ICLR 2025 论文列表

Large language models (LLMs) have achieved remarkable performance on vari- ous natural language tasks. However, they are trained on static corpora and their knowledge can become outdated quickly in the fast-changing world. This moti- vates the development of knowledge editing methods designed to update certain knowledge in LLMs without changing unrelated others. To make selective edits, previous efforts often sought to update a small amount of parameters in some spe- cific layer(s) of a LLM. Nonetheless, in challenging scenarios, they still fall short in making successful edits while preserving knowledge irrelevant to the updates simultaneously, resulting in a notable editing-locality trade-off. In this work, we question if the trade-offs are caused by the fact that parameter-based updates have a global effect, i.e., edited parameters affect all inputs indiscriminately. In light of this, we explore the feasibility of representation fine-tuning, which applied some linear update to a few representations in a learned subspace, for knowledge edit- ing. While being effective to enhance an LLM’s general ability as demonstrated in the previous work, we theoretically show that this linear update imposes a tension in editing-locality trade-off. Subsequently, BaFT is proposed to break the linear- ity. BaFT computes a weight for each basis that spans a dimension of the subspace based on the input representation. This input-dependent weighting mechanism al- lows BaFT to manage different types of knowledge in an adaptive way, thereby achieving a better editing-locality trade-off. Experiments on three LLMs with five editing benchmarks in diverse scenarios show the superiority of our method.

The Maximal Update Parametrization ($\mu$P) aims to make the optimal hyperparameters (HPs) of a model independent of its size, allowing them to be swept using a cheap proxy model rather than the full-size target model. We present a new scheme, u-$\mu$P, which improves upon $\mu$P by combining it with Unit Scaling, a method for designing models that makes them easy to train in low-precision. The two techniques have a natural affinity: $\mu$P ensures that the scale of activations is independent of model size, and Unit Scaling ensures that activations, weights and gradients begin training with a scale of one. This synthesis opens the door to a simpler scheme, whose default values are near-optimal. This in turn facilitates a more efficient sweeping strategy, with u-$\mu$P models reaching a lower loss than comparable $\mu$P models and working out-of-the-box in FP8.

Large language models (LLMs) encounter computational challenges during long-sequence inference, especially in the attention pre-filling phase, where the complexity grows quadratically with the prompt length. Previous efforts to mitigate these challenges have relied on fixed sparse attention patterns or identifying sparse attention patterns based on limited cases. However, these methods lacked the flexibility to efficiently adapt to varying input demands. In this paper, we introduce FlexPrefill, a Flexible sparse Pre-filling mechanism that dynamically adjusts sparse attention patterns and computational budget in real-time to meet the specific requirements of each input and attention head. The flexibility of our method is demonstrated through two key innovations: 1) Query-Aware Sparse Pattern Determination: By measuring Jensen-Shannon divergence, this component adaptively switches between query-specific diverse attention patterns and predefined attention patterns. 2) Cumulative-Attention Based Index Selection: This component dynamically selects query-key indexes to be computed based on different attention patterns, ensuring the sum of attention scores meets a predefined threshold. FlexPrefill adaptively optimizes the sparse pattern and sparse ratio of each attention head based on the prompt, enhancing efficiency in long-sequence inference tasks. Experimental results show significant improvements in both speed and accuracy over prior methods, providing a more flexible and efficient solution for LLM inference.

In spite of the great potential of large language models (LLMs) across various tasks, their deployment on resource-constrained devices remains challenging due to their excessive computational and memory demands. Quantization has emerged as an effective solution by storing weights in reduced precision. However, utilizing low precisions (i.e.~2/3-bit) to substantially alleviate the memory-boundedness of LLM decoding, still suffers from prohibitive performance drop. In this work, we argue that existing approaches fail to explore the diversity in computational patterns, redundancy, and sensitivity to approximations of the different phases of LLM inference, resorting to a uniform quantization policy throughout. Instead, we propose a novel phase-aware method that selectively allocates precision during different phases of LLM inference, achieving both strong context extraction during prefill and efficient memory bandwidth utilization during decoding. To further address the memory-boundedness of the decoding phase, we introduce Progressive Mixed-Precision Decoding (PMPD), a technique that enables the gradual lowering of precision deeper in the generated sequence, together with a spectrum of precision-switching schedulers that dynamically drive the precision-lowering decisions in either task-adaptive or prompt-adaptive manner. Extensive evaluation across diverse language tasks shows that when targeting Nvidia GPUs, PMPD achieves 1.4$-$12.2$\times$ speedup in matrix-vector multiplications over fp16 models, while when targeting an LLM-optimized NPU, our approach delivers a throughput gain of 3.8$-$8.0$\times$ over fp16 models and up to 1.54$\times$ over uniform quantization approaches while preserving the output quality.

