Awesome-LVLM-Hallucination

Even though the world has seen the imersive capabilities of large vision language models, particularly in zero-shot inference, such models struggle with hallucinations, which can be referred to as the generation of text with information that is not present in the visual input. Lots of research work is going on to tackle this problem, such as hallucinated objects, inaccurate attributes and relationships, unfaithful descriptions, and so on. Possible reasons behind this could be language prior, insufficient visual context, biases and misinformation in the training dataset, and lot more.

This repository will provide an organized list of state-of-the-art research papers, relevant code, and a brief description related to hallucinations of the Large-Vision-Language Model (LVLM), also known as the Multimodal Large Language Model (MLLM).

The main intention of this project is to provide a platform where all the research work in the field of hallucination in LVLMs is accessed in a constructive way. If you have any suggestions for intersecting work within this field, kindly contribute them by raising an open issue. I am looking forward to fruitful discussion and learning!

---

• Awesome LVLM Hallucination
  • Hallucination Evaluation Benchmark
  • Hallucination Detection
  • Hallucination Mitigation
  • Survey

---

Evaluation-Benchmark

1. CHAIR: Object Hallucination in Image Captioning (EMNLP 2018)

   • Introduce problem of object hallucination on MSCOCO image captioning task
   • CHAIR metrics [built upon unique 80 MSCOCO dataset objects]

2. POPE: Evaluating Object Hallucination in Large Vision-Language Models (EMNLP 2023)

   • Object existence hallucination [Yes/No]
   • Random, Popular and Adversial settings on MSCOCO dataset

3. MME: A Comprehensive Evaluation Benchmark for Multimodal Large Language Models (23 June, 2023)

   • MME benchmark covers the evaluation of MLLM's perception and cognition abilities
   • Perception (Coase-Grained): 4; Perception (Fine-Grained): 5; Perception (OCR): 1; Cognition (Reasoning): 4; [Total 14 subtasks]
   • Answer in Yes/No format for easy evaluation & 30 advanced MLLMs are benchmarked

4. M-HalDetect: Detecting and Preventing Hallucinations in Large Vision Language Models (AAAI 2024)

   • Hallucination detection dataset with fine-grained annotations [accurate, inaccurate and analysis]
   • Fine-grained Direct Preference Optimization (FDPO) technique and reward model dataset
   • High correlation of reward model score with human evaluation

5. HaELM: Evaluation and Analysis of Hallucination in Large Vision-Language Models (29 August, 2023)

   • Discussed LVLMs tendency to response as 'Yes' to judgement type queries
   • Use of ChatGPT to collect hallucination data via iterative prompt modification
   • Open-source LLM trained over this dataset for evaluation of LVLM's response
   • Evaluation results on various LVLMs, Generation length and Top-K of sampling

6. CIEM: Contrastive Instruction Evaluation Method for Better Instruction Tuning (NeurIPS 2023 Workshop)

   • Automatic construction of question-answer pair with based on dataset with caption annotation using ChatGPT [Yes/No QA pair] and automatic pipeline for evaluation
   • Constractive instruction tuning (CIT) with Factual and Constractive QA pairs with Chain-of-Thought (CoT) justification

7. CAST: Cross-modal Alignment Similarity Test for Vision Language Models (17 September, 2024)

   • Proposed CAST as a way to measure the self-consistency of LVLMs across different modalities.
   • This works in two stage, in the first stage the models generate similarities/true statements comparing two inputs, and in the second stage the model judges its own output for truthfulness.

8. MMHAL-BENCH: Aligning Large Multimodal Models with Factually Augmented RLHF (25 September, 2023)

   • Introduced novel algorithm called Factually Augmented RLHF (Fact-RLHF) to alleviate the reward hacking phenomenon in RLHF
   • Developed evaluation benchmark MMHAL-BENCH with a special focus on penalizing hallucinations
   • Trained a LLM with RLHF (Llava-RLHF) which shows improved multimodal alignment

9. LRV (GAVIE): Mitigating Hallucination in Large Multi-Modal Models via Robust Instruction Tuning (29 September, 2023)

   • LRV-Instruction - positive and negative robust instruction tuning dataset with 400k visual instructions (16 tasks)
   • Negative instruction semantics: (a) Nonexistent Object Manipulation (b) Existent Object Manipulation (c) Knowledge Manipulation
   • GPT4-Assisted Visual Instruction Evaluation (GAVIE)

10. NOPE: Negative Object Presence Evaluation (NOPE) to Measure Object Hallucination in Vision-Language Models (09 October, 2023)

    • VQA diagnostic benchmark to measure object hallucination with use of 'Negative Pompt' based questions
    • LLM based generation of 29.5k synthetic negative pronoum (none, no one, nobody. nowhere, neither) dataset
    • Finding: tendency of VLMs to hallucinate more on data with higher lexical diversity, more scene relavent objects (co-occurance) and large answer copes.

11. HallusionBench: An Advanced Diagnostic Suite for Entangled Language Hallucination and Visual Illusion in Large Vision-Language Models (CVPR 2024)

    • Language Hallucination + Visual Illusion: 1129 VQA paired with total 346 images
    • It includes topics such as food, math, geometry, statistics, geography, sports, cartoon, famous illusions, movie, meme, etc. and formats such as including logo, poster, figure, charts, table, map, consecutive images, etc.

