Di Zhang

Hai Zhang, Hang Yu, Junqiao Zhao et al.

Designing and deriving effective model-based reinforcement learning (MBRL) algorithms with a performance improvement guarantee is challenging, mainly attributed to the high coupling between model learning and policy optimization. Many prior methods that rely on return discrepancy to guide model learning ignore the impacts of model shift, which can lead to performance deterioration due to excessive model updates. Other methods use performance difference bound to explicitly consider model shift. However, these methods rely on a fixed threshold to constrain model shift, resulting in a heavy dependence on the threshold and a lack of adaptability during the training process. In this paper, we theoretically derive an optimization objective that can unify model shift and model bias and then formulate a fine-tuning process. This process adaptively adjusts the model updates to get a performance improvement guarantee while avoiding model overfitting. Based on these, we develop a straightforward algorithm USB-PO (Unified model Shift and model Bias Policy Optimization). Empirical results show that USB-PO achieves state-of-the-art performance on several challenging benchmark tasks.

Xiao Zhang, Hai Zhang, Hongtu Zhou et al.

Safety is one of the main challenges in applying reinforcement learning to realistic environmental tasks. To ensure safety during and after training process, existing methods tend to adopt overly conservative policy to avoid unsafe situations. However, overly conservative policy severely hinders the exploration, and makes the algorithms substantially less rewarding. In this paper, we propose a method to construct a boundary that discriminates safe and unsafe states. The boundary we construct is equivalent to distinguishing dead-end states, indicating the maximum extent to which safe exploration is guaranteed, and thus has minimum limitation on exploration. Similar to Recovery Reinforcement Learning, we utilize a decoupled RL framework to learn two policies, (1) a task policy that only considers improving the task performance, and (2) a recovery policy that maximizes safety. The recovery policy and a corresponding safety critic are pretrained on an offline dataset, in which the safety critic evaluates upper bound of safety in each state as awareness of environmental safety for the agent. During online training, a behavior correction mechanism is adopted, ensuring the agent to interact with the environment using safe actions only. Finally, experiments of continuous control tasks demonstrate that our approach has better task performance with less safety violations than state-of-the-art algorithms.

Xinlong Chen, Yuanxing Zhang, Chongling Rao et al.

The training of controllable text-to-video (T2V) models relies heavily on the alignment between videos and captions, yet little existing research connects video caption evaluation with T2V generation assessment. This paper introduces VidCapBench, a video caption evaluation scheme specifically designed for T2V generation, agnostic to any particular caption format. VidCapBench employs a data annotation pipeline, combining expert model labeling and human refinement, to associate each collected video with key information spanning video aesthetics, content, motion, and physical laws. VidCapBench then partitions these key information attributes into automatically assessable and manually assessable subsets, catering to both the rapid evaluation needs of agile development and the accuracy requirements of thorough validation. By evaluating numerous state-of-the-art captioning models, we demonstrate the superior stability and comprehensiveness of VidCapBench compared to existing video captioning evaluation approaches. Verification with off-the-shelf T2V models reveals a significant positive correlation between scores on VidCapBench and the T2V quality evaluation metrics, indicating that VidCapBench can provide valuable guidance for training T2V models. The project is available at https://github.com/VidCapBench/VidCapBench.

Hao Yi, Qingyang Li, Yulan Hu et al.

Recently, enhancing the numerical and logical reasoning capability of Large Language Models (LLMs) has emerged as a research hotspot. Existing methods face several limitations: inference-phase techniques (e.g., Chain of Thoughts) rely on prompt selection and the pretrained knowledge; sentence-level Supervised Fine-Tuning (SFT) and Direct Preference Optimization (DPO) struggle with step-wise mathematical correctness and depend on stronger models distillation or human annotations; while Reinforcement Learning (RL) approaches incur high GPU memory costs and unstable training. To address these, we propose \textbf{S}elf-training framework integrating \textbf{P}rocess \textbf{P}reference learning using \textbf{D}ynamic value margin (SPPD). SPPD leverages a process-based Markov Decision Process (MDP) and Bellman optimality equation to derive \textbf{dynamic value margin} on step-level preference optimization, which employs tree-based self-sampling on model responses \textbf{without any distillation} from other models. Furthermore, we theoretically prove that SPPD is \textbf{equivalent to on-policy policy gradient methods} under reward constraints. Experiments on 7B-scale models demonstrate superior performance across in-domain and out-domain mathematical benchmarks. We open-source our code at \href{https://anonymous.4open.science/r/SSDPO-D-DCDD}{https://anonymous.4open.science/r/SPPD-DCDD}.

Di Zhang, Wei Liu, Qian Tan et al.

Large language models (LLMs) have made impressive progress in chemistry applications. However, the community lacks an LLM specifically designed for chemistry. The main challenges are two-fold: firstly, most chemical data and scientific knowledge are stored in structured databases, which limits the model's ability to sustain coherent dialogue when used directly. Secondly, there is an absence of objective and fair benchmark that encompass most chemistry tasks. Here, we introduce ChemLLM, a comprehensive framework that features the first LLM dedicated to chemistry. It also includes ChemData, a dataset specifically designed for instruction tuning, and ChemBench, a robust benchmark covering nine essential chemistry tasks. ChemLLM is adept at performing various tasks across chemical disciplines with fluid dialogue interaction. Notably, ChemLLM achieves results comparable to GPT-4 on the core chemical tasks and demonstrates competitive performance with LLMs of similar size in general scenarios. ChemLLM paves a new path for exploration in chemical studies, and our method of incorporating structured chemical knowledge into dialogue systems sets a new standard for developing LLMs in various scientific fields. Codes, Datasets, and Model weights are publicly accessible at https://hf.co/AI4Chem

Luozhou Wang, Ziyang Mai, Guibao Shen et al.

