ICLR 2025 论文列表

Post-training KV-Cache compression methods typically either sample a subset of effectual tokens or quantize the data into lower numerical bit width. However, these methods cannot exploit redundancy in the hidden dimension of the KV tenors. This paper presents a hidden dimension compression approach called Palu, a KV-Cache compression framework that utilizes low-rank projection to reduce inference-time LLM memory usage. Palu decomposes the linear layers into low-rank matrices, caches compressed intermediate states, and reconstructs the full keys and values on the fly. To improve accuracy, compression rate, and efficiency, Palu further encompasses (1) a medium-grained low-rank decomposition scheme, (2) an efficient rank search algorithm, (3) low-rank-aware quantization compatibility enhancements, and (4) an optimized GPU kernel with matrix fusion. Extensive experiments with popular LLMs show that Palu compresses KV-Cache by 50% while maintaining strong accuracy and delivering up to 1.89× speedup on the RoPE-based attention module. When combined with quantization, Palu’s inherent quantization-friendly design yields small to negligible extra accuracy degradation while saving additional memory than quantization-only methods and achieving up to 2.91× speedup for the RoPE-based attention. Moreover, it maintains comparable or even better accuracy (up to 1.19 lower perplexity) compared to quantization-only methods. These results demonstrate Palu’s superior capability to effectively address the efficiency and memory challenges of LLM inference posed by KV-Cache. Our code is publicly available at: https://github.com/shadowpa0327/Palu.

The long-context capability of the Large Language Models (LLM) has made significant breakthroughs, but \textit{the maximum supported context length in length extrapolation} remains a critical bottleneck limiting their practical applications. The constraint of context length in LLMs arises from the self-attention mechanism, which cannot effectively and efficiently capture the semantic relationships within infinitely long contexts via the limited pre-trained positional information and attention scope. In this work, we propose \textbf{ReAttention}, a training-free approach enabling LLM based on the self-attention mechanism to support an infinite context with a finite attention scope under sufficient memory resources. ReAttention performs the position-agnostic top-$k$ attention before the ordinary position-aware self-attention, freeing LLMs from the length extrapolation issue. We validate the performance of ReAttention on the LongBench, L-Eval, and InfiniteBench and demonstrate that it is on par with traditional methods. Furthermore, we also apply ReAttention on mainstream LLMs, including LLaMA3.1-8B and Mistral-v0.3-7B, enabling them to support context lengths of at least 1M and even expanding the context length of LLaMA3.2-3B-chat by 128$\times$ to 4M without any further training in Needle-In-A-Haystack tests. We also improve the efficiency of ReAttention with Triton and achieve an efficient extrapolation without additional overhead. The code is available at \url{https://github.com/OpenMOSS/ReAttention}.

Unsupervised meta-learning aims to learn feature representations from unsupervised datasets that can transfer to downstream tasks with limited labeled data. In this paper, we propose a novel approach to unsupervised meta-learning that leverages the generalization abilities of in-context learning observed in transformer architectures. Our method reframes meta-learning as a sequence modeling problem, enabling the transformer encoder to learn task context from support images and utilize it to predict query images. At the core of our approach lies the creation of diverse tasks generated using a combination of data augmentations and a mixing strategy that challenges the model during training while fostering generalization to unseen tasks at test time. Experimental results on benchmark datasets showcase the superiority of our approach over existing unsupervised meta-learning baselines, establishing it as the new state-of-the-art. Remarkably, our method achieves competitive results with supervised and self-supervised approaches, underscoring its efficacy in leveraging generalization over memorization.

In this work, we explore the mechanism of in-context learning (ICL) on out-of-distribution (OOD) tasks that were not encountered during training. To achieve this, we conduct synthetic experiments where the objective is to learn OOD mathematical functions through ICL using a GPT-2 model. We reveal that Transformers may struggle to learn OOD task functions through ICL. Specifically, ICL performance resembles implementing a function within the pretraining hypothesis space and optimizing it with gradient descent based on the in-context examples. Additionally, we investigate ICL's well-documented ability to learn unseen abstract labels in context. We demonstrate that such ability only manifests in the scenarios without distributional shifts and, therefore, may not serve as evidence of new-task-learning ability. Furthermore, we assess ICL's performance on OOD tasks when the model is pretrained on multiple tasks. Both empirical and theoretical analyses demonstrate the existence of the \textbf{low-test-error preference} of ICL, where it tends to implement the pretraining function that yields low test error in the testing context. We validate this through numerical experiments. This new theoretical result, combined with our empirical findings, elucidates the mechanism of ICL in addressing OOD tasks.