12. FAITHSCORE: Evaluating Hallucinations in Large Vision-Language Models (02 November, 2023)

    • Reference-free and fine-grained evaluation metric
    • 1. Recognizer : LLM is used for descriptive content identification of LVLM's prediction
    • 2. Decomposer : LLM is used to generate atomic facts based on recognizer's output
    • 3. Verifier : Visual Entailment Model (e.g. OFA) is used to verify atomic facts with input image

13. Bingo: Holistic Analysis of Hallucination in GPT-4V(ision): Bias and Interference Challenges (07 November, 2023)

    • Total 308 Images and 370 QA Pairs
    • Bias category: Region, OCR and Factual
    • Interferance catogary: Image-to-Image and Text-to-Image

14. AMBER: An LLM-free Multi-dimensional Benchmark for MLLMs Hallucination Evaluation (13 November, 2023)

    • LLM free evaluation of hallucination using AMBER benchmark
    • Evaluation of hallucination for generative and discriminative task using AMBERSCORE metric (covers existence, attributes and relation types of hallucination)
    • Includes hallucinatory target objects (more likely to be imagined by LVLMs)

15. RAH-Bench: Mitigating Hallucination in Visual Language Models with Visual Supervision (27 Novemebr, 2023)

    • Introduce fine-grained vision instruction dataset named RAI-30K (built upon panoptic scene graph dataset (PSG))
    • RAH-BENCH vision hallucination evaluation benchmark (3 types: Categorial, Relation and Attribute Hallucination)
    • False Positive Rates as evaluation metric

16. Behind the Magic, MERLIM: Multi-modal Evaluation Benchmark for Large Image-Language Models (03 Decemeber, 2023)

    • Proposed a novel test-bed to evaluate IT-LVLMs (Instruction Tuning Large Vision and Language models) on core computer vision tasks
    • Observed poor performance of IT-LVLMs with multiple failure cases in visual grounding
    • Identify problems with IT-LVLMSs like generation of hallucinatory events and sensitivity to the input query

17. CCEval: HallE-Switch: Controlling Object Hallucination in Large Vision Language Models (03 Decemebr, 2023)

    • Suggest an approach to control object existence hallucination in detailed captions of LVLM
    • Introduced CCEval which is a GPT-4 assisted evaluation method for detailed captioning (Metrics: CHAIR(i&s), Coverage, Average Length, Average Objects)
    • Detailed investigation on LVLM's component that might imfluence hallucination such as alignment of language decoder, volume of instruction data, resolution of input image and so on
    • Introduced a controlling parameters over LLMs (HallE-Control) to condition the inference of objects

18. FGHE: Mitigating Fine-Grained Hallucination by Fine-Tuning Large Vision-Language Models with Caption Rewrites (04 December, 2023)

    • Dealing with fine-grained object hallucination with ReCaption framework
    • Two stage frame work : 1) Caption generation with help of ChatGPT 2) Finetuning LVLMs on generated captions
    • Inroduced Fine-Grained Object Hallucination Evaluation (FGHE) which similar to POPE. (manually annotted 50 images with 200 binary questions with type multi-object, attributes and behaviour)

19. OpenCHAIR: Mitigating Open-Vocabulary Caption Hallucinations (06 Decemeber, 2023)

    • soon

20. CorrelationQA: The Instinctive Bias: Spurious Images lead to Hallucination in MLLMs (06 February, 2024)

    • soon

21. ViGoR: Improving Visual Grounding of Large Vision Language Models with Fine-Grained Reward Modeling (09 February, 2024)

    • soon

22. VQAv2-IDK: Visually Dehallucinative Instruction Generation: Know What You Don’t Know (15 February, 2024)

    • soon

23. MHaluBench: Unified Hallucination Detection for Multimodal Large Language Models (20 February, 2024)

    • soon

24. MAD-Bench: How Easy is It to Fool Your Multimodal LLMs? An Empirical Analysis on Deceptive Prompts (20 February, 2024)

    • soon

25. VHTest: Visual Hallucinations of Multi-modal Large Language Models (22 February, 2024)

    • soon

26. Hal-Eval: A Universal and Fine-grained Hallucination Evaluation Framework for Large Vision Language Models (24 February, 2024)

    • soon

27. Evaluating and Mitigating Number Hallucinations in Large Vision-Language Models: A Consistency Perspective (03 March, 2024)

    • soon

28. ** EvalDial**: Mitigating Dialogue Hallucination for Large Multi-modal Models via Adversarial Instruction Tuning (15 March, 2024)

    • soon

29. IVL-Hallu: PhD: A Prompted Visual Hallucination Evaluation Dataset (17 March, 2024)

    • soon

30. Unsolvable Problem Detection: Evaluating Trustworthiness of Vision Language Models (29 March, 2024)

    • soon

31. ALOHa: A New Measure for Hallucination in Captioning Models (3 April, 2024)

    • soon

32. VALOR-EVAL: Holistic Coverage and Faithfulness Evaluation of Large Vision-Language Models (22 April, 2024)

    • soon

33. THRONE: An Object-based Hallucination Benchmark for the Free-form Generations of Large Vision-Language Models (08 May, 2024)

