Long-context language models (LCLMs) have exhibited impressive capabilities in long-context understanding tasks. Among these, long-context referencing -- a crucial task that requires LCLMs to attribute items of interest to specific parts of long-context data -- remains underexplored. To bridge this gap, this paper proposes Referencing Evaluation for Long-context Language Models (Ref-Long), a novel benchmark designed to assess the long-context referencing capability of LCLMs. Specifically, Ref-Long requires LCLMs to identify the indexes of documents that reference a specific key, emphasizing contextual relationships between the key and the documents over simple retrieval. Based on the task design, we construct three subsets ranging from synthetic to realistic scenarios to form the Ref-Long benchmark. Experimental results of 13 LCLMs reveal significant shortcomings in long-context referencing, even among advanced models like GPT-4o. To further investigate these challenges, we conduct comprehensive analyses, including human evaluations, task format adjustments, fine-tuning experiments, and error analyses, leading to several key insights. Our data and code can be found in https://github. com/wujunjie1998/Ref-Long.
Language models handle increasingly long contexts for tasks such as book summarization, but this leads to growing memory costs for the key-value (KV) cache. Many prior works have proposed ways of discarding KVs from memory, but their approaches are tailored to favorable settings, obscuring caveats like high peak memory and performance degradation, and a fair comparison between methods is difficult. In this paper, we propose the *KV footprint* as a unified metric, which accounts for both the amount of KV entries stored and their lifespan in memory. We evaluate methods based on the smallest footprint they attain while preserving performance in both long-context understanding and generation, with context lengths of up to 128K tokens. This metric reveals the high peak memory of prior KV eviction methods. One class of methods -- *post-fill eviction* -- has a high footprint due to being incompatible with eviction during pre-filling. We adapt these methods to be able to evict KVs during pre-filling, achieving substantially lower KV footprints. We then turn to *recency eviction* methods, wherein we propose PruLong, an end-to-end optimization method for learning which attention heads need to retain the full KV cache and which do not. PruLong saves memory while preserving long-context performance, achieving 12% smaller KV footprint than prior methods while retaining performance in challenging recall tasks. Our paper clarifies the complex tangle of long-context inference methods and paves the way for future development to minimize the KV footprint.
Long-context understanding is crucial for many NLP applications, yet transformers struggle with efficiency due to the quadratic complexity of self-attention. Sparse attention methods alleviate this cost but often impose static, predefined masks, failing to capture heterogeneous attention patterns. This results in suboptimal token interactions, limiting adaptability and retrieval accuracy in long-sequence tasks. This work introduces a dynamic sparse attention mechanism that assigns adaptive masks at the attention-map level, preserving heterogeneous patterns across layers and heads. Unlike existing approaches, our method eliminates the need for fine-tuning and predefined mask structures while maintaining computational efficiency. By learning context-aware attention structures, it achieves high alignment with full-attention models, ensuring minimal performance degradation while reducing memory and compute overhead. This approach provides a scalable alternative to full attention, enabling the practical deployment of large-scale Large Language Models (LLMs) without sacrificing retrieval performance. DAM is available at: https://github.com/HanzhiZhang-Ulrica/DAM.
Speculative decoding is a widely adopted technique for accelerating inference in large language models (LLMs), but its performance degrades on long inputs due to increased attention cost and reduced draft accuracy. We introduce SpecExtend, a drop-in enhancement that improves the performance of speculative decoding on long sequences without any additional training. SpecExtend integrates efficient attention mechanisms such as FlashAttention and Hybrid Tree Attention into both the draft and target models, reducing latency across all stages. To improve draft accuracy and speed, we propose Cross-model Retrieval, a novel KV cache update strategy that uses the target model's attention scores to dynamically select relevant context for the draft model. Extensive evaluations on three long-context understanding datasets show that SpecExtend accelerates standard tree-based speculative decoding by up to 2.22x for inputs up to 16K tokens, providing an effective solution for speculative decoding of long sequences. The code is available at https://github.com/jycha98/SpecExtend .
