ICLR 2025 论文列表

Sampling from generative models has become a crucial tool for applications like data synthesis and augmentation. Diffusion, Flow Matching and Continuous Normalising Flows have shown effectiveness across various modalities, and rely on latent variables for generation. For experimental design or creative applications that require more control over the generation process, it has become common to manipulate the latent variable directly. However, existing approaches for performing such manipulations (e.g. interpolation or forming low-dimensional representations) only work well in special cases or are network or data-modality specific. We propose Latent Optimal Linear combinations (LOL) as a general-purpose method to form linear combinations of latent variables that adhere to the assumptions of the generative model. As LOL is easy to implement and naturally addresses the broader task of forming any linear combinations, e.g. the construction of subspaces of the latent space, LOL dramatically simplifies the creation of expressive low-dimensional representations of high-dimensional objects.

Policy Mirror Descent (PMD) is a popular framework in reinforcement learning, serving as a unifying perspective that encompasses numerous algorithms. These algorithms are derived through the selection of a mirror map and enjoy finite-time convergence guarantees. Despite its popularity, the exploration of PMD's full potential is limited, with the majority of research focusing on a particular mirror map---namely, the negative entropy---which gives rise to the renowned Natural Policy Gradient (NPG) method. It remains uncertain from existing theoretical studies whether the choice of mirror map significantly influences PMD's efficacy. In our work, we conduct empirical investigations to show that the conventional mirror map choice (NPG) often yields less-than-optimal outcomes across several standard benchmark environments. Using evolutionary strategies, we identify more efficient mirror maps that enhance the performance of PMD. We first focus on a tabular environment, i.e.\ Grid-World, where we relate existing theoretical bounds with the performance of PMD for a few standard mirror maps and the learned one. We then show that it is possible to learn a mirror map that outperforms the negative entropy in more complex environments, such as the MinAtar suite. Additionally, we demonstrate that the learned mirror maps generalize effectively to different tasks by testing each map across various other environments.

The \textbf{i}nexact \textbf{s}tochastic \textbf{p}roximal \textbf{p}oint \textbf{a}lgorithm (isPPA) is popular for solving stochastic composite optimization problems with many applications in machine learning. While the convergence theory of the (inexact) PPA has been well established, the known convergence guarantees of isPPA require restrictive assumptions. In this paper, we establish the stability and almost sure convergence of isPPA under mild assumptions, where smoothness and (restrictive) strong convexity of the objective function are not required. Imposing a local Lipschitz condition on component functions and a quadratic growth condition on the objective function, we establish last-iterate iteration complexity bounds of isPPA regarding the distance to the solution set and the Karush–Kuhn–Tucker (KKT) residual. Moreover, we show that the established iteration complexity bounds are tight up to a constant by explicitly analyzing the bounds for the regularized Fr\'echet mean problem. We further validate the established convergence guarantees of isPPA by numerical experiments.

In time-series classification, interpretable models can bring additional insights but be outperformed by deep models since human-understandable features have limited expressivity and flexibility. In this work, we present InterpGN, a framework that integrates an interpretable model and a deep neural network. Within this framework, we introduce a novel gating function design based on the confidence of the interpretable expert, preserving interpretability for samples where interpretable features are significant while also identifying samples that require additional expertise. For the interpretable expert, we incorporate shapelets to effectively model shape-level features for time-series data. We introduce a variant of Shapelet Transforms to build logical predicates using shapelets. Our proposed model achieves comparable performance with state-of-the-art deep learning models while additionally providing interpretable classifiers for various benchmark datasets. We further show that our models improve on quantitative shapelet quality and interpretability metrics over existing shapelet-learning formulations. Finally, we show that our models can integrate additional advanced architectures and be applied to real-world tasks beyond standard benchmarks such as the MIMIC-III and time series extrinsic regression datasets.

