	text-only	47.57	47.45	45.33	46.37

Appendix

Table 10: Ablation studies for components of decoding logits and model sizes. For LLaVA-1.5 on MSCOCO of POPE- random , the performance drops significantly with only multimodal or text-only logits. Additionally, the performance gap between NoLan and other decoding methods increases as the model size grows.

Dataset	POPE	Model	Decoding	Accuracy $\uparrow$	Precision	Recall	F1 Score $\uparrow$
MSCOCO	Random	LLaVA1.5-7B	Regular	$83.29$	$92.13$	$72.80$	$81.33$
			NoLan-Base	${86.50}$	$96.68$	$75.60$	${84.85}$
			NoLan-Plus	${87.00}$	$97.27$	$76.13$	${85.42}$
		LLaVA1.5-13B	Regular	$83.31$	$91.46$	$73.48$	$81.49$
			NoLan-Base	87.37	95.61	78.33	86.11
			NoLan-Plus	88.70	$96.03$	$80.73$	87.72

Appendix

Table 11: Ablation study for LVLMs’ model sizes on MSCOCO of POPE. Scaling up LVLM model sizes does not significantly mitigate object hallucinations. In contrast, NoLan consistently enhances model performance.

Dataset	Model	Function	Random		Popular		Adversarial
Dataset	Model	Function	Accuracy $\uparrow$	F1 Score $\uparrow$	Accuracy $\uparrow$	F1 Score $\uparrow$	Accuracy $\uparrow$	F1 Score $\uparrow$	All $\uparrow$
GQA	LLaVA1.5	Sigmoid	88.53	87.83	84.53	84.13	81.47	81.54	84.67
	LLaVA1.5	Tanh	88.57	87.88	84.57	84.31	81.50	81.88	84.79
	Qwen-VL	Sigmoid	86.83	86.44	82.20	82.42	80.10	80.76	83.13
	Qwen-VL	Tanh	87.27	86.99	83.20	83.62	80.20	81.05	83.72
	InstructBLIP	Sigmoid	84.47	84.62	77.90	79.45	74.60	77.03	79.68

Appendix

Table 12: Results of NoLan-Plus’s variants on POPE ( 40 ) . Sigmoid refers to the use of the Sigmoid function as the processing term, while Tanh denotes the use of the Tanh function for the same purpose.The best performances within each setting are bolded.

Dataset	Decoding	Random		Popular		Adversarial
Dataset	Decoding	Accuracy $\uparrow$	F1 Score $\uparrow$	Accuracy $\uparrow$	F1 Score $\uparrow$	Accuracy $\uparrow$	F1 Score $\uparrow$
GQA	Regular	79.65	80.56	73.87	76.42	70.56	74.12
	VCD	83.69	84.16	78.57	80.17	75.08	77.53
	ICD	85.10	85.29	78.50	80.87	75.17	77.65
	NoLan-Base (Ours)	85.63	85.04	79.60	80.01	76.97	77.99
	NoLan-Plus (Ours)	86.13	85.23	81.13	80.92	78.10	78.41

Appendix

Table 13: Results of InstructBLIP on POPE ( 40 ) . Regular decoding denotes direct sampling, VCD ( 33 ) indicates sampling from visual contrastive distribution, ICD ( 77 ) expresses using Instruction Contrastive Decoding, while methods prefixed with NoLan refers to sampling from our proposed contrastive distribution

p_{\mathrm{nolan}}

. The best performances within each setting are bolded.

Model	Rec	OCR	Know	Gen	Spat	Math	Total
LLaVA1.5-7B ( 50 )	–	–	–	–	–	–	31.1
LLaVA1.5-7B NoLan-Base	36.2	26.5	21.3	23.3	33.5	7.7	33.0 $\pm$ 0.1
LLaVA1.5-7B NoLan-Plus	38.0	25.7	18.8	24.2	31.8	7.7	33.3 $\pm$ 0.2
LLaVA1.5-13B ( 50 )	–	–	–	–	–	–	36.1
LLaVA1.5-13B NoLan-Base	42.2	29.8	27.3	28.2	35.2	14.2	37.6 $\pm$ 0.2
LLaVA1.5-13B NoLan-Plus	41.8	31.4	24.9	26.0	36.6	15.4	38.3 $\pm$ 0.2

Appendix

Table 14: MM-Vet ( 86 ) evaluation results regarding each core VL capability. All the numbers are presented in % and the full score is 100%. Our NoLan can improve performance for different models.

