Pith. sign in

REVIEW 3 major objections 4 minor 2 cited by

Nullu: Mitigating Object Hallucinations in Large Vision-Language Models via HalluSpace Projection

T0 review · 3 major / 4 minor · reviewed 2026-08-11 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read By projecting MLP weights onto the null space of a low-rank HalluSpace learned from paired truthful and hallucinated captions, Nullu suppresses object hallucinations in large vision-language models with no training and no extra inference…

desk verdict Empirically the zero-cost patch works across LVLMs, but the HalluSpace story is unproven and the interpretability test is circular. read the letter →

arxiv 2412.13817 v3 pith:3HYSDWZM submitted 2024-12-18 cs.CV

Le Yang , Ziwei Zheng , Boxu Chen , Zhengyu Zhao , Chenhao Lin , Chao Shen This is my paper · ORCID
classification cs.CV
keywords objecthallucinationlargevision-languagemodelsHalluSpacenullprojectionsingularvaluedecompositionweighteditingLLMpriorCHAIR
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper introduces Nullu, a training-free method for reducing object hallucinations in large vision-language models (LVLMs). The authors' central claim is that hallucinated and truthful responses differ along a low-rank subspace of the model's hidden-feature space, which they call the HalluSpace, and that projecting the model's MLP weights onto the null space of that subspace suppresses hallucinated objects in generated captions. Because the edited weights are re-loaded into the model, the method adds no inference-time computation. The paper reports consistent reductions in hallucination metrics across LLaVA-1.5, MiniGPT-4, and mPLUG-Owl2, together with preserved or improved scores on general multimodal benchmarks. The significance, if the claim holds, is that a simple weight edit can remove a known failure mode of LVLMs without retraining or extra decoding steps.

What carries the argument

The HalluSpace is the central object: a rank-$k$ subspace spanned by the top-$k$ right singular vectors of the matrix of differences between hallucinated and truthful hidden features, extracted per transformer layer. The carrying mechanism is the null-space projection operator $(I - \mathbf{V}_{k}\mathbf{V}_{k}^\top)$, applied to the MLP weight matrix of selected layers. Under the paper's feature-decomposition assumption, hidden activations split into truthful contexts, hallucination biases, and noise; the top-$k$ SVD directions align with the hallucination bias subspace, so the projection removes exactly those biases. The edited weights are reparameterized back into the model, which is why the method costs nothing extra at inference.

What would settle it

Compute, on a held-out set, how much variance of truthful object-related features projects onto the HalluSpace: if the norm of the truthful projection is comparable to the hallucinated projection, null-space projection is erasing factual content. A simpler behavioral test: compare CHAIR/POPE before and after editing on images containing rare objects; if rare-object accuracy drops while common-object accuracy holds, the HalluSpace overlaps with genuine knowledge.

Watch

Extended reading notes

Core claim

The paper's core discovery is that object hallucination in LVLMs can be attacked directly in weight space. Using paired captions of the same image, one truthful and one hallucinated, the authors extract hidden features at each transformer layer and compute the difference matrix $E_\ell = X_\ell^+ - X_\ell^-$. The top-$k$ right singular vectors of $E_\ell$ define the low-rank HalluSpace $\mathbf{V}_{\ell,k}$, which the paper argues captures the language-model priors that drive hallucinated object mentions. Replacing the MLP weights with $W^{\mathrm{ed}}_\ell = (I - \mathbf{V}_{\ell,k}\mathbf{V}_{\ell,k}^\top) W^{\mathrm{org}}_\ell$ projects input features away from that subspace, filtering out hallucination-prone directions while leaving truthful content intact. Empirically, the paper shows this single edit lowers CHAIR scores, raises POPE/OPOPE accuracy and F-score across three model families, and even improves perception and recognition scores on the MME benchmark.

Load-bearing premise

The load-bearing premise is that the top directions of the difference between hallucinated and truthful hidden features point exactly at the hallucination bias and are orthogonal to every direction that carries truthful content, so zeroing them removes hallucinations without erasing factual knowledge.