    • soon

34. MRHal-Bench: Automated Multi-level Preference for MLLMs (18 May, 2024)

    • soon

35. VLind-Bench: Measuring Language Priors in Large Vision-Language Models (13 June, 2024)

    • soon

36. MMRel: A Relation Understanding Dataset and Benchmark in the MLLM Era (13 June, 2024)

    • soon

37. Med-HallMark: Detecting and Evaluating Medical Hallucinations in Large Vision Language Models (14 June, 2024)

    • Medical field hallucination benchmark
    • MediHall Score - evaluation metric

38. AUTOHALLUSION: Automatic Generation of Hallucination Benchmarks for Vision-Language Models (16 June, 2024)

    • soon

39. MFC-Bench: Benchmarking Multimodal Fact-Checking with Large Vision-Language Models (17 June, 2024)

    • soon

40. CHAIR-MEN: Does Object Grounding Really Reduce Hallucination of Large Vision-Language Models? (20 June, 2024)

    • soon

41. R-BENCH: Evaluating and Analyzing Relationship Hallucinations in Large Vision-Language Models (24 June, 2024) (ICML2024)

    • Introduce an evaluation benchmark to tackle relation type of hallucination
    • soon

42. HQH: Evaluating the Quality of Hallucination Benchmarks for Large Vision-Language Models (24 June, 2024)

    • Propose a framework called Hallucination benchmark Quality Measurement (HQM) to assess the quality of existing hallucination benchmarks
    • soon

43. VideoHallucer: Evaluating Intrinsic and Extrinsic Hallucinations in Large Video-Language Models (24 June, 2024)

    • soon

44. MMHalSnowball: Investigating and Mitigating the Multimodal Hallucination Snowballing in Large Vision-Language Models (30 June, 2024)

    • soon

45. MedVH: Towards Systematic Evaluation of Hallucination for Large Vision Language Models in the Medical Context (03 July, 2024)

    • soon

46. ROPE: Multi-Object Hallucination in Vision-Language Models (08 July, 2024)

    • Deals with multi-object hallucinations and their cause
    • Introduce Recognition-based Ob�ject Probing Evaluation (ROPE) for assessing multi-object hallucination
    • In-depth analysis of hallucinatory behaviors

47. BEAF: Observing BEfore-AFter Changes to Evaluate Hallucination in Vision-language Models (18 July, 2024) (ECCV 2024)

    • Proposed a hallucination evaluation benchmark called BEfore-After (BEAF)
    • New metrics introduced: True Understanding (TU), IGnorance (IG), StuBbornness (SB), and InDecision (ID)

48. HaloQuest: A Visual Hallucination Dataset for Advancing Multimodal Reasoning (22 July, 2024) (ECCV 2024)

    • Introduced a novel VQA dataset for VLM evaluation
    • soon

49. MMINSTRUCT: A High-Quality Multi-Modal Instruction Tuning Dataset with Extensive Diversity (22 July, 2024)

    • Introduced high-quality and diverse visual instruction tuning dataset
    • Claims SOTA performance of MMINSTRUCT finetuned LLava-1.5 on 10 out of 12 famous benchmarks

50. Hallu-PI: Evaluating Hallucination in Multi-modal Large Language Models within Perturbed Inputs (02 August, 2024)

    • Constructed hallucination evaluation benchmark with perturbed inputs with 7 different purturbed scenarios
    • 12 SOTA MLLMs are benchmarked

51. Reefknot: A Comprehensive Benchmark for Relation Hallucination Evaluation, Analysis and Mitigation in Multimodal Large Language Models (18th August, 2024)

    • Introduced a benchmark to evaluate relation hallucination which further catogarized in to Perceptive and Cognitice type
    • 3 evaluation tasks: Yes/No, MCQ, VQA
    • code and dataset will be released after paper's acceptance

52. Pfram: Understanding Multimodal Hallucination with Parameter-Free Representation Alignment (02 September, 2024)

    • soon

53. ODE: Open-Set Evaluation of Hallucinations in Multimodal Large Language Models (14 September, 2024)

    • soon

54. LLSAVisionQA: Explore the Hallucination on Low-level Perception for MLLMs (15 September, 2024)

    • soon

55. CAST: Cross-modal Alignment Similarity Test for Vision Language Models (17 September, 2024)

    • soon

56. JourneyBench: Challenging One-Stop Vision-Language Understanding Benchmark of Generated Images (25 September, 2024)

    • soon

57. FIHA: Autonomous Hallucination Evaluation in Vision-Language Models with Davidson Scene Graphs (20 September, 2024)

    • code: here
    • soon

58. EventHallusion: Diagnosing Event Hallucinations in Video LLMs (25 September, 2024)

    • soon

59. TUBench: Benchmarking Large Vision-Language Models on Trustworthiness with Unanswerable Questions (05 October, 2024)

    • soon

60. LongHalQA: Long-Context Hallucination Evaluation for MultiModal Large Language Models (15 October, 2024)

    • soon

61. MM-SY: Have the VLMs Lost Confidence? A Study of Sycophancy in VLMs (15 October, 2024)

    • soon (code and benchmark)

62. Magnifier Prompt: Tackling Multimodal Hallucination via Extremely Simple Instructions (15 October, 2024)

    • soon

63. DeCo: MLLM can see? Dynamic Correction Decoding for Hallucination Mitigation (15 October, 2024)

    • decoding technique
    • soon

64. The Curse of Multi-Modalities: Evaluating Hallucinations of Large Multimodal Models across Language, Visual, and Audio (16 October, 2024)

    • soon

65. Trust but Verify: Programmatic VLM Evaluation in the Wild (17 October, 2024)

    • project_page
    • soon

66. Tri-HE: Unified Triplet-Level Hallucination Evaluation for Large Vision-Language Models (03 November, 2024)

    • soon

67. H-POPE: Hierarchical Polling-based Probing Evaluation of Hallucinations in Large Vision-Language Models (06 November, 2024)

    • soon

68. VIDHAL: Benchmarking Temporal Hallucinations in Vision LLMs (25 November 2024)

    • perfromance evaluation on video / frames
    • soon

69. Up to Date (25th December, 2024) and SOTA research work loading...

Note: 'soon' will be replaced with brief description!

