Score as Action: Fine-Tuning Diffusion Generative Models by Continuous-time Reinforcement Learning

Reinforcement learning from human feedback (RLHF), which aligns a diffusion model with input prompt, has become a crucial step in building reliable generative AI models. Most works in this area use a discrete-time formulation, which is prone to induced errors, and often not applicable to models with higher-order/black-box solvers. The objective of this study is to develop a disciplined approach to fine-tune diffusion models using continuous-time RL, formulated as a stochastic control problem with a reward function that aligns the end result (terminal state) with input prompt. The key idea is to treat score matching as controls or actions, and thereby making connections to policy optimization and regularization in continuous-time RL. To carry out this idea, we lay out a new policy optimization framework for continuous-time RL, and illustrate its potential in enhancing the value networks design space via leveraging the structural property of diffusion models. We validate the advantages of our method by experiments in downstream tasks of fine-tuning large-scale Text2Image models of Stable Diffusion v1.5.

Mobile-Agent-E: Self-Evolving Mobile Assistant for Complex Tasks

Smartphones have become indispensable in modern life, yet navigating complex tasks on mobile devices often remains frustrating. Recent advancements in large multimodal model (LMM)-based mobile agents have demonstrated the ability to perceive and act in mobile environments. However, current approaches face significant limitations: they fall short in addressing real-world human needs, struggle with reasoning-intensive and long-horizon tasks, and lack mechanisms to learn and improve from prior experiences. To overcome these challenges, we introduce Mobile-Agent-E, a hierarchical multi-agent framework capable of self-evolution through past experience. By hierarchical, we mean an explicit separation of high-level planning and low-level action execution. The framework comprises a Manager, responsible for devising overall plans by breaking down complex tasks into subgoals, and four subordinate agents--Perceptor, Operator, Action Reflector, and Notetaker--which handle fine-grained visual perception, immediate action execution, error verification, and information aggregation, respectively. Mobile-Agent-E also features a novel self-evolution module which maintains a persistent long-term memory comprising Tips and Shortcuts. Tips are general guidance and lessons learned from prior tasks on how to effectively interact with the environment. Shortcuts are reusable, executable sequences of atomic operations tailored for specific subroutines. The inclusion of Tips and Shortcuts facilitates continuous refinement in performance and efficiency. Alongside this framework, we introduce Mobile-Eval-E, a new benchmark featuring complex mobile tasks requiring long-horizon, multi-app interactions. Empirical results show that Mobile-Agent-E achieves a 22% absolute improvement over previous state-of-the-art approaches across three foundation model backbones. Project page: https://x-plug.github.io/MobileAgent.

DISC: Plug-and-Play Decoding Intervention with Similarity of Characters for Chinese Spelling Check

One key characteristic of the Chinese spelling check (CSC) task is that incorrect characters are usually similar to the correct ones in either phonetics or glyph. To accommodate this, previous works usually leverage confusion sets, which suffer from two problems, i.e., difficulty in determining which character pairs to include and lack of probabilities to distinguish items in the set. In this paper, we propose a light-weight plug-and-play DISC (i.e., decoding intervention with similarity of characters) module for CSC models.DISC measures phonetic and glyph similarities between characters and incorporates this similarity information only during the inference phase. This method can be easily integrated into various existing CSC models, such as ReaLiSe, SCOPE, and ReLM, without additional training costs. Experiments on three CSC benchmarks demonstrate that our proposed method significantly improves model performance, approaching and even surpassing the current state-of-the-art models.

