As large language models (LLMs) become increasingly advanced, their ability to exhibit compositional generalization---the capacity to combine learned skills in novel ways not encountered during training---has garnered significant attention. This type of generalization, particularly in scenarios beyond training data, is also of great interest in the study of AI safety and alignment. A recent study introduced the Skill-Mix evaluation, where models are tasked with composing a short paragraph demonstrating the use of a specified $k$-tuple of language skills. While small models struggled with composing even with $k=3$, larger models like GPT-4 performed reasonably well with $k=5$ and $6$.In this paper, we employ a setup akin to Skill-Mix to evaluate the capacity of smaller models to learn compositional generalization from examples. Utilizing a diverse set of language skills---including rhetorical, literary, reasoning, theory of mind, and common sense---GPT was used to generate text samples that exhibit random subsets of $k$ skills. Subsequent fine-tuning of 7B and 13B parameter models on these combined skill texts, for increasing values of $k$, revealed the following findings: (1) Training on combinations of $k=2$ and $3$ skills results in noticeable improvements in the ability to compose texts with $k=4$ and $5$ skills, despite models never having seen such examples during training. (2) When skill categories are split into training and held-out groups, models significantly improve at composing texts with held-out skills during testing despite having only seen training skills during fine-tuning, illustrating the efficacy of the training approach even with previously unseen skills.This study also suggests that incorporating skill-rich (potentially synthetic) text into training can substantially enhance the compositional capabilities of models.
论文检索
输入标题、作者或关键词，从 14,273 篇学术成果中精准定位
Provably Mitigating Overoptimization in RLHF: Your SFT Loss is Implicitly an Adversarial Regularizer
Generative Models
Aligning generative models with human preference via RLHF typically suffers from overoptimization, where an imperfectly learned reward model can misguide the generative model to output even undesired responses. We investigate this problem in a principled manner by identifying the source of the issue as the distributional shift and uncertainty of human preference in dataset. To mitigate overoptimization, we first propose a theoretical algorithm which optimizes the policy against an adversarially chosen reward model, one that simultaneously minimizes its MLE loss and a reward penalty term. The penalty pessimistically biases the uncertain rewards so as to prevent the policy from choosing actions with spursiouly high proxy rewards, resulting in provable sample efficiency of the algorithm under a partial coverage style condition. Moving from theory to practice, the proposed algorithm further enjoys an equivalent but surprisingly easy to implement form. With a clever usage of the equivalence between reward models and the corresponding optimal policy, the algorithm features a simple objective that combines (i) a preference optimization loss that directly aligns the policy with human preference, and (ii) a supervised learning loss which explicitly imitates the policy with a baseline distribution. In the context of aligning large language models (LLM), this objective fuses the direct preference optimization (DPO) loss with the supervised fune-tuning (SFT) loss to help mitigate the overoptimization towards undesired responses, for which we name the algorithm Regularized Preference Optimization (RPO).Experiments of aligning LLMs demonstrate the improved performance of our method when compared with DPO baselines. Our work sheds light on the interplay between preference optimization and SFT in tuning LLMs with both theoretical guarantees and empirical evidence.
Reinforcement Learning · Deep RL
Current LLM alignment techniques use pairwise human preferences at a sample level, and as such, they do not imply an alignment on the distributional level. We propose in this paper Alignment via Optimal Transport (AOT), a novel method for distributional preference alignment of LLMs. AOT aligns LLMs on unpaired preference data by making the reward distribution of the positive samples stochastically dominant in the first order on the distribution of negative samples. We introduce a convex relaxation of this first-order stochastic dominance and cast it as an optimal transport problem with a smooth and convex cost. Thanks to the one-dimensional nature of the resulting optimal transport problem and the convexity of the cost, it has a closed-form solution via sorting on empirical measures. We fine-tune LLMs with this AOT objective, which enables alignment by penalizing the violation of the stochastic dominance of the reward distribution of the positive samples on the reward distribution of the negative samples. We analyze the sample complexity of AOT by considering the dual of the OT problem and show that it converges at the parametric rate. Empirically, we show on a diverse set of alignment datasets and LLMs that AOT leads to state-of-the-art models in the 7B family of models when evaluated with Open LLM Benchmarks and AlpacaEval. Code for $\mathsf{AOT}$ is available in the Hugging Face TRL library \url{https://ibm.biz/AOT_TRL}.