Editorial extensions

If this is right

  • Object hallucination can be reduced by a closed-form weight edit computed from paired truthful and hallucinated captions, with no fine-tuning and no extra inference cost.
  • The HalluSpace carries language-model priors, so the mechanism offers a parameter-space route to debias LVLMs at their source rather than patching outputs.
  • Performance on general benchmarks such as MME and LLaVA-Bench is preserved or improved, meaning the edit does not trade away general capability for hallucination control.
  • The formal link to DPO indicates that the edit approximates one step of preference optimization, which offers a unified view of weight-editing and alignment-based hallucination mitigation.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A natural extension the paper does not explore: the same paired-data, SVD, null-space recipe should apply to other systematic failure modes of LLMs/LVLMs, such as toxicity, sycophancy, or stylistic bias, whenever one can construct contrastive truthful/harmful text pairs.
  • The orthogonality assumption could be tested directly by measuring how much variance of factual object information projects onto the HalluSpace; if rare object categories are suppressed more than common ones, the edit is removing content along with hallucination.
  • Because the edited weights are fixed, the method cannot adapt per image; a dynamic variant that updates the projection from the current image's features might close the remaining gap with adaptive decoding methods, at some inference cost.
  • The method's success on small models suggests a scaling prediction: as LVLMs get larger and their language priors stronger, the HalluSpace rank needed to control hallucination may grow, and null-space projection might need to be applied to more layers.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper proposes Nullu, a training-free method for mitigating object hallucination (OH) in large vision-language models (LVLMs). Nullu constructs paired truthful and hallucinated text prompts for the same images, extracts hidden features at each LLM layer, forms the difference matrix E_l = X^+_l - X^-_l, and computes its top-k right singular vectors V_k, called HalluSpace. The MLP weights at selected layers are then edited as W_ed = (I - V_k V_k^T) W_org, which projects input features into the null space of HalluSpace. The edited weights can be reloaded into the model, so the method adds no inference-time cost. Experiments on LLaVA-1.5, MiniGPT-4, and mPLUG-Owl2 report consistent improvements on CHAIR, POPE/OPOPE, and preserved or improved performance on MME and LLaVA-Bench. The paper also offers interpretability evidence by decoding HalluSpace directions and draws a conceptual connection to DPO.

Significance. If the reported empirical results are reliable, Nullu is a practically attractive contribution: it is simple, training-free, requires no extra inference cost, and shows consistent gains across three different LVLM families on externally defined benchmarks such as CHAIR and OPOPE. The paper also ships code, which is a strength. However, the mechanistic claim that HalluSpace isolates a low-rank 'hallucination bias' subspace is not established: the derivation in Eq. (5) relies on an untested orthogonality assumption, and the validation in §4.5 is circular. In addition, key hyperparameters appear to be selected on the CHAIR evaluation set itself, which weakens the headline CHAIR numbers. These issues do not invalidate the empirical method, but they do mean the paper currently overstates what is demonstrated.

major comments (3)
  1. [§4.5 and Supp. §12] The derivation of Eq. (5) is not justified. Starting from x_i = t_i + h_i + noise, the edited output is x_i(I - V_k V_k^T)W_org = t_i W_org + h_i W_org + noise·W_org - (t_i + h_i + noise) V_k V_k^T W_org. The paper claims that the projection removes only the hallucinated bias term, but this requires the additional condition (t_i + noise) V_k ≈ 0, i.e., that the top singular vectors of E = X^+ - X^- are orthogonal to truthful contexts and noise. This orthogonality is never measured. The SVD of E only identifies directions of maximal difference between the two feature sets; if E carries variance from general linguistic or positional variation, V_k will overlap with truthful features, and the projection will suppress factual content. The decoding results in Supp. Table 5 lend concrete substance to this concern: the top tokens associated with V_k include words such as 'background', 'middle', and 'left', which are not hallucinated objects. The authors should directly quantify the overlap between V_k and the truthful feature subspace (for example, report ||X^- V_k||_F / ||X^-||_F across layers, compare against random subspaces of the same rank), and should show that the projection's effect on general benchmarks is not simply a generic low-rank perturbation.
  2. [§4.5, Table 3 and §3.2] The validation that 'HalluSpace represents hallucination biases' is circular. The authors select 100 CHAIR samples where Nullu successfully mitigates OH, then compute difference vectors e_i between raw and edited LLaVA features. But for any input x, the difference introduced by the edit is x - x(I - V_k V_k^T) = x V_k V_k^T, which by construction lies in span(V_k). Therefore the finding that these difference vectors have larger projection norms onto V_k than random vectors is a mathematical consequence of the edit operation, not evidence that V_k encodes hallucination-related semantics. A proper test would use held-out paired data from the construction in §3.1 and ask whether the difference vectors X^+ - X^- (without any editing) align with V_k more than with random subspaces, or whether the projection norm of hallucinated prompts onto V_k exceeds that of truthful prompts. As written, this experiment supports only the trivial statement that a projection removes the component it projects away.
  3. [§4.5 and Supp. §12] The central empirical claim is partially compromised by hyperparameter selection on the evaluation benchmark. Section 3.2 states that the layer set {l} and rank k are determined via a standard hyper-parameter sweep, and Table 3 in §4.5 shows this sweep is performed on CHAIR with LLaVA-1.5-7B. The final configuration (layers 16-32, k=4) is then used to produce the headline CHAIR results in Table 1, so those numbers are a selected optimum rather than an unbiased estimate of the method's performance. The same concern applies to the per-model choices in Supp. §9 (k=8 for MiniGPT-4, k=32 on CHAIR for mPLUG-Owl2). The authors should either use a separate validation split for hyperparameter selection and report the held-out evaluation numbers, or explicitly characterize the CHAIR results as tuned on the evaluation set. This is important because the method's main practical claim is empirical, and the current reporting makes the magnitude of the improvement hard to assess.
minor comments (4)
  1. [§3.3 and Supp. Table 5] The paper says it selects only 'the most representative results' for interpreting V_k, but does not define the selection criterion. Since the full decoding in Supp. Table 5 shows high-frequency positional words rather than hallucinated object names, the authors should clarify how representativeness was judged and whether the interpretation changes if other directions are examined.
  2. [§3.4, Eq. (7)] The DPO connection is conceptual and the notation is compressed: in the first equality of Eq. (7), o_{y_i^+}(x_i^+)^T - o_{y_i^-}(x_i^-)^T is rewritten by adding and subtracting o_{y_i^+}(x_i^-)^T, but the second equality then contains (o_{y_i^+} - o_{y_i^-})(x_i^-)^T; the two displayed expressions are consistent only if the reader infers the intermediate algebraic step. Please spell out the rearrangement.
  3. [§3.1] It is not described how the hidden features x^+_{i,l} and x^-_{i,l} are aligned when averaging over the token dimension. Since the truthful and hallucinated captions will generally have different lengths, the paper should state whether features are averaged over all tokens, over the last token, or via some other alignment, because this affects the construction of E_l and the resulting SVD directions.
  4. [Figure 3] Figure 3(c) reports word-frequency statistics for Nullu, LLM priors, LLaVA, and VCD, but no numerical axis or table is provided. This makes the claimed overlap between HalluSpace-decoded words and LLM priors difficult to verify. A small table with the frequencies of the specific overlapping words would be more informative.