Detection

1. FPDO - Reward Model: Detecting and Preventing Hallucinations in Large Vision Language Models (AAAI 2024)
   • M-HalDetect - Hallucination detection dataset with fine-grained annotations [accurate, inaccurate and analysis]
   • Fine-grained Direct Preference Optimization (FDPO) technique and reward model trained on introduced dataset
   • High correlation of reward model score with human evaluation
2. HaELM: Evaluation and Analysis of Hallucination in Large Vision-Language Models (29 August, 2023)
   • Discussed LVLMs tendency to response as 'Yes' to judgement type queries
   • Use of ChatGPT to collect hallucination data via iterative prompt modification
   • Open-source LLM trained over this dataset for evaluation of LVLM's response
   • Evaluation results on various LVLMs, Generation length and Top-K of sampling
3. HallE-Switch: Controlling Object Hallucination in Large Vision Language Models (3 October, 2023)
   • Suggest an approach to control object existence hallucination in detailed captions of LVLM
   • Introduced CCEval which is a GPT-4 assisted evaluation method for detailed captioning (Metrics: CHAIR(i&s), Coverage, Average Length, Average Objects)
   • Detailed investigation on LVLM's component that might imfluence hallucination such as alignment of language decoder, volume of instruction data, resolution of input image and so on
   • Introduced a controlling parameters over LLMs (HallE-Control) to condition the inference of objects
4. HalluciDoctor: Mitigating Hallucinatory Toxicity in Visual Instruction Data (22 November, 2023)
   • Investigates hallucination toxicity in already existing visual instruction dataset
   • Proposed HalluciDoctor method for automatic elimination of such toxicity
   • Generation of more counterfactual instruction data with help of HalluciDoctor to improve LVLMs' resistance to hallucination
5. LogicCheckGPT: Logical Closed Loop: Uncovering Object Hallucinations in Large Vision-Language Models (18 february, 2024)
   • Postprocessing output description of LVLMs
   • 5 steps logical loop procedure such as
     • Object extraction, Object-to-Attribute inquiring, Attribute-to-Object inquiring, Logic closed llop check and Hallucination detection and mitigation
   • Experimental analysis on POPE and MME benchmark
6. UNIHD: Unified Hallucination Detection for Multimodal Large Language Models (20 February, 2024)
   • Introduce a meta evaluation benchmark called MHALUBENCH
   • Introduce a framework named UNIHD which detect modality-conflicting hallucinations at various levels such as object, attribute, and scene-text, as well as fact-conflicting hallucinations
7. Detecting and Mitigating Hallucination in Large Vision Language Models via Fine-Grained AI Feedback (22 April, 2024)
   • Use of GPT-4/GPT-4v to generate fine-grained feedback for hallucination detection and detection (by supervised finetuning (SFT) of LVLM)
   • Propose automatic pipeline for preference dataset construction
   • Hallucination Severity Aware Direct Prefential Optimization (HSA-DPO) is introduced for mitigation of LVLM's hallucination
8. MetaToken: Detecting Hallucination in Image Descriptions by Meta Classification (29 May, 2024)
   • Really cool approach
   • Lightweight method for hallucination detection
9. Image Textualization: An Automatic Framework for Creating Accurate and Detailed Image Descriptions (11 June, 2024)
   • soon
10. MediHallDetector: Detecting and Evaluating Medical Hallucinations in Large Vision Language Models (14 June, 2024)
    • Medical field hallucination detection
    • soon
11. Pelican: Correcting Hallucination in Vision-LLMs via Claim Decomposition and Program of Thought Verification (02 July, 2024)
    • soon
12. SUQ: Reference-free Hallucination Detection for Large Vision-Language Models (11 August, 2024)
    • Concluded that Supervised Uncertainity Quantification (SUQ) outperforms other reference-free hallucination detection technique such as Uncertainity-based methods and Consistency-based methods
    • An example of supervised Uncertainity Quantification method --> METATOKEN paper
    • soon
13. Pre-Training Multimodal Hallucination Detectors with Corrupted Grounding Data (30 August, 2024)
    • Proposed an approach to create corrupted grounding data which can be used to pre-train MLM hallucination detector
    • soon
14. LLMs Can Check Their Own Results to Mitigate Hallucinations in Traffic Understanding Tasks (19 September, 2024)
    • soon
15. TLDR: Token-Level Detective Reward Model for Large Vision Language Models (07 October, 2024)
    • soon
16. RadFlag: A Black-Box Hallucination Detection Method for Medical Vision Language Models (01 November, 2024)
    • soon
17. VL-Uncertainty: Detecting Hallucination in Large Vision-Language Model via Uncertainty Estimation (18 November, 2024)
    • soon
18. DHCP: Detecting Hallucinations by Cross-modal Attention Pattern in Large Vision-Language Models (27 November, 2024)
    • soon
19. Beyond Logit Lens: Contextual Embeddings for Robust Hallucination Detection & Grounding in VLMs (28 November, 2024)
    • soon
20. Up to Date (25th December, 2024) and SOTA research work loading...

Note: 'soon' will be replaced with brief description!