Sequence modeling is currently dominated by causal transformer architectures that use softmax self-attention. Although widely adopted, transformers require scaling memory and compute linearly during inference. A recent stream of work linearized the softmax operation, resulting in powerful recurrent neural network (RNN) models with constant memory and compute costs such as DeltaNet, Mamba or xLSTM. These models can be unified by noting that their recurrent layer dynamics can all be derived from an in-context regression objective, approximately optimized through an online learning rule. Here, we join this line of work and introduce a numerically stable, chunkwise parallelizable version of the recently proposed Mesa layer (von Oswald et al., 2024), and study it in language modeling at the billion-parameter scale. This layer again stems from an in-context loss, but which is now minimized to optimality at every time point using a fast conjugate gradient solver. Through an extensive suite of experiments, we show that optimal test-time training enables reaching lower language modeling perplexity and higher downstream benchmark performance than previous RNNs, especially on tasks requiring long context understanding. This performance gain comes at the cost of additional flops spent during inference time. Our results are therefore intriguingly related to recent trends of increasing test-time compute to improve performance -- here by spending compute to solve sequential optimization problems within the neural network itself.
Post-training compression reduces the computational and memory costs of large language models (LLMs), enabling resource-efficient deployment. However, existing compression benchmarks only focus on language modeling (e.g., perplexity) and natural language understanding tasks (e.g., GLUE accuracy), ignoring the agentic capabilities - workflow, tool use/function call, long-context understanding and real-world application. We introduce the Agent Compression Benchmark (ACBench), the first comprehensive benchmark for evaluating how compression impacts LLMs' agentic abilities. ACBench spans (1) 12 tasks across 4 capabilities (e.g., WorfBench for workflow generation, Needle-in-Haystack for long-context retrieval), (2) quantization (GPTQ, AWQ) and pruning (Wanda, SparseGPT), and (3) 15 models, including small (Gemma-2B), standard (Qwen2.5 7B-32B), and distilled reasoning LLMs (DeepSeek-R1-Distill). Our experiments reveal compression tradeoffs: 4-bit quantization preserves workflow generation and tool use (1%-3% drop) but degrades real-world application accuracy by 10%-15%. We introduce ERank, Top-k Ranking Correlation and Energy to systematize analysis. ACBench provides actionable insights for optimizing LLM compression in agentic scenarios. The code can be found in https://github.com/pprp/ACBench.
Long Context Understanding (LCU) is a critical area for exploration in current large language models (LLMs). However, due to the inherently lengthy nature of long-text data, existing LCU benchmarks for LLMs often result in prohibitively high evaluation costs, like testing time and inference expenses. Through extensive experimentation, we discover that existing LCU benchmarks exhibit significant redundancy, which means the inefficiency in evaluation. In this paper, we propose a concise data compression method tailored for long-text data with sparse information characteristics. By pruning the well-known LCU benchmark LongBench, we create MiniLongBench. This benchmark includes only 237 test samples across six major task categories and 21 distinct tasks. Through empirical analysis of over 60 LLMs, MiniLongBench achieves an average evaluation cost reduced to only 4.5% of the original while maintaining an average rank correlation coefficient of 0.97 with LongBench results. Therefore, our MiniLongBench, as a low-cost benchmark, holds great potential to substantially drive future research into the LCU capabilities of LLMs. See https://github.com/MilkThink-Lab/MiniLongBench for our code, data and tutorial.
Although the context length of large language models (LLMs) has increased to millions of tokens, evaluating their effectiveness beyond needle-in-a-haystack approaches has proven difficult. We argue that novels provide a case study of subtle, complicated structure and long-range semantic dependencies often over 128k tokens in length. Inspired by work on computational novel analysis, we release the Too Long, Didn't Model (TLDM) benchmark, which tests a model's ability to report plot summary, storyworld configuration, and elapsed narrative time. We find that none of seven tested frontier LLMs retain stable understanding beyond 64k tokens. Our results suggest language model developers must look beyond "lost in the middle" benchmarks when evaluating model performance in complex long-context scenarios. To aid in further development we release the TLDM benchmark together with reference code and data.