Circularity Check

1 steps flagged · score 4.0 of 10

Main benchmark results are externally validated and independent, but the Section 4.5 validation of HalluSpace is self-confirming: the alignment it measures is created by the projection itself.

  1. self definitional [Section 4.5, 'Does HalluSpace represent the hallucination biases?' (extended in Supplementary Section 12)]
    "To evaluate this, we select 100 test samples from CHAIR where Nullu successfully mitigates OH issues. We compute difference vectors ei for each sample between the raw and edited LLaVA features. Moreover, we generate 100 random vectors ri as a comparison baseline. ... Since the selected test samples were successfully edited to avoid OH, this evidence indicates that HalluSpace captures directions in the feature space associated with hallucinations in LVLMs."

    Given Eq. (3), W_ed = (I - VV^T)W_org, the edited hidden features are xW_ed = xW_org - xVV^TW_org, so the 'difference vector between raw and edited features' equals xVV^TW_org (or xVV^T before the output readout). Its projection onto V is nonzero whenever x has any component along V, and it points along the same V that defines the edit; the large ||sigma|| compared with random isotropic vectors only reflects that real inputs are not orthogonal to V. Thus the claimed evidence that HalluSpace captures hallucination is entailed by the projection construction, not by independent measurement. Selecting only the 100 samples where Nullu 'successfully mitigates' OH further guarantees the edit changed the output, amplifying the by-construction alignment.

full rationale

The central empirical claim that Nullu reduces object hallucination is checked against external benchmarks (CHAIR, POPE, OPOPE, MME, LLaVA-Bench) with baseline comparisons; those results are not derived from the HalluSpace definition by construction, so the main method claim is not circular. The choice of layers {ell} and rank k on CHAIR itself (Table 3) is an experimental-design / overfitting concern, not a circularity. The orthogonality assumption behind Eq. (5) is asserted rather than verified, and the decoded vocabulary of V contains positional tokens, which weakens the mechanistic story but is not a circular reduction. The genuine circular step is the Section 4.5 interpretability test: the difference between raw and edited features is produced by projecting along V, so the measured alignment with V is a mathematical consequence of the edit itself, independent of whether V semantically encodes hallucination. This does not invalidate the benchmark gains, but it means the paper's mechanistic interpretation of HalluSpace is not established by that experiment.

Assumptions & free parameters 2 free parameters · 5 assumptions · 1 invented entities

The central claim rests mainly on the representativeness of the paired data, the SVD subspace identification, and the unverified orthogonality between HalluSpace and truthful directions. The rank k and layer choices are free parameters tuned on the evaluation benchmarks. No physical entity is postulated; HalluSpace is a data-derived mathematical construct.