Miscellaneous Aspects of Machine Learning · Transfer, Multitask and Meta-learning
Meta-learning offers a promising avenue for few-shot learning (FSL), enabling models to glean a generalizable feature embedding through episodic training on synthetic FSL tasks in a source domain. Yet, in practical scenarios where the target task diverges from that in the source domain, meta-learning based method is susceptible to over-fitting. To overcome this, we introduce a novel framework, Meta-Exploiting Frequency Prior for Cross-Domain Few-Shot Learning, which is crafted to comprehensively exploit the cross-domain transferable image prior that each image can be decomposed into complementary low-frequency content details and high-frequency robust structural characteristics. Motivated by this insight, we propose to decompose each query image into its high-frequency and low-frequency components, and parallel incorporate them into the feature embedding network to enhance the final category prediction. More importantly, we introduce a feature reconstruction prior and a prediction consistency prior to separately encourage the consistency of the intermediate feature as well as the final category prediction between the original query image and its decomposed frequency components. This allows for collectively guiding the network's meta-learning process with the aim of learning generalizable image feature embeddings, while not introducing any extra computational cost in the inference phase. Our framework establishes new state-of-the-art results on multiple cross-domain few-shot learning benchmarks.
Robotics
Data-driven autonomous driving motion generation tasks are frequently impacted by the limitations of dataset size and the domain gap between datasets, which precludes their extensive application in real-world scenarios. To address this issue, we introduce SMART, a novel autonomous driving motion generation paradigm that models vectorized map and agent trajectory data into discrete sequence tokens. These tokens are then processed through a decoder-only transformer architecture to train for the next token prediction task across spatial-temporal series. This GPT-style method allows the model to learn the motion distribution in real driving scenarios. SMART achieves state-of-the-art performance across most of the metrics on the generative Sim Agents challenge, ranking 1st on the leaderboards of Waymo Open Motion Dataset (WOMD), demonstrating remarkable inference speed. Moreover, SMART represents the generative model in the autonomous driving motion domain, exhibiting zero-shot generalization capabilities: Using only the NuPlan dataset for training and WOMD for validation, SMART achieved a competitive score of 0.72 on the Sim Agents challenge. Lastly, we have collected over 1 billion motion tokens from multiple datasets, validating the model's scalability. These results suggest that SMART has initially emulated two important properties: scalability and zero-shot generalization, and preliminarily meets the needs of large-scale real-time simulation applications. We have released all the code to promote the exploration of models for motion generation in the autonomous driving field. The source code is available at https://github.com/rainmaker22/SMART.
Computer Vision
In image editing, Denoising Diffusion Implicit Models (DDIM) inversion has become a widely adopted method and is extensively used in various image editing approaches. The core concept of DDIM inversion stems from the deterministic sampling technique of DDIM, which allows the DDIM process to be viewed as an Ordinary Differential Equation (ODE) process that is reversible. This enables the prediction of corresponding noise from a reference image, ensuring that the restored image from this noise remains consistent with the reference image. Image editing exploits this property by modifying the cross-attention between text and images to edit specific objects while preserving the remaining regions. However, in the DDIM inversion, using the $t-1$ time step to approximate the noise prediction at time step $t$ introduces errors between the restored image and the reference image. Recent approaches have modeled each step of the DDIM inversion process as finding a fixed-point problem of an implicit function. This approach significantly mitigates the error in the restored image but lacks theoretical support regarding the existence of such fixed points. Therefore, this paper focuses on the study of fixed points in DDIM inversion and provides theoretical support. Based on the obtained theoretical insights, we further optimize the loss function for the convergence of fixed points in the original DDIM inversion, improving the visual quality of the edited image. Finally, we extend the fixed-point based image editing to the application of unsupervised image dehazing, introducing a novel text-based approach for unsupervised dehazing.
How did humanity coax mathematics from the aether? We explore the Platonic view that mathematics can be discovered from its axioms---a game of conjecture and proof. We describe an agent that jointly learns to pose challenging problems for itself (conjecturing) and solve them (theorem proving). Given a mathematical domain axiomatized in dependent type theory, we first combine methods for constrained decoding and type-directed synthesis to sample valid conjectures from a language model. Our method guarantees well-formed conjectures by construction, even as we start with a randomly initialized model. We use the same model to represent a policy and value function for guiding proof search. Our agent targets generating hard but provable conjectures --- a moving target, since its own theorem proving ability also improves as it trains. We propose novel methods for hindsight relabeling on proof search trees to significantly improve the agent's sample efficiency in both tasks. Experiments on 3 axiomatic domains (propositional logic, arithmetic and group theory) demonstrate that our agent can bootstrap from only the axioms, self-improving in generating true and challenging conjectures and in finding proofs.
