the Institute of Intelligent Machines, Chinese Academy of Sciences
Abstract:Mixture of Experts (MoE) architectures have become a key approach for scaling large language models, with growing interest in extending them to multimodal tasks. Existing methods to build multimodal MoE models either incur high training costs or suffer from degraded language capabilities when adapting pretrained models. To address this, we propose Soft ModalityAware Routing (SMAR), a novel regularization technique that uses Kullback Leibler divergence to control routing probability distributions across modalities, encouraging expert specialization without modifying model architecture or heavily relying on textual data. Experiments on visual instruction tuning show that SMAR preserves language ability at 86.6% retention with only 2.5% pure text, outperforming baselines while maintaining strong multimodal performance. Our approach offers a practical and efficient solution to balance modality differentiation and language capabilities in multimodal MoE models.
Abstract:Large language models (LLMs) exhibit remarkable capabilities in handling natural language tasks; however, they may struggle to consistently follow complex instructions including those involve multiple constraints. Post-training LLMs using supervised fine-tuning (SFT) is a standard approach to improve their ability to follow instructions. In addressing complex instruction following, existing efforts primarily focus on data-driven methods that synthesize complex instruction-output pairs for SFT. However, insufficient attention allocated to crucial sub-contexts may reduce the effectiveness of SFT. In this work, we propose transforming sequentially structured input instruction into multiple parallel instructions containing subcontexts. To support processing this multi-input, we propose MISO (Multi-Input Single-Output), an extension to currently dominant decoder-only transformer-based LLMs. MISO introduces a mixture-of-contexts paradigm that jointly considers the overall instruction-output alignment and the influence of individual sub-contexts to enhance SFT effectiveness. We apply MISO fine-tuning to complex instructionfollowing datasets and evaluate it with standard LLM inference. Empirical results demonstrate the superiority of MISO as a fine-tuning method for LLMs, both in terms of effectiveness in complex instruction-following scenarios and its potential for training efficiency.
Abstract:Subject-driven image generation aims to synthesize novel scenes that faithfully preserve subject identity from reference images while adhering to textual guidance, yet existing methods struggle with a critical trade-off between fidelity and efficiency. Tuning-based approaches rely on time-consuming and resource-intensive subject-specific optimization, while zero-shot methods fail to maintain adequate subject consistency. In this work, we propose FreeGraftor, a training-free framework that addresses these limitations through cross-image feature grafting. Specifically, FreeGraftor employs semantic matching and position-constrained attention fusion to transfer visual details from reference subjects to the generated image. Additionally, our framework incorporates a novel noise initialization strategy to preserve geometry priors of reference subjects for robust feature matching. Extensive qualitative and quantitative experiments demonstrate that our method enables precise subject identity transfer while maintaining text-aligned scene synthesis. Without requiring model fine-tuning or additional training, FreeGraftor significantly outperforms existing zero-shot and training-free approaches in both subject fidelity and text alignment. Furthermore, our framework can seamlessly extend to multi-subject generation, making it practical for real-world deployment. Our code is available at https://github.com/Nihukat/FreeGraftor.
Abstract:Flexible electronic skins that simultaneously sense touch and bend are desired in several application areas, such as to cover articulated robot structures. This paper introduces a flexible tactile sensor based on Electrical Impedance Tomography (EIT), capable of simultaneously detecting and measuring contact forces and flexion of the sensor. The sensor integrates a magnetic hydrogel composite and utilizes EIT to reconstruct internal conductivity distributions. Real-time estimation is achieved through the one-step Gauss-Newton method, which dynamically updates reference voltages to accommodate sensor deformation. A convolutional neural network is employed to classify interactions, distinguishing between touch, bending, and idle states using pre-reconstructed images. Experimental results demonstrate an average touch localization error of 5.4 mm (SD 2.2 mm) and average bending angle estimation errors of 1.9$^\circ$ (SD 1.6$^\circ$). The proposed adaptive reference method effectively distinguishes between single- and multi-touch scenarios while compensating for deformation effects. This makes the sensor a promising solution for multimodal sensing in robotics and human-robot collaboration.
Abstract:Large language models (LLMs) based agent systems have made great strides in real-world applications beyond traditional NLP tasks. This paper proposes a new LLM-powered Multi-Agent System (LLM-MAS) benchmark, Collab-Overcooked, built on the popular Overcooked-AI game with more applicable and challenging tasks in interactive environments. Collab-Overcooked extends existing benchmarks from two novel perspectives. First, it provides a multi-agent framework supporting diverse tasks and objectives and encourages collaboration through natural language communication. Second, it introduces a spectrum of process-oriented evaluation metrics to assess the fine-grained collaboration capabilities of different LLM agents, a dimension often overlooked in prior work. We conduct extensive experiments over 10 popular LLMs and show that, while the LLMs present a strong ability in goal interpretation, there is a significant discrepancy in active collaboration and continuous adaption that are critical for efficiently fulfilling complicated tasks. Notably, we highlight the strengths and weaknesses in LLM-MAS and provide insights for improving and evaluating LLM-MAS on a unified and open-sourced benchmark. Environments, 30 open-ended tasks, and an integrated evaluation package are now publicly available at https://github.com/YusaeMeow/Collab-Overcooked.
