| Total: 3178
Large Language Models (LLMs) and Vision Language Models (VLMs) possess extensive knowledge and exhibit promising reasoning abilities, however, they still struggle to perform well in complex, dynamic environments. Real-world tasks require handling intricate interactions, advanced spatial reasoning, long-term planning, and continuous exploration of new strategies—areas in which we lack effective methodologies for comprehensively evaluating these capabilities. To address this gap, we introduce AgentQuest, a novel benchmark designed to assess the agentic capabilities of LLMs and VLMs through a diverse set of challenging games. Our benchmark incorporates a range of existing reinforcement learning environments with varying levels of difficulty, including tasks that are solvable by non-expert humans in seconds to extremely challenging ones that may take years to master (e.g., the NetHack Learning Environment). We devise fine-grained metrics to measure performance and conduct an extensive evaluation of several popular open-source and closed-source LLMs and VLMs. Our findings indicate that while current models achieve partial success in the easier games, they struggle significantly with more challenging tasks. Notably, we observe severe deficiencies in vision-based decision-making, as models perform worse when visual representations of the environments are provided. We release AgentQuest as an open and user-friendly benchmark to facilitate future research and development in the agentic community.
Masked transformer models for class-conditional image generation have become a compelling alternative to diffusion models. Typically comprising two stages - an initial VQGAN model for transitioning between latent space and image space, and a subsequent Transformer model for image generation within latent space - these frameworks offer promising avenues for image synthesis. In this study, we present two primary contributions: Firstly, an empirical and systematic examination of VQGANs, leading to a modernized VQGAN. Secondly, a novel embedding-free generation network operating directly on bit tokens - a binary quantized representation of tokens with rich semantics. The first contribution furnishes a transparent, reproducible, and high-performing VQGAN model, enhancing accessibility and matching the performance of current state-of-the-art methods while revealing previously undisclosed details. The second contribution demonstrates that embedding-free image generation using bit tokens achieves a new state-of-the-art FID of 1.52 on the ImageNet $256\times256$ benchmark, with a compact generator model of mere 305M parameters. The code for this project is available on https://github.com/markweberdev/maskbit.
The widespread deployment of sensing devices leads to a surge in data for spatio-temporal forecasting applications such as traffic flow, air quality, and wind energy. Although spatio-temporal graph neural networks (STGNNs) have achieved success in modeling various static spatio-temporal forecasting scenarios, real-world spatio-temporal data are typically received in a streaming manner, and the network continuously expands with the installation of new sensors. Thus, spatio-temporal forecasting in streaming scenarios faces dual challenges: the inefficiency of retraining models over newly-arrived data and the detrimental effects of catastrophic forgetting over long-term history. To address these challenges, we propose a novel prompt tuning-based continuous forecasting method, **_EAC_**, following two fundamental tuning principles guided by empirical and theoretical analysis: _**e**xpand **a**nd **c**ompress_, which effectively resolve the aforementioned problems with lightweight tuning parameters. Specifically, we integrate the base STGNN with a continuous prompt pool, utilizing stored prompts (\ie, few learnable parameters) in memory, and jointly optimize them with the base STGNN. This method ensures that the model sequentially learns from the spatio-temporal data stream to accomplish tasks for corresponding periods. Extensive experimental results on multiple real-world datasets demonstrate the multi-faceted superiority of **_EAC_** over the state-of-the-art baselines, including effectiveness, efficiency, universality, etc.