This work provides the first theoretical analysis of training transformers to solve complex problems by recursively generating intermediate states, analogous to fine-tuning for chain-of-thought (CoT) reasoning. We consider training a one-layer transformer to solve the fundamental $k$-parity problem, extending the work on RNNs by \citet{Wies23}. We establish three key results: (1) any finite-precision gradient-based algorithm, without intermediate supervision, requires substantial iterations to solve parity with finite samples. (2) In contrast, when intermediate parities are incorporated into the loss function, our model can learn parity in one gradient update when aided by \emph{teacher forcing}, where ground-truth labels of the reasoning chain are provided at each generation step. (3) Even without teacher forcing, where the model must generate CoT chains end-to-end, parity can be learned efficiently if augmented data is employed to internally verify the soundness of intermediate steps. Our findings, supported by numerical experiments, show that task decomposition and stepwise reasoning naturally arise from optimizing transformers with CoT; moreover, self-consistency checking can improve multi-step reasoning ability, aligning with empirical studies of CoT.

Canonicalization, a popular method for generating invariant or equivariant function classes from arbitrary function sets, involves initial data projection onto a reduced input space subset, followed by applying any learning method to the projected dataset. Despite recent research on the expressive power and continuity of functions represented by canonicalization, its generalization capabilities remain less explored. This paper addresses this gap by theoretically examining the generalization benefits and sample complexity of canonicalization, comparing them with group averaging, another popular technique for creating invariant or equivariant function classes. Our findings reveal two distinct regimes where canonicalization may outperform or underperform compared to group averaging, with precise quantification of this phase transition in terms of sample size, group action characteristics, and a newly introduced concept of alignment. To the best of our knowledge, this study represents the first theoretical exploration of such behavior, offering insights into the relative effectiveness of canonicalization and group averaging under varying conditions.

Large Language Models (LLMs) have revolutionized the field of natural language processing, achieving unprecedented performance across a variety of applications. However, their increased computational and memory demands present significant challenges, especially when handling long sequences. This paper focuses on the long-context scenario, addressing the inefficiencies in KV cache memory consumption during inference. Unlike existing approaches that optimize the memory based on the sequence length, we identify substantial redundancy in the channel dimension of the KV cache, as indicated by an uneven magnitude distribution and a low-rank structure in the attention weights. In response, we propose ThinK, a novel query-dependent KV cache pruning method designed to minimize attention weight loss while selectively pruning the least significant channels. Our approach not only maintains or enhances model accuracy but also achieves a reduction in KV cache memory costs by over 20\% compared with vanilla KV cache eviction and quantization methods. For instance, ThinK integrated with KIVI can achieve $2.8\times$ peak memory reduction while maintaining nearly the same quality, enabling a batch size increase from 4$\times$ (with KIVI alone) to 5$\times$ when using a single GPU. Extensive evaluations on the LLaMA and Mistral models across various long-sequence datasets verified the efficiency of ThinK. Our code has been made available at https://github.com/SalesforceAIResearch/ThinK.

The fusion of Large Language Models (LLMs) with vision models is pioneering new possibilities in user-interactive vision-language tasks. A notable application is reasoning segmentation, where models generate pixel-level segmentation masks by comprehending implicit meanings in human instructions. However, seamless human-AI interaction demands more than just object-level recognition; it requires understanding both objects and the functions of their detailed parts, particularly in multi-target scenarios. For example, when instructing a robot to \textit{“turn on the TV"}, there could be various ways to accomplish this command. Recognizing multiple objects capable of turning on the TV, such as the TV itself or a remote control (multi-target), provides more flexible options and aids in finding the optimized scenario. Furthermore, understanding specific parts of these objects, like the TV's button or the remote's button (part-level), is important for completing the action. Unfortunately, current reasoning segmentation datasets predominantly focus on a single target object-level reasoning, which limits the detailed recognition of an object's parts in multi-target contexts. To address this gap, we construct a large-scale dataset called Multi-target and Multi-granularity Reasoning (MMR). MMR comprises 194K complex and implicit instructions that consider multi-target, object-level, and part-level aspects, based on pre-existing image-mask sets. This dataset supports diverse and context-aware interactions by hierarchically providing object and part information. Moreover, we propose a straightforward yet effective framework for multi-target, object-level, and part-level reasoning segmentation. Experimental results on MMR show that the proposed method can reason effectively in multi-target and multi-granularity scenarios, while the existing reasoning segmentation model still has room for improvement. The dataset is available at \url{https://github.com/jdg900/MMR}.

