arxiv: 2604.07914 · v1 · submitted 2026-04-09 · 💻 cs.CV · cs.AI

Recognition: no theorem link

Mitigating Entangled Steering in Large Vision-Language Models for Hallucination Reduction

Yuanhong Zhang , Zhaoyang Wang , Xin Zhang , Weizhan Zhang , Joey Tianyi Zhou

Authors on Pith no claims yet

Pith reviewed 2026-05-10 18:12 UTC · model grok-4.3

classification 💻 cs.CV cs.AI

keywords hallucination mitigationvision-language modelslatent space steeringentangled signalsplug-and-play frameworkmultimodal generation

0 comments

The pith

MESA reduces hallucinations in vision-language models by selective latent intervention that leaves token distributions unchanged.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper claims that hallucinations arise because steering signals meant to suppress inconsistent outputs become entangled with the model's normal generation processes, so existing fixes shorten responses or shift probabilities. MESA solves this by performing targeted latent-space adjustments that act only on hallucination-relevant parts of the response while leaving the rest of the token distribution intact. A sympathetic reader would care because this separation promises hallucination reduction that can be added to existing models without forcing users to accept truncated, altered, or less natural outputs.

Core claim

MESA is a plug-and-play framework that performs controlled and selective latent intervention for hallucination mitigation, targeting hallucination-relevant responses while preserving the model's original token distribution and thereby reducing hallucinations without compromising generation behavior.

What carries the argument

Selective latent intervention that isolates and suppresses hallucination signals from the model's intrinsic generation signals.

If this is right

Hallucination rates drop across both generative and discriminative vision-language benchmarks.
Output length and token probabilities remain closer to the unmodified model than with prior steering methods.
The same framework works across multiple families of large vision-language models.
The method requires no retraining and can be inserted at inference time.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The separation of signals suggests hallucination is a distinct, addressable component rather than an unavoidable byproduct of multimodal generation.
The same isolation technique could be tested on other model failures such as factual errors or safety violations.
Real-time interactive applications may benefit most because users would no longer trade response quality for reduced hallucination.

Load-bearing premise

Hallucination signals can be isolated from normal generation signals in latent space without side effects on token distribution or output length.

What would settle it

Measuring token distributions and output lengths on the same prompts before and after MESA intervention and finding either a statistically significant shift or no reduction in hallucination rates on standard benchmarks.

Figures

Figures reproduced from arXiv: 2604.07914 by Joey Tianyi Zhou, Weizhan Zhang, Xin Zhang, Yuanhong Zhang, Zhaoyang Wang.

**Figure 2.** Figure 2: Analysis of latent space steering effects on hallucination and generation behavior. (Left) Relationship between [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Figure 3: Overview of MESA. MESA decomposes hallucination mitigation into three offline stages and an inference-time [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

**Figure 5.** Figure 5: Ablation of key parameters. (a) Effect of differ [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗

**Figure 6.** Figure 6: Further analysis of MESA. (a)-(b) Generation behavior on CHAIR with LLaVA-v1.5: MESA preserves the EOS margin and [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗

**Figure 7.** Figure 7: The throughput (tested on NVIDIA A800) v.s. [PITH_FULL_IMAGE:figures/full_fig_p011_7.png] view at source ↗

**Figure 8.** Figure 8: Comparison of token probability distributions and representative examples under normal and hallucination-inducing [PITH_FULL_IMAGE:figures/full_fig_p015_8.png] view at source ↗

**Figure 9.** Figure 9: Illustration of hallucination correction comparisons. Text highlighted in red bold denotes explicit hallucinations or [PITH_FULL_IMAGE:figures/full_fig_p015_9.png] view at source ↗

read the original abstract

Large Vision-Language Models (LVLMs) have achieved remarkable success across cross-modal tasks but remain hindered by hallucinations, producing textual outputs inconsistent with visual content. Existing methods mitigate hallucinations but often alter generation behavior, resulting in shorter outputs and shifted token distributions, especially in latent space steering approaches. We identify that this issue stems from entangled steering signals, where suppressing hallucinations inadvertently disrupts the model's intrinsic generation behavior. To address this, we propose MESA, an effective plug-and-play framework that performs controlled and selective latent intervention for hallucination mitigation. Specifically, MESA targets hallucination-relevant responses while preserving the model's original token distribution, enabling effective hallucination reduction without compromising generation behavior. Extensive experiments across diverse generative and discriminative benchmarks demonstrate that MESA consistently reduces hallucinations while better preserving generation behavior, outperforming prior methods across multiple LVLM families.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

