ICML 2025 论文列表

Large Language Models (LLMs) are increasingly being deployed on edge devices for long-context settings, creating a growing need for fast and efficient long-context inference. In these scenarios, the Key-Value (KV) cache is the primary bottleneck in terms of both GPU memory and latency, as the full KV cache must be loaded for each decoding step. While speculative decoding is a widely accepted technique to accelerate autoregressive decoding, existing methods often struggle to achieve significant speedups due to inefficient KV cache optimization strategies and result in low acceptance rates. To address these challenges, we propose a novel self-speculative decoding framework, QuantSpec, where the draft model shares the architecture of the target model but employs a hierarchical 4-bit quantized KV cache and 4-bit quantized weights for acceleration. QuantSpec maintains high acceptance rates ($>$90\%) and reliably provides consistent end-to-end speedups upto $\sim2.5\times$, outperforming other self-speculative decoding methods that use sparse KV cache for long-context LLM inference. QuantSpec also reduces the memory requirements by $\sim 1.3\times$ compared to these alternatives.

The emergence of long-context large language models (LLMs) offers a promising alternative to traditional retrieval-augmented generation (RAG) for processing extensive documents. However, the computational overhead of long-context inference presents significant efficiency challenges. While Speculative Decoding (SD) traditionally accelerates inference using smaller draft models, its effectiveness diminishes substantially in long-context scenarios due to memory-bound KV cache operations. We introduce Retrieval-Augmented Speculative Decoding (RAPID), which leverages RAG for both accelerating and enhancing generation quality in long-context inference. RAPID introduces the RAG drafter—a draft LLM operating on shortened retrieval contexts—to speculate on the generation of long-context target LLMs. Our approach enables a new paradigm where same-scale or even larger LLMs can serve as RAG drafters while maintaining computational efficiency. To fully leverage the potentially superior capabilities from stronger RAG drafters, we develop an inference-time knowledge transfer that enriches the target distribution by RAG. Extensive experiments on the LLaMA-3.1 and Qwen2.5 backbones demonstrate that RAPID effectively integrates the strengths of both RAG and long-context LLMs, achieving significant performance improvements (e.g., from 39.33 to 42.83 on InfiniteBench for LLaMA-3.1-8B) with more than 2$\times$ speedups for long-context inference. Our analyses also reveal the robustness of RAPID across various context lengths and retrieval quality.

While state-of-the-art LLMs have demonstrated great promise of using long Chains-of-Thought (CoT) to boost reasoning, scaling it up to more challenging problems is fundamentally limited by suboptimal memory usage — intermediate computations accumulate indefinitely in context even no longer needed for future thoughts. We introduce PENCIL, which incorporates a novel reduction mechanism into the autoregressive generation process that recursively clean up intermediate thoughts based on patterns learned from training. By alternately generating and erasing, PENCIL can think deeper to solve harder problems using shorter context and less computes. Empirically, for example, we demonstrate PENCIL with a small 25M-parameter transformer and 2048 context length solves Einstein's puzzle — a task that challenges much larger models like GPT-4. Theoretically, we prove PENCIL can perform universal efficient computation by simulating any Turing machines with optimal time and space complexity, and thus can solve arbitrary computable tasks that are otherwise intractable for vanilla CoT.

Traditional approaches to adapting multi-modal large language models (MLLMs) to new tasks have relied heavily on fine-tuning. This paper introduces Efficient Multi-Modal Long Context Learning (EMLoC), a novel training-free alternative that embeds demonstration examples directly into the model input. EMLoC offers a more efficient, flexible, and scalable solution for task adaptation. Because extremely lengthy inputs introduce prohibitive computational and memory overhead, EMLoC contributes a chunk-wise compression mechanism combined with layer-wise adaptive pruning. It condenses long-context multimodal inputs into compact, task-specific memory representations. By adaptively pruning tokens at each layer under a Jensen-Shannon divergence constraint, our method achieves a dramatic reduction in inference complexity without sacrificing performance. This approach is the first to seamlessly integrate compression and pruning techniques for multi-modal long-context learning, offering a scalable and efficient solution for real-world applications. Extensive experiments on diverse vision-language benchmarks demonstrate that EMLoC achieves performance on par with or superior to naive long-context approaches. Our results highlight the potential of EMLoC as a groundbreaking framework for efficient and flexible adaptation of multi-modal models in resource-constrained environments. Codes are publicly available at https://github.com/Zehong-Ma/EMLoC.