Mitigation

1. ObjMLM: Plausible May Not Be Faithful: Probing Object Hallucination in Vision-Language Pre-training (10 February 2023)

   • Deals with object hallucination problem of VLMs
   • Discuss the influence of various Vision Language Pretraining (VLP) objective (ITM, ITC and ICLM) and Image encoding methods (region-based, grid-based, and patch-based) on object hallucination
   • Introduce novel VLP objective ObjMLM to mitigate object hallucination

2. MMCoT: Multimodal Chain-of-Thought Reasoning in Language Models (17 February 2023)

   • Two stage framework by finetuning language models to perform Multimodal chain-of-thoughts (CoT) which incorporates language (text) and vision (images) modalities
   • Claims state-of-the-art performance of model under 1 billion parameters on ScienceQA benchmark
   • Multimodal-CoT has the merits of mitigating hallucination and enhancing convergence speed

3. LRV-GAVIE: Mitigating Hallucination in Large Multi-Modal Models via Robust Instruction Tuning (26 June, 2023)

   • LRV-Instruction - positive and negative robust instruction tuning dataset with 400k visual instructions (16 tasks)
   • Negative instruction semantics: (a) Nonexistent Object Manipulation (b) Existent Object Manipulation (c) Knowledge Manipulation
   • GPT4-Assisted Visual Instruction Evaluation (GAVIE)

4. LLaVA-RLHF: Aligning Large Multimodal Models with Factually Augmented RLHF (25 September, 2023)

   • Introduced novel algorithm called Factually Augmented RLHF (Fact-RLHF) to alleviate the reward hacking phenomenon in RLHF
   • Developed evaluation benchmark MMHAL-BENCH with a special focus on penalizing hallucinations
   • Trained a LLM with RLHF (Llava-RLHF) which shows improved multimodal alignment

5. LURE: Analyzing and Mitigating Object Hallucination in Large Vision-Language Models (01 October, 2023)

   • Introduced LURE framework which is lightweight and compatible post-hoc approach for rectifying object hallucination in LVLMs
   • Statstical analysis of Co-occurence of objects, object uncertainity and object position in generated description which might correlate with object hallucination
   • Uncertain objects are put as placeholder with tokens while training LURE and while infernece (for revision)
   • Really popular method

6. HallE-Switch: Controlling Object Hallucination in Large Vision Language Models (3 October, 2023)

   • Suggest an approach to control object existence hallucination in detailed captions of LVLM
   • Introduced CCEval which is a GPT-4 assisted evaluation method for detailed captioning (Metrics: CHAIR(i&s), Coverage, Average Length, Average Objects)
   • Detailed investigation on LVLM's component that might imfluence hallucination such as alignment of language decoder, volume of instruction data, resolution of input image and so on
   • Introduced a controlling parameters over LLMs (HallE-Control) to condition the inference of objects

7. Woodpecker: Hallucination Correction for Multimodal Large Language Models (24 October, 2023)

   • Really popular method
   • Training free, post-hoc method to mitigate hallucination (but computationally expensive!!)
   • 5 steps framework:
     1. Key concept extraction from LVLM's output
     2. Formulation of questions based on key concepts
     3. Visual Knowledge validation (use of open-source object detector + pretrained VQA model)
     4. Visual claim generation (use of fix sentence templates + QA to claim model)
     5. Hallucination Correction (use LLM to correct LVLM's response)

8. VOLCANO: Mitigating Multimodal Hallucination through Self-Feedback Guided Revision (14 November, 2023)

   • soon

9. HalluciDoctor: Mitigating Hallucinatory Toxicity in Visual Instruction Data (22 November, 2023)

   • Investigates hallucination toxicity in already existing visual instruction dataset
   • Proposed HalluciDoctor method for automatic elimination of such toxicity
   • Generation of more counterfactual instruction data with help of HalluciDoctor to improve LVLMs' resistance to hallucination

10. RAH-Bench: Mitigating Hallucination in Visual Language Models with Visual Supervision (27 Novemebr, 2023)

    • soon

11. HA-DPO: Beyond Hallucinations: Enhancing LVLMs through Hallucination-Aware Direct Preference Optimization (28 November, 2023)

    • soon

12. VCD: Mitigating Object Hallucinations in Large Vision-Language Models through Visual Contrastive Decoding (28 November, 2023)

    • Decoding strategy
    • soon

13. OPERA: Alleviating Hallucination in Multi-Modal Large Language Models via Over-Trust Penalty and Retrospection-Allocation (CVPR 2024)

    • soon

14. FGHE: Mitigating Fine-Grained Hallucination by Fine-Tuning Large Vision-Language Models with Caption Rewrites (04 December, 2023)

    • soon

15. RLHF-V: Towards Trustworthy MLLMs via Behavior Alignment from Fine-grained Correctional Human Feedback (01 December, 2023)

    • fine-grained refined DPO!
    • soon

16. MOCHa: Mitigating Open-Vocabulary Caption Hallucinations (06 December 2023)

    • soon

17. HACL: Hallucination Augmented Contrastive Learning for Multimodal Large Language Model (12 December 2023)

    • soon

18. SILKIE: Preference Distillation for Large Visual Language Models (17 December, 2023)

    • propose VLFeedback dataset for DPO
    • soon

19. KAM-CoT: Knowledge Augmented Multimodal Chain-of-Thoughts Reasoning (23 January, 2024)

    • soon

20. Enhancing Multimodal Large Language Models with Vision Detection Models: An Empirical Study (31 January, 2024)

    • soon

21. ViGoR: Improving Visual Grounding of Large Vision Language Models with Fine-Grained Reward Modeling (09 February, 2024)