MESA gives a selective latent intervention that cuts hallucinations in LVLMs without the usual drops in output length or token distribution shifts, and the multi-benchmark experiments support the claim.

read the letter

The main point is that this paper isolates hallucination-related signals in the latent space of vision-language models and steers only those, leaving the rest of the generation process alone. Prior steering methods often bleed over and shorten responses or change probabilities, but MESA claims to avoid that by design. The experiments test this across several LVLM families on both generative and discriminative benchmarks, and the results show lower hallucination rates while output statistics stay closer to the base model. That breadth of testing is the strongest part of the work; it moves beyond single-model demos and gives a clearer picture of where the method holds up. The plug-and-play framing also makes it easy to try without retraining, which is practical for people already running these models. The soft spot is that the description of how the selective intervention actually separates the signals stays fairly high-level. Without more detail on the exact identification step or edge cases where entanglement might not separate cleanly, it is hard to judge how far the gains will generalize beyond the reported setups. Still, the numbers do not show obvious side effects, so the central claim does not look circular or overfitted. This paper is for researchers who work on reliable multimodal generation and want an add-on fix rather than a full retrain. It is solid enough on the experimental side to deserve a serious referee, even if some reviewers will want tighter analysis of the disentanglement step.

Referee Report

0 major / 3 minor

Summary. The manuscript proposes MESA, a plug-and-play framework to mitigate hallucinations in Large Vision-Language Models (LVLMs) by identifying entangled steering signals that disrupt generation behavior in prior latent-space methods. MESA performs controlled, selective latent interventions targeting hallucination-relevant responses while preserving the model's original token distribution and output characteristics. The central claim is supported by experiments across generative and discriminative benchmarks on multiple LVLM families, showing reduced hallucinations without the side effects of shorter outputs or shifted distributions.

Significance. If the reported results hold under the described conditions, the work is significant for providing a practical, non-disruptive approach to hallucination mitigation in LVLMs. By addressing the entanglement issue directly and emphasizing preservation of generation behavior, it improves upon existing steering techniques that often trade off output quality. The multi-benchmark, multi-family evaluation and plug-and-play design enhance its potential utility and reproducibility.

minor comments (3)

Abstract: The claim of 'consistent' outperformance would be strengthened by including one or two specific quantitative metrics (e.g., hallucination rate reduction on a named benchmark) rather than qualitative statements alone.
Section 4 (Experiments): Verify that token-distribution preservation is measured with the same metrics (e.g., KL divergence or perplexity) across all compared methods to ensure fair side-effect analysis.
Notation: Ensure the selective intervention operator is defined with explicit input/output dimensions in the method section to avoid ambiguity when readers implement the plug-and-play module.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for their positive summary of our work on MESA and for recommending minor revision. The assessment correctly captures the core contribution of controlled selective latent intervention to reduce hallucinations while preserving token distributions and generation behavior. As the report contains no specific major comments, we have no points requiring detailed rebuttal or clarification.

Circularity Check

0 steps flagged

No significant circularity in derivation chain

full rationale

The paper introduces MESA as a plug-and-play framework for selective latent intervention targeting hallucination signals in LVLMs while preserving original token distributions and generation behavior. The abstract and available text describe the approach as an independent addition motivated by identifying entangled steering signals, with claims supported by experimental results on multiple benchmarks and model families. No equations, fitted parameters presented as predictions, self-citations as load-bearing premises, or ansatzes that reduce the central result to a redefinition of inputs are present. The derivation remains self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The approach rests on the domain assumption that hallucination signals can be disentangled from normal generation behavior in latent space; no free parameters or invented entities are explicitly introduced in the abstract.

axioms (1)

domain assumption Hallucinations in LVLMs stem from entangled steering signals that disrupt intrinsic generation behavior when suppressed.
Directly stated as the identified root cause in the abstract.

pith-pipeline@v0.9.0 · 5449 in / 1087 out tokens · 32620 ms · 2026-05-10T18:12:01.214679+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

55 extracted references · 17 canonical work pages · 9 internal anchors

[1]

David H Ackley, Geoffrey E Hinton, and Terrence J Sejnowski. 1985. A learning algorithm for Boltzmann machines.Cognitive science9, 1 (1985), 147–169

1985
[2]