In this work, we present a novel approach for motion customization in video generation, addressing the widespread gap in the exploration of motion representation within video generative models. Recognizing the unique challenges posed by the spatiotemporal nature of video, our method introduces Motion Embeddings, a set of explicit, temporally coherent embeddings derived from a given video. These embeddings are designed to integrate seamlessly with the temporal transformer modules of video diffusion models, modulating self-attention computations across frames without compromising spatial integrity. Our approach provides a compact and efficient solution to motion representation, utilizing two types of embeddings: a Motion Query-Key Embedding to modulate the temporal attention map and a Motion Value Embedding to modulate the attention values. Additionally, we introduce an inference strategy that excludes spatial dimensions from the Motion Query-Key Embedding and applies a differential operation to the Motion Value Embedding, both designed to debias appearance and ensure the embeddings focus solely on motion. Our contributions include the introduction of a tailored motion embedding for customization tasks and a demonstration of the practical advantages and effectiveness of our method through extensive experiments.

Wei Chen, Xin Yan, Bin Wen et al.

Although multimodal large language models (MLLMs) exhibit remarkable reasoning capabilities on complex multimodal understanding tasks, they still suffer from the notorious hallucination issue: generating outputs misaligned with obvious visual or factual evidence. Currently, training-based solutions, like direct preference optimization (DPO), leverage paired preference data to suppress hallucinations. However, they risk sacrificing general reasoning capabilities due to the likelihood displacement. Meanwhile, training-free solutions, like contrastive decoding, achieve this goal by subtracting the estimated hallucination pattern from a distorted input. Yet, these handcrafted perturbations (e.g., add noise to images) may poorly capture authentic hallucination patterns. To avoid these weaknesses of existing methods, and realize robust hallucination mitigation (i.e., maintaining general reasoning performance), we propose a novel framework: Decoupling Contrastive Decoding (DCD). Specifically, DCD decouples the learning of positive and negative samples in preference datasets, and trains separate positive and negative image projections within the MLLM. The negative projection implicitly models real hallucination patterns, which enables vision-aware negative images in the contrastive decoding inference stage. Our DCD alleviates likelihood displacement by avoiding pairwise optimization and generalizes robustly without handcrafted degradation. Extensive ablations across hallucination benchmarks and general reasoning tasks demonstrate the effectiveness of DCD, i.e., it matches DPO's hallucination suppression while preserving general capabilities and outperforms the handcrafted contrastive decoding methods.

Hao Yi, Qingyang Li, Yulan Hu et al.

High-quality video-text preference data is crucial for Multimodal Large Language Models (MLLMs) alignment. However, existing preference data is very scarce. Obtaining VQA preference data for preference training is costly, and manually annotating responses is highly unreliable, which could result in low-quality pairs. Meanwhile, AI-generated responses controlled by temperature adjustment lack diversity. To address these issues, we propose a high-quality VQA preference dataset, called \textit{\textbf{M}ultiple \textbf{M}ultimodal \textbf{A}rtificial \textbf{I}ntelligence \textbf{P}reference Datasets in \textbf{V}QA} (\textbf{MMAIP-V}), which is constructed by sampling from the response distribution set and using an external scoring function for response evaluation. Furthermore, to fully leverage the preference knowledge in MMAIP-V and ensure sufficient optimization, we propose \textit{\textbf{Iter}ative \textbf{W}eak-to-\textbf{S}trong \textbf{R}einforcement \textbf{L}earning from \textbf{AI} \textbf{F}eedback for video MLLMs} (\textbf{Iter-W2S-RLAIF}), a framework that gradually enhances MLLMs' alignment capabilities by iteratively updating the reference model and performing parameter extrapolation. Finally, we propose an unbiased and information-complete evaluation scheme in VQA evaluation. Experiments demonstrate that MMAIP-V is beneficial for MLLMs in preference learning and Iter-W2S-RLAIF fully exploits the alignment information in MMAIP-V. We believe that the proposed automatic VQA preference data generation pipeline based on AI feedback can greatly promote future work in the MLLMs alignment. \textbf{Code and dataset are available} \href{https://anonymous.4open.science/r/MMAIP-V_Iter-W2S-RLAIF-702F}{MMAIP-V\_Iter-W2S-RLAIF-702F}.

Kaiwen Wei, Jingyuan Zhang, Hongzhi Zhang et al.

Large Language Models (LLMs) exhibit remarkable generative capabilities, enabling the generation of valuable information. Despite these advancements, previous research found that LLMs sometimes struggle with adhering to specific constraints (e.g., in specific place or at specific time), at times even overlooking them, which leads to responses that are either too generic or not fully satisfactory. Existing approaches attempted to address this issue by decomposing or rewriting input instructions, yet they fall short in adequately emphasizing specific constraints and in unlocking the underlying knowledge (e.g., programming within the context of software development). In response, this paper proposes a simple yet effective method named Chain-of-Specificity (CoS). Specifically, CoS iteratively emphasizes the specific constraints in the input instructions, unlocks knowledge within LLMs, and refines responses. Experiments conducted on publicly available and self-build complex datasets demonstrate that CoS outperforms existing methods in enhancing generated content especially for the specificity. Besides, as the number of specific constraints increase, other baselines falter, while CoS still performs well. Moreover, we show that distilling responses generated by CoS effectively enhances the ability of smaller models to follow the constrained instructions. Resources of this paper will be released for further research.