Skip Tuning: Pre-trained Vision-Language Models are Effective and Efficient Adapters Themselves

Prompt tuning (PT) has long been recognized as an effective and efficient paradigm for transferring large pre-trained vision-language models (VLMs) to downstream tasks by learning a tiny set of context vectors. Nevertheless, in this work, we reveal that freezing the parameters of VLMs during learning the context vectors neither facilitates the transferability of pre-trained knowledge nor improves the memory and time efficiency significantly. Upon further investigation, we find that reducing both the length and width of the feature-gradient propagation flows of the full fine-tuning (FT) baseline is key to achieving effective and efficient knowledge transfer. Motivated by this, we propose Skip Tuning, a novel paradigm for adapting VLMs to downstream tasks. Unlike existing PT or adapter-based methods, Skip Tuning applies Layer-wise Skipping (LSkip) and Class-wise Skipping (CSkip) upon the FT baseline without introducing extra context vectors or adapter modules. Extensive experiments across a wide spectrum of benchmarks demonstrate the superior effectiveness and efficiency of our Skip Tuning over both PT and adapter-based methods. Code: https://github.com/Koorye/SkipTuning.

Learning to Solve Domain-Specific Calculation Problems with Knowledge-Intensive Programs Generator

Domain Large Language Models (LLMs) are developed for domain-specific tasks based on general LLMs. But it still requires professional knowledge to facilitate the expertise for some domain-specific tasks. In this paper, we investigate into knowledge-intensive calculation problems. We find that the math problems to be challenging for LLMs, when involving complex domain-specific rules and knowledge documents, rather than simple formulations of terminologies. Therefore, we propose a pipeline to solve the domain-specific calculation problems with Knowledge-Intensive Programs Generator more effectively, named as KIPG. It generates knowledge-intensive programs according to the domain-specific documents. For each query, key variables are extracted, then outcomes which are dependent on domain knowledge are calculated with the programs. By iterative preference alignment, the code generator learns to improve the logic consistency with the domain knowledge. Taking legal domain as an example, we have conducted experiments to prove the effectiveness of our pipeline, and extensive analysis on the modules. We also find that the code generator is also adaptable to other domains, without training on the new knowledge.

Streamlined Federated Unlearning: Unite as One to Be Highly Efficient

Recently, the enactment of "right to be forgotten" laws and regulations has imposed new privacy requirements on federated learning (FL). Researchers aim to remove the influence of certain data from the trained model without training from scratch through federated unlearning (FU). While current FU research has shown progress in enhancing unlearning efficiency, it often results in degraded model performance upon achieving the goal of data unlearning, necessitating additional steps to recover the performance of the unlearned model. Moreover, these approaches also suffer from many shortcomings such as high consumption of computational and storage resources. To this end, we propose a streamlined federated unlearning approach (SFU) aimed at effectively removing the influence of target data while preserving the model's performance on the retained data without degradation. We design a practical multi-teacher system that achieves both target data influence removal and model performance preservation by guiding the unlearned model through several distinct teacher models. SFU is both computationally and storage-efficient, highly flexible, and generalizable. We conducted extensive experiments on both image and text benchmark datasets. The results demonstrate that SFU significantly improves time and communication efficiency compared to the benchmark retraining method and significantly outperforms existing state-of-the-art (SOTA) methods. Additionally, we verified the effectiveness of SFU using the backdoor attack.

A More Advanced Group Polarization Measurement Approach Based on LLM-Based Agents and Graphs

Group polarization is an important research direction in social media content analysis, attracting many researchers to explore this field. Therefore, how to effectively measure group polarization has become a critical topic. Measuring group polarization on social media presents several challenges that have not yet been addressed by existing solutions. First, social media group polarization measurement involves processing vast amounts of text, which poses a significant challenge for information extraction. Second, social media texts often contain hard-to-understand content, including sarcasm, memes, and internet slang. Additionally, group polarization research focuses on holistic analysis, while texts is typically fragmented. To address these challenges, we designed a solution based on a multi-agent system and used a graph-structured Community Sentiment Network (CSN) to represent polarization states. Furthermore, we developed a metric called Community Opposition Index (COI) based on the CSN to quantify polarization. Finally, we tested our multi-agent system through a zero-shot stance detection task and achieved outstanding results. In summary, the proposed approach has significant value in terms of usability, accuracy, and interpretability.