free parameters (2)
  • Editing layers {l} = LLaVA-1.5, MiniGPT-4, mPLUG-Owl2 all use layers 16 to 32
    Selected via hyperparameter sweep on CHAIR (Table 3) and used across models; not derived from theory.
  • Rank k of HalluSpace = 4 (LLaVA-1.5 CHAIR and POPE), 8 (MiniGPT-4), 32 (mPLUG-Owl2 CHAIR), 16 (mPLUG-Owl2 POPE)
    Per-model and per-dataset tuning on CHAIR and POPE; affects the strength of the projection and is a fitted value, not a prediction.
assumptions (5)
  • ad hoc to paper Hidden features decompose into truthful contexts, hallucinated biases, and noise, with HalluSpace aligned with the hallucinated bias subspace.
    Assumed in Eq. (4) and supplementary Eq. (11) to justify Eq. (5); adopted heuristically from prior work and not proven for LVLMs.
  • domain assumption Top-k right singular vectors of E = X^+ - X^- span exactly the hallucination bias subspace.
    Used to define HalluSpace in Section 3.2; assumes paired-data differences are dominated by hallucination semantics rather than caption style, length, or object co-occurrence artifacts.
  • domain assumption GPT-3.5 and LURE generated hallucinated descriptions are representative of the object hallucinations that occur at test time.
    Paired data construction in Section 3.1 follows LURE; if synthetic hallucinations differ from natural object hallucination, the subspace will be misaligned.
  • domain assumption Projecting MLP weights to the null space of V does not remove truthful or general-knowledge directions.
    Required for the claim that Nullu preserves general performance; supported only empirically, with no direct orthogonality check between V and truthful directions.
  • standard math Standard linear algebra: (I - V V^T) is the orthogonal projection onto the null space of V.
    Proof in supplementary Section 6; uncontroversial.
invented entities (1)
  • HalluSpace (low-rank subspace V_l,k) independent evidence
    purpose: Captures directions in feature space that distinguish hallucinated from truthful representations, used to edit MLP weights.
    The subspace is defined from paired data and has observable effects on outputs, and its projection norms are compared on benchmark samples. However, the direct evidence that it specifically represents hallucination is partly circular because the validation vectors are produced by the same projection.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Nullu: Mitigating Object Hallucinations in Large Vision-Language Models via HalluSpace Projection." pith.science (2026). https://pith.science/paper/3HYSDWZM

@misc{pith2026241213817,
  author       = {Pith},
  title        = {Pith review of: Nullu: Mitigating Object Hallucinations in Large Vision-Language Models via HalluSpace Projection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3HYSDWZM}},
  note         = {Machine review of arXiv:2412.13817}
}
read the original abstract

Recent studies have shown that large vision-language models (LVLMs) often suffer from the issue of object hallucinations (OH). To mitigate this issue, we introduce an efficient method that edits the model weights based on an unsafe subspace, which we call HalluSpace in this paper. With truthful and hallucinated text prompts accompanying the visual content as inputs, the HalluSpace can be identified by extracting the hallucinated embedding features and removing the truthful representations in LVLMs. By orthogonalizing the model weights, input features will be projected into the Null space of the HalluSpace to reduce OH, based on which we name our method Nullu. We reveal that HalluSpaces generally contain prior information in the large language models (LLMs) applied to build LVLMs, which have been shown as essential causes of OH in previous studies. Therefore, null space projection suppresses the LLMs' priors to filter out the hallucinated features, resulting in contextually accurate outputs. Experiments show that our method can effectively mitigate OH across different LVLM families without extra inference costs and also show strong performance in general LVLM benchmarks. Code is released at https://github.com/Ziwei-Zheng/Nullu.

Figures

Figures reproduced from arXiv: 2412.13817 by the authors.

Figure 1
Figure 1. An illustration of Nullu. (a) In the editing phase, Nullu [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. An overview of Nullu, which identifies the HalluSpaces to edit model weights for LVLMs. (a) The paired truthful and hallucinated [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. The relation between Nullu and other debiasing methods [ [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: (a) The throughput (tested on NVIDIA RTX 4090) v.s. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: MME full set evaluation results on LLaVA-1.5. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Illustration of hallucination correction by our proposed Nullu with two samples from LLaVA-Bench using LLaVA-1.5-7B. [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: (a) The illustration of experiments. (b) The mean of [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: The illustration of difference vectors and random vectors [PITH_FULL_IMAGE:figures/full_fig_p016_8.png]
Figure 9
Figure 9. Figure 9: The generated website of Nullu [PITH_FULL_IMAGE:figures/full_fig_p018_9.png]
Figure 10
Figure 10. Figure 10: More examples of Nullu on LLaVA-Bench [PITH_FULL_IMAGE:figures/full_fig_p021_10.png]
Figure 11
Figure 11. Figure 11: More examples of Nullu on LLaVA-Bench [PITH_FULL_IMAGE:figures/full_fig_p022_11.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. GrAInS: Gradient-based Attribution for Inference-Time Steering of LLMs and VLMs

    cs.CL 2025-07 conditional novelty 6.0 of 10

    GrAInS uses Integrated Gradients to identify the most influential tokens, then builds layer-wise steering vectors that improve truthfulness, reduce hallucination, and preserve general capabilities in LLMs and VLMs.

  2. Not Just Text: Uncovering Vision Modality Typographic Threats in Image Generation Models

    cs.CV 2024-12 conditional novelty 6.0 of 10

    Typography inserted into input images can manipulate CLIP-guided image generation models to produce harmful or biased content, and existing text-focused defenses do not catch it.