Despite the success of Transformers, handling longer contexts remains challenging due to the limited length generalization and quadratic complexity of self-attention, which often requires post-training with a larger attention window, significantly increasing computational and memory costs. In this paper, we propose a novel attention mechanism based on dynamic context, Grouped Cross Attention (GCA), which can generalize to 1000 $\times$ the pre-training context length while maintaining the ability to access distant information with a constant attention window size. For a given input sequence, we split it into chunks and use each chunk to retrieve top-$k$ relevant past chunks for subsequent text generation. Specifically, unlike most previous works that use an off-the-shelf retriever, our key innovation allows the retriever to learn how to retrieve past chunks that better minimize the auto-regressive loss of subsequent tokens in an end-to-end manner, which adapts better to causal language models. Such a mechanism accommodates retrieved chunks with a fixed-size attention window to achieve long-range information access, significantly reducing computational and memory costs during training and inference. Experiments show that GCA-based models achieve near-perfect accuracy in passkey retrieval for 16M context lengths, which is $1000 \times$ the training length.

Large Language Models (LLMs) have demonstrated remarkable in-context learning (ICL) capabilities. In this study, we explore a surprising phenomenon related to ICL: LLMs can perform multiple, computationally distinct ICL tasks simultaneously, during a single inference call, a capability we term task superposition". We provide empirical evidence of this phenomenon across various LLM families and scales and show that this phenomenon emerges even if we train the model to in-context learn one task at a time. We offer theoretical explanations that this capability is well within the expressive power of transformers. We also explore how LLMs internally compose task vectors during superposition. Furthermore, we show that larger models can solve more ICL tasks in parallel, and better calibrate their output distribution. Our findings offer insights into the latent capabilities of LLMs, further substantiate the perspective of "LLMs as superposition of simulators", and raise questions about the mechanisms enabling simultaneous task execution.

The Conditional Sequence Modeling (CSM) paradigm, benefiting from the transformer's powerful distribution modeling capabilities, has demonstrated considerable promise in offline Reinforcement Learning (RL) tasks. Depending on the task's nature, it is crucial to carefully balance the interplay between inherent local features and long-term dependencies in Markov decision trajectories to mitigate potential performance degradation and unnecessary computational overhead. In this paper, we propose Decision Mixer (DM), which addresses the conflict between features of different scales in the modeling process from the perspective of dynamic integration. Drawing inspiration from conditional computation, we design a plug-and-play dynamic token selection mechanism to ensure the model can effectively allocate attention to different features based on task characteristics. Additionally, we employ an auxiliary predictor to alleviate the short-sightedness issue in the autoregressive sampling process. DM achieves state-of-the-art performance on various standard RL benchmarks while requiring significantly fewer computational resources, offering a viable solution for building efficient and scalable RL foundation models. Code is available at here.

Long-range graph tasks --- those dependent on interactions between `distant' nodes --- are an open problem in graph neural network research. Real-world benchmark tasks, especially the Long Range Graph Benchmark, have become popular for validating the long-range capability of proposed architectures. However, this is an empirical approach that lacks both robustness and theoretical underpinning; a more principled characterization of the long-range problem is required. To bridge this gap, we formalize long-range interactions in graph tasks, introduce a **range measure** for operators on graphs, and validate it with synthetic experiments. We then leverage our measure to examine commonly used tasks and architectures, and discuss to what extent they are, in fact, long-range. We believe our work advances efforts to define and address the long-range problem on graphs, and that our range measure will aid evaluation of new datasets and architectures.