When reading a document, glancing at the spatial layout of a document is an initial step to understand it roughly. Traditional document layout analysis (DLA) methods, however, offer only a superficial parsing of documents, focusing on basic instance detection and often failing to capture the nuanced spatial and logical relationships between instances. These limitations hinder DLA-based models from achieving a gradually deeper comprehension akin to human reading. In this work, we propose a novel graph-based Document Structure Analysis (gDSA) task. This task requires that model not only detects document elements but also generates spatial and logical relations in form of a graph structure, allowing to understand documents in a holistic and intuitive manner. For this new task, we construct a relation graph-based document structure analysis dataset(GraphDoc) with 80K document images and 4.13M relation annotations, enabling training models to complete multiple tasks like reading order, hierarchical structures analysis, and complex inter-element relationship inference. Furthermore, a document relation graph generator (DRGG) is proposed to address the gDSA task, which achieves performance with 57.6% at $mAP_g$@$0.5$ for a strong benchmark baseline on this novel task and dataset. We hope this graphical representation of document structure can mark an innovative advancement in document structure analysis and understanding. The new dataset and code will be made publicly available.
Positional Encodings (PEs) are a critical component of Transformer-based Large Language Models (LLMs), providing the attention mechanism with important sequence-position information. One of the most popular types of encoding used today in LLMs are Rotary Positional Encodings (RoPE), that rotate the queries and keys based on their relative distance. A common belief is that RoPE is useful because it helps to decay token dependency as relative distance increases. In this work, we argue that this is unlikely to be the core reason. We study the internals of a trained Gemma 7B model to understand how RoPE is being used at a mechanical level. We find that Gemma learns to use RoPE to construct robust `positional' attention patterns by exploiting the highest frequencies. We also find that, in general, Gemma greatly prefers to use the lowest frequencies of RoPE, which we suspect are used to carry semantic information. We mathematically prove interesting behaviours of RoPE and conduct experiments to verify our findings, proposing a modification of RoPE that fixes some highlighted issues and improves performance. We believe that this work represents an interesting step in better understanding PEs in LLMs, which we believe holds crucial value for scaling LLMs to large sizes and context lengths.
Conventional image compression methods are inadequate for intelligent analysis, as they overemphasize pixel-level precision while neglecting semantic significance and the interaction among multiple tasks. This paper introduces a Taskonomy-Aware Multi-Task Compression framework comprising (1) inter-coherent task grouping, which organizes synergistic tasks into shared representations to improve multi-task accuracy and reduce encoding volume, and (2) a conditional entropy-based directed acyclic graph (DAG) that captures causal dependencies among grouped representations. By leveraging parent representations as contextual priors for child representations, the framework effectively utilizes cross-task information to improve entropy model accuracy. Experiments on diverse vision tasks, including Keypoint 2D, Depth Zbuffer, Semantic Segmentation, Surface Normal, Edge Texture, and Autoencoder, demonstrate significant bitrate-performance gains, validating the method’s capability to reduce system entropy uncertainty. These findings underscore the potential of leveraging representation disentanglement, synergy, and causal modelling for compact representation learning, enabling efficient multi-task compression in intelligent systems. Code will be available.
Out-of-Distribution (OOD) detection, aiming to distinguish outliers from known categories, has gained prominence in practical scenarios. Recently, the advent of vision-language models (VLM) has heightened interest in enhancing OOD detection for VLM through few-shot tuning. However, existing methods mainly focus on optimizing global prompts, ignoring refined utilization of local information with regard to outliers. Motivated by this, we freeze global prompts and introduce Local-Prompt, a novel coarse-to-fine tuning paradigm to emphasize regional enhancement with local prompts. Our method comprises two integral components: global prompt guided negative augmentation and local prompt enhanced regional regularization. The former utilizes frozen, coarse global prompts as guiding cues to incorporate negative augmentation, thereby leveraging local outlier knowledge. The latter employs trainable local prompts and a regional regularization to capture local information effectively, aiding in outlier identification. We also propose regional-related metric to empower the enrichment of OOD detection. Moreover, since our approach explores enhancing local prompts only, it can be seamlessly integrated with trained global prompts during inference to boost the performance. Comprehensive experiments demonstrate the effectiveness and potential of our method. Notably, our method reduces average FPR95 by 5.17% against state-of-the-art method in 4-shot tuning on challenging ImageNet-1k dataset, even outperforming 16-shot results of previous methods.