Long-context understanding poses significant challenges in natural language processing, particularly for real-world dialogues characterized by speech-based elements, high redundancy, and uneven information density. Although large language models (LLMs) achieve impressive results on existing benchmarks, these datasets fail to reflect the complexities of such texts, limiting their applicability to practical scenarios. To bridge this gap, we construct the first spoken long-text dataset, derived from live streams, designed to reflect the redundancy-rich and conversational nature of real-world scenarios. We construct tasks in three categories: retrieval-dependent, reasoning-dependent, and hybrid. We then evaluate both popular LLMs and specialized methods to assess their ability to understand long-contexts in these tasks. Our results show that current methods exhibit strong task-specific preferences and perform poorly on highly redundant inputs, with no single method consistently outperforming others. We propose a new baseline that better handles redundancy in spoken text and achieves strong performance across tasks. Our findings highlight key limitations of current methods and suggest future directions for improving long-context understanding. Finally, our benchmark fills a gap in evaluating long-context spoken language understanding and provides a practical foundation for developing real-world e-commerce systems. The code and benchmark are available at https://github.com/Yarayx/livelongbench.
State space models (SSMs) have emerged as an efficient alternative to Transformer models for language modeling, offering linear computational complexity and constant memory usage as context length increases. However, despite their efficiency in handling long contexts, recent studies have shown that SSMs, such as Mamba models, generally underperform compared to Transformers in long-context understanding tasks. To address this significant shortfall and achieve both efficient and accurate long-context understanding, we propose LongMamba, a training-free technique that significantly enhances the long-context capabilities of Mamba models. LongMamba builds on our discovery that the hidden channels in Mamba can be categorized into local and global channels based on their receptive field lengths, with global channels primarily responsible for long-context capability. These global channels can become the key bottleneck as the input context lengthens. Specifically, when input lengths largely exceed the training sequence length, global channels exhibit limitations in adaptively extend their receptive fields, leading to Mamba's poor long-context performance. The key idea of LongMamba is to mitigate the hidden state memory decay in these global channels by preventing the accumulation of unimportant tokens in their memory. This is achieved by first identifying critical tokens in the global channels and then applying token filtering to accumulate only those critical tokens. Through extensive benchmarking across synthetic and real-world long-context scenarios, LongMamba sets a new standard for Mamba's long-context performance, significantly extending its operational range without requiring additional training. Our code is available at https://github.com/GATECH-EIC/LongMamba.
We present Kimi-VL, an efficient open-source Mixture-of-Experts (MoE) vision-language model (VLM) that offers advanced multimodal reasoning, long-context understanding, and strong agent capabilities - all while activating only 2.8B parameters in its language decoder (Kimi-VL-A3B). Kimi-VL demonstrates strong performance across challenging domains: as a general-purpose VLM, Kimi-VL excels in multi-turn agent tasks (e.g., OSWorld), matching flagship models. Furthermore, it exhibits remarkable capabilities across diverse challenging vision language tasks, including college-level image and video comprehension, OCR, mathematical reasoning, and multi-image understanding. In comparative evaluations, it effectively competes with cutting-edge efficient VLMs such as GPT-4o-mini, Qwen2.5-VL-7B, and Gemma-3-12B-IT, while surpassing GPT-4o in several key domains. Kimi-VL also advances in processing long contexts and perceiving clearly. With a 128K extended context window, Kimi-VL can process diverse long inputs, achieving impressive scores of 64.5 on LongVideoBench and 35.1 on MMLongBench-Doc. Its native-resolution vision encoder, MoonViT, further allows it to see and understand ultra-high-resolution visual inputs, achieving 83.2 on InfoVQA and 34.5 on ScreenSpot-Pro, while maintaining lower computational cost for common tasks. Building upon Kimi-VL, we introduce an advanced long-thinking variant: Kimi-VL-Thinking. Developed through long chain-of-thought (CoT) supervised fine-tuning (SFT) and reinforcement learning (RL), this model exhibits strong long-horizon reasoning capabilities. It achieves scores of 61.7 on MMMU, 36.8 on MathVision, and 71.3 on MathVista while maintaining the compact 2.8B activated LLM parameters, setting a new standard for efficient multimodal thinking models. Code and models are publicly accessible at https://github.com/MoonshotAI/Kimi-VL.