Miscellaneous Aspects of Machine Learning · Transfer, Multitask and Meta-learning
Zero-shot graph machine learning, especially with graph neural networks (GNNs), has garnered significant interest due to the challenge of scarce labeled data. While methods like self-supervised learning and graph prompt learning have been extensively explored, they often rely on fine-tuning with task-specific labels, limiting their effectiveness in zero-shot scenarios. Inspired by the zero-shot capabilities of instruction-fine-tuned large language models (LLMs), we introduce a novel framework named Token Embedding-Aligned Graph Language Model (TEA-GLM) that leverages LLMs as cross-dataset and cross-task zero-shot learners for graph machine learning. Concretely, we pretrain a GNN, aligning its representations with token embeddings of an LLM. We then train a linear projector that transforms the GNN's representations into a fixed number of graph token embeddings without tuning the LLM. A unified instruction is designed for various graph tasks at different levels, such as node classification (node-level) and link prediction (edge-level). These design choices collectively enhance our method's effectiveness in zero-shot learning, setting it apart from existing methods. Experiments show that our graph token embeddings help the LLM predictor achieve state-of-the-art performance on unseen datasets and tasks compared to other methods using LLMs as predictors. Our code is available at https://github.com/W-rudder/TEA-GLM.
Interpretability
Most currently deployed LLMs undergo continuous training or additional finetuning. By contrast, most research into LLMs' internal mechanisms focuses on models at one snapshot in time (the end of pre-training), raising the question of whether their results generalize to real-world settings. Existing studies of mechanisms over time focus on encoder-only or toy models, which differ significantly from most deployed models. In this study, we track how model mechanisms, operationalized as circuits, emerge and evolve across 300 billion tokens of training in decoder-only LLMs, in models ranging from 70 million to 2.8 billion parameters. We find that task abilities and the functional components that support them emerge consistently at similar token counts across scale. Moreover, although such components may be implemented by different attention heads over time, the overarching algorithm that they implement remains. Surprisingly, both these algorithms and the types of components involved therein tend to replicate across model scale. Finally, we find that circuit size correlates with model size and can fluctuate considerably over time even when the same algorithm is implemented. These results suggest that circuit analyses conducted on small models at the end of pre-training can provide insights that still apply after additional training and over model scale.
Time Series
Large pre-trained models excel in zero/few-shot learning for language and vision tasks but face challenges in multivariate time series (TS) forecasting due to diverse data characteristics. Consequently, recent research efforts have focused on developing pre-trained TS forecasting models. These models, whether built from scratch or adapted from large language models (LLMs), excel in zero/few-shot forecasting tasks. However, they are limited by slow performance, high computational demands, and neglect of cross-channel and exogenous correlations. To address this, we introduce Tiny Time Mixers (TTM), a compact model (starting from 1M parameters) with effective transfer learning capabilities, trained exclusively on public TS datasets. TTM, based on the light-weight TSMixer architecture, incorporates innovations like adaptive patching, diverse resolution sampling, and resolution prefix tuning to handle pre-training on varied dataset resolutions with minimal model capacity. Additionally, it employs multi-level modeling to capture channel correlations and infuse exogenous signals during fine-tuning. TTM outperforms existing popular benchmarks in zero/few-shot forecasting by (4-40\%), while reducing computational requirements significantly. Moreover, TTMs are lightweight and can be executed even on CPU-only machines, enhancing usability and fostering wider adoption in resource-constrained environments. The model weights for reproducibility and research use are available at https://huggingface.co/ibm/ttm-research-r2/, while enterprise-use weights under the Apache license can be accessed as follows: the initial TTM-Q variant at https://huggingface.co/ibm-granite/granite-timeseries-ttm-r1, and the latest variants (TTM-B, TTM-E, TTM-A) weights are available at https://huggingface.co/ibm-granite/granite-timeseries-ttm-r2. The source code for the TTM model along with the usage scripts are available at https://github.com/ibm-granite/granite-tsfm/tree/main/tsfm_public/models/tinytimemixer
Structured state-space models (SSMs) are gaining popularity as effective foundational architectures for sequential data, demonstrating outstanding performance across a diverse set of domains alongside desirable scalability properties. Recent developments show that if the linear recurrence powering SSMs allows for a selectivity mechanism leveraging multiplicative interactions between inputs and hidden states (e.g. Mamba, GLA, Hawk/Griffin, HGRN2), then the resulting architecture can surpass attention-powered foundation models trained on text in both accuracy and efficiency, at scales of billion parameters. In this paper, we give theoretical grounding to the selectivity mechanism, often linked to in-context learning, using tools from Rough Path Theory. We provide a framework for the theoretical analysis of generalized selective SSMs, fully characterizing their expressive power and identifying the gating mechanism as the crucial architectural choice. Our analysis provides a closed-form description of the expressive powers of modern SSMs, such as Mamba, quantifying theoretically the drastic improvement in performance from the previous generation of models, such as S4. Our theory not only motivates the success of modern selective state-space models, but also provides a solid framework to understand the expressive power of future SSM variants. In particular, it suggests cross-channel interactions could play a vital role in future improvements.