Reference graph

Works this paper leans on

67 extracted references · 41 canonical work pages · cited by 2 Pith papers

  1. [1]

    Gpt-4 technical report

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ah- mad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774 ,

    arXiv

  2. [2]

    Qwen technical report

    Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. 2

    arXiv 2023

  3. [3]

    Qwen-vl: A frontier large vision-language model with versatile abilities

    Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 2

    arXiv 2023

  4. [4]

    Hallucination of multimodal large language models: A survey

    Zechen Bai, Pichao Wang, Tianjun Xiao, Tong He, Zongbo Han, Zheng Zhang, and Mike Zheng Shou. Hallucination of multimodal large language models: A survey. arXiv preprint arXiv:2404.18930, 2024. 1, 2

    arXiv 2024

  5. [5]

    Let there be a clock on the beach: Reducing object halluci- nation in image captioning

    Ali Furkan Biten, Llu ´ıs G ´omez, and Dimosthenis Karatzas. Let there be a clock on the beach: Reducing object halluci- nation in image captioning. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision , pages 1381–1390, 2022. 2

    work page 2022

  6. [6]

    Shikra: Unleashing multi- modal llm’s referential dialogue magic

    Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multi- modal llm’s referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 2

    arXiv 2023

  7. [7]

    Dress: Instructing large vision-language models to align and interact with humans via natural lan- guage feedback

    Yangyi Chen, Karan Sikka, Michael Cogswell, Heng Ji, and Ajay Divakaran. Dress: Instructing large vision-language models to align and interact with humans via natural lan- guage feedback. In IEEE Conf. Comput. Vis. Pattern Recog., pages 14239–14250, 2024. 2

    work page 2024

  8. [8]

    Internvl: Scaling up vision founda- tion models and aligning for generic visual-linguistic tasks

    Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision founda- tion models and aligning for generic visual-linguistic tasks. In IEEE Conf. Comput. Vis. Pattern Recog. , pages 24185– 24198, 2024. 2

    work page 2024

Show all 67 references
  1. [9]

    Halc: Object hallucination reduc- tion via adaptive focal-contrast decoding

    Zhaorun Chen, Zhuokai Zhao, Hongyin Luo, Huaxiu Yao, Bo Li, and Jiawei Zhou. Halc: Object hallucination reduc- tion via adaptive focal-contrast decoding. In Int. Conf. Ma- chine Learn., 2024. 1, 2, 5, 6, 7, 14, 16

    2024

  2. [10]

    Gonzalez, Ion Stoica, and Eric P

    Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang- hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong- hao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with 90%* chatgpt quality, 2023. 2, 6

    2023

  3. [11]

    Dola: Decoding by con- trasting layers improves factuality in large language models

    Yung-Sung Chuang, Yujia Xie, Hongyin Luo, Yoon Kim, James R Glass, and Pengcheng He. Dola: Decoding by con- trasting layers improves factuality in large language models. In Int. Conf. Learn. Represent., 2023. 6, 7, 14

    2023

  4. [12]

    Bert: Pre-training of deep bidirectional transformers for language understanding

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 2

    2018 arXiv

  5. [13]

    Multi-modal hal- lucination control by visual information grounding

    Alessandro Favero, Luca Zancato, Matthew Trager, Sid- dharth Choudhary, Pramuditha Perera, Alessandro Achille, Ashwin Swaminathan, and Stefano Soatto. Multi-modal hal- lucination control by visual information grounding. In IEEE Conf. Comput. Vis. Pattern Recog. , pages 14303–14312,

  6. [14]

    Beam search strategies for neural machine translation

    Markus Freitag and Yaser Al-Onaizan. Beam search strategies for neural machine translation. arXiv preprint arXiv:1702.01806, 2017. 6, 7

    2017 arXiv

  7. [15]

    Mme: A comprehensive evaluation benchmark for multimodal large language models

    Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Zhenyu Qiu, Wei Lin, Jinrui Yang, Xiawu Zheng, et al. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394, 2023. 5

    2023 arXiv

  8. [16]

    Transformer feed-forward layers are key-value memories

    Mor Geva, Roei Schuster, Jonathan Berant, and Omer Levy. Transformer feed-forward layers are key-value memories. In Adv. Neural Inform. Process. Syst., pages 5484–5495, 2021. 4

    2021

  9. [17]

    Detecting and preventing hallucinations in large vision language models

    Anisha Gunjal, Jihan Yin, and Erhan Bas. Detecting and preventing hallucinations in large vision language models. In AAAI, pages 18135–18143, 2024. 2

    2024

  10. [18]

    Opera: Alleviating hallucination in multi- modal large language models via over-trust penalty and retrospection-allocation

    Qidong Huang, Xiaoyi Dong, Pan Zhang, Bin Wang, Con- ghui He, Jiaqi Wang, Dahua Lin, Weiming Zhang, and Nenghai Yu. Opera: Alleviating hallucination in multi- modal large language models via over-trust penalty and retrospection-allocation. In IEEE Conf. Comput. Vis. Pattern Re...

