ICLR 2025 论文列表

To deliver coherent and personalized experiences in long-term conversations, existing approaches typically perform retrieval augmented response generation by constructing memory banks from conversation history at either the turn-level, session-level, or through summarization techniques. In this paper, we explore the impact of different memory granularities and present two key findings: (1) Both turn-level and session-level memory units are suboptimal, affecting not only the quality of final responses, but also the accuracy of the retrieval process. (2) The redundancy in natural language introduces noise, hindering precise retrieval. We demonstrate that *LLMLingua-2*, originally designed for prompt compression to accelerate LLM inference, can serve as an effective denoising method to enhance memory retrieval accuracy. Building on these insights, we propose **SeCom**, a method that constructs a memory bank with topical segments by introducing a conversation **Se**gmentation model, while performing memory retrieval based on **Com**pressed memory units. Experimental results show that **SeCom** outperforms turn-level, session-level, and several summarization-based methods on long-term conversation benchmarks such as *LOCOMO* and *Long-MT-Bench+*. Additionally, the proposed conversation segmentation method demonstrates superior performance on dialogue segmentation datasets such as *DialSeg711*, *TIAGE*, and *SuperDialSeg*.

Differentially private diffusion models (DPDMs) harness the remarkable generative capabilities of diffusion models while enforcing differential privacy (DP) for sensitive data. However, existing DPDM training approaches often suffer from significant utility loss, large memory footprint, and expensive inference cost, impeding their practical uses. To overcome such limitations, we present RAPID: Retrieval Augmented PrIvate Diffusion model, a novel approach that integrates retrieval augmented generation (RAG) into DPDM training. Specifically, RAPID leverages available public data to build a knowledge base of sample trajectories; when training the diffusion model on private data, RAPID computes the early sampling steps as queries, retrieves similar trajectories from the knowledge base as surrogates, and focuses on training the later sampling steps in a differentially private manner. Extensive evaluation using benchmark datasets and models demonstrates that, with the same privacy guarantee, RAPID significantly outperforms state-of-the-art approaches by large margins in generative quality, memory footprint, and inference cost, suggesting that retrieval-augmented DP training represents a promising direction for developing future privacy-preserving generative models. The code is available at: https://github.com/TanqiuJiang/RAPID

The memory and computational demands of Key-Value (KV) cache present significant challenges for deploying long-context language models. Previous approaches attempt to mitigate this issue by selectively dropping tokens, which irreversibly erases critical information that might be needed for future queries. In this paper, we propose a novel compression technique for KV cache that preserves all token information. Our investigation reveals that: i) Most attention heads primarily focus on the local context; ii) Only a few heads, denoted as retrieval heads, can essentially pay attention to all input tokens. These key observations motivate us to use separate caching strategy for attention heads.Therefore, we propose RazorAttention, a training-free KV cache compression algorithm, which maintains a full cache for these crucial retrieval heads and discards the remote tokens in non-retrieval heads. Furthermore, we introduce a novel mechanism involving a “compensation token” to further recover the information in the dropped tokens. Extensive evaluations across a diverse set of large language models (LLMs) demonstrate that RazorAttention achieves a reduction in KV cache size by over 70% without noticeable impacts on performance. Additionally, RazorAttention is compatible with FlashAttention, rendering it an efficient and plug-and-play solution that enhances LLM inference efficiency without overhead or retraining of the original model.

This paper proposes Decomposed Retrieval of Constraints (DRoC), a novel framework aimed at enhancing large language models (LLMs) in exploiting solvers to tackle vehicle routing problems (VRPs) with intricate constraints. While LLMs have shown promise in solving simple VRPs, their potential in addressing complex VRP variants is still suppressed, due to the limited embedded internal knowledge that is required to accurately reflect diverse VRP constraints. Our approach mitigates the issue by integrating external knowledge via a novel retrieval-augmented generation (RAG) approach. More specifically, the DRoC decomposes VRP constraints, externally retrieves information relevant to each constraint, and synergistically combines internal and external knowledge to benefit the program generation for solving VRPs. The DRoC also allows LLMs to dynamically select between RAG and self-debugging mechanisms, thereby optimizing program generation without the need for additional training. Experiments across 48 VRP variants exhibit the superiority of DRoC, with significant improvements in the accuracy rate and runtime error rate delivered by the generated programs. The DRoC framework has the potential to elevate LLM performance in complex optimization tasks, fostering the applicability of LLMs in industries such as transportation and logistics.

