ICLR 2025 Papers

Recent advancements in large language models (LLMs) with billions of parameters have improved performance in various applications, but their inference processes demand significant energy and computational resources. In contrast, the human brain, with approximately 86 billion neurons, is much more energy-efficient than LLMs with similar parameters. Inspired by this, we redesign 7$\sim$70 billion parameter LLMs using bio-plausible spiking mechanisms, emulating the efficient behavior of the human brain. We propose the first spiking large language model, SpikeLLM. Coupled with the proposed model, two essential approaches are proposed to improve spike training efficiency: Generalized Integrate-and-Fire (GIF) neurons to compress spike length from $T$ to $\frac{T}{L} \log_2 L$ bits, and an Optimal Brain Spiking framework to divide outlier channels and allocate different $T$ for GIF neurons, which further compresses spike length to approximate $log_2T$ bits. The necessity of spike-driven LLM is proved by comparison with quantized LLMs with similar operations. In the OmniQuant pipeline, SpikeLLM reduces 11.01\% WikiText2 perplexity and improves 2.55\% accuracy of common scene reasoning on a LLAMA-7B W4A4 model. In the GPTQ pipeline, SpikeLLM achieves direct additive in linear layers, significantly exceeding PB-LLMs. Our code is publicly available at https://github.com/Xingrun-Xing2/SpikeLLM.

The impact of communication on decision-making systems has been extensively studied under the assumption of dedicated communication channels. We instead consider communicating through actions, where the message is embedded into the actions of an agent which interacts with the environment in a Markov decision process (MDP) framework. We conceptualize the MDP environment as a finite-state channel (FSC), where the actions of the agent serve as the channel input, while the states of the MDP observed by another agent (i.e., receiver) serve as the channel output. Here, we treat the environment as a communication channel over which the agent communicates through its actions, while at the same time, trying to maximize its reward. We first characterize the optimal information theoretic trade-off between the average reward and the rate of reliable communication in the infinite-horizon regime. Then, we propose a novel framework to design a joint control/coding policy, termed Act2Comm, which seamlessly embeds messages into actions. From a communication perspective, Act2Comm functions as a learning-based channel coding scheme for non-differentiable FSCs under input-output constraints. From a control standpoint, Act2Comm learns an MDP policy that incorporates communication capabilities, though at the cost of some control performance. Overall, Act2Comm effectively balances the dual objectives of control and communication in this environment. Experimental results validate Act2Comm's capability to enable reliable communication while maintaining a certain level of control performance.

Beyond high-fidelity image synthesis, diffusion models have recently exhibited promising results in dense visual perception tasks. However, most existing work treats diffusion models as a standalone component for perception tasks, employing them either solely for off-the-shelf data augmentation or as mere feature extractors. In contrast to these isolated and thus sub-optimal efforts, we introduce an integrated, versatile, diffusion-based framework, Diff-2-in-1, that can simultaneously handle both multi-modal data generation and dense visual perception, through a unique exploitation of the diffusion-denoising process. Within this framework, we further enhance discriminative visual perception via multi-modal generation, by utilizing the denoising network to create multi-modal data that mirror the distribution of the original training set. Importantly, Diff-2-in-1 optimizes the utilization of the created diverse and faithful data by leveraging a novel self-improving learning mechanism. Comprehensive experimental evaluations validate the effectiveness of our framework, showcasing consistent performance improvements across various discriminative backbones and high-quality multi-modal data generation characterized by both realism and usefulness. Our project website is available at https://zsh2000.github.io/diff-2-in-1.github.io/.

Consistency models have recently been introduced to accelerate the generation speed of diffusion models by directly predicting the solution (data) of the probability flow ODE (PF ODE) from initial noise. However, the training of consistency models requires learning to map all intermediate points along PF ODE trajectories to their corresponding endpoints. This task is much more challenging than the ultimate objective of one-step generation, which only concerns the PF ODE's noise-to-data mapping. We empirically find that this training paradigm limits the one-step generation performance of consistency models. To address this issue, we generalize consistency training to the truncated time range, which allows the model to ignore denoising tasks at earlier time steps and focus its capacity on generation. We propose a new parameterization of the consistency function and a two-stage training procedure that prevent the truncated-time training from collapsing to a trivial solution. Experiments on CIFAR-10 and ImageNet $64\times64$ datasets show that our method achieves better one-step and two-step FIDs than the state-of-the-art consistency models such as iCT-deep, using more than 2$\times$ smaller networks.