    2024

  11. [19]

    Visual hallucinations of multi-modal large language models

    Wen Huang, Hongbin Liu, Minxin Guo, and Neil Zhenqiang Gong. Visual hallucinations of multi-modal large language models. arXiv preprint arXiv:2402.14683, 2024. 2

    2024 arXiv

  12. [20]

    Hallucination augmented contrastive learn- ing for multimodal large language model

    Chaoya Jiang, Haiyang Xu, Mengfan Dong, Jiaxing Chen, Wei Ye, Ming Yan, Qinghao Ye, Ji Zhang, Fei Huang, and Shikun Zhang. Hallucination augmented contrastive learn- ing for multimodal large language model. In IEEE Conf. Comput. Vis. Pattern Recog., pages 27036–27046, 2024. 1, 2

    2024

  13. [21]

    Fgaif: Aligning large vision- language models with fine-grained ai feedback

    Liqiang Jing and Xinya Du. Fgaif: Aligning large vision- language models with fine-grained ai feedback. arXiv preprint arXiv:2404.05046, 2024. 2

    2024 arXiv

  14. [22]

    Exposing and mitigating spurious correlations for cross-modal retrieval

    Jae Myung Kim, A Koepke, Cordelia Schmid, and Zeynep Akata. Exposing and mitigating spurious correlations for cross-modal retrieval. In IEEE Conf. Comput. Vis. Pattern Recog., pages 2584–2594, 2023. 1, 2

    2023

  15. [23]

    Mitigating object hal- lucinations in large vision-language models through visual contrastive decoding

    Sicong Leng, Hang Zhang, Guanzheng Chen, Xin Li, Shijian Lu, Chunyan Miao, and Lidong Bing. Mitigating object hal- lucinations in large vision-language models through visual contrastive decoding. In IEEE Conf. Comput. Vis. Pattern Recog., pages 13872–13882, 2024. 1, 2, 4, 5, 6...

    2024

  16. [24]

    Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation

    Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In Int. Conf. Machine Learn., pages 12888–12900, 2022. 2

    2022

  17. [25]

    Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models

    Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In Int. Conf. Machine Learn., pages 19730–19742, 2023. 2

    2023

  18. [26]

    Inference-time intervention: Elic- iting truthful answers from a language model

    Kenneth Li, Oam Patel, Fernanda Vi ´egas, Hanspeter Pfister, and Martin Wattenberg. Inference-time intervention: Elic- iting truthful answers from a language model. Advances in Neural Information Processing Systems, 36, 2024. 3

    2024

  19. [27]

    Evaluating object hallucination in large vision-language models

    Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Pro- cessing, pages 292–305, 2023. 1, 5, 14, 16

    2023

  20. [28]

    Microsoft coco: Common objects in context

    Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Doll´ar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision–ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceeding...

    2014

  21. [29]

    Mitigating hallucination in large multi-modal models via robust instruction tuning

    Fuxiao Liu, Kevin Lin, Linjie Li, Jianfeng Wang, Yaser Ya- coob, and Lijuan Wang. Mitigating hallucination in large multi-modal models via robust instruction tuning. arXiv preprint arXiv:2306.14565, 2023. 1, 2

    2023 arXiv

  22. [30]

    Improved baselines with visual instruction tuning

    Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 2, 4, 5, 8, 18

    2023 arXiv

  23. [31]

    Visual instruction tuning

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Adv. Neural Inform. Process. Syst., 36, 2024. 1, 2, 6

    2024

  24. [32]

    Reducing hallucina- tions in large vision-language models via latent space steer- ing

    Sheng Liu, Haotian Ye, and James Zou. Reducing hallucina- tions in large vision-language models via latent space steer- ing. In Int. Conf. Learn. Represent., 2024. 3

    2024

  25. [33]

    Paying more atten- tion to image: A training-free method for alleviating halluci- nation in lvlms

    Shi Liu, Kecheng Zheng, and Wei Chen. Paying more atten- tion to image: A training-free method for alleviating halluci- nation in lvlms. In Eur. Conf. Comput. Vis., 2024. 2, 5

    2024

  26. [34]

    Roberta: A robustly optimized bert pretraining approach

    Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettle- moyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692,

    1907 arXiv

  27. [35]

    Introducing meta llama 3: The most capable openly available llm to date

    AI Meta. Introducing meta llama 3: The most capable openly available llm to date. Meta AI Blog (accessed 2024–04–20).,

    2024

  28. [36]

    Direct preference optimization: Your language model is secretly a reward model

    Rafael Rafailov, Archit Sharma, Eric Mitchell, Christo- pher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. Adv. Neural Inform. Process. Syst., 36, 2024. 2, 5, 7, 12

    2024

  29. [37]

    Object hallucination in image cap- tioning

    Anna Rohrbach, Lisa Anne Hendricks, Kaylee Burns, Trevor Darrell, and Kate Saenko. Object hallucination in image cap- tioning. arXiv preprint arXiv:1809.02156, 2018. 2, 4, 5, 16

    2018 arXiv

  30. [38]

    Aligning large multi- modal models with factually augmented rlhf

    Zhiqing Sun, Sheng Shen, Shengcao Cao, Haotian Liu, Chunyuan Li, Yikang Shen, Chuang Gan, Liang-Yan Gui, Yu-Xiong Wang, Yiming Yang, et al. Aligning large multi- modal models with factually augmented rlhf. arXiv preprint arXiv:2309.14525, 2023. 2