The ever-increasing fine-tuning cost of large-scale pre-trained models gives rise to the importance of dataset pruning, which aims to reduce dataset size while maintaining task performance.However, existing dataset pruning methods require training on the entire dataset, which is impractical for large-scale pre-trained models.In this paper, we propose a straightforward, novel, and training-free hardness score named Distorting-based Learning Complexity (DLC), to identify informative images and instructions from the downstream dataset efficiently.Our method is motivated by the observation that easy samples learned faster can also be learned with fewer parameters.Specifically, we define the Learning Complexity to quantify sample hardness and utilize a lightweight weights masking process for fast estimation, instead of the costly SGD optimization.Based on DLC, we further design a flexible under-sampling strategy with randomness (dubbed FlexRand), replacing the top-K strategy, to alleviate the severe subset distribution shift.Extensive experiments with downstream image and instructions dataset pruning benchmarks demonstrate the effectiveness and efficiency of the proposed approach.In the images pruning benchmark, DLC significantly reduces the pruning time by 35$\times$ while establishing state-of-the-art performance with FlexRand.
Mamba-based architectures have shown to be a promising new direction for deep learning models owing to their competitive performance and sub-quadratic deployment speed. However, current Mamba multi-modal large language models (MLLM) are insufficient in extracting visual features, leading to imbalanced cross-modal alignment between visual and textural latents, negatively impacting performance on multi-modal tasks. In this work, we propose Empowering Multi-modal Mamba with Structural and Hierarchical Alignment (EMMA), which enables the MLLM to extract fine-grained visual information. Specifically, we propose a pixel-wise alignment module to autoregressively optimize the learning and processing of spatial image-level features along with textual tokens, enabling structural alignment at the image level. In addition, to prevent the degradation of visual information during the cross-model alignment process, we propose a multi-scale feature fusion (MFF) module to combine multi-scale visual features from intermediate layers, enabling hierarchical alignment at the feature level. Extensive experiments are conducted across a variety of multi-modal benchmarks. Our model shows lower latency than other Mamba-based MLLMs and is nearly four times faster than transformer-based MLLMs of similar scale during inference. Due to better cross-modal alignment, our model exhibits lower degrees of hallucination and enhanced sensitivity to visual details, which manifests in superior performance across diverse multi-modal benchmarks. Code will be provided.
As a neural network's depth increases, it can achieve high generalization performance. However, training deep networks is challenging due to gradient and signal propagation issues. To address these challenges, extensive theoretical research and various methods have been introduced. Despite these advances, effective weight initialization methods for tanh neural networks remain underexplored. This paper presents a novel weight initialization method for Neural Networks with tanh activation function. Based on an analysis of the fixed points of the function $\tanh(ax)$, our proposed method aims to determine values of $a$ that mitigate activation saturations. A series of experiments on various classification datasets and Physics-Informed Neural Networks demonstrate that the proposed method outperforms Xavier initialization methods (with or without normalization) in terms of robustness to network size variations, data efficiency, and convergence speed.
We study a class of decision-making problems with one-sided feedback, where outcomes are only observable for specific actions. A typical example is bank loans, where the repayment status is known only if a loan is approved and remains undefined if rejected. In such scenarios, conventional approaches to causal decision evaluation and learning from observational data are not directly applicable. In this paper, we introduce a novel value function to evaluate decision rules that addresses the issue of undefined counterfactual outcomes. Without assuming no unmeasured confounders, we establish the identification of the value function using shadow variables. Furthermore, leveraging semiparametric theory, we derive the efficiency bound for the proposed value function and develop efficient methods for decision evaluation and learning. Numerical experiments and a real-world data application demonstrate the empirical performance of our proposed methods.