Biological and artificial neural networks develop internal representations that enable them to perform complex tasks. In artificial networks, the effectiveness of these models relies on their ability to build task specific representation, a process influenced by interactions among datasets, architectures, initialization strategies, and optimization algorithms. Prior studies highlight that different initializations can place networks in either a lazy regime, where representations remain static, or a rich/feature learning regime, where representations evolve dynamically. Here, we examine how initialization influences learning dynamics in deep linear neural networks, deriving exact solutions for lambda-balanced initializations-defined by the relative scale of weights across layers. These solutions capture the evolution of representations and the Neural Tangent Kernel across the spectrum from the rich to the lazy regimes. Our findings deepen the theoretical understanding of the impact of weight initialization on learning regimes, with implications for continual learning, reversal learning, and transfer learning, relevant to both neuroscience and practical applications.

Policy gradient methods have become a staple of any single-agent reinforcement learning toolbox, due to their combination of desirable properties: iterate convergence, efficient use of stochastic trajectory feedback, and theoretically-sound avoidance of importance sampling corrections. In multi-agent imperfect-information settings (extensive-form games), however, it is still unknown whether the same desiderata can be guaranteed while retaining theoretical guarantees. Instead, sound methods for extensive-form games rely on approximating \emph{counterfactual} values (as opposed to Q values), which are incompatible with policy gradient methodologies. In this paper, we investigate whether policy gradient can be safely used in two-player zero-sum imperfect-information extensive-form games (EFGs). We establish positive results, showing for the first time that a policy gradient method leads to provable best-iterate convergence to a regularized Nash equilibrium in self-play.

Parameter-efficient fine-tuning (PEFT) has revolutionized supervised fine-tuning, where LoRA and its variants gain the most popularity due to their low training costs and zero inference latency. However, LoRA tuning not only injects knowledgeable features but also noisy hallucination during fine-tuning, which hinders the utilization of tunable parameters with the increasing LoRA rank. In this work, we first investigate in-depth the redundancies among LoRA parameters with substantial empirical studies. Aiming to resemble the learning capacity of high ranks from the findings, we set up a new fine-tuning framework, \textbf{P}arameter-\textbf{Re}dundant \textbf{F}ine-\textbf{T}uning (\preft), which follows the vanilla LoRA tuning process but is required to reduce redundancies before merging LoRA parameters back to pre-trained models. Based on this framework, we propose \textbf{No}ise reduction with \textbf{R}eserved \textbf{M}ajority~(\norm), which decomposes the LoRA parameters into majority parts and redundant parts with random singular value decomposition. The major components are determined by the proposed \search method, specifically employing subspace similarity to confirm the parameter groups that share the highest similarity with the base weight. By employing \norm, we enhance both the learning capacity and benefits from larger ranks, which consistently outperforms both LoRA and other \preft-based methods on various downstream tasks, such as general instruction tuning, math reasoning and code generation. Code is available at \url{https://github.com/pixas/NoRM}.

Large Language Models (LLMs) have greatly pushed forward advancements in natural language processing, yet their high memory and computational demands hinder practical deployment. Binarization, as an effective compression technique, can shrink model weights to just 1 bit, significantly reducing the high demands on computation and memory. However, current binarization methods struggle to narrow the distribution gap between binarized and full-precision weights, while also overlooking the column deviation in LLM weight distribution. To tackle these issues, we propose ARB-LLM, a novel 1-bit post-training quantization (PTQ) technique tailored for LLMs. To narrow the distribution shift between binarized and full-precision weights, we first design an alternating refined binarization (ARB) algorithm to progressively update the binarization parameters, which significantly reduces the quantization error. Moreover, considering the pivot role of calibration data and the column deviation in LLM weights, we further extend ARB to ARB-X and ARB-RC. In addition, we refine the weight partition strategy with column-group bitmap (CGB), which further enhance performance. Equipping ARB-X and ARB-RC with CGB, we obtain ARB-LLM$_{\text{X}}$ and ARB-LLM$ _{\text{RC}} $ respectively, which significantly outperform state-of-the-art (SOTA) binarization methods for LLMs. As a binary PTQ method, our ARB-LLM$ _{\text{RC}} $ is the first to surpass FP16 models of the same size. Code: https://github.com/ZHITENGLI/ARB-LLM.

