# Too Late to Recall: Explaining the Two-Hop Problem in Multimodal Knowledge Retrieval

arXiv: https://arxiv.org/abs/2512.03276 · HTML: https://mlsanity.com/paper/2512.03276/

## Key facts

- Cited by 7 papers in the map (data through 14 Oct 2026).
- Placed in MLLM training and visual instruction tuning, one of 16 directions in Multimodal LLMs / vision-language models.
- Builds on 29 of its 37 references that are in the map.
- First posted on arXiv 2 Dec 2025.

## Placed in

- Multimodal LLMs / vision-language models / [MLLM training and visual instruction tuning](https://mlsanity.com/mllm/mllm-training-and-visual-instruction-tuning/)
- Mechanistic interpretability / [Interpreting vision and multimodal models](https://mlsanity.com/mechinterp/interpreting-vision-and-multimodal-models/)

## Builds on (most cited first, 29 of 29)

- [An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale](https://mlsanity.com/paper/2010.11929/) (2020)
- [LLaMA: Open and Efficient Foundation Language Models](https://mlsanity.com/paper/2302.13971/) (2023)
- [Visual Instruction Tuning](https://mlsanity.com/paper/2304.08485/) (2023)
- [Improved Baselines with Visual Instruction Tuning](https://mlsanity.com/paper/2310.03744/) (2023)
- [Flamingo: a Visual Language Model for Few-Shot Learning](https://mlsanity.com/paper/2204.14198/) (2022)
- [Locating and Editing Factual Associations in GPT](https://mlsanity.com/paper/2202.05262/) (2022)
- [InternVL: Scaling up Vision Foundation Models and Aligning for Generic Visual-Linguistic Tasks](https://mlsanity.com/paper/2312.14238/) (2023)
- [VILA: On Pre-training for Visual Language Models](https://mlsanity.com/paper/2312.07533/) (2023)
- [Dissecting Recall of Factual Associations in Auto-Regressive Language Models](https://mlsanity.com/paper/2304.14767/) (2023)
- [WIT: Wikipedia-based Image Text Dataset for Multimodal Multilingual Machine Learning](https://mlsanity.com/paper/2103.01913/) (2021)
- [Interpreting CLIP's Image Representation via Text-Based Decomposition](https://mlsanity.com/paper/2310.05916/) (2023)
- [Do Large Language Models Latently Perform Multi-Hop Reasoning?](https://mlsanity.com/paper/2402.16837/) (2024)
- [Towards Interpreting Visual Information Processing in Vision-Language Models](https://mlsanity.com/paper/2410.07149/) (2024)
- [Interpreting and Editing Vision-Language Representations to Mitigate Hallucinations](https://mlsanity.com/paper/2410.02762/) (2024)
- [Hopping Too Late: Exploring the Limitations of Large Language Models on Multi-Hop Queries](https://mlsanity.com/paper/2406.12775/) (2024)
- [The Semantic Hub Hypothesis: Language Models Share Semantic Representations Across Languages and Modalities](https://mlsanity.com/paper/2411.04986/) (2024)
- [Multimodal Neurons in Pretrained Text-Only Transformers](https://mlsanity.com/paper/2308.01544/) (2023)
- [Summing Up the Facts: Additive Mechanisms Behind Factual Recall in LLMs](https://mlsanity.com/paper/2402.07321/) (2024)
- [Steering CLIP's vision transformer with sparse autoencoders](https://mlsanity.com/paper/2504.08729/) (2025)
- [Same Task, Different Circuits: Disentangling Modality-Specific Mechanisms in VLMs](https://mlsanity.com/paper/2506.09047/) (2025)
- [Visual-Text Cross Alignment: Refining the Similarity Score in Vision-Language Models](https://mlsanity.com/paper/2406.02915/) (2024)
- [LLaVA-MORE: A Comparative Study of LLMs and Visual Backbones for Enhanced Visual Instruction Tuning](https://mlsanity.com/paper/2503.15621/) (2025)
- [WildHallucinations: Evaluating Long-form Factuality in LLMs with Real-World Entity Queries](https://mlsanity.com/paper/2407.17468/) (2024)
- [How Visual Representations Map to Language Feature Space in Multimodal LLMs](https://mlsanity.com/paper/2506.11976/) (2025)
- [Memory Injections: Correcting Multi-Hop Reasoning Failures during Inference in Transformer-Based Language Models](https://mlsanity.com/paper/2309.05605/) (2023)
- [Prisma: An Open Source Toolkit for Mechanistic Interpretability in Vision and Video](https://mlsanity.com/paper/2504.19475/) (2025)
- [Performance Gap in Entity Knowledge Extraction Across Modalities in Vision Language Models](https://mlsanity.com/paper/2412.14133/) (2024)
- [Interpreting and Controlling Vision Foundation Models via Text Explanations](https://mlsanity.com/paper/2310.10591/) (2023)
- [TG-LLaVA: Text Guided LLaVA via Learnable Latent Embeddings](https://mlsanity.com/paper/2409.09564/) (2024)

## Built on by (most cited first, 7 of 7)

- [The Topological Trouble With Transformers](https://mlsanity.com/paper/2604.17121/) (2026)
- [Uncovering Entity Identity Confusion in Multimodal Knowledge Editing](https://mlsanity.com/paper/2605.06096/) (2026)
- [Failing to See or Failing to Know? Attributing Errors in Vision-Language Models](https://mlsanity.com/paper/2607.04683/) (2026)
- [Slow to See, Slow to Suppress: Understanding the Effects of Modality in Context-Memory Conflicts](https://mlsanity.com/paper/2609.00293/) (2026)
- [Structured Layout Priors for Robust Out-of-Distribution Visual Document Understanding](https://mlsanity.com/paper/2605.19866/) (2026)
- [Using OCR Heads to Verbalize Image Semantics](https://mlsanity.com/paper/2609.18823/) (2026)
- [Generalizable Model Editing Could Be More Than Direct Generalization](https://mlsanity.com/paper/2610.15707/) (2026)

## Abstract (arXiv)

Training vision language models (VLMs) aims to align visual representations from a vision encoder with the textual representations of a pretrained large language model (LLM). However, many VLMs exhibit reduced factual recall performance compared to their LLM backbones, raising the question of how effective multimodal fine-tuning is at extending existing mechanisms within the LLM to visual inputs. We argue that factual recall based on visual inputs requires VLMs to solve a two-hop problem: (1) forming entity representations from visual inputs, and (2) recalling associated factual knowledge based on these entity representations. By benchmarking 14 VLMs with various architectures (LLaVA, Native, Cross-Attention), sizes (7B-124B parameters), and training setups on factual recall tasks against their original LLM backbone models, we find that 11 of 14 models exhibit factual recall degradation. We select three models with high and two models with low performance degradation, and use attribution patching, activation patching, and probing to show that degraded VLMs struggle to use the existing factual recall circuit of their LLM backbone, because they resolve the first hop too late in the computation. In contrast, high-performing VLMs resolve entity representations early enough to reuse the existing factual recall mechanism. Finally, we demonstrate two methods to recover performance: patching entity representations from the LLM backbone into the VLM, and prompting with chain-of-thought reasoning. Our results highlight that the speed of early entity resolution critically determines how effective VLMs are in using preexisting LLM mechanisms. More broadly, our work illustrates how mechanistic analysis can explain and unveil systematic failures in multimodal alignment.