Graph neural networks (GNNs) have shown considerable promise in computational chemistry. However, the limited availability of molecular data raises concerns regarding GNNs' ability to effectively capture the fundamental principles of physics and chemistry, which constrains their generalization capabilities. To address this challenge, we introduce a novel self-supervised approach termed Equivariant Masked Position Prediction (EMPP), grounded in intramolecular potential and force theory. Unlike conventional attribute masking techniques, EMPP formulates a nuanced position prediction task that is more well-defined and enhances the learning of quantum mechanical features. EMPP also bypasses the approximation of the Gaussian mixture distribution commonly used in denoising methods, allowing for more accurate acquisition of physical properties. Experimental results indicate that EMPP significantly enhances performance of advanced molecular architectures, surpassing state-of-the-art self-supervised approaches. Our code is released in https://github.com/ajy112/EMPP.

We discover a theoretical connection between explanation estimation and distribution compression that significantly improves the approximation of feature attributions, importance, and effects. While the exact computation of various machine learning explanations requires numerous model inferences and becomes impractical, the computational cost of approximation increases with an ever-increasing size of data and model parameters. We show that the standard i.i.d. sampling used in a broad spectrum of algorithms for post-hoc explanation leads to an approximation error worthy of improvement. To this end, we introduce Compress Then Explain (CTE), a new paradigm of sample-efficient explainability. It relies on distribution compression through kernel thinning to obtain a data sample that best approximates its marginal distribution. CTE significantly improves the accuracy and stability of explanation estimation with negligible computational overhead. It often achieves an on-par explanation approximation error 2-3x faster by using fewer samples, i.e. requiring 2-3x fewer model evaluations. CTE is a simple, yet powerful, plug-in for any explanation method that now relies on i.i.d. sampling.

In the field of image editing, three core challenges persist: controllability, background preservation, and efficiency. Inversion-based methods rely on time-consuming optimization to preserve the features of the initial images, which results in low efficiency due to the requirement for extensive network inference. Conversely, inversion-free methods lack theoretical support for background similarity, as they circumvent the issue of maintaining initial features to achieve efficiency. As a consequence, none of these methods can achieve both high efficiency and background consistency. To tackle the challenges and the aforementioned disadvantages, we introduce PostEdit, a method that incorporates a posterior scheme to govern the diffusion sampling process. Specifically, a corresponding measurement term related to both the initial features and Langevin dynamics is introduced to optimize the estimated image generated by the given target prompt. Extensive experimental results indicate that the proposed PostEdit achieves state-of-the-art editing performance while accurately preserving unedited regions. Furthermore, the method is both inversion- and training-free, necessitating approximately 1.5 seconds and 18 GB of GPU memory to generate high-quality results.

Recent advancements in Transformer-based large language models (LLMs) have set new standards in natural language processing. However, the classical softmax attention incurs significant computational costs, leading to a $O(T)$ complexity for per-token generation, where $T$ represents the context length. This work explores reducing LLMs' complexity while maintaining performance by introducing Rodimus and its enhanced version, Rodimus$+$. Rodimus employs an innovative data-dependent tempered selection (DDTS) mechanism within a linear attention-based, purely recurrent framework, achieving significant accuracy while drastically reducing the memory usage typically associated with recurrent models. This method exemplifies semantic compression by maintaining essential input information with fixed-size hidden states. Building on this, Rodimus$+$ combines Rodimus with the innovative Sliding Window Shared-Key Attention (SW-SKA) in a hybrid approach, effectively leveraging the complementary semantic, token, and head compression techniques. Our experiments demonstrate that Rodimus$+$-1.6B, trained on 1 trillion tokens, achieves superior downstream performance against models trained on more tokens, including Qwen2-1.5B and RWKV6-1.6B, underscoring its potential to redefine the accuracy-efficiency balance in LLMs. Model code and pre-trained checkpoints are open-sourced at https://github.com/codefuse-ai/rodimus.

We present Self-MoE, an approach that transforms a monolithic LLM into a compositional, modular system of self-specialized experts, named MiXSE (MiXture of Self-specialized Experts). Our approach leverages self-specialization, which constructs expert modules using self-generated synthetic data, each equipping a shared base LLM with distinct domain-specific capabilities, activated via self-optimized routing. This allows for dynamic and capability-specific handling of various target tasks, enhancing overall capabilities, without extensive human-labeled data and added parameters. Our empirical results reveal that specializing LLMs may exhibit potential trade-offs in performances on non-specialized tasks. On the other hand, our Self-MoE demonstrates substantial improvements (6.5%p on average) over the base LLM across diverse benchmarks such as knowledge, reasoning, math, and coding. It also consistently outperforms other methods, including instance merging and weight merging, while offering better flexibility and interpretability by design with semantic experts and routing. Our findings highlight the critical role of modularity, the applicability of Self-MoE to multiple base LLMs, and the potential of self-improvement in achieving efficient, scalable, and adaptable systems.