Trustworthy Machine Learning
Unsupervised out-of-distribution (OOD) detection aims to identify out-of-domain data by learning only from unlabeled In-Distribution (ID) training samples, which is crucial for developing a safe real-world machine learning system. Current reconstruction-based method provides a good alternative approach, by measuring the reconstruction error between the input and its corresponding generative counterpart in the pixel/feature space. However, such generative methods face the key dilemma, $i.e.$, improving the reconstruction power of the generative model, while keeping compact representation of the ID data. To address this issue, we propose the diffusion-based layer-wise semantic reconstruction approach for unsupervised OOD detection. The innovation of our approach is that we leverage the diffusion model's intrinsic data reconstruction ability to distinguish ID samples from OOD samples in the latent feature space. Moreover, to set up a comprehensive and discriminative feature representation, we devise a multi-layer semantic feature extraction strategy. Through distorting the extracted features with Gaussian noises and applying the diffusion model for feature reconstruction, the separation of ID and OOD samples is implemented according to the reconstruction errors. Extensive experimental results on multiple benchmarks built upon various datasets demonstrate that our method achieves state-of-the-art performance in terms of detection accuracy and speed.
Statistical Methods
This paper introduces a conformal inference method to evaluate uncertainty in classification by generating prediction sets with valid coverage conditional on adaptively chosen features. These features are carefully selected to reflect potential model limitations or biases. This can be useful to find a practical compromise between efficiency---by providing informative predictions---and algorithmic fairness---by ensuring equalized coverage for the most sensitive groups. We demonstrate the validity and effectiveness of this method on simulated and real data sets.
Theory · Everything Else
Approximating the action of a matrix function $f(\vec{A})$ on a vector $\vec{b}$ is an increasingly important primitive in machine learning, data science, and statistics, with applications such as sampling high dimensional Gaussians, Gaussian process regression and Bayesian inference, principle component analysis, and approximating Hessian spectral densities.Over the past decade, a number of algorithms enjoying strong theoretical guarantees have been proposed for this task.Many of the most successful belong to a family of algorithms called Krylov subspace methods.Remarkably, a classic Krylov subspace method, called the Lanczos method for matrix functions (Lanczos-FA), frequently outperforms newer methods in practice. Our main result is a theoretical justification for this finding: we show that, for a natural class of rational functions, Lanczos-FA matches the error of the best possible Krylov subspace method up to a multiplicative approximation factor. The approximation factor depends on the degree of $f(x)$'s denominator and the condition number of $\vec{A}$, but not on the number of iterations $k$. Our result provides a strong justification for the excellent performance of Lanczos-FA, especially on functions that are well approximated by rationals, such as the matrix square root.
Reinforcement Learning · Batch Offline
Offline safe reinforcement learning (RL) aims to train a policy that satisfies con- straints using a pre-collected dataset. Most current methods struggle with the mismatch between imperfect demonstrations and the desired safe and rewarding performance. In this paper, we mitigate this issue from a data-centric perspective and introduce OASIS (cOnditionAl diStributIon Shaping), a new paradigm in offline safe RL designed to overcome these critical limitations. OASIS utilizes a conditional diffusion model to synthesize offline datasets, thus shaping the data dis- tribution toward a beneficial target domain. Our approach makes compliance with safety constraints through effective data utilization and regularization techniques to benefit offline safe RL training. Comprehensive evaluations on public benchmarks and varying datasets showcase OASIS’s superiority in benefiting offline safe RL agents to achieve high-reward behavior while satisfying the safety constraints, out- performing established baselines. Furthermore, OASIS exhibits high data efficiency and robustness, making it suitable for real-world applications, particularly in tasks where safety is imperative and high-quality demonstrations are scarce. More details are available at the website https://sites.google.com/view/saferl-oasis/home.