We present Layout-Your-3D, a framework that allows controllable and compositional 3D generation from text prompts. Existing text-to-3D methods often struggle to generate assets with plausible object interactions or require tedious optimization processes. To address these challenges, our approach leverages 2D layouts as a blueprint to facilitate precise and plausible control over 3D generation. Starting with a 2D layout provided by a user or generated from a text description, we first create a coarse 3D scene using a carefully designed initialization process based on efficient reconstruction models. To enforce coherent global 3D layouts and enhance the quality of instance appearances, we propose a collision-aware layout optimization process followed by instance-wise refinement. Experimental results demonstrate that Layout-Your-3D yields more reasonable and visually appealing compositional 3D assets while significantly reducing the time required for each prompt. Additionally, Layout-Your-3D can be easily applicable to downstream tasks, such as 3D editing and object insertion.

Zeroth-order optimization (ZO) is a memory-efficient strategy for fine-tuning Large Language Models using only forward passes. However, applying ZO fine-tuning in memory-constrained settings such as mobile phones and laptops remains challenging since these settings often involve weight quantization, while ZO requires full-precision perturbation and update. In this study, we address this limitation by combining static sparse ZO fine-tuning with quantization. Our approach transfers a small, static subset (0.1%) of "sensitive" parameters from pre-training to downstream tasks, focusing fine-tuning on this sparse set of parameters. The remaining untuned parameters are quantized, reducing memory demands. Our proposed workflow enables efficient ZO fine-tuning of an Llama2-7B model on a GPU device with less than 8GB of memory while outperforming full model ZO fine-tuning performance and in-context learning.

Molecule-and-text cross-modal representation learning has emerged as a promising direction for enhancing the quality of molecular representation, thereby improving performance in various scientific fields. However, most approaches employ a global alignment approach to learn the knowledge from different modalities that may fail to capture fine-grained information, such as molecule-and-text fragments and stereoisomeric nuances, which is crucial for downstream tasks. Furthermore, it is incapable of modeling such information using a similar global alignment strategy due to the lack of annotations about the fine-grained fragments in the existing dataset. In this paper, we propose Atomas, a hierarchical molecular representation learning framework that jointly learns representations from SMILES strings and text. We design a Hierarchical Adaptive Alignment model to automatically learn the fine-grained fragment correspondence between two modalities and align these representations at three semantic levels. Atomas's end-to-end training framework supports understanding and generating molecules, enabling a wider range of downstream tasks. Atomas achieves superior performance across 12 tasks on 11 datasets, outperforming 11 baseline models thus highlighting the effectiveness and versatility of our method. Scaling experiments further demonstrate Atomas’s robustness and scalability. Moreover, visualization and qualitative analysis, validated by human experts, confirm the chemical relevance of our approach. Codes are released on ~\url{https://github.com/yikunpku/Atomas}.