Model	Rec Know Gen	Rec	OCR Spat	OCR Spat Math	Rec Spat	OCR	OCR Math	Rec Know	Rec OCR Know Gen	Rec OCR Gen Spat	Rec OCR Spat	Rec OCR	OCR Know Spat	Rec Know Spat	OCR Gen Spat	Rec OCR Spat Math	Total
LLaVA1.5-7B ( 50 )	–	–	–	–	–	–	–	–	–	–	–	–	–	–	–	–	31.1
LLaVA1.5-7B NoLan-Base	21.1	62.2	34.6	14.3	65.8	40.8	0.0	27.8	16.8	43.8	14.3	50.0	33.3	0.0	37.0	0.0	33.0 $\pm$ 0.1
LLaVA1.5-7B NoLan-Plus	22.3	68.9	29.8	14.3	66.7	40.8	0.0	5.6	52.2	12.5	14.3	75.0	16.7	0.0	20.0	0.0	33.3 $\pm$ 0.2
LLaVA1.5-13B ( 50 )	–	–	–	–	–	–	–	–	–	–	–	–	–	–	–	–	36.1
LLaVA1.5-13B NoLan-Base	27.5	70.0	28.5	26.4	58.3	47.3	0.0	26.7	12.5	54.2	17.1	87.5	50.0	50.0	12.0	0.0	37.6 $\pm$ 0.2
LLaVA1.5-13B NoLan-Plus	23.5	73.5	34.7	28.6	58.3	53.3	0.0	38.9	51.2	16.2	14.3	75.0	16.7	50.0	40.0	0.0	38.3 $\pm$ 0.2

Appendix

Table 15: MM-Vet ( 86 ) evaluation results regarding each capability integration. Our NoLan can improve model performance for different models.

Decoding	MMHalBench
Decoding	Overall Score $\uparrow$	Hallucination Rate $\downarrow$	Attribute	Adversarial	Comparison	Counting	Relation	Environment	Holistic	Other
Regular	1.55	76%	1.33	0	1.83	1.17	2.00	2.58	1.67	1.83
NoLan-Base	1.85	75%	3.42	1.58	1.42	1.42	1.75	3.33	0.83	1.08
NoLan-Plus	2.29	68%	3.42	3.25	1.33	1.58	1.58	3.83	1.33	2.00

Appendix

Table 16: Results on MMHalBench ( 65 ) for different decoding strategies. NoLan variants improve the overall score and decrease the hallucination rate.

Decoding	HallusionBench
Decoding	qAcc	fAcc	easyaAcc	hardaAcc	aAcc
Regular	14.2857	15.6069	37.1429	38.8372	43.4898
NoLan-Base	15.1648	17.9191	45.0549	35.5814	46.5899
NoLan-Plus	18.6813	19.6532	43.9560	40.6977	47.4756

Appendix

Table 17: Results on HallusionBench ( 22 ) for different decoding strategies. NoLan variants improve overall accuracy (aAcc) and category-specific metrics.

Decoding	MMBench
Decoding	Overall	AR	CP	FP-C	FP-S	LR	RR
Regular	63.4	77.6	70.0	57.7	68.0	33.2	56.2
NoLan-Base	64.6	76.0	77.1	56.7	66.3	33.0	53.6
NoLan-Plus	65.8	74.7	77.5	55.1	67.1	38.7	60.2

Appendix

Table 18: CircularEval results on MMBench ( 55 ) test set (L-2 abilities). NoLan variants improve overall and category-specific metrics.

Decoding	MathVision
Decoding	ALL	Alg	AnaG	Ari	CombG	Comb	Cnt	DescG	GrphT	Log	Angle	Area	Len	SolG	Stat	Topo	TransG
Random Chance	7.17	1.50	11.90	7.10	9.70	4.80	6.00	22.10	1.10	7.60	0.60	9.40	6.70	8.20	8.60	13.00	7.10
Regular	8.52	7.00	7.10	10.70	7.10	4.80	10.50	7.70	10.00	9.20	15.60	10.20	9.80	5.30	8.60	4.40	4.80
NoLan-Base	9.34	5.22	4.76	5.71	11.36	7.14	8.96	13.46	14.44	7.56	13.29	10.60	10.02	6.15	17.24	4.35	10.71
NoLan-Plus	9.84	6.96	8.33	7.14	11.04	6.55	5.97	17.31	17.78	7.56	13.29	8.60	9.80	9.43	13.79	13.04	13.10

Appendix

Table 19: Results on MathVision ( 74 ) for different decoding strategies. NoLan variants improve overall and most sub-categories, such as Algebra (Alg), Geometry (e.g., Angle, Area), and Logical reasoning (Log).

Decoding	MSCOCO of POPE- random
Decoding	Accuracy	F1 Score
Regular	83.29	81.33
OPERA ( 26 )	–	85.40
PAI ( 53 )	86.33	85.89
NoLan-Base (Ours)	87.80	85.60
NoLan-Plus (Ours)	88.80	86.70

Appendix

Table 20: Comparison between NoLan and attention-based methods

	MSCOCO of POPE- random
Decoding	Accuracy	Precision	Recall	F1 Score
Qwen2-VL-2B
Regular	71.27	76.76	61.00	67.98
NoLan-Base	73.93	85.83	57.33	68.75
NoLan-Plus	77.67	79.64	74.33	76.90
Qwen2-VL-7B

Appendix

Table 21: Results of Qwen2-VL ( 75 ) and Qwen2.5-VL ( 5 ) on POPE ( 40 ) .