Wenbin An, Feng Tian, Sicong Leng, Jiahao Nie, Haonan Lin, QianYing Wang, Ping Chen, Xiaoqin Zhang, and Shijian Lu. 2025. Mitigating object hallucinations in large vision-language models with assembly of global and local attention. In Proceedings of the Computer Vision and Pattern Recognition Conference. 29915– 29926

2025
[3]

Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. 2023. Qwen technical report.arXiv preprint arXiv:2309.16609(2023)

work page internal anchor Pith review Pith/arXiv arXiv 2023
[4]

Zechen Bai, Pichao Wang, Tianjun Xiao, Tong He, Zongbo Han, Zheng Zhang, and Mike Zheng Shou. 2024. Hallucination of multimodal large language models: A survey.arXiv preprint arXiv:2404.18930(2024)

work page internal anchor Pith review arXiv 2024
[5]

Jiawei Chen, Dingkang Yang, Tong Wu, Yue Jiang, Xiaolu Hou, Mingcheng Li, Shunli Wang, Dongling Xiao, Ke Li, and Lihua Zhang. 2024. Detecting and evaluating medical hallucinations in large vision language models.arXiv preprint arXiv:2406.10185(2024)

work page internal anchor Pith review Pith/arXiv arXiv 2024
[6]

Junzhe Chen, Tianshu Zhang, Shiyu Huang, Yuwei Niu, Linfeng Zhang, Lijie Wen, and Xuming Hu. 2025. Ict: Image-object cross-level trusted intervention for mitigating object hallucination in large vision-language models. InProceedings of the Computer Vision and Pattern Recognition Conference. 4209–4221

2025
[7]

Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al . 2024. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 24185–24198

2024
[8]

Wei-Lin Chiang, Zhuohan Li, Ziqing Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. 2023. Vicuna: An open-source chatbot impressing gpt-4 with 90%* chatgpt quality.See https://vicuna. lmsys. org (accessed 14 April 2023)2, 3 (2023), 6

2023
[9]

Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Se- bastian Gehrmann, et al. 2023. Palm: Scaling language modeling with pathways. Journal of machine learning research24, 240 (2023), 1–113

2023
[10]

Hongyuan Dong, Jiawen Li, Bohong Wu, Jiacong Wang, Yuan Zhang, and Haoyuan Guo. 2024. Benchmarking and improving detail image caption.arXiv preprint arXiv:2405.19092(2024)

work page arXiv 2024
[11]

Angela Fan, Mike Lewis, and Yann Dauphin. 2018. Hierarchical neural story generation. InProceedings of the 56th Annual Meeting of the Association for Com- putational Linguistics (Volume 1: Long Papers). 889–898

2018
[12]

Fabrizio Gilardi, Meysam Alizadeh, and Maël Kubli. 2023. ChatGPT outperforms crowd workers for text-annotation tasks.Proceedings of the National Academy of Sciences120, 30 (2023), e2305016120

2023
[13]

Iryna Hartsock and Ghulam Rasool. 2024. Vision-language models for medical report generation and visual question answering: A review.Frontiers in artificial intelligence7 (2024), 1430984

2024
[14]

Ari Holtzman, Jan Buys, Li Du, Maxwell Forbes, and Yejin Choi. 2019. The curious case of neural text degeneration.arXiv preprint arXiv:1904.09751(2019)

work page internal anchor Pith review arXiv 2019
[15]

Jakub Hoscilowicz and Artur Janicki. 2025. Adversarial Confusion Attack: Dis- rupting Multimodal Large Language Models.arXiv preprint arXiv:2511.20494 (2025)

work page arXiv 2025
[16]

MD Zakir Hossain, Ferdous Sohel, Mohd Fairuz Shiratuddin, and Hamid Laga
[17]

A comprehensive survey of deep learning for image captioning.ACM Computing Surveys (CsUR)51, 6 (2019), 1–36

2019
[18]

Yiyang Huang, Liang Shi, Yitian Zhang, Yi Xu, and Yun Fu. 2025. SHIELD: Suppressing Hallucinations In LVLM Encoders via Bias and Vulnerability Defense. arXiv preprint arXiv:2510.16596(2025)

work page arXiv 2025
[19]

Drew A Hudson and Christopher D Manning. 2019. Gqa: A new dataset for real- world visual reasoning and compositional question answering. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 6700–6709

2019
[20]

Ngoc Dung Huynh, Mohamed Reda Bouadjenek, Sunil Aryal, Imran Razzak, and Hakim Hacid. 2025. Visual question answering: from early developments to recent advances–a survey.arXiv preprint arXiv:2501.03939(2025)

work page arXiv 2025
[21]