The performance of large language models (LLMs) is closely linked to their underlying size, leading to ever-growing networks and hence slower inference. Speculative decoding has been proposed as a technique to accelerate autoregressive generation, leveraging a fast draft model to propose candidate tokens, which are then verified in parallel based on their likelihood under the target model. While this approach guarantees to reproduce the target output, it incurs a substantial penalty: many high-quality draft tokens are rejected, even when they represent objectively valid continuations. Indeed, we show that even powerful draft models such as GPT-4o, as well as human text cannot achieve high acceptance rates under the standard verification scheme. This severely limits the speedup potential of current speculative decoding methods, as an early rejection becomes overwhelmingly likely when solely relying on alignment of draft and target. We thus ask the following question: Can we adapt verification to recognize correct, but non-aligned replies? To this end, we draw inspiration from the LLM-as-a-judge framework, which demonstrated that LLMs are able to rate answers in a versatile way. We carefully design a dataset coined TokenCourt to elicit the same capability in the target model by training a compact module on top of the embeddings to produce ``judgements" of the current continuation. We showcase our strategy on the Llama-3.1 family, where our 8B/405B-Judge achieves a speedup of $9\times$ over Llama-405B, while maintaining its quality on a large range of benchmarks. These benefits remain present even in optimized inference frameworks, where our method reaches up to $141$ tokens/s for 8B/70B-Judge and $129$ tokens/s for 8B/405B on $2$ and $8$ H100s respectively.

Self-improvement is a mechanism in Large Language Model (LLM) pre-training, post-training and test-time inference. We explore a framework where the model verifies its own outputs, filters or reweights data based on this verification, and distills the filtered data. Despite several empirical successes, a fundamental understanding is still lacking. In this work, we initiate a comprehensive, modular and controlled study on LLM self-improvement. We provide a mathematical formulation for self-improvement, which is largely governed by a quantity which we formalize as the **generation-verification gap**. Through experiments with various model families and tasks, we discover a scaling phenomenon of self-improvement -- a variant of the generation-verification gap scales monotonically with the model pre-training flops. We also examine when self-improvement is possible, an iterative self-improvement procedure, and ways to improve its performance. Our findings not only advance understanding of LLM self-improvement with practical implications, but also open numerous avenues for future research into its capabilities and boundaries.

Because it is difficult to precisely specify complex objectives, reinforcement learning policies are often optimized using proxy reward functions that only approximate the true goal. However, optimizing proxy rewards frequently leads to reward hacking: the optimized reward function ceases to be a good proxy and the resulting policy performs poorly with respect to the unspecified true reward. Principled solutions to reward hacking have been impeded by the lack of a good definition for the problem. To address this gap, we introduce a definition of reward hacking based on the correlation between proxy and true rewards for states and actions seen by a “reference policy” that breaks down under optimization. We show that this definition captures reward hacking behavior across several realistic settings, including in reinforcement learning from human feedback (RLHF). Using our formulation, we show theoretically that regularization to the reference policy can effectively prevent reward hacking. While the current practice in RLHF applies a KL penalty between action distributions for this purpose, our theory suggests regularizing the χ2 divergence between the policies’ occupancy measures can be more effective. We intuitively show the benefits of this type of regularization and demonstrate that it better mitigates reward hacking in practice across four realistic settings, including RLHF. Our code is available at https://github.com/cassidylaidlaw/orpo.

The depth of neural networks is a critical factor for their capability, with deeper models often demonstrating superior performance. Motivated by this, significant efforts have been made to enhance layer aggregation - reusing information from previous layers to better extract features at the current layer, to improve the representational power of deep neural networks. However, previous works have primarily addressed this problem from a discrete-state perspective which is not suitable as the number of network layers grows. This paper novelly treats the outputs from layers as states of a continuous process and considers leveraging the state space model (SSM) to design the aggregation of layers in very deep neural networks. Moreover, inspired by its advancements in modeling long sequences, the Selective State Space Models (S6) is employed to design a new module called Selective State Space Model Layer Aggregation (S6LA). This module aims to combine traditional CNN or transformer architectures within a sequential framework, enhancing the representational capabilities of state-of-the-art vision networks. Extensive experiments show that S6LA delivers substantial improvements in both image classification and detection tasks, highlighting the potential of integrating SSMs with contemporary deep learning techniques.