Prior methods propose to offset the escalating costs of modern foundation models by dropping specific parts of their contexts with hand-designed rules, while attempting to preserve their original performance. We overcome this trade-off with Neural Attention Memory Models (NAMMs), introducing a learned network for memory management that improves both the performance and efficiency of transformers. We evolve NAMMs atop pre-trained transformers to provide different latent contexts focusing on the most relevant information for individual layers and attention heads. NAMMs are universally applicable to any model using self-attention as they condition exclusively on the values in the produced attention matrices. Learning NAMMs on a small set of problems, we achieve substantial performance improvements across multiple long-context benchmarks while cutting the model's input contexts up to a fraction of the original sizes. We show the generality of our conditioning enables zero-shot transfer of NAMMs trained only on language to entirely new transformer architectures even across input modalities, with their benefits carrying over to vision and reinforcement learning.

High-quality sentence embeddings are fundamental in many natural language processing (NLP) tasks, such as semantic textual similarity (STS) and retrieval-augmented generation (RAG). However, most existing methods leverage fixed-length sentence embeddings from full-layer language models, which lack the scalability to accommodate the diverse available resources across various applications. Viewing this gap, we propose a novel sentence embedding model Espresso Sentence Embeddings (ESE) with two learning processes. First, the learn-to-express process encodes more salient representations to shallow layers. Second, the learn-to-compress process compacts essential features into the initial dimensions using Principal Component Analysis (PCA). This way, ESE can scale model depth via the former process and embedding size via the latter. Extensive experiments on STS and RAG suggest that ESE can effectively produce high-quality sentence embeddings with less model depth and embedding size, enhancing inference efficiency. The code is available at https://github.com/SeanLee97/AnglE/blob/main/README_ESE.md.

Recent large language model (LLM)-driven chat assistant systems have integrated memory components to track user-assistant chat histories, enabling more accurate and personalized responses. However, their long-term memory capabilities in sustained interactions remain underexplored. We introduce LongMemEval, a comprehensive benchmark designed to evaluate five core long-term memory abilities of chat assistants: information extraction, multi-session reasoning, temporal reasoning, knowledge updates, and abstention. With 500 meticulously curated questions embedded within freely scalable user-assistant chat histories, LongMemEval presents a significant challenge to existing long-term memory systems, with commercial chat assistants and long-context LLMs showing a 30% accuracy drop on memorizing information across sustained interactions. We then present a unified framework that breaks down the long-term memory design into three stages: indexing, retrieval, and reading. Built upon key experimental insights, we propose several memory design optimizations including session decomposition for value granularity, fact-augmented key expansion for indexing, and time-aware query expansion for refining the search scope. Extensive experiments show that these optimizations greatly improve both memory recall and downstream question answering on LongMemEval. Overall, our study provides valuable resources and guidance for advancing the long-term memory capabilities of LLM-based chat assistants, paving the way toward more personalized and reliable conversational AI. Our benchmark and code are publicly available at https://github.com/xiaowu0162/LongMemEval.

Large language models (LLMs) have revolutionized numerous applications, yet their deployment remains challenged by memory constraints on local devices. While scaling laws have enhanced LLM capabilities, the primary bottleneck has shifted from $\textit{capability}$ to $\textit{availability}$, emphasizing the need for efficient memory management. Traditional compression methods, such as quantization, often require predefined compression ratios and separate compression processes for each setting, complicating deployment in variable memory environments. In this paper, we introduce $\textbf{BitStack}$, a novel, training-free weight compression approach that enables megabyte-level trade-offs between memory usage and model performance. By leveraging weight decomposition, BitStack can dynamically adjust the model size with minimal transmission between running memory and storage devices. Our approach iteratively decomposes weight matrices while considering the significance of each parameter, resulting in an approximately 1-bit per parameter residual block in each decomposition iteration. These blocks are sorted and stacked in storage as basic transmission units, with different quantities loaded based on current memory availability. Extensive experiments across a wide range of tasks demonstrate that, despite offering fine-grained size control, BitStack consistently matches or surpasses strong quantization baselines, particularly at extreme compression ratios. To the best of our knowledge, this is the first decomposition-based method that effectively bridges the gap to practical compression techniques like quantization. Code is available at https://github.com/xinghaow99/BitStack.