This paper introduces LLM-Streamline, a pioneer work on layer pruning for large language models (LLMs). It is based on the observation that different layers have varying impacts on hidden states, enabling the identification of less important layers to be pruned. LLM-Streamline comprises two parts: layer pruning, which removes consecutive layers with the lowest importance based on target sparsity, and layer replacement, a novel module that trains a lightweight network to replace the pruned layers to mitigate performance loss. Additionally, a new metric called stability is proposed to address the limitations of the widely used accuracy metric in evaluating model compression. Experiments show that LLM-Streamline outperforms both previous and concurrent state-of-the-art pruning methods in terms of both performance and training efficiency. Our code is available at \href{https://github.com/RUCKBReasoning/LLM-Streamline}{this repository}.

In-context Learning (ICL) empowers large language models (LLMs) to swiftly adapt to unseen tasks at inference-time by prefixing a few demonstration examples before queries. Despite its versatility, ICL incurs substantial computational and memory overheads compared to zero-shot learning and is sensitive to the selection and order of demonstration examples. In this work, we introduce \textbf{Implicit In-context Learning} (I2CL), an innovative paradigm that reduces the inference cost of ICL to that of zero-shot learning with minimal information loss. I2CL operates by first generating a condensed vector representation, namely a context vector, extracted from the demonstration examples. It then conducts an inference-time intervention through injecting a linear combination of the context vector and query activations back into the model’s residual streams. Empirical evaluation on nine real-world tasks across three model architectures demonstrates that I2CL achieves few-shot level performance at zero-shot inference cost, and it exhibits robustness against variations in demonstration examples. Furthermore, I2CL facilitates a novel representation of ``task-ids'', enhancing task similarity detection and fostering effective transfer learning. We also perform a comprehensive analysis and ablation study on I2CL, offering deeper insights into its internal mechanisms. Code is available at https://github.com/LzVv123456/I2CL.

Large Language Models (LLMs) have significantly advanced AI with their exceptional performance across a wide range of tasks. However, their extensive computational requirements restrict their use on devices with limited resources. While recent compression methods based on low-rank matrices show potential solutions, they often suffer from significant loss of accuracy or introduce substantial overhead in parameters and inference time. In this paper, we introduce Modular De- composition (MoDeGPT), a new, efficient, and structured compression framework that overcomes these limitations. MoDeGPT jointly decomposes pairs of consecu- tive subcomponents within Transformer blocks, reduces hidden dimensions through output reconstruction on a larger structural scale than conventional low-rank meth- ods, and repurposes three classical matrix decomposition algorithms—Nyström approximation, CR decomposition, and SVD—to ensure bounded errors in our novel decomposition approach. Our experiments show that MoDeGPT, without relying on backward propagation, consistently matches or surpasses the performance of prior techniques that depend on gradient information, while achieving a 98% reduction in compute costs when compressing a 13B-parameter model. On LLaMA-2/3 and OPT models, MoDeGPT retains 90-95% of zero-shot performance with compression rates of 25-30%. The compression process can be completed on a single GPU in a few hours, boosting inference throughput by up to 46%.

Providing Large Language Models with relevant contextual knowledge at inference time has been shown to greatly improve the quality of their generations. This is often achieved by prepending informative passages of text, or 'contexts', retrieved from external knowledge bases to their input. However, processing additional contexts online incurs significant computation costs that scale with their length. State Space Models (SSMs) offer a promising solution by allowing a database of contexts to be mapped onto fixed-dimensional states from which to start the generation. A key challenge arises when attempting to leverage information present across multiple contexts, since there is no straightforward way to condition generation on multiple independent states in existing SSMs. To address this, we leverage a simple mathematical relation derived from SSM dynamics to compose multiple states into one that efficiently approximates the effect of concatenating raw context tokens. Since the temporal ordering of contexts can often be uninformative, we enforce permutation-invariance by efficiently averaging states obtained via our composition algorithm across all possible context orderings. We evaluate our resulting method on WikiText and MSMARCO in both zero-shot and fine-tuned settings, and show that we can match the strongest performing baseline while enjoying on average $5.4\times$ speedup.