Unsupervised anomaly detection (UAD) has important applications in diverse fields such as manufacturing industry and medical diagnosis. In the past decades, although numerous insightful and effective UAD methods have been proposed, it remains a huge challenge to tune the hyper-parameters of each method and select the most appropriate method among many candidates for a specific dataset, due to the absence of labeled anomalies in the training phase of UAD methods and the high diversity of real datasets. In this work, we aim to address this challenge, so as to make UAD more practical and reliable. We propose two internal evaluation metrics, \textit{relative-top-median} and \textit{expected-anomaly-gap}, and one semi-internal evaluation metric, \textit{normalized pseudo discrepancy} (NPD), as surrogate functions of the expected model performance on unseen test data. For instance, NPD measures the discrepancy between the anomaly scores of a validation set drawn from the training data and a validation set drawn from an isotropic Gaussian. NPD is simple and hyper-parameter-free and is able to compare different UAD methods, and its effectiveness is theoretically analyzed. We integrate the three metrics with Bayesian optimization to effectively optimize the hyper-parameters of UAD models. Extensive experiments on 38 datasets show the effectiveness of our methods.
Learning from offline data without interacting with the environment is a promising way to fully leverage the intelligent decision-making capabilities of multi-agent reinforcement learning (MARL). Previous approaches have primarily focused on developing learning techniques, such as conservative methods tailored to MARL using limited offline data. However, these methods often overlook the temporal relationships across different timesteps and spatial relationships between teammates, resulting in low learning efficiency in imbalanced data scenarios. To comprehensively explore the data structure of MARL and enhance learning efficiency, we propose Multi-Agent offline coordination via Diffusion-based Trajectory Stitching (MADiTS), a novel diffusion-based data augmentation pipeline that systematically generates trajectories by stitching high-quality coordination segments together. MADiTS first generates trajectory segments using a trained diffusion model, followed by applying a bidirectional dynamics constraint to ensure that the trajectories align with environmental dynamics. Additionally, we develop an offline credit assignment technique to identify and optimize the behavior of underperforming agents in the generated segments. This iterative procedure continues until a satisfactory augmented episode trajectory is generated within the predefined limit or is discarded otherwise. Empirical results on imbalanced datasets of multiple benchmarks demonstrate that MADiTS significantly improves MARL performance.
Unlike traditional vision-only models, vision language models (VLMs) offer an intuitive way to access visual content through language prompting by combining a large language model (LLM) with a vision encoder. However, both the LLM and the vision encoder come with their own set of biases, cue preferences, and shortcuts, which have been rigorously studied in uni-modal models. A timely question is how such (potentially misaligned) biases and cue preferences behave under multi-modal fusion in VLMs. As a first step towards a better understanding, we investigate a particularly well-studied vision-only bias - the texture vs. shape bias and the dominance of local over global information. As expected, we find that VLMs inherit this bias to some extent from their vision encoders. Surprisingly, the multi-modality alone proves to have important effects on the model behavior, i.e., the joint training and the language querying change the way visual cues are processed. While this direct impact of language-informed training on a model's visual perception is intriguing, it raises further questions on our ability to actively steer a model's output so that its prediction is based on particular visual cues of the user's choice. Interestingly, VLMs have an inherent tendency to recognize objects based on shape information, which is different from what a plain vision encoder would do. Further active steering towards shape-based classifications through language prompts is however limited. In contrast, active VLM steering towards texture-based decisions through simple natural language prompts is often more successful.