SymDPO: Boosting In-Context Learning of Large Multimodal Models with Symbol Demonstration Direct Preference Optimization

As language models continue to scale, Large Language Models (LLMs) have exhibited emerging capabilities in In-Context Learning (ICL), enabling them to solve language tasks by prefixing a few in-context demonstrations (ICDs) as context. Inspired by these advancements, researchers have extended these techniques to develop Large Multimodal Models (LMMs) with ICL capabilities. However, existing LMMs face a critical issue: they often fail to effectively leverage the visual context in multimodal demonstrations and instead simply follow textual patterns. This indicates that LMMs do not achieve effective alignment between multimodal demonstrations and model outputs. To address this problem, we propose Symbol Demonstration Direct Preference Optimization (SymDPO). Specifically, SymDPO aims to break the traditional paradigm of constructing multimodal demonstrations by using random symbols to replace text answers within instances. This forces the model to carefully understand the demonstration images and establish a relationship between the images and the symbols to answer questions correctly. We validate the effectiveness of this method on multiple benchmarks, demonstrating that with SymDPO, LMMs can more effectively understand the multimodal context within examples and utilize this knowledge to answer questions better.

OccLoff: Learning Optimized Feature Fusion for 3D Occupancy Prediction

3D semantic occupancy prediction is crucial for finely representing the surrounding environment, which is essential for ensuring the safety in autonomous driving. Existing fusion-based occupancy methods typically involve performing a 2D-to-3D view transformation on image features, followed by computationally intensive 3D operations to fuse these with LiDAR features, leading to high computational costs and reduced accuracy. Moreover, current research on occupancy prediction predominantly focuses on designing specific network architectures, often tailored to particular models, with limited attention given to the more fundamental aspect of semantic feature learning. This gap hinders the development of more transferable methods that could enhance the performance of various occupancy models. To address these challenges, we propose OccLoff, a framework that Learns to Optimize Feature Fusion for 3D occupancy prediction. Specifically, we introduce a sparse fusion encoder with entropy masks that directly fuses 3D and 2D features, improving model accuracy while reducing computational overhead. Additionally, we propose a transferable proxy-based loss function and an adaptive hard sample weighting algorithm, which enhance the performance of several state-of-the-art methods. Extensive evaluations on the nuScenes and SemanticKITTI benchmarks demonstrate the superiority of our framework, and ablation studies confirm the effectiveness of each proposed module.

VLA-3D: A Dataset for 3D Semantic Scene Understanding and Navigation

With the recent rise of Large Language Models (LLMs), Vision-Language Models (VLMs), and other general foundation models, there is growing potential for multimodal, multi-task embodied agents that can operate in diverse environments given only natural language as input. One such application area is indoor navigation using natural language instructions. However, despite recent progress, this problem remains challenging due to the spatial reasoning and semantic understanding required, particularly in arbitrary scenes that may contain many objects belonging to fine-grained classes. To address this challenge, we curate the largest real-world dataset for Vision and Language-guided Action in 3D Scenes (VLA-3D), consisting of over 11.5K scanned 3D indoor rooms from existing datasets, 23.5M heuristically generated semantic relations between objects, and 9.7M synthetically generated referential statements. Our dataset consists of processed 3D point clouds, semantic object and room annotations, scene graphs, navigable free space annotations, and referential language statements that specifically focus on view-independent spatial relations for disambiguating objects. The goal of these features is to aid the downstream task of navigation, especially on real-world systems where some level of robustness must be guaranteed in an open world of changing scenes and imperfect language. We benchmark our dataset with current state-of-the-art models to obtain a performance baseline. All code to generate and visualize the dataset is publicly released, see https://github.com/HaochenZ11/VLA-3D. With the release of this dataset, we hope to provide a resource for progress in semantic 3D scene understanding that is robust to changes and one which will aid the development of interactive indoor navigation systems.