How humans can effectively and efficiently acquire images has always been a perennial question. A classic solution is *text-to-image retrieval* from an existing database; however, the limited database typically lacks creativity. By contrast, recent breakthroughs in *text-to-image generation* have made it possible to produce attractive and counterfactual visual content, but it faces challenges in synthesizing knowledge-intensive images. In this work, we rethink the relationship between text-to-image generation and retrieval, proposing a *unified* framework for both tasks with one single Large Multimodal Model (LMM). Specifically, we first explore the intrinsic discriminative abilities of LMMs and introduce an efficient generative retrieval method for text-to-image retrieval in a training-free manner. Subsequently, we unify generation and retrieval autoregressively and propose an autonomous decision mechanism to choose the best-matched one between generated and retrieved images as the response to the text prompt. To standardize the evaluation of unified text-to-image generation and retrieval, we construct TIGeR-Bench, a benchmark spanning both creative and knowledge-intensive domains. Extensive experiments on TIGeR-Bench and two retrieval benchmarks, *i.e.*, Flickr30K and MS-COCO, demonstrate the superiority of our proposed framework.

Reinforcement learning with human feedback (RLHF) fine-tunes a pretrained large language model (LLM) using user preference data, enabling it to generate content aligned with human preferences. However, due to privacy concerns, users may be reluctant to share sensitive preference data. To address this, we propose utilizing Federated Learning (FL) techniques, allowing large-scale preference collection from diverse real-world users without requiring them to transmit data to a central server. Our federated RLHF methods (i.e., FedBis and FedBiscuit) encode each client’s preferences into binary selectors and aggregate them to capture common preferences. In particular, FedBiscuit overcomes key challenges, such as preference heterogeneity and reward hacking, through innovative solutions like grouping clients with similar preferences to reduce heterogeneity and using multiple binary selectors to enhance LLM output quality. To evaluate the performance of the proposed methods, we establish the first federated RLHF benchmark with a heterogeneous human preference dataset. Experimental results show that by integrating the LLM with aggregated client preferences, FedBis and FedBiscuit significantly enhance the professionalism and readability of the generated content.

Reinforcement Learning (RL) agents have demonstrated their potential across various robotic tasks. However, they still heavily rely on human-engineered reward functions, requiring extensive trial-and-error and access to target behavior information, often unavailable in real-world settings. This paper introduces REDS: REward learning from Demonstration with Segmentations, a novel reward learning framework that leverages action-free videos with minimal supervision. Specifically, REDS employs video demonstrations segmented into subtasks from diverse sources and treats these segments as ground-truth rewards. We train a dense reward function conditioned on video segments and their corresponding subtasks to ensure alignment with ground-truth reward signals by minimizing the Equivalent-Policy Invariant Comparison distance. Additionally, we employ contrastive learning objectives to align video representations with subtasks, ensuring precise subtask inference during online interactions. Our experiments show that REDS significantly outperforms baseline methods on complex robotic manipulation tasks in Meta-World and more challenging real-world tasks, such as furniture assembly in FurnitureBench, with minimal human intervention. Moreover, REDS facilitates generalization to unseen tasks and robot embodiments, highlighting its potential for scalable deployment in diverse environments.

Recent advancements in large language models have significantly improved their context windows, yet challenges in effective long-term memory management remain. We introduce MemTree, an algorithm that leverages a dynamic, tree-structured memory representation to optimize the organization, retrieval, and integration of information, akin to human cognitive schemas. MemTree organizes memory hierarchically, with each node encapsulating aggregated textual content, corresponding semantic embeddings, and varying abstraction levels across the tree's depths. Our algorithm dynamically adapts this memory structure by computing and comparing semantic embeddings of new and existing information to enrich the model’s context-awareness. This approach allows MemTree to handle complex reasoning and extended interactions more effectively than traditional memory augmentation methods, which often rely on flat lookup tables. Evaluations on benchmarks for multi-turn dialogue understanding and document question answering show that MemTree significantly enhances performance in scenarios that demand structured memory management.