We propose a novel formulation for dictionary learning with an overcomplete dictionary, i.e., when the number of atoms is larger than the dimension of the dictionary. The proposed formulation consists of a weighted sum of $\ell_1$ norms of the rows of the sparse coefficient matrix plus the log of the matrix volume of the dictionary matrix. The main contribution of this work is to show that this novel formulation guarantees global identifiability of the overcomplete dictionary, under a mild condition that the sparse coefficient matrix satisfies a strong scattering condition in the hypercube. Furthermore, if every column of the coefficient matrix is sparse and the dictionary guarantees $\ell_1$ recovery, then the coefficient matrix is identifiable as well. This is a major breakthrough for not only dictionary learning but also general matrix factorization models as identifiability is guaranteed even when the latent dimension is higher than the ambient dimension. We also provide a probabilistic analysis and show that if the sparse coefficient matrix is generated from the widely adopted sparse-Gaussian model, then the $m\times k$ overcomplete dictionary is globally identifiable if the sample size is bigger than a constant times $(k^2/m)\log(k^2/m)$, where $k$ is the number of atoms in the dictionary, with overwhelming probability. Finally, we propose an algorithm based on alternating minimization to solve the new proposed formulation.
The large-scale training of multi-modal models on data scraped from the web has shown outstanding utility in infusing these models with the required world knowledge to perform effectively on multiple downstream tasks. However, one downside of scraping data from the web can be the potential sacrifice of the benchmarks on which the abilities of these models are often evaluated. To safeguard against test data contamination and to truly test the abilities of these foundation models we propose LiveXiv: A scalable evolving live benchmark based on scientific ArXiv papers. LiveXiv accesses domain-specific manuscripts at any given timestamp and proposes to automatically generate visual question-answer pairs (VQA). This is done without any human-in-the-loop, using the multi-modal content in the manuscripts, like graphs, charts, and tables. Moreover, we introduce an efficient evaluation approach that estimates the performance of all models on the evolving benchmark using evaluations of only a subset of models. This significantly reduces the overall evaluation cost. We benchmark multiple open and proprietary Large Multi-modal Models (LMMs) on the first version of our benchmark, showing its challenging nature and exposing the models’ true abilities, avoiding contamination. Lastly, in our commitment to high quality, wehave collected and evaluated a manually verified subset. By comparing its overall results to our automatic annotations, we have found that the performance variance is indeed minimal (<2.5%). Our dataset is available online anonymously on HuggingFace.
The learning dynamics of neural networks—in particular, how parameters change over time during training—describe how data, architecture, and algorithm interact in time to produce a trained neural network model. Characterizing these dynamics, in general, remains an open problem in machine learning, but, handily, restricting the setting allows careful empirical studies and even analytical results. In this blog post, we review approaches to analyzing the learning dynamics of nonlinear neural networks, focusing on a particular setting known as *teacher-student* that permits an explicit analytical expression for the generalization error of a nonlinear neural network trained with online gradient descent. We provide an accessible mathematical formulation of this analysis and a `JAX` codebase to implement simulation of the analytical system of ordinary differential equations alongside neural network training in this setting. We conclude with a discussion of how this analytical paradigm has been used to investigate generalization in neural networks and beyond.
Compositional generalization (the ability to respond correctly to novel combinations of familiar components) is thought to be a cornerstone of intelligent behavior. Compositionally structured (e.g. disentangled) representations are essential for this; however, the conditions under which they yield compositional generalization remain unclear. To address this gap, we present a general theory of compositional generalization in kernel models with fixed, compositionally structured representations, a tractable framework for characterizing the impact of dataset statistics on generalization. We find that these models are constrained to adding up values assigned to each combination of components seen during training ("conjunction-wise additivity"). This imposes fundamental restrictions on the set of tasks compositionally structured kernel models can learn, in particular preventing them from transitively generalizing equivalence relations. Even for compositional tasks that they can in principle learn, we identify novel failure modes in compositional generalization that arise from biases in the training data and affect important compositional building blocks such as symbolic addition and context dependence (memorization leak and shortcut bias). Finally, we empirically validate our theory, showing that it captures the behavior of deep neural networks (convolutional networks, residual networks, and Vision Transformers) trained on a set of compositional tasks with similarly structured data. Ultimately, this work provides a theoretical perspective on how statistical structure in the training data can affect compositional generalization, with implications for how to identify and remedy failure modes in deep learning models.