Transformer models have demonstrated exceptional performance across a wide range of applications. Though forming the foundation of Transformer models, the dot-product attention does not scale well to long-context data since its time requirement grows quadratically with context length. In this work, we propose Radar, a training-free approach that accelerates inference by dynamically searching for the most important context tokens. For any pre-trained Transformer, Radar can reduce the decoding time complexity without training or heuristically evicting tokens. Moreover, we provide theoretical justification for our approach, demonstrating that Radar can reliably identify the most important tokens with high probability. We conduct extensive comparisons with the previous methods on a wide range of tasks. The results demonstrate that Radar achieves the state-of-the-art performance across different architectures with reduced time complexity, offering a practical solution for efficient long-context processing of Transformers. The code is publicly available at https://github.com/BorealisAI/radar-decoding.

Online and offline $\mathtt{RLHF}$ methods, such as $\mathtt{PPO}$ and $\mathtt{DPO}$, have been highly successful in aligning AI with human preferences. Despite their success, however, these methods suffer from fundamental limitations: $\mathbf{(a)}$ Models trained with $\mathtt{RLHF}$ can learn from mistakes or negative examples through RL mechanism or contrastive loss during training. However, at inference time, they lack an innate self-improvement mechanism for error corrections. $\mathbf{(b)}$ The optimal solution of existing methods is highly task-dependent, making it difficult for them to generalize to new tasks. To address these challenges, we propose Self-Improving Robust Preference Optimization ($\mathtt{SRPO}$), a practical and mathematically principled offline $\mathtt{RLHF}$ framework. The key idea behind $\mathtt{SRPO}$ is to cast the problem of learning from human preferences as a self-improvement process, mathematically formulated as a min-max objective that jointly optimizes a self-improvement policy and a generative policy in an adversarial fashion. Crucially, the solution for this optimization problem is independent of the training task, which makes it robust to its changes. We then show that this objective can be reformulated as a non-adversarial offline loss, which can be efficiently optimized using standard supervised learning techniques at scale. To demonstrate $\mathtt{SRPO}$’s effectiveness, we evaluate it using AI Win-Rate (WR) against human (GOLD) completions. When tested on the XSum dataset, $\mathtt{SRPO}$ outperforms $\mathtt{DPO}$ by a margin of $\mathbf{15}$% after $5$ self-revisions, achieving an impressive $\mathbf{90}$% WR. Moreover, on the challenging Arena-Hard prompts, $\mathtt{SRPO}$ outperforms both $\mathtt{DPO}$ and $\mathtt{IPO}$ (by $\mathbf{4}$% without revision and $\mathbf{6}$% after a single revision), reaching a $\mathbf{56}$% WR against against $\mathtt{Llama-3.1-8B-Instruct}$.

We study the problem of learning from truncated samples: instead of observing samples from some underlying population $p^\ast$, we observe only the examples that fall in some survival set $S \subset \mathbb{R}^d$ whose probability mass (measured with respect to $p^\ast$) is at least $\alpha$. Assuming membership oracle access to the truncation set $S$, prior works obtained algorithms for the case where $p^\ast$ is Gaussian or more generally an exponential family with strongly convex likelihood --- albeit with a super-polynomial dependency on the (inverse) survival mass $1/\alpha$ both in terms of runtime and in number of oracle calls to the set $S$. In this work we design a new learning method with runtime and query complexity polynomial in $1/\alpha$. Our result significantly improves over the prior works by focusing on efficiently solving the underlying optimization problem using a general purpose optimization algorithm with minimal assumptions.