    • soon

22. SKIP \N: A Simple Method to Reduce Hallucination in Large Vision-Language Models (12 February, 2024)

    • soon

23. MARINE: Mitigating Object Hallucination in Large Vision-Language Models via Classifier-Free Guidance (13 February, 2024)

    • soon

24. IDK-Instructions: Visually Dehallucinative Instruction Generation: Know What You Don’t Know (15 February, 2024)

    • soon

25. EFUF: Efficient Fine-grained Unlearning Framework for Mitigating Hallucinations in Multimodal Large Language Models (15 February, 2024)

    • soon

26. LogicCheckGPT: Logical Closed Loop: Uncovering Object Hallucinations in Large Vision-Language Models (18 february, 2024)

    • soon

27. POVID: Aligning Modalities in Vision Large Language Models via Preference Fine-tuning (18 february, 2024)

    • soon

28. Less is More: Mitigating Multimodal Hallucination from an EOS Decision Perspective (22 February, 2024)

    • decoding strategy
    • soon

29. CGD: Seeing is Believing: Mitigating Hallucination in Large Vision-Language Models via CLIP-Guided Decoding (23 February, 2024)

    • decoding strategy
    • soon

30. IBD: Alleviating Hallucinations in Large Vision-Language Models via Image-Biased Decoding (28 February, 2024)

    • decoding strategy
    • soon

31. HALC: Object Hallucination Reduction via Adaptive Focal-Contrast Decoding (01 March, 2024)

    • Decodig strategy to tackle object hallucination
    • soon

32. Evaluating and Mitigating Number Hallucinations in Large Vision-Language Models: A Consistency Perspective (03 March, 2024)

    • number hallucination
    • soon

33. AIT: Mitigating Dialogue Hallucination for Large Multi-modal Models via Adversarial Instruction Tuning (15 March, 2024)

    • soon

34. DVP: What if...?: Counterfactual Inception to Mitigate Hallucination Effects in Large Multimodal Models (20 March, 2024)

    • soon

35. M3ID: Multi-Modal Hallucination Control by Visual Information Grounding (20 March, 2024)

    • decoding strategy
    • soon

36. PENSIEVE: Retrospect-then-Compare Mitigates Visual Hallucination (21 March, 2024)

    • decoding strategy
    • soon

37. ESREAL: Exploiting Semantic Reconstruction to Mitigate Hallucinations in Vision-Language Models (26 March, 2024)

    • soon

38. ICD: Mitigating Hallucinations in Large Vision-Language Models with Instruction Contrastive Decoding (27 March, 2024)

    • decoding strategy
    • soon

39. FGAIF: Aligning Large Vision-Language Models with Fine-grained AI Feedback (07 April, 2024)

    • soon

40. Prescribing the Right Remedy: Mitigating Hallucinations in Large Vision-Language Models via Targeted Instruction Tuning (16 April, 2024)

    • soon

41. FACT: Teaching MLLMs with Faithful, Concise and Transferable Rationales (17 April, 2024)

    • soon

42. TVP: Exploring the Transferability of Visual Prompting for Multimodal Large Language Models (17 April, 2024)

    • soon

43. TextSquare: Scaling up Text-Centric Visual Instruction Tuning (19 April, 2024)

    • soon

44. HSA-DPO: Detecting and Mitigating Hallucination in Large Vision Language Models via Fine-Grained AI Feedback (22 April, 2024)

    • Use of GPT-4/GPT-4v to generate fine-grained feedback for hallucination detection and detection (by supervised finetuning (SFT) of LVLM)
    • Propose automatic pipeline for preference dataset construction
    • Hallucination Severity Aware Direct Prefential Optimization (HSA-DPO) is introduced for mitigation of LVLM's hallucination

45. Visual Fact Checker: Enabling High-Fidelity Detailed Caption Generation (30 April - CVPR 2024)

    • soon

46. CSR: Calibrated Self-Rewarding Vision Language Models (23 May, 2024)

    • soon

47. HIO: Alleviating Hallucinations in Large Vision-Language Models through Hallucination-Induced Optimization (24 May, 2024)

    • soon

48. VDGD: Mitigating LVLM Hallucinations in Cognitive Prompts by Bridging the Visual Perception Gap (24 May, 2024)

    • soon

49. RLAIF-V: Aligning MLLMs through Open-Source AI Feedback for Super GPT-4V Trustworthines (27 May, 2024)

    • soon

50. AvisC: Don’t Miss the Forest for the Trees: Attentional Vision Calibration for Large Vision Language Models (28 May, 2024)

    • decoding strategy

51. RITUAL: Random Image Transformations as a Universal Anti-hallucination Lever in LVLMs (28 May, 2024)

    • soon

52. HALVA: Mitigating Object Hallucination via Data Augmented Contrastive Tuning (28 May, 2024)

    • decoding strategy
    • will publish code soon

53. NoiseBoost: Alleviating Hallucination with Noise Perturbation for Multimodal Large Language Models (30 May, 2024)

    • soon

54. CODE: Contrasting Self-generated Description to Combat Hallucination in Large Multi-modal Model (04 June, 2024)

    • soon

55. mDPO: Conditional Preference Optimization for Multimodal Large Language Models (17 June, 2024)

    • soon

56. DBD: Do More Details Always Introduce More Hallucinations in LVLM-based Image Captioning? (18 June, 2024)

    • Introduce novel decoding technique called Differentiated Beam Decoding (DBD)
    • soon