Efficient generative inference in Large Language Models (LLMs) is impeded by the growing memory demands of Key-Value (KV) cache, especially for longer sequences. Traditional KV Cache eviction strategies, which discard less critical KV-pairs based on attention scores, often degrade generation quality, leading to issues such as context loss or hallucinations. To address this, we introduce **D**ynamic **D**iscriminative **O**perations ($\mathbf{D_2 O}$), a novel method that optimizes KV cache size dynamically and discriminatively at two levels without fine-tuning, while preserving essential context. At **layer-level**, by observing the varying densities of attention weights between shallow and deep layers, we dynamically determine which layers should avoid excessive eviction via our proposed ***dynamic allocation strategy*** to minimize information loss. At **token-level**, for the eviction strategy in each layer, $\mathbf{D_2 O}$ innovatively incorporates a ***compensation mechanism*** that maintains a similarity threshold to re-discriminate the importance of currently discarded tokens, determining whether they should be recalled and merged with similar tokens. Extensive experiments on various benchmarks and LLM architectures have shown that $\mathbf{D_2 O}$ not only achieves significant memory savings and enhances inference throughput by more than 3$\times$ but also maintains high-quality long-text generation.

Panoptic segmentation, which unifies semantic and instance segmentation into a single task, has witnessed considerable success on predefined tasks. However, traditional methods tend to struggle with catastrophic forgetting and poor generalization when learning from a continuous stream of new tasks. While continual learning aims to mitigate these challenges, our study reveals that existing continual panoptic segmentation (CPS) methods often suffer from efficiency or scalability issues. To address these limitations, we propose an efficient adaptation framework that incorporates attentive self-distillation and dual-decoder prediction fusion to efficiently preserve prior knowledge while facilitating model generalization. Specifically, we freeze the majority of model weights, enabling a shared forward pass between the teacher and student models during distillation. Attentive self-distillation then adaptively distills useful knowledge from the old classes without being distracted from non-object regions, which effectively enhances knowledge retention. Additionally, query-level fusion (QLF) is devised to seamlessly integrate the output of the dual decoders without incurring scale inconsistency. Our method achieves state-of-the-art performance on ADE20K and COCO benchmarks. Code is available at https://github.com/Ze-Yang/ADAPT.

Data Shapley offers a principled framework for attributing the contribution of data within machine learning contexts. However, the traditional notion of Data Shapley requires re-training models on various data subsets, which becomes computationally infeasible for large-scale models. Additionally, this retraining-based definition cannot evaluate the contribution of data for a specific model training run, which may often be of interest in practice. This paper introduces a novel concept, In-Run Data Shapley, which eliminates the need for model retraining and is specifically designed for assessing data contribution for a particular model of interest. In-Run Data Shapley calculates the Shapley value for each gradient update iteration and accumulates these values throughout the training process. We present several techniques that allow the efficient scaling of In-Run Data Shapley to the size of foundation models. In its most optimized implementation, our method adds negligible runtime overhead compared to standard model training. This dramatic efficiency improvement makes it possible to perform data attribution for the foundation model pretraining stage. We present several case studies that offer fresh insights into pretraining data's contribution and discuss their implications for copyright in generative AI and pretraining data curation.

Generating high-quality 3D objects from textual descriptions remains a challenging problem due to high computational costs, the scarcity of 3D data, and the complexity of 3D representations. We introduce Geometry Image Diffusion (GIMDiffusion), a novel Text-to-3D model that utilizes geometry images to efficiently represent 3D shapes using 2D images, thereby avoiding the need for complex 3D-aware architectures. By integrating a Collaborative Control mechanism, we exploit the rich 2D priors of existing Text-to-Image models, such as Stable Diffusion, to achieve strong generalization despite limited 3D training data. This allows us to use only high-quality training data while retaining compatibility with guidance techniques such as IPAdapter. GIMDiffusion enables the generation of 3D assets at speeds comparable to current Text-to-Image models, without being restricted to manifold meshes during either training or inference. We simultaneously generate a UV unwrapping for the objects, consisting of semantically meaningful parts as well as internal structures, enhancing both usability and versatility.

Self-supervised learning (SSL) offers a compelling solution to the challenge of extensive labeled data requirements in traditional supervised learning. With the proven success of Vision Transformers (ViTs) in supervised tasks, there is increasing interest in adapting them for SSL frameworks. However, the high computational demands of SSL pose substantial challenges, particularly on resource-limited platforms like edge devices, despite its ability to achieve high accuracy without labeled data. Recent studies in supervised learning have shown that token pruning can reduce training costs by removing less informative tokens without compromising accuracy. However, SSL’s dual-branch encoders make traditional single-branch pruning strategies less effective, as they fail to account for the critical cross-branch similarity information, leading to reduced accuracy in SSL. To this end, we introduce SimPrune, a novel token pruning strategy designed for ViTs in SSL. SimPrune leverages cross-branch similarity information to efficiently prune tokens, retaining essential semantic information across dual branches. Additionally, we incorporate a difficulty-aware pruning strategy to further enhance SimPrune's effectiveness. Experimental results show that our proposed approach effectively reduces training computation while maintaining accuracy. Specifically, our approach offers 24\% savings in training costs compared to SSL baseline, without sacrificing accuracy.