Recently, recurrent models such as state space models and linear attention have become popular due to their linear complexity in the sequence length. Thanks to their recurrent nature, in principle they can process arbitrarily long sequences, but their performance sometimes drops considerably beyond their training context lengths---i.e. they fail to length generalize. In this work, we provide comprehensive empirical and theoretical analysis to support the \textit{unexplored states hypothesis}, which posits that models fail to length generalize when during training they are only exposed to a limited subset of the distribution of all \textit{attainable} states (i.e. states that would be attained if the recurrence was applied to long sequences). Furthermore, we investigate simple training interventions that aim to increase the coverage of the states that the model is trained on, e.g. by initializing the state with Gaussian noise or with the final state of a different input sequence. With only 500 post-training steps ($\sim 0.1\%$ of the pre-training budget), these interventions enable length generalization for sequences that are orders of magnitude longer than the training context (e.g. $2k\longrightarrow 128k$) and show improved performance in long context tasks, thus presenting a simple and efficient way to enable robust length generalization in general recurrent models.

In-context learning has become a standard approach for utilizing language models. However, selecting and processing suitable demonstration examples can be challenging and time-consuming, especially when dealing with large numbers of them. We propose Iterative Vectors (IVs), a technique that explores activation space to enhance in-context performance by simulating gradient updates during inference. IVs extract and iteratively refine activation-based meta-gradients, applying them during inference without requiring backpropagation at any stage. We evaluate IVs across various tasks using four popular models and observe significant improvements. Our findings suggest that in-context activation steering is a promising direction, opening new avenues for future research.

The long-standing dominance of gradient-boosted decision trees on tabular data is currently challenged by tabular foundation models using In-Context Learning (ICL): setting the training data as context for the test data and predicting in a single forward pass without parameter updates. While TabPFNv2 foundation model excels on tables with up to 10K samples, its alternating column- and row-wise attentions make handling large training sets computationally prohibitive. So, can ICL be effectively scaled and deliver a benefit for larger tables? We introduce TabICL, a tabular foundation model for classification, pretrained on synthetic datasets with up to 60K samples and capable of handling 500K samples on affordable resources. This is enabled by a novel two-stage architecture: a column-then-row attention mechanism to build fixed-dimensional embeddings of rows, followed by a transformer for efficient ICL. Across 200 classification datasets from the TALENT benchmark, TabICL is on par with TabPFNv2 while being systematically faster (up to 10 times), and significantly outperforms all other approaches. On 53 datasets with over 10K samples, TabICL surpasses both TabPFNv2 and CatBoost, demonstrating the potential of ICL for large data. Pretraining code, inference code, and pre-trained models are available at https://github.com/soda-inria/tabicl.

Recent large language models (LLMs) support long contexts ranging from 128K to 1M tokens. A popular method for evaluating these capabilities is the needle-in-a-haystack (NIAH) test, which involves retrieving a "needle" (relevant information) from a "haystack" (long irrelevant context). Extensions of this approach include increasing distractors, fact chaining, and in-context reasoning. However, in these benchmarks, models can exploit existing literal matches between the needle and haystack to simplify the task. To address this, we introduce NoLiMa, a benchmark extending NIAH with a carefully designed needle set, where questions and needles have minimal lexical overlap, requiring models to infer latent associations to locate the needle within the haystack. We evaluate 13 popular LLMs that claim to support contexts of at least 128K tokens. While they perform well in short contexts (<1K), performance degrades significantly as context length increases. At 32K, for instance, 11 models drop below 50% of their strong short-length baselines. Even GPT-4o, one of the top-performing exceptions, experiences a reduction from an almost-perfect baseline of 99.3% to 69.7%. Our analysis suggests these declines stem from the increased difficulty the attention mechanism faces in longer contexts when literal matches are absent, making it harder to retrieve relevant information. Even models enhanced with reasoning capabilities or CoT prompting struggle to maintain performance in long contexts. We publicly release the dataset and evaluation code at https://github.com/adobe-research/NoLiMa.