Self-evaluation using large language models (LLMs) has proven valuable not only in benchmarking but also methods like reward modeling, constitutional AI, and self-refinement. But new biases are introduced due to the same LLM acting as both the evaluator and the evaluatee. One such bias is self-preference, where an LLM evaluator scores its own outputs higher than others’ while human annotators consider them of equal quality. But do LLMs actually recognize their own outputs when they give those texts higher scores, or is it just a coincidence? In this paper, we investigate if self-recognition capability contributes to self-preference. We discover that, out of the box, LLMs such as GPT-4 and Llama 2 have non-trivial accuracy at distinguishing themselves from other LLMs and humans. By finetuning LLMs, we discover a linear correlation between self-recognition capability and the strength of self-preference bias; using controlled experiments, we show that the causal explanation resists straightforward confounders. We discuss how self-recognition can interfere with unbiased evaluations and AI safety more generally.
Online Learning
Vision-Language Models (VLMs) such as CLIP have yielded unprecedented performances for zero-shot image classification, yet their generalization capability may still be seriously challenged when confronted to domain shifts. In response, we present Weight Average Test-Time Adaptation (WATT) of CLIP, a new approach facilitating full test-time adaptation (TTA) of this VLM. Our method employs a diverse set of templates for text prompts, augmenting the existing framework of CLIP. Predictions are utilized as pseudo labels for model updates, followed by weight averaging to consolidate the learned information globally. Furthermore, we introduce a text ensemble strategy, enhancing the overall test performance by aggregating diverse textual cues.Our findings underscore the effectiveness of WATT across diverse datasets, including CIFAR-10-C, CIFAR-10.1, CIFAR-100-C, VisDA-C, and several other challenging datasets, effectively covering a wide range of domain shifts. Notably, these enhancements are achieved without the need for additional model transformations or trainable modules. Moreover, compared to other TTA methods, our approach can operate effectively with just a single image. The code is available at: https://github.com/Mehrdad-Noori/WATT.
Trustworthy Machine Learning
Although 3D point cloud recognition has achieved substantial progress on standard benchmarks, the typical models are vulnerable to point cloud corruptions, leading to security threats in real-world applications. To improve the corruption robustness, various data augmentation methods have been studied, but they are mainly limited to the spatial domain. As the point cloud has low information density and significant spatial redundancy, it is challenging to analyze the effects of corruptions. In this paper, we focus on the frequency domain to observe the underlying structure of point clouds and their corruptions. Through graph Fourier transform (GFT), we observe a correlation between the corruption robustness of point cloud recognition models and their sensitivity to different frequency bands, which is measured by the GFT spectrum of the model’s Jacobian matrix. To reduce the sensitivity and improve the corruption robustness, we propose Frequency Adversarial Training (FAT) that adopts frequency-domain adversarial examples as data augmentation to train robust point cloud recognition models against corruptions. Theoretically, we provide a guarantee of FAT on its out-of-distribution generalization performance. Empirically, we conduct extensive experiments with various network architectures to validate the effectiveness of FAT, which achieves the new state-of-the-art results.
Incomplete multi-view clustering aims to learn complete correlations among samples by leveraging complementary information across multiple views for clustering. Anchor-based methods further establish sample-level similarities for representative anchor generation, effectively addressing scalability issues in large-scale scenarios. Despite efficiency improvements, existing methods overlook the misguidance in anchors learning induced by partial missing samples, i.e., the absence of samples results in shift of learned anchors, further leading to sub-optimal clustering performance. To conquer the challenges, our solution involves a cross-view reconstruction strategy that not only alleviate the anchor shift problem through a carefully designed cross-view learning process, but also reconstructs missing samples in a way that transcends the limitations imposed by convex combinations. By employing affine combinations, our method explores areas beyond the convex hull defined by anchors, thereby illuminating blind spots in the reconstruction of missing samples. Experimental results on four benchmark datasets and three large-scale datasets validate the effectiveness of our proposed method.
The Janus Problem is a common issue in SDS-based text-to-3D methods. Due to view encoding approach and 2D diffusion prior guidance, the 3D representation model tends to learn content with higher certainty from each perspective, leading to view inconsistency. In this work, we first model and analyze the problem, visualizing the specific causes of the Janus Problem, which are associated with discrete view encoding and shared priors in 2D lifting. Based on this, we further propose the LCGen method, which guides text-to-3D to obtain different priors with different certainty from various viewpoints, aiding in view-consistent generation. Experiments have proven that our LCGen method can be directly applied to different SDS-based text-to-3D methods, alleviating the Janus Problem without introducing additional information, increasing excessive training burden, or compromising the generation effect.