Transformers are deep architectures that define ``in-context mappings'' which enable predicting new tokens based on a given set of tokens (such as a prompt in NLP applications or a set of patches for a vision transformer). In this work, we study in particular the ability of these architectures to handle an arbitrarily large number of context tokens. To mathematically, uniformly address their expressivity, we consider the case that the mappings are conditioned on a context represented by a probability distribution of tokens which becomes discrete for a finite number of these. The relevant notion of smoothness then corresponds to continuity in terms of the Wasserstein distance between these contexts. We demonstrate that deep transformers are universal and can approximate continuous in-context mappings to arbitrary precision, uniformly over compact token domains. This result implies, as a special case, that transformers are universal approximators for continuous permutation-invariant mappings over a fixed number of tokens. It also establishes the universal approximation capability of transformers for certain in-context learning tasks, demonstrating in particular their ability to perform regression within context. A key aspect of our results, compared to existing findings, is that for a fixed precision, a single transformer can operate on an arbitrary (even infinite) number of tokens. Additionally, it operates with a fixed embedding dimension of tokens (this dimension does not increase with precision) and a fixed number of heads (proportional to the dimension). The use of MLPs between multi-head attention layers is also explicitly controlled. We consider both unmasked attentions (as used for the vision transformer) and masked causal attentions (as used for NLP and time series applications). We tackle the causal setting leveraging a space-time lifting to analyze causal attention as a mapping over probability distributions of tokens.

The ever-increasing size of large language models (LLMs) presents significant challenges for deployment due to their heavy computational and memory requirements. Current model pruning techniques attempt to alleviate these issues by relying heavily on external calibration datasets to determine which parameters to prune or compress, thus limiting their flexibility and scalability across different compression ratios. Moreover, these methods often cause severe performance degradation, particularly in downstream tasks, when subjected to higher compression rates. In this paper, we propose *PruneNet*, a novel model compression method that addresses these limitations by reformulating model pruning as a policy learning process. PruneNet decouples the pruning process from the model architecture, eliminating the need for calibration datasets. It learns a stochastic pruning policy to assess parameter importance solely based on intrinsic model properties while preserving the spectral structure to minimize information loss. PruneNet can compress the LLaMA-2-7B model in just 15 minutes, achieving over 80\% retention of its zero-shot performance with a 30\% compression ratio, outperforming existing methods that retain only 75\% performance. Furthermore, on complex multitask language understanding tasks, PruneNet demonstrates its robustness by preserving up to 80\% performance of the original model, proving itself a superior alternative to conventional structured compression techniques.

Long context compression is a critical research problem due to its significance in reducing the high computational and memory costs associated with LLMs. In this paper, we propose Activation Beacon, a plug-in module for transformer-based LLMs that targets effective, efficient, and flexible compression of long contexts. To achieve this, our method introduces the following technical designs. 1) We directly compress the activations (i.e. keys and values at every layer), rather than leveraging soft prompts to relay information (which constitute a major bottleneck to encapsulate the complex information within long contexts). 2) We tailor the compression workflow, where each fine-grained input unit is progressively compressed, enabling high-quality compression and efficient computation during both training and inference. 3) We train the model through compression-based auto-regression, making full use of plain texts and instructional data to optimize the model's compression performance. 4) During training, we randomly sample a compression ratio at each step, teaching the model to support a wide range of compression configurations. Extensive evaluations are conducted on various long-context tasks whose lengths (e.g., 128K) may far exceed the maximum training length (20K), such as document understanding, few-shot learning, and Needle-in-a-Haystack. Whilst existing methods struggle to handle these challenging tasks, Activation Beacon maintains a comparable performance to the uncompressed baseline across various scenarios, achieving a 2x acceleration in inference time and an 8x reduction of memory costs for KV cache.