Scientific problem-solving involves synthesizing information while applying expert knowledge. We introduce CURIE, a scientific long-Context Understanding,Reasoning and Information Extraction benchmark to measure the potential of Large Language Models (LLMs) in scientific problem-solving and assisting scientists in realistic workflows. This benchmark introduces ten challenging tasks with a total of 580 problems and solution pairs curated by experts in six disciplines - materials science, condensed matter physics, quantum computing, geospatial analysis, biodiversity, and proteins - covering both experimental and theoretical work-flows in science. We evaluate a range of closed and open LLMs on tasks in CURIE which requires domain expertise, comprehension of long in-context information,and multi-step reasoning. While Gemini Flash 2.0 and Claude-3 show consistent high comprehension across domains, the popular GPT-4o and command-R+ fail dramatically on protein sequencing tasks. With the best performance at 32% there is much room for improvement for all models. We hope that insights gained from CURIE can guide the future development of LLMs in sciences. Evaluation code and data are in https://github.com/google/curie
Recently, Probabilistic Language-Image Pre-Training (ProLIP) has been proposed to tackle the multiplicity issue of vision-language (VL) tasks. Despite their success in probabilistic representation learning at a scale, the ProLIP models cannot handle long context texts longer than 64 context length, which limits their ability to capture rich contextual information from longer text sequences. To address this issue, this paper proposes a fine-tuning strategy for ProLIP to accept longer texts, e.g., 256 text tokens. Experimental results on Urban-1k and the DataComp evaluation suite show that the proposed LongProLIP recipe can improve understanding of long contexts while minimizing the negative effect of fine-tuning.We also observe a trade-off between the long context understanding (measured by Urban-1k) and general zero-shot capability (measured by evaluation datasets by DataComp). Code is available at https://github.com/naver-ai/prolip
Many applications of large language models (LLMs) require long-context understanding, but models continue to struggle with such tasks. We hypothesize that conventional next-token prediction training could contribute to this, because each token is assigned equal weight. Yet, intuitively, the amount of context needed to predict the next token accurately varies greatly across different data. To reflect this, we propose various novel token-weighting schemes that assign different weights to each training token in the loss, thereby generalizing existing works. For this, we categorize token-weighting methods using a two-step framework which compares the confidences of a long-context and short-context model to score tokens. We evaluate all methods on multiple long-context understanding tasks and show that non-uniform loss weights are helpful to improve the long-context abilities of LLMs. Different short-context models can be used effectively for token scoring, including models that are much smaller than the long-context model that is trained. All in all, this work contributes to a better understanding of the trade-offs long-context language modeling faces and provides guidelines for model steering via loss-weighting based on empirical evidence. The code can be found on Github.
Answering complex, long-context questions remains a major challenge for large language models (LLMs) as it requires effective question clarifications and context retrieval. We propose Agentic Long-Context Understanding (AgenticLU), a framework designed to enhance an LLM's understanding of such queries by integrating targeted self-clarification with contextual grounding within an agentic workflow. At the core of AgenticLU is Chain-of-Clarifications (CoC), where models refine their understanding through self-generated clarification questions and corresponding contextual groundings. By scaling inference as a tree search where each node represents a CoC step, we achieve 97.8% answer recall on NarrativeQA with a search depth of up to three and a branching factor of eight. To amortize the high cost of this search process to training, we leverage the preference pairs for each step obtained by the CoC workflow and perform two-stage model finetuning: (1) supervised finetuning to learn effective decomposition strategies, and (2) direct preference optimization to enhance reasoning quality. This enables AgenticLU models to generate clarifications and retrieve relevant context effectively and efficiently in a single inference pass. Extensive experiments across seven long-context tasks demonstrate that AgenticLU significantly outperforms state-of-the-art prompting methods and specialized long-context LLMs, achieving robust multi-hop reasoning while sustaining consistent performance as context length grows.
Evaluating large language models' (LLMs) long-context understanding capabilities remains challenging. We present SCALAR (Scientific Citation-based Live Assessment of Long-context Academic Reasoning), a novel benchmark that leverages academic papers and their citation networks. SCALAR features automatic generation of high-quality ground truth labels without human annotation, controllable difficulty levels, and a dynamic updating mechanism that prevents data contamination. Using ICLR 2025 papers, we evaluate 8 state-of-the-art LLMs, revealing key insights about their capabilities and limitations in processing long scientific documents across different context lengths and reasoning types. Our benchmark provides a reliable and sustainable way to track progress in long-context understanding as LLM capabilities evolve.