Ziwei Ji, Nayeon Lee, Rita Frieske, Tiezheng Yu, Dan Su, Yan Xu, Etsuko Ishii, Ye Jin Bang, Andrea Madotto, and Pascale Fung. 2023. Survey of hallucination in natural language generation.ACM computing surveys55, 12 (2023), 1–38

2023
[22]

Junhwan Kim, Fabio Pellacini, et al. 2002. Jigsaw image mosaics.ACM Transac- tions on Graphics21, 3 (2002), 657–664

2002
[23]

Sicong Leng, Hang Zhang, Guanzheng Chen, Xin Li, Shijian Lu, Chunyan Miao, and Lidong Bing. 2024. Mitigating object hallucinations in large vision-language models through visual contrastive decoding. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 13872–13882

2024
[24]

Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. 2023. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. InInternational conference on machine learning. PMLR, 19730–19742

2023
[25]

Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. 2022. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. InInternational conference on machine learning. PMLR, 12888–12900

2022
[26]

Li Li, Jiawei Peng, Huiyi Chen, Chongyang Gao, and Xu Yang. 2024. How to configure good in-context sequence for visual question answering. InProceedings Conference acronym ’XX, June 03–05, 2018, Woodstock, NY Yuanhong Zhang, Zhaoyang Wang, Xin Zhang, Weizhan Zhang, and Joey Tianyi Zhou of the IEEE/CVF Conference on Computer Vision and Pattern Recognition....

2024
[27]

Ming Li, Keyu Chen, Ziqian Bi, Ming Liu, Xinyuan Song, Zekun Jiang, Tianyang Wang, Benji Peng, Qian Niu, Junyu Liu, et al. 2024. Surveying the mllm landscape: A meta-review of current surveys.arXiv preprint arXiv:2409.18991(2024)

work page arXiv 2024
[28]

Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. 2023. Evaluating object hallucination in large vision-language models. InProceedings of the 2023 conference on empirical methods in natural language processing. 292–305

2023
[29]

Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C Lawrence Zitnick. 2014. Microsoft coco: Common objects in context. InEuropean conference on computer vision. Springer, 740–755

2014
[30]

Fuxiao Liu, Kevin Lin, Linjie Li, Jianfeng Wang, Yaser Yacoob, and Lijuan Wang
[31]

Mitigating hallucination in large multi-modal models via robust instruction tuning.arXiv preprint arXiv:2306.14565(2023)

work page internal anchor Pith review arXiv 2023
[32]

Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. 2024. Improved baselines with visual instruction tuning. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 26296–26306

2024
[33]

Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. 2023. Visual in- struction tuning.Advances in neural information processing systems36 (2023), 34892–34916

2023
[34]

Hanchao Liu, Wenyuan Xue, Yifei Chen, Dapeng Chen, Xiutian Zhao, Ke Wang, Liping Hou, Rongjun Li, and Wei Peng. 2024. A survey on hallucination in large vision-language models.arXiv preprint arXiv:2402.00253(2024)

work page internal anchor Pith review arXiv 2024
[35]

Sheng Liu, Haotian Ye, and James Zou. 2025. Reducing hallucinations in large vision-language models via latent space steering. InThe Thirteenth International Conference on Learning Representations

2025
[36]

Ilya Loshchilov and Frank Hutter. 2017. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101(2017)

work page internal anchor Pith review Pith/arXiv arXiv 2017
[37]

Andrzej Maćkiewicz and Waldemar Ratajczak. 1993. Principal components analysis (PCA).Computers & Geosciences19, 3 (1993), 303–342

1993
[38]

Nikolay Mikhaylovskiy. 2025. Zipf’s and Heaps’ Laws for Tokens and LLM- generated Texts.Findings of the Association for Computational Linguistics: EMNLP 2025(2025), 15469–15481

2025
[39]

Yassine Ouali, Adrian Bulat, Brais Martinez, and Georgios Tzimiropoulos. 2024. Clip-dpo: Vision-language models as a source of preference for fixing hallucina- tions in lvlms. InEuropean Conference on Computer Vision. Springer, 395–413

2024
[40]

Anna Rohrbach, Lisa Anne Hendricks, Kaylee Burns, Trevor Darrell, and Kate Saenko. 2018. Object hallucination in image captioning. InProceedings of the 2018 Conference on Empirical Methods in Natural Language Processing. 4035–4045

2018
[41]