57. AGLA: Mitigating Object Hallucinations in Large Vision-Language Models with Assembly of Global and Local Attention (18 June, 2024)

    • Introduce AGLA, a training-free and plug-and-play decoding framework
    • soon

58. Residual Visual Decoding: Investigating and Mitigating the Multimodal Hallucination Snowballing in Large Vision-Language Models (30 June, 2024)

    • decoding method
    • Soon

59. BDHS: UNDERSTANDING ALIGNMENT IN MULTIMODAL LLMS: A COMPREHENSIVE STUDY (02 July, 2024)

    • soon

60. REVERIE: Reflective Instruction Tuning: Mitigating Hallucinations in Large Vision-Language Models (16 July, 2024) (ECCV 2024)

    • Introduced novel reflective instruction tuning to incorporate rationales into visual instruction tuning
    • Proposed large-scale instruction tuning dataset called REVERIE

61. VACoDe: Visual Augmented Contrastive Decoding (26 July, 2024)

    • decoding strategy using various visual augmentation
    • analysed effect of various visual augmentation on LVLMs performance and introduced an algorithm to select the most suitable augmentation for constractive decoding for input image
    • soon

62. PAI: Paying More Attention to Image: A Training-Free Method for Alleviating Hallucination in LVLMs (31 July, 2024) (ECCV 2024)

    • soon

63. MHR: Mitigating Multilingual Hallucination in Large Vision-Language Models (01 August, 2024)

    • soon

64. ARA: Alleviating Hallucination in Large Vision-Language Models with Active Retrieval Augmentation (01 August, 2024)

    • RAG for LVLMs for mitigating hallucination
    • soon

65. SID: Self-Introspective Decoding: Alleviating Hallucinations for Large Vision-Language Models (04 August, 2024)

    • Decoding strategy
    • Rethink constractuve decoding (CD) methods in LVLMs for hallucination mitigation
    • soon

66. LCD: Mitigating Hallucinations in Large Vision-Language Models (LVLMs) via Language-Contrastive Decoding (06 August, 2024)

    • decoding strategy to mitigate object hallucination
    • soon

67. Detect-then-Calibrate: A Comprehensive Benchmark for Relation Hallucination Evaluation, Analysis and Mitigation in Multimodal Large Language Models (18 August, 2024)

    • Proposed a novel detect-then-calibrate method to detect and mitigate hallucination
    • throshold based hallucination identification
    • hallucination rate as metric to calculate final metric called R_score

68. CLIP-DPO: Vision-Language Models as a Source of Preference for Fixing Hallucinations in LVLMs (19 August, 2024)

    • Do not require additiona training or external dataset or esemble of external LVLMs such as GPT-4
    • Use of CLIP model to prepare positive-negative pairs for DPO
    • Claims far better performance then similar work - HA-DPO with very few training data samples

69. LQCD: Towards Analyzing and Mitigating Sycophancy in Large Vision-Language Models (21 August, 2024)

    • Deals with Sycophancy in LVLMs which exists due to negative prompting
    • Introduce decoding strategy for improving LVLM's robustness toward sycophancy

70. RoVRM: A Robust Visual Reward Model Optimized via Auxiliary Textual Preference Data (22 August, 2024)

    • Introduced Robust Visula Reward model (RoVRM) to improve human-preference alignment in LVLMs
    • 3 stage progressive training and optimal transport-based preference data selection approaches to train RoVRM
    • Seemless integration with arbitrary ranking-based alignment techniques, such as direct preference optimization (DPO)

71. ConVis: Contrastive Decoding with Hallucination Visualization for Mitigating Hallucinations in Multimodal Large Language Models (25 August, 2024)

    • constractive decoding method
    • use of text-to-image (T2I) model for constractive decoding and mitigate hallucination
    • Claimed that experimental investigation on 5 benchmarks showing superior performance compared to existing techniques for hallucination mitigation

72. See or Guess: Counterfactually Regularized Image Captioning (29 August, 2024)

    • soon

73. Look, Compare, Decide: Alleviating Hallucination in Large Vision-Language Models via Multi-View Multi-Path Reasoning (30 August, 2024)

    • multi-path certainity based decoding
    • soon

74. FaithD2T Generating Faithful and Salient Text from Multimodal Data (06 September, 2024)

    • soon

75. RBD: Mitigating Hallucination in Visual-Language Models via Re-Balancing Contrastive Decoding (10 September, 2024)

    • Decoding strategy
    • soon

76. PACU: Effectively Enhancing Vision Language Large Models by Prompt Augmentation and Caption Utilization (22 September, 2024)

    • soon

77. Dentist: A Unified Hallucination Mitigation Framework for Large Vision-Language Models (24 September, 2024)

    • soon

78. TCD: Diagnosing Event Hallucinations in Video LLMs (25 September, 2024)

    • soon

79. HELPD: Mitigating Hallucination of LVLMs by Hierarchical Feedback Learning with Vision-enhanced Penalty Decoding (30 September, 2024)

    • extension of OPERA paper with vision enhanced penalty decoding
    • soon

80. PROJECTAWAY: Interpreting and Editing Vision-Language Representations to Mitigate Hallucinations (03 October, 2024)

    • soon

81. OHD-Caps: Investigating and Mitigating Object Hallucinations in Pretrained Vision-Language (CLIP) Models (04 October, 2024)

    • soon

82. LOOK TWICE BEFORE YOU ANSWER: Memory-Space Visual Retracing for Hallucination Mitigation in Multimodal Large Language Models (04 October, 2024)

