2000 character limit reached
Why do LLaVA Vision-Language Models Reply to Images in English?
Published 2 Jul 2024 in cs.CL and cs.CV | (2407.02333v1)
Abstract: We uncover a surprising multilingual bias occurring in a popular class of multimodal vision-LLMs (VLMs). Including an image in the query to a LLaVA-style VLM significantly increases the likelihood of the model returning an English response, regardless of the language of the query. This paper investigates the causes of this loss with a two-pronged approach that combines extensive ablation of the design space with a mechanistic analysis of the models' internal representations of image and text inputs. Both approaches indicate that the issue stems in the language modelling component of the LLaVA model. Statistically, we find that switching the language backbone for a bilingual LLM has the strongest effect on reducing this error. Mechanistically, we provide compelling evidence that visual inputs are not mapped to a similar space as text ones, and that intervening on intermediary attention layers can reduce this bias. Our findings provide important insights to researchers and engineers seeking to understand the crossover between multimodal and multilingual spaces, and contribute to the goal of developing capable and inclusive VLMs for non-English contexts.
- Yi: Open foundation models by 01.ai. Preprint, arXiv:2403.04652.
- Michael Andersland. 2024. Amharic llama and llava: Multimodal llms for low resource languages. arXiv preprint arXiv:2403.06354.
- Refusal in llms is mediated by a single direction. AI Alignment Forum.
- VisIT-Bench: A Benchmark for Vision-Language Instruction Following Inspired by Real-World Use. arXiv preprint arXiv:2308.06595.
- Terra Blevins and Luke Zettlemoyer. 2022. Language contamination helps explains the cross-lingual capabilities of English pretrained models. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 3563–3574, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.
- InstructPix2Pix: Learning to Follow Image Editing Instructions. arXiv preprint. ArXiv:2211.09800 [cs].
- MaXM: Towards Multilingual Visual Question Answering. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 2667–2682, Singapore.
- Unsupervised cross-lingual representation learning at scale. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 8440–8451, Online. Association for Computational Linguistics.
- XNLI: Evaluating cross-lingual sentence representations. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 2475–2485, Brussels, Belgium. Association for Computational Linguistics.
- BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171–4186, Minneapolis, Minnesota. Association for Computational Linguistics.
- Using imperfect surrogates for downstream inference: Design-based supervised learning for social science applications of large language models. In Advances in Neural Information Processing Systems, volume 36, pages 68589–68601. Curran Associates, Inc.
- MAGMA – multimodal augmentation of generative models through adapter-based finetuning. In Findings of the Association for Computational Linguistics: EMNLP 2022, pages 2416–2428, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.
- Scaling and evaluating sparse autoencoders. arXiv preprint arXiv:2406.04093.
- Llava-gemma: Accelerating multimodal foundation models with a compact language model. Preprint, arXiv:2404.01331.
- Evaluating the elementary multilingual capabilities of large language models with multiq. Preprint, arXiv:2403.03814.
- Prismatic vlms: Investigating the design space of visually-conditioned language models. In International Conference on Machine Learning (ICML).
- GlotLID: Language identification for low-resource languages. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 6155–6218, Singapore. Association for Computational Linguistics.
- ViLT: Vision-and-Language Transformer Without Convolution or Region Supervision. arXiv preprint. ArXiv:2102.03334 [cs, stat].
- Similarity of neural network representations revisited. In International conference on machine learning, pages 3519–3529. PMLR.
- The Open Images Dataset v4: Unified Image Classification, Object Detection, and Visual Relationship Detection at Scale. International journal of computer vision, 128(7):1956–1981.
- Okapi: Instruction-tuned large language models in multiple languages with reinforcement learning from human feedback. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 318–327, Singapore. Association for Computational Linguistics.
- BLIP-2: Bootstrapping Language-Image Pre-training with Frozen Image Encoders and Large Language Models. arXiv preprint. ArXiv:2301.12597 [cs].
- Visual Instruction Tuning. In Advances in Neural Information Processing Systems, volume 36, pages 34892–34916, New Orleans, LT, USA. Curran Associates, Inc.
- PALO: A Polyglot Large Multimodal Model for 5B People. arXiv preprint arXiv:2402.14818.
- MAPL: Parameter-Efficient Adaptation of Unimodal Pre-Trained Models for Vision-Language Few-Shot Prompting. In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 2523–2548, Dubrovnik, Croatia. Association for Computational Linguistics.
- Umap: Uniform manifold approximation and projection for dimension reduction. arXiv preprint arXiv:1802.03426.
- Linearly Mapping from Image to Text Space. arXiv preprint. ArXiv:2209.15162 [cs].
- OpenAI. 2024. Gpt-4o: Documentation. Accessed: 2024-06-15.
- Dinov2: Learning robust visual features without supervision. Preprint, arXiv:2304.07193.
- Björn Plüster. 2023. LeoLM: Igniting German-Language LLM Research. Accessed: 2024-06-15.
- Learning Transferable Visual Models From Natural Language Supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 8748–8763. PMLR. ISSN: 2640-3498.
- A Survey of Cross-lingual Word Embedding Models. Journal of Artificial Intelligence Research, 65:569–631.
- Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2556–2565, Melbourne, Australia. Association for Computational Linguistics.
- X-llava: Optimizing bilingual large vision-language alignment. arXiv preprint arXiv:2403.11399.
- SkunkworksAI. 2023. Bakllava-1. Hugging Face model repository.
- What is missing in multilingual visual reasoning and how to fix it. arXiv preprint arXiv:2403.01404.
- No language left behind: Scaling human-centered machine translation. Preprint, arXiv:2207.04672.
- Scaling monosemanticity: Extracting interpretable features from claude 3 sonnet. Transformer Circuits Thread.
- Crossmodal-3600: A Massively Multilingual Multimodal Evaluation Dataset. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 715–729, Abu Dhabi, United Arab Emirates.
- Llama 2: Open foundation and fine-tuned chat models. Preprint, arXiv:2307.09288.
- SimVLM: Simple Visual Language Model Pretraining with Weak Supervision. arXiv preprint. ArXiv:2108.10904 [cs].
- Do llamas work in english? on the latent language of multilingual transformers. Preprint, arXiv:2402.10588.
- Wildchat: 1m chatgpt interaction logs in the wild. arXiv preprint arXiv:2405.01470.
- Judging llm-as-a-judge with mt-bench and chatbot arena. Preprint, arXiv:2306.05685.
Paper Prompts
Sign up for free to create and run prompts on this paper using GPT-5.
Top Community Prompts
Collections
Sign up for free to add this paper to one or more collections.