Decoding	Seconds per Token $\downarrow$	Memory Usage (GB, 50 tokens) $\downarrow$
Regular	0.4579	13.57
VCD	0.7537	15.09
VDD	0.7359	15.09
NoLan-Base	0.6075	13.59
NoLan-Plus	0.6277	13.59

Appendix

Table 22: Inference efficiency comparison of contrastive decoding strategies.

Figure 5: More examples from LLaVA-Bench of our proposed NoLan-Plus for hallucination corrections. Hallucinated objects from LVLM’s regular decoding are highlighted in red.

Figure 6: More examples from LLaVA-Bench of our proposed NoLan-Plus for hallucination corrections. Hallucinated objects from LVLM’s regular decoding are highlighted in red.

Object Hallucination in VLMs

Object hallucination in vision-language models is the production of object claims unsupported by visual evidence.

See all 128 papers

Same Attention, Different Truths: Put Logit-Lens over Visual Attention to Detect and Mitigate LVLM Object Hallucination
Aug 7, 2026Zichuan Wang, Songlin Yang, Bo Peng +4Hallucination Detection in VLMs Vision-Language Models
School of Artificial Intelligence, University of Chinese Academy of Sciences · 3Hong Kong University of Science and Technology · 2New Laboratory of Pattern Recognition, Institute of Automation, Chinese Academy of Science
FADE: Mitigating Hallucinations by Reducing Language-Prior Dominance in Large Vision-Language Models
Jun 28, 2026Yichen Guo, Kai Tang, Fenglai Lin +5Vision-Language Models Large Vision-Language Models
1Nanyang Technological University · 2State Key Laboratory of Multimedia Information Processing, School of Computer Science, Peking University · 3Tsinghua University
Mitigating Object Hallucinations in Vision-Language Models through Region-Aware Attention Recalibration
May 24, 2026Yuanzhi Xu, Qian Gao, Jun Fan +4VLM Robustness Vision-Language Models
Qilu University of Technology (Shandong Academy of Sciences) Jinan, China · China Telecom Digital Intelligence Technology Co, Ltd Jinan, China · Shenyang Aerospace University Shenyang, China +1

Aug 7, 2026J/K move · Enter open · S save

Same Attention, Different Truths: Put Logit-Lens over Visual Attention to Detect and Mitigate LVLM Object Hallucination

Zichuan Wang, Songlin Yang, Bo Peng, Zhenchen Tang+3

School of Artificial Intelligence, University of Chinese Academy of Sciences · 3Hong Kong University of Science and Technology · 2New Laboratory of Pattern Recognition, Institute of Automation, Chinese Academy of Science

Hallucination Detection in VLMs Vision-Language Models VLM Hallucination Large Vision-Language Models Object Hallucination in VLMs VLM Hallucination Mitigation

Open paper AskPDF arXiv

Large Vision-Language Models (LVLMs) often suffer from object hallucination, generating objects that are absent from the image. Prior work largely attributes this to insufficient visual attention. However, we find that both real and hallucinated objects receive equally strong visual attention in the model's mid-to-late layers, suggesting that the key issue may not be how much the model attends, but what it attends to and why. To this end, we decode the visual features of high-attention regions using Logit Lens, and observe that regions corresponding to real objects can be correctly decoded to the target object tokens, whereas those for hallucinated objects cannot. Building on this, we identify two hallucination mechanisms: (i) visual uncertainty, triggered by semantically similar or confusable regions; masking these regions eliminates the hallucination. (ii) contextual prior, triggered by strong co-occurrence priors; even when the initially attended region is masked, the hallucination persists and attention drifts to other regions. Based on these findings, we propose a simple yet effective training-free Detect-Mitigate framework comprising a Logit-Lens Consistency Check to detect hallucination and targeted remedies: High-Attention Regions Masking (HARM) for visual uncertainty hallucination, and Visual Evidence Enhanced Decoding (VEED) for contextual prior hallucination. Our approach achieves state-of-the-art results on multiple hallucination benchmarks. Code will be available.

Abstract

Object Hallucination in VLMs

Explore similar work

Same Attention, Different Truths: Put Logit-Lens over Visual Attention to Detect and Mitigate LVLM Object Hallucination

FADE: Mitigating Hallucinations by Reducing Language-Prior Dominance in Large Vision-Language Models

Mitigating Object Hallucinations in Vision-Language Models through Region-Aware Attention Recalibration