Adaptive Dense Reward: Understanding the Gap Between Action and Reward Space in Alignment

Reinforcement Learning from Human Feedback (RLHF) has proven highly effective in aligning Large Language Models (LLMs) with human preferences. However, the original RLHF typically optimizes under an overall reward, which can lead to a suboptimal learning process. This limitation stems from RLHF's lack of awareness regarding which specific tokens should be reinforced or suppressed. Moreover, conflicts in supervision can arise, for instance, when a chosen response includes erroneous tokens, while a rejected response contains accurate elements. To rectify these shortcomings, increasing dense reward methods, such as step-wise and token-wise RLHF, have been proposed. However, these existing methods are limited to specific tasks (like mathematics). In this paper, we propose the ``Adaptive Message-wise RLHF'' method, which robustly applies to various tasks. By defining pivot tokens as key indicators, our approach adaptively identifies essential information and converts sample-level supervision into fine-grained, subsequence-level supervision. This aligns the density of rewards and action spaces more closely with the information density of the input. Experiments demonstrate that our method can be integrated into various training methods, significantly mitigating hallucinations and catastrophic forgetting problems while outperforming other methods on multiple evaluation metrics. Our method improves the success rate on adversarial samples by 10\% compared to the sample-wise approach and achieves a 1.3\% improvement on evaluation benchmarks such as MMLU, GSM8K, and HumanEval et al.

FlatQuant: Flatness Matters for LLM Quantization

Recently, quantization has been widely used for the compression and acceleration of large language models~(LLMs). Due to the outliers in LLMs, it is crucial to flatten weights and activations to minimize quantization error with the equally spaced quantization points. Prior research explores various pre-quantization transformations to suppress outliers, such as per-channel scaling and Hadamard transformation. However, we observe that these transformed weights and activations can still remain steep and outspread. In this paper, we propose FlatQuant (Fast and Learnable Affine Transformation), a new post-training quantization approach to enhance flatness of weights and activations. Our approach identifies optimal affine transformations tailored to each linear layer, calibrated in hours via a lightweight objective. To reduce runtime overhead, we apply Kronecker decomposition to the transformation matrices, and fuse all operations in FlatQuant into a single kernel. Extensive experiments show that FlatQuant sets up a new state-of-the-art quantization benchmark. For instance, it achieves less than $\textbf{1}\%$ accuracy drop for W4A4 quantization on the LLaMA-3-70B model, surpassing SpinQuant by $\textbf{7.5}\%$. For inference latency, FlatQuant reduces the slowdown induced by pre-quantization transformation from 0.26x of QuaRot to merely $\textbf{0.07x}$, bringing up to $\textbf{2.3x}$ speedup for prefill and $\textbf{1.7x}$ speedup for decoding, respectively. Code is available at: \url{https://github.com/ruikangliu/FlatQuant}.

Kendall's $τ$ Coefficient for Logits Distillation

Knowledge distillation typically employs the Kullback-Leibler (KL) divergence to constrain the student model's output to match the soft labels provided by the teacher model exactly. However, sometimes the optimization direction of the KL divergence loss is not always aligned with the task loss, where a smaller KL divergence could lead to erroneous predictions that diverge from the soft labels. This limitation often results in suboptimal optimization for the student. Moreover, even under temperature scaling, the KL divergence loss function tends to overly focus on the larger-valued channels in the logits, disregarding the rich inter-class information provided by the multitude of smaller-valued channels. This hard constraint proves too challenging for lightweight students, hindering further knowledge distillation. To address this issue, we propose a plug-and-play ranking loss based on Kendall's $\tau$ coefficient, called Rank-Kendall Knowledge Distillation (RKKD). RKKD balances the attention to smaller-valued channels by constraining the order of channel values in student logits, providing more inter-class relational information. The rank constraint on the top-valued channels helps avoid suboptimal traps during optimization. We also discuss different differentiable forms of Kendall's $\tau$ coefficient and demonstrate that the proposed ranking loss function shares a consistent optimization objective with the KL divergence. Extensive experiments on the CIFAR-100 and ImageNet datasets show that our RKKD can enhance the performance of various knowledge distillation baselines and offer broad improvements across multiple teacher-student architecture combinations.