Humans excel at reusing prior knowledge to address new challenges and developing skills while solving problems. This paradigm becomes increasingly popular in the development of autonomous agents, as it develops systems that can self-evolve in response to new challenges like human beings. However, previous methods suffer from limited training efficiency when expanding new skills and fail to fully leverage prior knowledge to facilitate new task learning. We propose Parametric Skill Expansion and Composition (PSEC), a new framework designed to iteratively evolve the agents' capabilities and efficiently address new challenges by maintaining a manageable skill library. This library can progressively integrate skill primitives as plug-and-play Low-Rank Adaptation (LoRA) modules in parameter-efficient finetuning, facilitating efficient and flexible skill expansion. This structure also enables the direct skill compositions in parameter space by merging LoRA modules that encode different skills, leveraging shared information across skills to effectively program new skills. Based on this, we propose a context-aware modular to dynamically activate different skills to collaboratively handle new tasks. Empowering diverse applications including multi-objective composition, dynamics shift, and continual policy shift, the results on D4RL, DSRL benchmarks, and the DeepMind Control Suite show that PSEC exhibits superior capacity to leverage prior knowledge to efficiently tackle new challenges, as well as expand its skill libraries to evolve the capabilities. Project website: https://ltlhuuu.github.io/PSEC/.

Scaling laws describe the relationship between the size of language models and their capabilities. Unlike prior studies that evaluate a model's capability via loss or benchmarks, we estimate information-theoretically the number of knowledge \emph{bits} a model stores. We focus on factual knowledge represented as tuples, such as (USA, capital, Washington D.C.) from a Wikipedia page. Through multiple controlled datasets, we establish that language models can and only can store \emph{2 bits of knowledge per parameter, even when quantized to int8}, and such knowledge can be flexibly extracted for downstream applications. More broadly, we present 12 results on how (1) training duration, (2) model architecture, (3) quantization, (4) sparsity constraints such as MoE, and (5) data signal-to-noise ratio affect a model's knowledge storage capacity.

Fine-tuning large language models (LLMs) is crucial for task specialization but often becomes resource-intensive due to redundant or uninformative data. Existing data selection methods typically rely either on computationally expensive gradient-based metrics or static embeddings that fail to adapt dynamically to the model’s evolving state, thus limiting their practical effectiveness. To address this, we propose DELIFT (Data Efficient Language model Instruction Fine-Tuning), leveraging a novel, computationally efficient utility metric inspired by In-Context Learning (ICL). Our ICL-based metric measures the informational value of each data sample by quantifying its effectiveness as an in-context example in improving model predictions for other samples, reflecting its actual contribution relative to the model’s current state. Integrated with tailored submodular optimization methods, DELIFT systematically selects diverse, informative subsets optimized specifically for each fine-tuning stage: instruction tuning, task-specific adaptation, and continual fine-tuning. Experimental results across multiple datasets and model scales show DELIFT reduces fine-tuning data requirements by up to 70% without compromising performance, consistently outperforming existing methods by up to 26% in effectiveness and efficiency.

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.

Large-scale pre-trained language models (PLMs) require significant computational resources to train from scratch on large volumes of data. But in the real world, emerging data from diverse sources may not be initially available for pre-training. Recent studies on lifelong learning have tried to solve this problem by exploring the use of model growth techniques to effectively incorporate new knowledge without the need for complete re-training. However, model growth approaches utilized have issues with growth operators that do not ensure strict function preservation or growth schedules that only include a few growth dimensions, reducing lifelong learning's effect. Furthermore, existing approaches often assume that emerging data has the same distribution as pre-training data, causing catastrophic forgetting of previously acquired knowledge. To address the aforementioned issues, we introduce LOIRE, a framework for lifelong learning that enables PLMs to effectively grow their capacity using incremental data. LOIRE employs growth operators for all feasible dimensions and a growth schedule to generate the optimal expansion sequence in the field of lifelong learning. Specifically, we present a novel plug-in layer growth operator with residual connections that skip the newly added layer during initial training while ensuring function preservation. We additionally propose an iterative distillation strategy for LOIRE that allows an intermediate model in the growth stages to switch between being a student and a teacher, reducing catastrophic forgetting during growth. Experiments show that LOIRE can reduce computational expenses by an average of 29.22\% while retaining equivalent or better downstream performance.