Dustin Schwenk, Apoorv Khandelwal, Christopher Clark, Kenneth Marino, and Roozbeh Mottaghi. 2022. A-okvqa: A benchmark for visual question answering using world knowledge. InEuropean conference on computer vision. Springer, 146–162

2022
[42]

Jingran Su, Jingfan Chen, Hongxin Li, Yuntao Chen, Li Qing, and Zhaoxiang Zhang. 2025. Activation steering decoding: Mitigating hallucination in large vision-language models through bidirectional hidden state intervention. InPro- ceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 12964–12974

2025
[43]

Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. 2023. Llama: Open and efficient foundation language models.arXiv preprint arXiv:2302.13971(2023)

work page internal anchor Pith review Pith/arXiv arXiv 2023
[44]

Junyang Wang, Yuhang Wang, Guohai Xu, Jing Zhang, Yukai Gu, Haitao Jia, Jiaqi Wang, Haiyang Xu, Ming Yan, Ji Zhang, et al. 2023. Amber: An llm-free multi-dimensional benchmark for mllms hallucination evaluation.arXiv preprint arXiv:2311.07397(2023)

work page arXiv 2023
[45]

Xintong Wang, Jingheng Pan, Liang Ding, and Chris Biemann. 2024. Mitigat- ing hallucinations in large vision-language models with instruction contrastive decoding. InFindings of the Association for Computational Linguistics: ACL 2024. 15840–15853

2024
[46]

Le Yang, Ziwei Zheng, Boxu Chen, Zhengyu Zhao, Chenhao Lin, and Chao Shen
[47]

InProceedings of the Computer Vision and Pattern Recognition Conference

Nullu: Mitigating object hallucinations in large vision-language models via halluspace projection. InProceedings of the Computer Vision and Pattern Recognition Conference. 14635–14645
[48]

Jiabo Ye, Haiyang Xu, Haowei Liu, Anwen Hu, Ming Yan, Qi Qian, Ji Zhang, Fei Huang, and Jingren Zhou. 2024. mplug-owl3: Towards long image- sequence understanding in multi-modal large language models.arXiv preprint arXiv:2408.04840(2024)

work page arXiv 2024
[49]

Hao Yin, Guangzong Si, and Zilei Wang. 2025. Clearsight: Visual signal en- hancement for object hallucination mitigation in multimodal large language models. InProceedings of the Computer Vision and Pattern Recognition Conference. 14625–14634

2025
[50]

Jianghao Yin, Qin Chen, Kedi Chen, Jie Zhou, Xingjiao Wu, and Liang He. 2026. Dynamic Multimodal Activation Steering for Hallucination Mitigation in Large Vision-Language Models.arXiv preprint arXiv:2602.21704(2026)

work page arXiv 2026
[51]

Tianyu Yu, Yuan Yao, Haoye Zhang, Taiwen He, Yifeng Han, Ganqu Cui, Jinyi Hu, Zhiyuan Liu, Hai-Tao Zheng, Maosong Sun, et al . 2024. Rlhf-v: Towards trustworthy mllms via behavior alignment from fine-grained correctional human feedback. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 13807–13816

2024
[52]

Tianyu Yu, Haoye Zhang, Qiming Li, Qixin Xu, Yuan Yao, Da Chen, Xiaoman Lu, Ganqu Cui, Yunkai Dang, Taiwen He, et al . 2025. Rlaif-v: Open-source ai feedback leads to super gpt-4v trustworthiness. InProceedings of the Computer Vision and Pattern Recognition Conference. 19985–19995

2025
[53]

Jingyi Zhang, Jiaxing Huang, Sheng Jin, and Shijian Lu. 2024. Vision-language models for vision tasks: A survey.IEEE transactions on pattern analysis and machine intelligence46, 8 (2024), 5625–5644

2024
[54]

Shanghang Zhang, Xiaohui Shen, Zhe Lin, Radomír Měch, Joao P Costeira, and José MF Moura. 2018. Learning to understand image blur. InProceedings of the IEEE conference on computer vision and pattern recognition. 6586–6595

2018
[55]

MiniGPT-4: Enhancing Vision-Language Understanding with Advanced Large Language Models

Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. 2023. Minigpt-4: Enhancing vision-language understanding with advanced large lan- guage models.arXiv preprint arXiv:2304.10592(2023). Mitigating Entangled Steering in Large Vision-Language Models for Hallucination Reduction Conference acronym ’XX, June 03–05, 2018, Woodstock, NY A Detail...

work page internal anchor Pith review arXiv 2023