The stochastic three points (STP) algorithm is a derivative-free optimization technique designed for unconstrained optimization problems in $\mathbb{R}^d$. In this paper, we analyze this algorithm for three classes of functions : smooth functions that may lack convexity, smooth convex functions, and smooth functions that are strongly convex. Our work provides the first almost sure convergence results of the STP algorithm, alongside some convergence results in expectation.For the class of smooth functions, we establish that the best gradient iterate of the STP algorithm converges almost surely to zero at a rate arbitrarily close to $o(\frac{1}{\sqrt{T}})$, where $T$ is the number of iterations. Furthermore, within the same class of functions, we establish both almost sure convergence and convergence in expectation of the final gradient iterate towards zero.For the class of smooth convex functions, we establish that $f(\theta^T)$ converges to $\inf_{\theta \in \mathbb{R}^d} f(\theta)$ almost surely at a rate arbitrarily close to $o(\frac{1}{T})$, and in expectation at a rate of $O(\frac{d}{T})$ where $d$ is the dimension of the space.Finally, for the class of smooth functions that are strongly convex, we establish that when step sizes are obtained by approximating the directional derivatives of the function, $f(\theta^T)$ converges to $\inf_{\theta \in \mathbb{R}^d} f(\theta)$ in expectation at a rate of $O((1-\frac{\mu}{dL})^T)$, and almost surely at a rate arbitrarily close to $o((1-\frac{\mu}{dL})^T)$, where $\mu$ and $L$are the strong convexity and smoothness parameters of the function.
Recently, Role-Playing Agents (RPAs) have garnered increasing attention for their potential to deliver emotional value and facilitate sociological research.However, existing studies are primarily confined to the textual modality, unable to simulate humans' multimodal perceptual capabilities.To bridge this gap, we introduce the concept of Multimodal Role-Playing Agents (MRPAs), and propose a comprehensive framework, MMRole, for their development and evaluation, which comprises a personalized multimodal dataset and a robust evaluation approach.Specifically, we construct a large-scale, high-quality dataset, MMRole-Data, consisting of 85 characters, 11K images, and 14K single or multi-turn dialogues.Additionally, we present a robust evaluation approach, MMRole-Eval, encompassing eight metrics across three dimensions, where a reward model is designed to score MRPAs with the constructed ground-truth data for comparison.Moreover, we develop the first specialized MRPA, MMRole-Agent.Extensive evaluation results demonstrate the improved performance of MMRole-Agent and highlight the primary challenges in developing MRPAs, emphasizing the need for enhanced multimodal understanding and role-playing consistency.The data, code, and models are all available at https://github.com/YanqiDai/MMRole.
Personalized text-to-image diffusion models have grown popular for their ability to efficiently acquire a new concept from user-defined text descriptions and a few images. However, in the real world, a user may wish to personalize a model on multiple concepts but one at a time, with no access to the data from previous concepts due to storage/privacy concerns. When faced with this continual learning (CL) setup, most personalization methods fail to find a balance between acquiring new concepts and retaining previous ones -- a challenge that *continual personalization* (CP) aims to solve. Inspired by the successful CL methods that rely on class-specific information for regularization, we resort to the inherent class-conditioned density estimates, also known as *diffusion classifier* (DC) scores, for CP of text-to-image diffusion models. Namely, we propose using DC scores for regularizing the parameter-space and function-space of text-to-image diffusion models.Using several diverse evaluation setups, datasets, and metrics, we show that our proposed regularization-based CP methods outperform the state-of-the-art C-LoRA, and other baselines. Finally, by operating in the replay-free CL setup and on low-rank adapters, our method incurs zero storage and parameter overhead, respectively, over the state-of-the-art.