Abstract:The study of mechanistic interpretability aims to reverse-engineer a model to explain its behaviors. While recent studies have focused on the static mechanism of a certain behavior, the training dynamics inside a model remain to be explored. In this work, we develop an interpretable method for fine-tuning and reveal the mechanism behind learning. We first propose the concept of node redundancy as an extension of intrinsic dimension and explain the idea behind circuit discovery from a fresh view. Based on the theory, we propose circuit-tuning, a two-stage algorithm that iteratively performs circuit discovery to mask out irrelevant edges and updates the remaining parameters responsible for a specific task. Experiments show that our method not only improves performance on a wide range of tasks but is also scalable while preserving general capabilities. We visualize and analyze the circuits before, during, and after fine-tuning, providing new insights into the self-organization mechanism of a neural network in the learning process.
Abstract:Large language models (LLMs) exhibit remarkable capabilities in natural language processing but face catastrophic forgetting when learning new tasks, where adaptation to a new domain leads to a substantial decline in performance on previous tasks. In this paper, we propose Controlled LoRA (CLoRA), a subspace regularization method on LoRA structure. Aiming to reduce the scale of output change while introduce minimal constraint on model capacity, CLoRA imposes constraint on the direction of updating matrix null space. Experimental results on commonly used LLM finetuning tasks reveal that CLoRA significantly outperforms existing LoRA subsequent methods on both in-domain and outdomain evaluations, highlighting the superority of CLoRA as a effective parameter-efficient finetuning method with catastrophic forgetting mitigating. Further investigation for model parameters indicates that CLoRA effectively balances the trade-off between model capacity and degree of forgetting.
Abstract:The customization of text-to-image models has seen significant advancements, yet generating multiple personalized concepts remains a challenging task. Current methods struggle with attribute leakage and layout confusion when handling multiple concepts, leading to reduced concept fidelity and semantic consistency. In this work, we introduce a novel training-free framework, Concept Conductor, designed to ensure visual fidelity and correct layout in multi-concept customization. Concept Conductor isolates the sampling processes of multiple custom models to prevent attribute leakage between different concepts and corrects erroneous layouts through self-attention-based spatial guidance. Additionally, we present a concept injection technique that employs shape-aware masks to specify the generation area for each concept. This technique injects the structure and appearance of personalized concepts through feature fusion in the attention layers, ensuring harmony in the final image. Extensive qualitative and quantitative experiments demonstrate that Concept Conductor can consistently generate composite images with accurate layouts while preserving the visual details of each concept. Compared to existing baselines, Concept Conductor shows significant performance improvements. Our method supports the combination of any number of concepts and maintains high fidelity even when dealing with visually similar concepts. The code and models are available at https://github.com/Nihukat/Concept-Conductor.
Abstract:Physics-Informed Neural Networks (PINNs) are a machine learning technique for solving partial differential equations (PDEs) by incorporating PDEs as loss terms in neural networks and minimizing the loss function during training. Tomographic imaging, a method to reconstruct internal properties from external measurement data, is highly complex and ill-posed, making it an inverse problem. Recently, PINNs have shown significant potential in computational fluid dynamics (CFD) and have advantages in solving inverse problems. However, existing research has primarily focused on semi-inverse Electrical Impedance Tomography (EIT), where internal electric potentials are accessible. The practical full inverse EIT problem, where only boundary voltage measurements are available, remains challenging. To address this, we propose a two-stage hybrid learning framework combining Convolutional Neural Networks (CNNs) and PINNs to solve the full inverse EIT problem. This framework integrates data-driven and model-driven approaches, combines supervised and unsupervised learning, and decouples the forward and inverse problems within the PINN framework in EIT. Stage I: a U-Net constructs an end-to-end mapping from boundary voltage measurements to the internal potential distribution using supervised learning. Stage II: a Multilayer Perceptron (MLP)-based PINN takes the predicted internal potentials as input to solve for the conductivity distribution through unsupervised learning.
Abstract:Instruction Fine-Tuning, a method enhancing pre-trained language models' capabilities from mere next-word prediction to complex instruction following, often employs a one-off training approach on diverse instruction dataset. However, this method may not effectively enhance models' adherence to instructions due to the simultaneous handling of varying instruction complexities. To address this, we propose a novel phased instruction fine-tuning (Phased IFT) method, grounded in the hypothesis of progressive alignment, which posits that the transition of a pre-trained language model from simple next-word prediction to sophisticated instruction following is a gradual learning process. Specifically, we obtain the score of difficulty for each instruction via GPT-4, stratify the instruction data into subsets of increasing difficulty, and sequentially uptrain on these subsets using the standard supervised loss. Through extensive experiments on the pre-trained models Llama-2 7B/13B, and Mistral-7B using the 52K Alpaca instruction data, we demonstrate that Phased IFT significantly surpasses traditional one-off instruction fine-tuning (One-off IFT) method in win rate, empirically validating the progressive alignment hypothesis. Our findings suggest that Phased IFT offers a simple yet effective pathway for elevating the instruction-following capabilities of pre-trained language models. Models and datasets from our experiments are freely available at https://github.com/xubuvd/PhasedSFT.