TextureDiffusion: Target Prompt Disentangled Editing for Various Texture Transfer

Recently, text-guided image editing has achieved significant success. However, existing methods can only apply simple textures like wood or gold when changing the texture of an object. Complex textures such as cloud or fire pose a challenge. This limitation stems from that the target prompt needs to contain both the input image content and <texture>, restricting the texture representation. In this paper, we propose TextureDiffusion, a tuning-free image editing method applied to various texture transfer. Initially, the target prompt is directly set to "<texture>", making the texture disentangled from the input image content to enhance texture representation. Subsequently, query features in self-attention and features in residual blocks are utilized to preserve the structure of the input image. Finally, to maintain the background, we introduce an edit localization technique which blends the self-attention results and the intermediate latents. Comprehensive experiments demonstrate that TextureDiffusion can harmoniously transfer various textures with excellent structure and background preservation.

Learning to Compress Contexts for Efficient Knowledge-based Visual Question Answering

Multimodal Large Language Models (MLLMs) have demonstrated great zero-shot performance on visual question answering (VQA). However, when it comes to knowledge-based VQA (KB-VQA), MLLMs may lack human commonsense or specialized domain knowledge to answer such questions and require obtaining necessary information from external knowledge sources. Previous works like Retrival-Augmented VQA-v2 (RAVQA-v2) focus on utilizing as much input information, such as image-based textual descriptions and retrieved knowledge, as possible to improve performance, but they all overlook the issue that with the number of input tokens increasing, inference efficiency significantly decreases, which contradicts the demands of practical applications. To address this issue, we propose Retrieval-Augmented MLLM with Compressed Contexts (RACC). RACC learns to compress and aggregate retrieved contexts, from which it generates a compact modulation in the form of Key-Value (KV) cache. This modulation is then used to adapt the downstream frozen MLLM, thereby achieving effective and efficient inference. RACC achieves a state-of-the-art (SOTA) performance of 62.9% on OK-VQA. Moreover, it significantly reduces inference latency by 22.0%-59.7% compared to the prominent RAVQA-v2. Abundant experiments show RACC's broad applicability. It is compatible with various off-the-shelf MLLMs and can also handle different knowledge sources including textual and multimodal documents.

ChartMoE: Mixture of Expert Connector for Advanced Chart Understanding

Automatic chart understanding is crucial for content comprehension and document parsing. Multimodal large language models (MLLMs) have demonstrated remarkable capabilities in chart understanding through domain-specific alignment and fine-tuning. However, the application of alignment training within the chart domain is still underexplored. To address this, we propose ChartMoE, which employs the mixture of expert (MoE) architecture to replace the traditional linear projector to bridge the modality gap. Specifically, we train multiple linear connectors through distinct alignment tasks, which are utilized as the foundational initialization parameters for different experts. Additionally, we introduce ChartMoE-Align, a dataset with over 900K chart-table-JSON-code quadruples to conduct three alignment tasks (chart-table/JSON/code). Combined with the vanilla connector, we initialize different experts in four distinct ways and adopt high-quality knowledge learning to further refine the MoE connector and LLM parameters. Extensive experiments demonstrate the effectiveness of the MoE connector and our initialization strategy, e.g., ChartMoE improves the accuracy of the previous state-of-the-art from 80.48% to 84.64% on the ChartQA benchmark.

Learn To Learn More Precisely

Meta-learning has been extensively applied in the domains of few-shot learning and fast adaptation, achieving remarkable performance. While Meta-learning methods like Model-Agnostic Meta-Learning (MAML) and its variants provide a good set of initial parameters for the model, the model still tends to learn shortcut features, which leads to poor generalization. In this paper, we propose the formal conception of "learn to learn more precisely", which aims to make the model learn precise target knowledge from data and reduce the effect of noisy knowledge, such as background and noise. To achieve this target, we proposed a simple and effective meta-learning framework named Meta Self-Distillation(MSD) to maximize the consistency of learned knowledge, enhancing the models' ability to learn precise target knowledge. In the inner loop, MSD uses different augmented views of the same support data to update the model respectively. Then in the outer loop, MSD utilizes the same query data to optimize the consistency of learned knowledge, enhancing the model's ability to learn more precisely. Our experiment demonstrates that MSD exhibits remarkable performance in few-shot classification tasks in both standard and augmented scenarios, effectively boosting the accuracy and consistency of knowledge learned by the model.