    • soon

83. DAMRO: Dive into the Attention Mechanism of LVLM to Reduce Object Hallucination (06 October, 2024)

    • decoding strategy
    • soon

84. CAUSALMM: Mitigating Modality Prior-Induced Hallucinations in Multimodal Large Language Models via Deciphering Attention Causality (07 October, 2024)

    • soon

85. FROM PIXELS TO TOKENS: Revisiting Object Hallucinations in Large Vision-Language Models (09 October, 2024)

    • soon

86. VHExpansion: Automatically Generating Visual Hallucination Test Cases for Multimodal Large Language Models (15 October, 2024)

    • soon

87. SGD: Mitigating Hallucinations in Large Vision-Language Models via Summary-Guided Decoding (17 October, 2024)

    • decoding technique
    • soon

88. Fine-Grained Verifiers: Preference Modeling as Next-token Prediction in Vision-Language Alignment (18 October, 2024)

    • soon

89. MFPO: Modality-Fair Preference Optimization for Trustworthy MLLM Alignment (20 October, 2024)

    • soon (code)

90. CCA: Mitigating Object Hallucination via Concentric Causal Attention (21 October, 2024)

    • soon

91. VTI: Reducing Hallucinations in Vision-Language Models via Latent Space Steering (22 October, 2024)

    • soon

92. V-DPO: Mitigating Hallucination in Large Vision Language Models viaVision-Guided Direct Preference Optimization (05 November, 2024)

    • soon

93. EAH: Seeing Clearly by Layer Two: Enhancing Attention Heads to Alleviate Hallucination in LVLMs (15 November, 2024)

    • soon

94. HDPO: Mitigating Hallucination in Multimodal Large Language Model via Hallucination-targeted Direct Preference Optimization (15 November, 2024)

    • soon

95. Thinking Before Looking: Improving Multimodal LLM Reasoning via Mitigating Visual Hallucination (15 November, 2024)

    • soon

96. CATCH: Complementary Adaptive Token-level Contrastive Decoding to Mitigate Hallucinations in LVLMs (19 November, 2024)

    • soon

97. Looking Beyond Text: Reducing Language bias in Large Vision-Language Models via Multimodal Dual-Attention and Soft-Image Guidance (21 November, 2024)

    • project page
    • soon

98. ICT: Image-Object Cross-Level Trusted Intervention for Mitigating Object Hallucination in Large Vision-Language Models (22 November, 2024)

    • code will be released soon
    • soon

99. VaLiD: Mitigating the Hallucination of Large Vision Language Models by Visual Layer Fusion Contrastive Decoding (24 November, 2024)

    • soon

100. Devils in Middle Layers of Large Vision-Language Models: Interpreting, Detecting and Mitigating Object Hallucinations via Attention Lens (23 November, 2024)

     • soon

101. TPO: A Topic-level Self-Correctional Approach to Mitigate Hallucinations in MLLMs (26 November, 2024)

     • soon

102. WhoBrings the Frisbee: Probing Hidden Hallucination Factors in Large Vision-Language Model via Causality Analysis (03 December, 2024)

     • soon

103. VisVM: Scaling Inference-Time Search with Vision Value Model for Improved Visual Comprehension (06 December, 2024)

     • soon

104. Verb Mirage: Unveiling and Assessing Verb Concept Hallucinations in Multimodal Large Language Models (06 December, 2024)

     • code will be published soon
     • soon

105. From Uncertainty to Trust: Enhancing Reliability in Vision-Language Models with Uncertainty-Guided Dropout Decoding (09 December, 2024)

     • soon

106. VCD Analysis: Delve into Visual Contrastive Decoding for Hallucination Mitigation of Large Vision-Language Models (09 December, 2024)

     • soon

107. DEHALL: Combating Multimodal LLM Hallucination via Bottom-Up Holistic Reasoning (15th December, 2024)

     • soon

108. Nullu: Mitigating Object Hallucinations in Large Vision-Language Models via HalluSpace Projection (18th December, 2024)

     • soon

109. VHD: Cracking the Code of Hallucination in LVLMs with Vision-aware Head Divergence (18th December, 2024)

     • soon

110. TPO: Token Preference Optimization with Self-Calibrated Visual-Anchored Rewards for Hallucination Mitigation (19th December, 2024)

     • soon

111. Toward Robust Hyper-Detailed Image Captioning: A Multiagent Approach and Dual Evaluation Metrics for Factuality and Coverage (20th December, 2024)

     • soon

112. VORD: Visual Ordinal Calibration for Mitigating Object Hallucinations in Large Vision-Language Models (20th December, 2024)

     • soon

113. SENA: Beyond Human Data: Aligning Multimodal Large Language Models by Iterative Self-Evolution (20th December, 2024)

     • soon

114. Up to Date (25th December, 2024) and SOTA research work loading...

Note: 'soon' will be replaced with brief description!

Survey

1. DEEP LEARNING APPROACHES ON IMAGE CAPTIONING: A REVIEW (22 August, 2023)
2. A Survey on Hallucination in Large Vision-Language Models (1 February, 2024)
3. Visual Hallucination: Definition, Quantification, and Prescriptive Remediations (26 March, 2024)
4. Hallucination of Multimodal Large Language Models: A Survey (29 April, 2024)
5. Unveiling Hallucination in Text, Image, Video, and Audio Foundation Models: A Comprehensive Survey (20 May, 2024)
6. Up to Date (25th December, 2024) and SOTA research work loading...