Remarkable progress in text-to-image diffusion models has brought a major concern about potentially generating images on inappropriate or trademarked concepts. Concept erasing has been investigated with the goals of deleting target concepts in diffusion models while preserving other concepts with minimal distortion. To achieve these goals, recent concept erasing methods usually fine-tune the cross-attention layers of diffusion models. In this work, we first show that merely updating the cross-attention layers in diffusion models, which is mathematically equivalent to adding linear modules to weights, may not be able to preserve diverse remaining concepts. Then, we propose a novel framework, dubbed Concept Pinpoint Eraser (CPE), by adding nonlinear Residual Attention Gates (ResAGs) that selectively erase (or cut) target concepts while safeguarding remaining concepts from broad distributions by employing an attention anchoring loss to prevent the forgetting. Moreover, we adversarially train CPE with ResAG and learnable text embeddings in an iterative manner to maximize erasing performance and enhance robustness against adversarial attacks. Extensive experiments on the erasure of celebrities, artistic styles, and explicit contents demonstrated that the proposed CPE outperforms prior arts by keeping diverse remaining concepts while deleting the target concepts with robustness against attack prompts. Code is available at https://github.com/Hyun1A/CPE.

We consider the problem of learning nonlinear dynamical systems from a single sample trajectory. While the least squares estimate (LSE) is commonly used for this task, it suffers from poor identification errors when the sample size is small or the model fails to capture the system's true dynamics. To overcome these limitations, we propose a robust LSE framework, which incorporates robust optimization techniques, and prove that it is equivalent to regularizing LSE using general Schatten $p$-norms. We provide non-asymptotic performance guarantees for linear systems, achieving an error rate of $\widetilde{\mathcal{O}}(1/\sqrt{T})$, and show that it avoids the curse of dimensionality, unlike state-of-the-art Wasserstein robust optimization models. Empirical results demonstrate substantial improvements in real-world system identification and online control tasks, outperforming existing methods.

Diffusion models excel at creating visually impressive images but often struggle to generate images with a specified topology. The Betti number, which represents the number of structures in an image, is a fundamental measure in topology. Yet, diffusion models fail to satisfy even this basic constraint. This limitation restricts their utility in applications requiring exact control, like robotics and environmental modeling. To address this, we propose TopoDiffusionNet (TDN), a novel approach that enforces diffusion models to maintain the desired topology. We leverage tools from topological data analysis, particularly persistent homology, to extract the topological structures within an image. We then design a topology-based objective function to guide the denoising process, preserving intended structures while suppressing noisy ones. Our experiments across four datasets demonstrate significant improvements in topological accuracy. TDN is the first to integrate topology with diffusion models, opening new avenues of research in this area.

Video understanding is a crucial next step for multimodal large language models (MLLMs). Various benchmarks are introduced for better evaluating the MLLMs. Nevertheless, current video benchmarks are still inefficient for evaluating video models during iterative development due to the high cost of constructing datasets and the difficulty in isolating specific skills. In this paper, we propose VideoNIAH (Video Needle in A Haystack), a benchmark construction framework through synthetic video generation. VideoNIAH decouples video content from their query-responses by inserting unrelated visual 'needles' into original videos. The framework automates the generation of query-response pairs using predefined rules, minimizing manual labor. The queries focus on specific aspects of video understanding, enabling more skill-specific evaluations. The separation between video content and the queries also allow for increased video variety and evaluations across different lengths. Utilizing VideoNIAH, we compile a video benchmark, VNBench, which includes tasks such as retrieval, ordering, and counting to evaluate three key aspects of video understanding: temporal perception, chronological ordering, and spatio-temporal coherence. We conduct a comprehensive evaluation of both proprietary and open-source models, uncovering significant differences in their video understanding capabilities across various tasks. Additionally, we perform an in-depth analysis of the test results and model configurations. Based on these findings, we provide some advice for improving video MLLM training, offering valuable insights to guide future research and model development.

Whether deep neural networks (DNNs) memorize the training data is a fundamental open question in understanding deep learning. A direct way to verify the memorization of DNNs is to reconstruct training data from DNNs’ parameters. Since parameters are gradually determined by data throughout training, characterizing training dynamics is important for reconstruction. Pioneering works rely on the linear training dynamics of shallow NNs with large widths, but cannot be extended to more practical DNNs which have non-linear dynamics. We propose Simulation of training Dynamics (SimuDy) to reconstruct training data from DNNs. Specifically, we simulate the training dynamics by training the model from the initial parameters with a dummy dataset, then optimize this dummy dataset so that the simulated dynamics reach the same final parameters as the true dynamics. By incorporating dummy parameters in the simulated dynamics, SimuDy effectively describes non-linear training dynamics. Experiments demonstrate that SimuDy significantly outperforms previous approaches when handling non-linear training dynamics, and for the first time, most training samples can be reconstructed from a trained ResNet’s parameters.