Effective decision-making in partially observable environments demands robust memory management. Despite their success in supervised learning, current deep-learning memory models struggle in reinforcement learning environments that are partially observable and long-term. They fail to efficiently capture relevant past information, adapt flexibly to changing observations, and maintain stable updates over long episodes. We theoretically analyze the limitations of existing memory models within a unified framework and introduce the Stable Hadamard Memory, a novel memory model for reinforcement learning agents. Our model dynamically adjusts memory by erasing no longer needed experiences and reinforcing crucial ones computationally efficiently. To this end, we leverage the Hadamard product for calibrating and updating memory, specifically designed to enhance memory capacity while mitigating numerical and learning challenges. Our approach significantly outperforms state-of-the-art memory-based methods on challenging partially observable benchmarks, such as meta-reinforcement learning, long-horizon credit assignment, and POPGym, demonstrating superior performance in handling long-term and evolving contexts.

Vision-language models (VLMs) are known to be susceptible to distribution shifts between pre-training data and test data, and test-time adaptation (TTA) methods for VLMs have been proposed to mitigate the detrimental impact of the distribution shifts. However, the existing methods solely rely on the internal knowledge encoded within the model parameters, which are constrained to pre-training data. To complement the limitation of the internal knowledge, we propose **Retrieval-Augmented-TTA (RA-TTA)** for adapting VLMs to test distribution using **external** knowledge obtained from a web-scale image database. By fully exploiting the bi-modality of VLMs, RA-TTA **adaptively** retrieves proper external images for each test image to refine VLMs' predictions using the retrieved external images, where fine-grained **text descriptions** are leveraged to extend the granularity of external knowledge. Extensive experiments on 17 datasets demonstrate that the proposed RA-TTA outperforms the state-of-the-art methods by 3.01-9.63\% on average.

Retrieval-Augmented Generation (RAG) has proven its effectiveness in mitigating hallucinations in Large Language Models (LLMs) by retrieving knowledge from external resources. To adapt LLMs for the RAG systems, current approaches use instruction tuning to optimize LLMs, improving their ability to utilize retrieved knowledge. This supervised fine-tuning (SFT) approach focuses on equipping LLMs to handle diverse RAG tasks using different instructions. However, it trains RAG modules to overfit training signals and overlooks the varying data preferences among agents within the RAG system. In this paper, we propose a Differentiable Data Rewards (DDR) method, which end-to-end trains RAG systems by aligning data preferences between different RAG modules. DDR works by collecting the rewards to optimize each agent in the RAG system with the rollout method, which prompts agents to sample some potential responses as perturbations, evaluates the impact of these perturbations on the whole RAG system, and subsequently optimizes the agent to produce outputs that improve the performance of the RAG system. Our experiments on various knowledge-intensive tasks demonstrate that DDR significantly outperforms the SFT method, particularly for LLMs with smaller-scale parameters that depend more on the retrieved knowledge. Additionally, DDR exhibits a stronger capability to align the data preference between RAG modules. The DDR method makes the generation module more effective in extracting key information from documents and mitigating conflicts between parametric memory and external knowledge. All codes are available at https://github.com/OpenMatch/RAG-DDR.

