Analyzing Finetuning Representation Shift for Multimodal LLMs Steering

• URL: http://arxiv.org/abs/2501.03012v2
• Date: Wed, 13 Aug 2025 13:42:57 GMT
• Title: Analyzing Finetuning Representation Shift for Multimodal LLMs Steering
• Authors: Pegah Khayatan, Mustafa Shukor, Jayneel Parekh, Arnaud Dapogny, Matthieu Cord,
• Abstract summary: We propose to map hidden states to interpretable visual and textual concepts.<n>This enables us to more efficiently compare certain semantic dynamics, such as the shift from an original and fine-tuned model.<n>We also demonstrate the use of shift vectors to capture these concepts changes.
• Score: 56.710375516257876
• License: http://creativecommons.org/licenses/by-nc-sa/4.0/
• Abstract: Multimodal LLMs (MLLMs) have reached remarkable levels of proficiency in understanding multimodal inputs. However, understanding and interpreting the behavior of such complex models is a challenging task, not to mention the dynamic shifts that may occur during fine-tuning, or due to covariate shift between datasets. In this work, we apply concept-level analysis towards MLLM understanding. More specifically, we propose to map hidden states to interpretable visual and textual concepts. This enables us to more efficiently compare certain semantic dynamics, such as the shift from an original and fine-tuned model, revealing concept alteration and potential biases that may occur during fine-tuning. We also demonstrate the use of shift vectors to capture these concepts changes. These shift vectors allow us to recover fine-tuned concepts by applying simple, computationally inexpensive additive concept shifts in the original model. Finally, our findings also have direct applications for MLLM steering, which can be used for model debiasing as well as enforcing safety in MLLM output. All in all, we propose a novel, training-free, ready-to-use framework for MLLM behavior interpretability and control. Our implementation is publicly available.