While large language models (LLMs) have shown strong general reasoning capabilities, their effectiveness in financial reasoning, which is crucial for real-world financial applications remains underexplored. In this study, we conduct a comprehensive evaluation of 24 state-of-the-art general and reasoning-focused LLMs across four complex financial reasoning tasks involving financial text, tabular data, and equations. We assess key capabilities such as numerical reasoning, tabular interpretation, financial terminology comprehension, long-context understanding, and equation-based problem solving. Our analysis reveals that while data quality and pretraining contribute to performance, general techniques like chain-of-thought (CoT) fine-tuning offer limited gains in financial tasks. To address this, we propose two domain-adapted models, Fino1-8B and Fino1-14B, trained with CoT fine-tuning and reinforcement learning using domain-specific reasoning paths. Our models are trained on a carefully curated dataset integrating high-quality examples from diverse sources, covering financial reports, tables, equations, and structured XBRL texts. Despite limited training data, they achieve an 7-9% performance improvement, outperforming several advanced LLMs, including GPT-o1, GPT-o3-mini, GPT-4.5, and comparable with DeepSeek models (V3 and R1), demonstrating strong practical value in resource, constrained scenarios. Our findings highlight the need for domain-specific adaptations in financial reasoning, and we release all datasets, models, and code for future research.
Previous multilingual benchmarks focus primarily on simple understanding tasks, but for large language models(LLMs), we emphasize proficiency in instruction following, reasoning, long context understanding, code generation, and so on. However, measuring these advanced capabilities across languages is underexplored. To address the disparity, we introduce BenchMAX, a multi-way multilingual evaluation benchmark that allows for fair comparisons of these important abilities across languages. To maintain high quality, three distinct native-speaking annotators independently annotate each sample within all tasks after the data was machine-translated from English into 16 other languages. Additionally, we present a novel translation challenge stemming from dataset construction. Extensive experiments on BenchMAX reveal varying effectiveness of core capabilities across languages, highlighting performance gaps that cannot be bridged by simply scaling up model size. BenchMAX serves as a comprehensive multilingual evaluation platform, providing a promising test bed to promote the development of multilingual language models. The dataset and code are publicly accessible.
Transformer-based Large Language Models (LLMs) struggle to process inputs exceeding their training context window, with performance degrading due to positional out-of-distribution (O.O.D.) that disrupt attention computations. Existing solutions, fine-tuning and training-free methods, are limited by computational inefficiency, attention logit outliers or loss of local positional information. To address this, we propose Greedy Attention Logit Interpolation (GALI), a training-free length extrapolation method that maximizes the utilization of pretrained positional intervals while avoiding attention logit outliers through attention logit interpolation. The result demonstrates that GALI consistently outperforms state-of-the-art training-free methods. Our findings reveal that LLMs interpret positional intervals unevenly within their training context window, suggesting that extrapolating within a smaller positional interval range yields superior results-even for short-context tasks. GALI represents a significant step toward resolving the positional O.O.D. challenge, enabling more reliable long-text understanding in LLMs. Our implementation of GALI, along with the experiments from our paper, is open-sourced at https://github.com/AcademyCityL/GALI.
Equipping large language models (LLMs) with latent-space memory has attracted increasing attention as they can extend the context window of existing language models. However, retaining information from the distant past remains a challenge. For example, MemoryLLM (Wang et al., 2024a), as a representative work with latent-space memory, compresses past information into hidden states across all layers, forming a memory pool of 1B parameters. While effective for sequence lengths up to 16k tokens, it struggles to retain knowledge beyond 20k tokens. In this work, we address this limitation by introducing M+, a memory-augmented model based on MemoryLLM that significantly enhances long-term information retention. M+ integrates a long-term memory mechanism with a co-trained retriever, dynamically retrieving relevant information during text generation. We evaluate M+ on diverse benchmarks, including long-context understanding and knowledge retention tasks. Experimental results show that M+ significantly outperforms MemoryLLM and recent strong baselines, extending knowledge retention from under 20k to over 160k tokens with similar GPU memory overhead.