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

• URL: http://arxiv.org/abs/2512.03276v1
• Date: Tue, 02 Dec 2025 22:31:43 GMT
• Title: Too Late to Recall: Explaining the Two-Hop Problem in Multimodal Knowledge Retrieval
• Authors: Constantin Venhoff, Ashkan Khakzar, Sonia Joseph, Philip Torr, Neel Nanda,
• Abstract summary: 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)<n>We show that 11 of 14 models exhibit factual recall degradation.<n>High-performing VLMs resolve entity representations early enough to reuse the existing factual recall mechanism.
• Score: 28.475457214103404
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: 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.

Related papers

• A Comprehensive Study on Visual Token Redundancy for Discrete Diffusion-based Multimodal Large Language Models [85.30893355216486]
  We study how visual token redundancy evolves with different dMLLM architectures and tasks.<n>Our study reveals that visual redundancy emerges only in from-scratch dMLLMs while handling long-answer tasks.<n>Layer-skipping is promising for accelerating AR-to-diffusion dMLLMs, whereas progressive or late-step pruning is more effective for from-scratch dMLLMs.
  arXiv  Detail & Related papers  (2025-11-19T04:13:36Z)
• Cross-Modal Attention Guided Unlearning in Vision-Language Models [16.460281156521646]
  Vision-Language Models (VLMs) have demonstrated immense capabilities in multi-modal understanding and inference tasks.<n>VLMs add a layer of complexity to this process, as the visual context in the query may also contain sensitive information in addition to the text.<n>We formulate Cross-Modal Attention Guided Unlearning (CAGUL), a lightweight and efficient VLM unlearning framework.
  arXiv  Detail & Related papers  (2025-10-08T21:21:59Z)
• Mimicking or Reasoning: Rethinking Multi-Modal In-Context Learning in Vision-Language Models [19.361686225381447]
  Vision-language models (VLMs) are widely assumed to exhibit in-context learning (ICL)<n>We propose a new MM-ICL with Reasoning pipeline that augments each demonstration with a generated rationale alongside the answer.
  arXiv  Detail & Related papers  (2025-06-09T16:55:32Z)
• Vad-R1: Towards Video Anomaly Reasoning via Perception-to-Cognition Chain-of-Thought [58.321044666612174]
  Vad-R1 is an end-to-end MLLM-based framework for Video Anomaly Reasoning.<n>We design a Perception-to-Cognition Chain-of-Thought (P2C-CoT) that simulates the human process of recognizing anomalies.<n>We also propose an improved reinforcement learning algorithm AVA-GRPO, which explicitly incentivizes the anomaly reasoning capability of MLLMs.
  arXiv  Detail & Related papers  (2025-05-26T12:05:16Z)
• Integrating Visual Interpretation and Linguistic Reasoning for Math Problem Solving [61.992824291296444]
  Current large vision-language models (LVLMs) typically employ a connector module to link visual features with text embeddings of large language models (LLMs)<n>This paper proposes a paradigm shift: instead of training end-to-end vision-language reasoning models, we advocate for developing a decoupled reasoning framework.
  arXiv  Detail & Related papers  (2025-05-23T08:18:00Z)
• Blind Spot Navigation: Evolutionary Discovery of Sensitive Semantic Concepts for LVLMs [24.76767896607915]
  Recent research suggests that models may be particularly sensitive to certain semantics in visual inputs, making them prone to errors.<n>Inspired by this, in this paper we conducted the first exploration on large vision-language models (LVLMs)<n>We found that LVLMs indeed are susceptible to hallucinations and various errors when facing specific semantic concepts in images.
  arXiv  Detail & Related papers  (2025-05-21T08:45:43Z)
• NoteLLM-2: Multimodal Large Representation Models for Recommendation [71.87790090964734]
  Large Language Models (LLMs) have demonstrated exceptional proficiency in text understanding and embedding tasks.<n>Their potential in multimodal representation, particularly for item-to-item (I2I) recommendations, remains underexplored.<n>We propose an end-to-end fine-tuning method that customizes the integration of any existing LLMs and vision encoders for efficient multimodal representation.
  arXiv  Detail & Related papers  (2024-05-27T03:24:01Z)
• Mitigating Object Hallucination in Large Vision-Language Models via Image-Grounded Guidance [51.30560006045442]
  Image-gRounded guIdaNcE (MARINE) is a framework that is both training-free and API-free.<n>MARINE effectively and efficiently reduces object hallucinations during inference by introducing image-grounded guidance to LVLMs.<n>Our framework's flexibility further allows for the integration of multiple vision models, enabling more reliable and robust object-level guidance.
  arXiv  Detail & Related papers  (2024-02-13T18:59:05Z)
• Prismatic VLMs: Investigating the Design Space of Visually-Conditioned Language Models [73.40350756742231]
  Visually-conditioned language models (VLMs) have seen growing adoption in applications such as visual dialogue, scene understanding, and robotic task planning. Despite the volume of new releases, key design decisions around image preprocessing, architecture, and optimization are under-explored.
  arXiv  Detail & Related papers  (2024-02-12T18:21:14Z)
• ViCor: Bridging Visual Understanding and Commonsense Reasoning with Large Language Models [27.5219975853389]
  We find that pre-trained vision-and-language models (VLMs) and large language models (LLMs) are good at different kinds of visual commonsense reasoning problems. For problems where the goal is to infer conclusions beyond image content,VLMs face difficulties, while LLMs, given sufficient visual evidence, can use commonsense to infer the answer well.
  arXiv  Detail & Related papers  (2023-10-09T17:10:35Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.