Related papers

• PlaM: Training-Free Plateau-Guided Model Merging for Better Visual Grounding in MLLMs [59.78917775399492]
  Multimodal instruction fine-tuning paradoxically degrades this text's reasoning capability.<n>We propose a training-free framework to mitigate this degradation.
  arXiv  Detail & Related papers  (2026-01-12T15:27:51Z)
• 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)
• Unlocking the Forgery Detection Potential of Vanilla MLLMs: A Novel Training-Free Pipeline [5.740204096484276]
  We propose Foresee, a training-free MLLM-based pipeline tailored for image forgery analysis.<n>Foresee employs a type-prior-driven strategy and utilizes a Flexible Feature Detector module to handle copy-move manipulations.<n>Our approach simultaneously achieves superior localization accuracy and provides more comprehensive textual explanations.
  arXiv  Detail & Related papers  (2025-11-17T14:49:57Z)
• From Insight to Exploit: Leveraging LLM Collaboration for Adaptive Adversarial Text Generation [3.75886080255807]
  We introduce two innovative attack frameworks designed to generate dynamic and adaptive adversarial examples.<n>We produce subtle and natural-looking adversarial inputs that preserve semantic similarity to the original text.<n>Our attacks evolve with the advancements in LLMs and demonstrate strong transferability acrossversa unknown to the attacker.
  arXiv  Detail & Related papers  (2025-11-05T02:27:56Z)
• From Generation to Generalization: Emergent Few-Shot Learning in Video Diffusion Models [65.0487600936788]
  Video Diffusion Models (VDMs) have emerged as powerful generative tools capable of synthesizing high-quality content.<n>We argue that VDMs naturally push to probe structured representations and an implicit understanding of the visual world.<n>Our method transforms each task into a visual transition, enabling the training of LoRA weights on short input- sequences.
  arXiv  Detail & Related papers  (2025-06-08T20:52:34Z)
• Reasoning-Aligned Perception Decoupling for Scalable Multi-modal Reasoning [95.44766931218896]
  Multi-modal large language models (MLLMs) still lag behind text-based reasoning.<n>We introduce Perception-Reasoning Decoupling, which modularizes the MLLM's reasoning component and makes it easily replaceable.<n>We propose a novel reinforcement learning algorithm called Visual Perception Optimization (VPO) to align the MLLM's perceptual output with the final reasoning task.
  arXiv  Detail & Related papers  (2025-06-05T02:28:07Z)
• Visual Instruction Bottleneck Tuning [11.675789515164269]
  multimodal large language models (MLLMs) suffer performance degradation when encountering unfamiliar queries under distribution shifts.<n>We take an alternative approach to enhance the generalization and robustness of MLLMs from a representation learning perspective.<n>Inspired by information bottleneck (IB) principle, we derive a variational lower bound of the IB for MLLMs and devise a practical implementation, Visual Instruction Bottleneck Tuning (Vittle)
  arXiv  Detail & Related papers  (2025-05-20T05:24:53Z)
• Bring Reason to Vision: Understanding Perception and Reasoning through Model Merging [32.70038648928894]
  Vision-Language Models (VLMs) combine visual perception with the general capabilities, such as reasoning, of Large Language Models (LLMs)<n>In this work, we explore to compose perception and reasoning through model merging that connects parameters of different models.<n>We find that perception capabilities are predominantly encoded in the early layers of the model, whereas reasoning is largely facilitated by the middle-to-late layers.
  arXiv  Detail & Related papers  (2025-05-08T17:56:23Z)
• Improving Reasoning Performance in Large Language Models via Representation Engineering [2.0099933815960256]
  We propose a representation engineering approach for large language models (LLMs)<n>Model activations are read from the residual stream of an LLM when processing a reasoning task.<n>We show that an LLM can, to a certain degree, be controlled to improve its perceived reasoning ability by modulating activations.
  arXiv  Detail & Related papers  (2025-04-28T04:58:43Z)
• Cross-model Transferability among Large Language Models on the Platonic Representations of Concepts [24.990218468155383]
  This work explores the intricate relationships between concept representations across different Large Language Models (LLMs)<n>We introduce a linear transformation method to bridge these representations, drawing an intriguing parallel to Plato's Allegory of the Cave.<n>A weak-to-strong transferability exists between LLM concept representations, whereby SVs extracted from smaller LLMs can effectively control the behavior of larger LLMs.
  arXiv  Detail & Related papers  (2025-01-02T11:56:59Z)
• Sparse autoencoders reveal selective remapping of visual concepts during adaptation [54.82630842681845]
  Adapting foundation models for specific purposes has become a standard approach to build machine learning systems. We develop a new Sparse Autoencoder (SAE) for the CLIP vision transformer, named PatchSAE, to extract interpretable concepts.
  arXiv  Detail & Related papers  (2024-12-06T18:59:51Z)
• The Mechanics of Conceptual Interpretation in GPT Models: Interpretative Insights [10.777646083061395]
  We introduce concept editing'', an innovative variation of knowledge editing that uncovers conceptualisation mechanisms within large language models. We analyse the Multi-Layer Perceptron (MLP), Multi-Head Attention (MHA), and hidden state components of transformer models. Our work highlights the complex, layered nature of semantic processing in LLMs and the challenges of isolating and modifying specific concepts within these models.
  arXiv  Detail & Related papers  (2024-08-05T18:50:08Z)
• CoMMIT: Coordinated Multimodal Instruction Tuning [90.1532838391285]
  Multimodal large language models (MLLMs) generally involve cooperative learning between a backbone LLM and a feature encoder of non-text input modalities.<n>In this paper, we analyze the MLLM instruction tuning from both theoretical and empirical perspectives.<n>We propose a Multimodal Balance Coefficient that enables quantitative measurement of the balance of learning.
  arXiv  Detail & Related papers  (2024-07-29T23:18:55Z)
• Multi-modal Auto-regressive Modeling via Visual Words [96.25078866446053]
  We propose the concept of visual tokens, which maps the visual features to probability distributions over Large Multi-modal Models' vocabulary. We further explore the distribution of visual features in the semantic space within LMM and the possibility of using text embeddings to represent visual information.
  arXiv  Detail & Related papers  (2024-03-12T14:58:52Z)
• Model Composition for Multimodal Large Language Models [71.5729418523411]
  We propose a new paradigm through the model composition of existing MLLMs to create a new model that retains the modal understanding capabilities of each original model. Our basic implementation, NaiveMC, demonstrates the effectiveness of this paradigm by reusing modality encoders and merging LLM parameters.
  arXiv  Detail & Related papers  (2024-02-20T06:38:10Z)
• Delving into Multi-modal Multi-task Foundation Models for Road Scene Understanding: From Learning Paradigm Perspectives [56.2139730920855]
  We present a systematic analysis of MM-VUFMs specifically designed for road scenes. Our objective is to provide a comprehensive overview of common practices, referring to task-specific models, unified multi-modal models, unified multi-task models, and foundation model prompting techniques. We provide insights into key challenges and future trends, such as closed-loop driving systems, interpretability, embodied driving agents, and world models.
  arXiv  Detail & Related papers  (2024-02-05T12:47:09Z)
• Veagle: Advancements in Multimodal Representation Learning [0.0]
  This paper introduces a novel approach to enhance the multimodal capabilities of existing models. Our proposed model Veagle, incorporates a unique mechanism inspired by the successes and insights of previous works. Our results indicate a improvement of 5-6 % in performance, with Veagle outperforming existing models by a notable margin.
  arXiv  Detail & Related papers  (2024-01-18T12:45:25Z)
• Improving Discriminative Multi-Modal Learning with Large-Scale Pre-Trained Models [51.5543321122664]
  This paper investigates how to better leverage large-scale pre-trained uni-modal models to enhance discriminative multi-modal learning. We introduce Multi-Modal Low-Rank Adaptation learning (MMLoRA)
  arXiv  Detail & Related papers  (2023-10-08T15:01:54Z)
• MultiViz: An Analysis Benchmark for Visualizing and Understanding Multimodal Models [103.9987158554515]
  MultiViz is a method for analyzing the behavior of multimodal models by scaffolding the problem of interpretability into 4 stages. We show that the complementary stages in MultiViz together enable users to simulate model predictions, assign interpretable concepts to features, perform error analysis on model misclassifications, and use insights from error analysis to debug models.
  arXiv  Detail & Related papers  (2022-06-30T18:42:06Z)

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.