Retrieval-augmented generation (RAG) has shown promising potential to enhance the accuracy and factuality of language models (LMs). However, imperfect retrievers or noisy corpora can introduce misleading or even erroneous information to the retrieved contents, posing a significant challenge to the generation quality. Existing RAG methods typically address this challenge by directly predicting final answers despite potentially noisy inputs, resulting in an implicit denoising process that is difficult to interpret and verify. On the other hand, the acquisition of explicit denoising supervision is often costly, involving significant human efforts. In this work, we propose InstructRAG, where LMs explicitly learn the denoising process through self-synthesized rationales --- First, we instruct the LM to explain how the ground-truth answer is derived from retrieved documents. Then, these rationales can be used either as demonstrations for in-context learning of explicit denoising or as supervised fine-tuning data to train the model. Compared to standard RAG approaches, InstructRAG requires no additional supervision, allows for easier verification of the predicted answers, and effectively improves generation accuracy. Experiments show InstructRAG consistently outperforms existing RAG methods in both training-free and trainable scenarios, achieving a relative improvement of 8.3% over the best baseline method on average across five knowledge-intensive benchmarks. Extensive analysis indicates that InstructRAG scales well with increased numbers of retrieved documents and consistently exhibits robust denoising ability even in out-of-domain datasets, demonstrating strong generalizability.

Symbolic regression is a key task in machine learning, aiming to discover mathematical expressions that best describe a dataset. While deep learning has increased interest in using neural networks for symbolic regression, many existing approaches rely on pre-trained models. These models require significant computational resources and struggle with regression tasks involving unseen functions and variables. A pre-training-free paradigm is needed to better integrate with search-based symbolic regression algorithms. To address these limitations, we propose a novel framework for symbolic regression that integrates evolutionary feature construction with a neural network, without the need for pre-training. Our approach adaptively generates symbolic trees that align with the desired semantics in real-time using a language model trained via online supervised learning, providing effective building blocks for feature construction. To mitigate hallucinations from the language model, we design a retrieval-augmented generation mechanism that explicitly leverages searched symbolic expressions. Additionally, we introduce a scale-invariant data augmentation technique that further improves the robustness and generalization of the model. Experimental results demonstrate that our framework achieves state-of-the-art accuracy across 25 regression algorithms and 120 regression tasks.

Augmenting LLMs with context leads to improved performance across many applications. Despite much research on Retrieval Augmented Generation (RAG) systems, an open question is whether errors arise because LLMs fail to utilize the context from retrieval or the context itself is insufficient to answer the query. To shed light on this, we develop a new notion of sufficient context, along with a method to classify instances that have enough information to answer the query. We then use sufficient context to analyze several models and datasets. By stratifying errors based on context sufficiency, we find that larger models with higher baseline performance (Gemini 1.5 Pro, GPT 4o, Claude 3.5) excel at answering queries when the context is sufficient, but often output incorrect answers instead of abstaining when the context is not. On the other hand, smaller models with lower baseline performance (Llama 3.1, Mistral 3, Gemma 2) hallucinate or abstain often, even with sufficient context. We further categorize cases when the context is useful, and improves accuracy, even though it does not fully answer the query and the model errs without the context. Building on our findings, we explore ways to reduce hallucinations in RAG systems, including a new selective generation method that leverages sufficient context information for guided abstention. Our method improves the fraction of correct answers among times where the model responds by 2--10% for Gemini, GPT, and Gemma. Code for our selective generation method and the prompts used in our autorater analysis are available on our [github](https://github.com/hljoren/sufficientcontext).

Large language models (LLMs) augmented with retrieval exhibit robust performance and extensive versatility by incorporating external contexts. However, the input length grows linearly in the number of retrieved documents, causing a dramatic increase in latency. In this paper, we propose a novel paradigm named Sparse RAG, which seeks to cut computation costs through sparsity. Specifically, Sparse RAG encodes retrieved documents in parallel, which eliminates latency introduced by long-range attention of retrieved documents. Then, LLMs selectively decode the output by only attending to highly relevant caches auto-regressively, which are chosen via prompting LLMs with special control tokens. It is notable that Sparse RAG combines the assessment of each individual document and the generation of the response into a single process. The designed sparse mechanism in a RAG system can facilitate the reduction of the number of documents loaded during decoding for accelerating the inference of the RAG system. Additionally, filtering out undesirable contexts enhances the model’s focus on relevant context, inherently improving its generation quality. Evaluation results on four datasets show that Sparse RAG can be used to strike an optimal balance between generation quality and computational efficiency, demonstrating its generalizability across tasks.