    2023 arXiv

  31. [39]

    Gemini: a family of highly capable multimodal models

    Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 1

    2023 arXiv

  32. [40]

    Llama: Open and efficient foundation language models

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timoth´ee Lacroix, Baptiste Rozi`ere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1

    2023 arXiv

  33. [41]

    Llama 2: Open foundation and fine-tuned chat models.arXiv preprint arXiv:2307.09288, 2023

    Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models.arXiv preprint arXiv:2307.09288, 2023. 2, 6

    2023 arXiv

  34. [42]

    Activation addition: Steering language models without opti- mization

    Alexander Matt Turner, Lisa Thiergart, Gavin Leech, David Udell, Juan J Vazquez, Ulisse Mini, and Monte MacDiarmid. Activation addition: Steering language models without opti- mization. arXiv e-prints, pages arXiv–2308, 2023. 3

    2023

  35. [43]

    Model editing as a robust and denoised variant of dpo: A case study on toxicity

    Rheeya Uppaal, Apratim Dey, Yiting He, Yiqiao Zhong, and Junjie Hu. Model editing as a robust and denoised variant of dpo: A case study on toxicity. In Int. Conf. Learn. Repre- sent., 2024. 3, 4, 5, 12

    2024

  36. [44]

    Model surgery: Modulating llm’s behavior via simple parameter editing

    Huanqian Wang, Yang Yue, Rui Lu, Jingxin Shi, Andrew Zhao, Shenzhi Wang, Shiji Song, and Gao Huang. Model surgery: Modulating llm’s behavior via simple parameter editing. arXiv preprint arXiv:2407.08770, 2024. 2

    2024 arXiv

  37. [45]

    Transformers: State-of-the-art natural language processing

    Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chau- mond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, R ´emi Louf, Morgan Funtowicz, et al. Transformers: State-of-the-art natural language processing. In Proceed- ings of the 2020 conference on empirical methods...

    2020

  38. [46]

    Logical closed loop: Uncov- ering object hallucinations in large vision-language models

    Junfei Wu, Qiang Liu, Ding Wang, Jinghao Zhang, Shu Wu, Liang Wang, and Tieniu Tan. Logical closed loop: Uncov- ering object hallucinations in large vision-language models. arXiv preprint arXiv:2402.11622, 2024. 2

    2024 arXiv

  39. [47]

    Evaluating and analyzing relationship hallucinations in large vision-language models

    Mingrui Wu, Jiayi Ji, Oucheng Huang, Jiale Li, Yuhang Wu, Xiaoshuai Sun, and Rongrong Ji. Evaluating and analyzing relationship hallucinations in large vision-language models. In Int. Conf. Machine Learn., pages 53553–53570, 2024. 2

    2024

  40. [48]

    mplug- owl3: Towards long image-sequence understanding in multi- modal large language models, 2024

    Jiabo Ye, Haiyang Xu, Haowei Liu, Anwen Hu, Ming Yan, Qi Qian, Ji Zhang, Fei Huang, and Jingren Zhou. mplug- owl3: Towards long image-sequence understanding in multi- modal large language models, 2024. 2

    2024

  41. [49]

    mplug- owl2: Revolutionizing multi-modal large language model with modality collaboration

    Qinghao Ye, Haiyang Xu, Jiabo Ye, Ming Yan, Anwen Hu, Haowei Liu, Qi Qian, Ji Zhang, and Fei Huang. mplug- owl2: Revolutionizing multi-modal large language model with modality collaboration. In IEEE Conf. Comput. Vis. Pattern Recog., pages 13040–13051, 2024. 1, 2, 6

    2024

  42. [50]

    Woodpecker: Hallucination correction for multimodal large language models

    Shukang Yin, Chaoyou Fu, Sirui Zhao, Tong Xu, Hao Wang, Dianbo Sui, Yunhang Shen, Ke Li, Xing Sun, and Enhong Chen. Woodpecker: Hallucination correction for multimodal large language models. arXiv preprint arXiv:2310.16045 ,

    arXiv

  43. [51]

    De- biasing large visual language models

    Yi-Fan Zhang, Weichen Yu, Qingsong Wen, Xue Wang, Zhang Zhang, Liang Wang, Rong Jin, and Tieniu Tan. De- biasing large visual language models. In Eur. Conf. Comput. Vis., 2024. 1, 2, 4

    2024

  44. [52]

    Diver-ct: Diversity-enhanced red teaming with relaxing constraints

    Andrew Zhao, Quentin Xu, Matthieu Lin, Shenzhi Wang, Yong-jin Liu, Zilong Zheng, and Gao Huang. Diver-ct: Diversity-enhanced red teaming with relaxing constraints. arXiv preprint arXiv:2405.19026, 2024. 2

    2024 arXiv

  45. [53]