Task Groupings Regularization: Data-Free Meta-Learning with Heterogeneous Pre-trained Models

Data-Free Meta-Learning (DFML) aims to derive knowledge from a collection of pre-trained models without accessing their original data, enabling the rapid adaptation to new unseen tasks. Current methods often overlook the heterogeneity among pre-trained models, which leads to performance degradation due to task conflicts. In this paper, we empirically and theoretically identify and analyze the model heterogeneity in DFML. We find that model heterogeneity introduces a heterogeneity-homogeneity trade-off, where homogeneous models reduce task conflicts but also increase the overfitting risk. Balancing this trade-off is crucial for learning shared representations across tasks. Based on our findings, we propose Task Groupings Regularization, a novel approach that benefits from model heterogeneity by grouping and aligning conflicting tasks. Specifically, we embed pre-trained models into a task space to compute dissimilarity, and group heterogeneous models together based on this measure. Then, we introduce implicit gradient regularization within each group to mitigate potential conflicts. By encouraging a gradient direction suitable for all tasks, the meta-model captures shared representations that generalize across tasks. Comprehensive experiments showcase the superiority of our approach in multiple benchmarks, effectively tackling the model heterogeneity in challenging multi-domain and multi-architecture scenarios.

MirrorGaussian: Reflecting 3D Gaussians for Reconstructing Mirror Reflections

3D Gaussian Splatting showcases notable advancements in photo-realistic and real-time novel view synthesis. However, it faces challenges in modeling mirror reflections, which exhibit substantial appearance variations from different viewpoints. To tackle this problem, we present MirrorGaussian, the first method for mirror scene reconstruction with real-time rendering based on 3D Gaussian Splatting. The key insight is grounded on the mirror symmetry between the real-world space and the virtual mirror space. We introduce an intuitive dual-rendering strategy that enables differentiable rasterization of both the real-world 3D Gaussians and the mirrored counterpart obtained by reflecting the former about the mirror plane. All 3D Gaussians are jointly optimized with the mirror plane in an end-to-end framework. MirrorGaussian achieves high-quality and real-time rendering in scenes with mirrors, empowering scene editing like adding new mirrors and objects. Comprehensive experiments on multiple datasets demonstrate that our approach significantly outperforms existing methods, achieving state-of-the-art results. Project page: https://mirror-gaussian.github.io/.

FREE: Faster and Better Data-Free Meta-Learning

Data-Free Meta-Learning (DFML) aims to extract knowledge from a collection of pre-trained models without requiring the original data, presenting practical benefits in contexts constrained by data privacy concerns. Current DFML methods primarily focus on the data recovery from these pre-trained models. However, they suffer from slow recovery speed and overlook gaps inherent in heterogeneous pre-trained models. In response to these challenges, we introduce the Faster and Better Data-Free Meta-Learning (FREE) framework, which contains: (i) a meta-generator for rapidly recovering training tasks from pre-trained models; and (ii) a meta-learner for generalizing to new unseen tasks. Specifically, within the module Faster Inversion via Meta-Generator, each pre-trained model is perceived as a distinct task. The meta-generator can rapidly adapt to a specific task in just five steps, significantly accelerating the data recovery. Furthermore, we propose Better Generalization via Meta-Learner and introduce an implicit gradient alignment algorithm to optimize the meta-learner. This is achieved as aligned gradient directions alleviate potential conflicts among tasks from heterogeneous pre-trained models. Empirical experiments on multiple benchmarks affirm the superiority of our approach, marking a notable speed-up (20$\times$) and performance enhancement (1.42\% $\sim$ 4.78\%) in comparison to the state-of-the-art.