Significant advances have been made in developing general-purpose embodied AI in environments like Minecraft through the adoption of LLM-augmented hierarchical approaches. While these approaches, which combine high-level planners with low-level controllers, show promise, low-level controllers frequently become performance bottlenecks due to repeated failures. In this paper, we argue that the primary cause of failure in many low-level controllers is the absence of an episodic memory system. To address this, we introduce MrSteve (Memory Recall Steve), a novel low-level controller equipped with Place Event Memory (PEM), a form of episodic memory that captures what, where, and when information from episodes. This directly addresses the main limitation of the popular low-level controller, Steve-1. Unlike previous models that rely on short-term memory, PEM organizes spatial and event-based data, enabling efficient recall and navigation in long-horizon tasks. Additionally, we propose an Exploration Strategy and a Memory-Augmented Task Solving Framework, allowing agents to alternate between exploration and task-solving based on recalled events. Our approach significantly improves task-solving and exploration efficiency compared to existing methods. We will release our code and demos on the project page: https://sites.google.com/view/mr-steve.

Large language models (LLMs) have shown remarkable capabilities, but still struggle with processing extensive contexts, limiting their ability to maintain coherence and accuracy over long sequences. In contrast, the human brain excels at organising and retrieving episodic experiences across vast temporal scales, spanning a lifetime. In this work, we introduce EM-LLM, a novel approach that integrates key aspects of human episodic memory and event cognition into LLMs with no fine-tuning, enabling them to handle practically infinite context lengths while maintaining computational efficiency. EM-LLM organises sequences of tokens into coherent episodic events using a combination of Bayesian surprise and graph-theoretic boundary refinement in an online fashion. When needed, these events are retrieved through a two-stage memory process, combining similarity-based and temporally contiguous retrieval for efficient, human-inspired access to relevant information. Experiments on the LongBench and $\infty$-Bench benchmarks demonstrate EM-LLM's superior performance, consistently outperforming the state-of-the-art retrieval model InfLLM across various baseline LLMs. In addition, EM-LLM outperforms its popular counterpart, RAG, in a wide range of tasks, while requiring similar resources. Notably, EM-LLM's performance even surpasses full-context models in most tasks, while successfully performing retrieval across 10 million tokens -- a scale computationally infeasible for such models. Finally, our analysis reveals strong correlations between EM-LLM's event segmentation and human-perceived events, suggesting parallels between this artificial system and its biological counterpart, thereby offering a novel computational framework for exploring human memory mechanisms.

Robot learning is witnessing a significant increase in the size, diversity, and complexity of pre-collected datasets, mirroring trends in domains such as natural language processing and computer vision. Many robot learning methods treat such datasets as multi-task expert data and learn a multi-task, generalist policy by training broadly across them. Notably, while these generalist policies can improve the average performance across many tasks, the performance of generalist policies on any one task is often suboptimal due to negative transfer between partitions of the data, compared to task-specific specialist policies. In this work, we argue for the paradigm of training policies during deployment given the scenarios they encounter: rather than deploying pre-trained policies to unseen problems in a zero-shot manner, we non-parametrically retrieve and train models directly on relevant data at test time. Furthermore, we show that many robotics tasks share considerable amounts of low-level behaviors and that retrieval at the "sub"-trajectory granularity enables significantly improved data utilization, generalization, and robustness in adapting policies to novel problems. In contrast, existing full-trajectory retrieval methods tend to underutilize the data and miss out on shared cross-task content. This work proposes STRAP, a technique for leveraging pre-trained vision foundation models and dynamic time warping to retrieve sub-sequences of trajectories from large training corpora in a robust fashion. STRAP outperforms both prior retrieval algorithms and multi-task learning methods in simulated and real experiments, showing the ability to scale to much larger offline datasets in the real world as well as the ability to learn robust control policies with just a handful of real-world demonstrations.