    A stitch in time saves nine: Small vlm is a precise guidance for accelerating large vlms

    Wangbo Zhao, Yizeng Han, Jiasheng Tang, Zhikai Li, Yib- ing Song, Kai Wang, Zhangyang Wang, and Yang You. A stitch in time saves nine: Small vlm is a precise guidance for accelerating large vlms. arXiv preprint arXiv:2412.03324 ,

    arXiv

  46. [54]

    Analyzing and mitigating object hallucination in large vision-language models

    Yiyang Zhou, Chenhang Cui, Jaehong Yoon, Linjun Zhang, Zhun Deng, Chelsea Finn, Mohit Bansal, and Huaxiu Yao. Analyzing and mitigating object hallucination in large vision-language models. arXiv preprint arXiv:2310.00754,

    arXiv

  47. [55]

    Minigpt-4: Enhancing vision-language understanding with advanced large language models

    Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mo- hamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 2, 6

    2023 arXiv

  48. [56]

    Ibd: Alleviating hallucinations in large vision- language models via image-biased decoding

    Lanyun Zhu, Deyi Ji, Tianrun Chen, Peng Xu, Jieping Ye, and Jun Liu. Ibd: Alleviating hallucinations in large vision- language models via image-biased decoding. In Int. Conf. Machine Learn., 2024. 2, 5

    2024

  49. [57]

    We want to proof that: any vector z ∈ RD in the null space of the v ∈ RD, (I −vv ⊤), is orthogonal to the vector v

    The derivation of null space Here, we give the details about obtaining the null space of the v. We want to proof that: any vector z ∈ RD in the null space of the v ∈ RD, (I −vv ⊤), is orthogonal to the vector v. Namely, we have v⊤z = 0, where v is the vector with norm 1. I is ...

  50. [58]

    Decoding information in HalluSpace layer Top Tokens 16 dynamic either further above background floor tables ...17 another notable left later others most tables ...18 nearby notable either tables group optional others ...19 notable middle either diverse background overall conce...

  51. [59]

    Factor component analysis The analysis is performed for each layer ℓ, and to avoid the notational burden, we will drop ℓ and focus on each layer separately

    Theoretical Analysis: How Nullu works? 8.1. Factor component analysis The analysis is performed for each layer ℓ, and to avoid the notational burden, we will drop ℓ and focus on each layer separately. We use the same notations with these in the main paper. Based on the heurist...

  52. [60]

    The over- all experimental setup is summarized in Table 6

    Implementation Details of LVLMs This section details the implementation of the evaluated LVLMs and the methods used for OH mitigation. The over- all experimental setup is summarized in Table 6. Unlike the standard greedy method, which selects the most probable token at each de...

  53. [61]

    Editing Layers ℓ: For all models, the editing layers are specified by ℓ ∈ range(16, 32)

  54. [62]

    We use the value 4 for LLaV A-1.5 on both CHAIR and POPE

    The Selected Top- k singular vector: The number of top- k singular vectors selected varies by model. We use the value 4 for LLaV A-1.5 on both CHAIR and POPE. Sim- ilarly, we use 8 for MiniGPT-4 on the evaluated two datasets. For mPLUG-Owl2, we use 32 on CHAIR and 16 on POPE

  55. [63]

    It is set to 3 for both LLaV A-1.5 and MiniGPT-4, while for mPLUG-Owl2, it is set to 1

    Num-beams: This parameter also differs across models. It is set to 3 for both LLaV A-1.5 and MiniGPT-4, while for mPLUG-Owl2, it is set to 1. For the comparison of Nullu with SOTAs methods specifically designed for OH mitigation, the evaluation code is built based on the publi...

  56. [64]

    Is there a/an {} in the image?

    POPE Settings and Additional Results Polling-based Object Probing Evaluation (POPE) [27], presents a streamlined approach to assess object hallucina- tion. POPE interacts directly with the examined LVLM, which distinguishes it from CHAIR. Within this bench- mark, LVLMs are que...

  57. [65]

    The baselines demonstrate consistent performance pat- terns, with Nullu uniformly improving the perceptual com- petencies of the LVLM model

    MME Numerical Results In Table 14, we present the performance of the edited LLaV A-1.5 baselines on the perception-related tasks of the MME benchmark. The baselines demonstrate consistent performance pat- terns, with Nullu uniformly improving the perceptual com- petencies of t...

  58. [66]

    The illustration of difference vectors and random vectors in the feature space

    Analysis about HalluSpace Figure 8. The illustration of difference vectors and random vectors in the feature space. This section provides a more comprehensive study about the question Does HalluSpace represent the hallucination Setting Model Method Accuracy Precision Recall F ...

  59. [67]

    {Response}

    LLaV A-Bench 13.1. Prompt for GPT-4V Aided Evaluation As we leverage LLaV A-Bench [30] to qualitatively evalu- ate the overall performance using GPT-4V Aided Evalua- tion5, in this section, we main describe the prompt used 5https://openai.com/research/gpt-4v-system-card for ev...

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.