Large language models (LLMs) have achieved remarkable success across diverse tasks, yet their inference processes are hindered by substantial time and energy demands due to single-token generation at each decoding step. While previous methods such as speculative decoding mitigate these inefficiencies by producing multiple tokens per step, each token is still generated by its single-token distribution, thereby enhancing speed without improving effectiveness. In contrast, our work simultaneously enhances inference speed and improves the output effectiveness. We consider multi-token joint decoding (MTJD), which generates multiple tokens from their joint distribution at each iteration, theoretically reducing perplexity and enhancing task performance. However, MTJD suffers from the high cost of sampling from the joint distribution of multiple tokens. Inspired by speculative decoding, we introduce multi-token assisted decoding (MTAD), a novel framework designed to accelerate MTJD. MTAD leverages a smaller auxiliary model to approximate the joint distribution of a larger model, incorporating a verification mechanism that not only ensures the accuracy of this approximation, but also improves the decoding efficiency over conventional speculative decoding. Theoretically, we demonstrate that MTAD closely approximates exact MTJD with bounded error. Empirical evaluations using Llama-2 and OPT models ranging from 13B to 70B parameters across various tasks reveal that MTAD reduces perplexity by 21.2% and improves downstream performance compared to standard single-token sampling. Furthermore, MTAD achieves a 1.42× speed-up and consumes 1.54× less energy than conventional speculative decoding methods. These results highlight MTAD’s ability to make multi-token joint decoding both effective and efficient, promoting more sustainable and high-performance deployment of LLMs.

Data analytics is essential for extracting valuable insights from data that can assist organizations in making effective decisions. We introduce InsightBench, a benchmark dataset with three key features. First, it consists of 100 datasets representing diverse business use cases such as finance and incident management, each accompanied by a carefully curated set of insights planted in the datasets. Second, unlike existing benchmarks focusing on answering single queries, InsightBench evaluates agents based on their ability to perform end-to-end data analytics, including formulating questions, interpreting answers, and generating a summary of insights and actionable steps. Third, we conducted comprehensive quality assurance to ensure that each dataset in the benchmark had clear goals and included relevant and meaningful questions and analysis. Furthermore, we implement a two-way evaluation mechanism using LLaMA-3 as an effective, open-source evaluator to assess agents’ ability to extract insights. We also propose AgentPoirot, our baseline data analysis agent capable of performing end-to-end data analytics. Our evaluation on InsightBench shows that AgentPoirot outperforms existing approaches (such as Pandas Agent) that focus on resolving single queries. We also compare the performance of open- and closed-source LLMs and various evaluation strategies. Overall, this benchmark serves as a testbed to motivate further development in comprehensive automated data analytics and can be accessed here: https://github.com/ServiceNow/insight-bench.

The problem of pre-training data detection for large language models (LLMs) has received growing attention due to its implications in critical issues like copyright violation and test data contamination. Despite improved performance, existing methods (including the state-of-the-art, Min-K%) are mostly developed upon simple heuristics and lack solid, reasonable foundations. In this work, we propose a novel and theoretically motivated methodology for pre-training data detection, named Min-K%++. Specifically, we present a key insight that training samples tend to be local maxima of the modeled distribution along each input dimension through maximum likelihood training, which in turn allow us to insightfully translate the problem into identification of local maxima. Then, we design our method accordingly that works under the discrete distribution modeled by LLMs, whose core idea is to determine whether the input forms a mode or has relatively high probability under the conditional categorical distribution. Empirically, the proposed method achieves new SOTA performance across multiple settings (evaluated with 5 families of 10 models and 2 benchmarks). On the WikiMIA benchmark, Min-K%++ outperforms the runner-up by 6.2% to 10.5% in detection AUROC averaged over five models. On the more challenging MIMIR benchmark, it consistently improves upon reference-free methods while performing on par with reference-based method that requires an extra reference model.