Self-supervised learning has become a cornerstone in computer vision, primarily divided into reconstruction-based methods like masked autoencoders (MAE) and discriminative methods such as contrastive learning (CL). Recent empirical observations reveal that MAE and CL capture different types of representations: CL tends to focus on global patterns, while MAE adeptly captures **both global and subtle local** information simultaneously. Despite a flurry of recent empirical investigations to shed light on this difference, theoretical understanding remains limited, especially on the dominant architecture **vision transformers** (ViTs). In this paper, to provide rigorous insights, we model the visual data distribution by considering two types of spatial features: dominant global features and comparatively minuscule local features, and study the impact of imbalance among these features. We analyze the training dynamics of one-layer softmax-based ViTs on both MAE and CL objectives using gradient descent. Our analysis shows that as the degree of feature imbalance varies, ViTs trained with the MAE objective effectively learn both global and local features to achieve near-optimal reconstruction, while the CL-trained ViTs favor predominantly global features, even under mild imbalance. These results provide a theoretical explanation for distinct behaviors of MAE and CL observed in empirical studies.
Large Language Models (LLMs) have become more prevalent in long-context applications such as interactive chatbots, document analysis, and agent workflows, but it is challenging to serve long-context requests with low latency and high throughput. Speculative decoding (SD) is a widely used technique to reduce latency losslessly, but the conventional wisdom suggests that its efficacy is limited to small batch sizes. In MagicDec, we show that surprisingly SD can achieve speedup even for a high throughput inference regime for moderate to long sequences. More interestingly, an intelligent drafting strategy can achieve better speedup with increasing batch size based on our rigorous analysis. MagicDec first identifies the bottleneck shifts with increasing batch size and sequence length, and uses these insights to deploy SD more effectively for high throughput inference. We leverage draft model with sparse KV cache to address the KV bottleneck, which scales with both sequence length and batch size. Additionally, we propose a theoretical model to select the optimal drafting strategy for maximum speedup. Our work highlights the broad applicability of speculative decoding in long-context serving, as it can enhance throughput and reduce latency without compromising accuracy. For moderate to long sequences, we demonstrate up to 2.51x speedup for LLaMA-3.1-8B when serving batch sizes ranging from 32 to 256 on various types of hardware and tasks.
Large Language Models (LLMs) have made significant strides in natural languageprocessing, and a precise understanding of the internal mechanisms driving theirsuccess is essential. In this work, we trace the trajectories of individual tokens as they pass through transformer blocks, and linearize the system along these trajectories through their Jacobian matrices. By examining the relationships between these Jacobians, we uncover a **transformer block coupling** phenomenon in a variety of LLMs, characterized by the coupling of their top singular vectors across tokens and depth. Our findings reveal that coupling *positively correlates* with model performance, and that this relationship is stronger than with other hyperparameters, namely parameter budget, model depth, and embedding dimension. We further investigate the emergence of these properties through training, noting the development of coupling, as well as an increase in linearity and layer-wise exponential growth in the token trajectories. These collective insights provide a novel perspective on the interactions between token embeddings, and prompt further approaches to study training and generalization in LLMs.
Generative foundation models are susceptible to implicit biases that can arise from extensive unsupervised training data. Such biases can produce suboptimal samples, skewed outcomes, and unfairness, with potentially serious consequences. Consequently, aligning these models with human ethics and preferences is an essential step toward ensuring their responsible and effective deployment in real-world applications. Prior research has primarily employed Reinforcement Learning from Human Feedback (RLHF) to address this problem, where generative models are fine-tuned with RL algorithms guided by a human-feedback-informed reward model. However, the inefficiencies and instabilities associated with RL algorithms frequently present substantial obstacles to the successful alignment, necessitating the development of a more robust and streamlined approach. To this end, we introduce a new framework, Reward rAnked FineTuning (RAFT), designed to align generative models effectively. Utilizing a reward model and a sufficient number of samples, our approach selects the high-quality samples, discarding those that exhibit undesired behavior, and subsequently enhancing the model by fine-tuning on these filtered samples. Our studies show that RAFT can effectively improve the model performance in both reward learning and other automated metrics in both large language models and diffusion models.