Quantong Qiu

Ruoxi Sun, Quantong Qiu, Juntao Li et al.

While Multimodal Large Language Models (MLLMs) demonstrate remarkable proficiency on complex vision-language tasks, the mechanisms by which they extract query-relevant visual features from complex, noisy contexts remain opaque. In this paper, we present an in-depth interpretability study that uncovers a profound structural property within MLLMs: functional sparsity in cross-modal retrieval. Leveraging a token-level metric termed Retrieval Attention Mass (RAM), we identify and characterize a highly specialized subset of attention heads, referred to as Context-aware Retrieval (CoRe) heads. Across diverse visual domains and model scales, we observe a clear functional division: CoRe heads act as dedicated information extractors, while most other heads distribute attention over broader contextual regions. Causal interventions further demonstrate the necessity of these specialized heads. Ablating only the top 5% of CoRe heads causes significant degradation in multimodal reasoning performance, whereas ablating lower-ranked heads has minimal effect. Moreover, acceleration experiments validate the utility of CoRe heads, showing that leveraging this localized sparsity significantly accelerates inference while maintaining robust task performance. Our findings reveal a structural principle of functional sparsity within MLLMs, refining the current understanding of mechanistic interpretability and laying a theoretical foundation that can inspire future architecture design and model optimization.

Quantong Qiu, Zhiyi Hong, Yi Yang et al.

The quadratic computational complexity of standard attention mechanisms presents a severe scalability bottleneck for LLMs in long-context scenarios. While hybrid attention mechanisms combining Full Attention (FA) and Sparse Attention (SA) offer a potential solution, existing methods typically rely on static allocation ratios that fail to accommodate the variable retrieval demands of different tasks. Furthermore, head-level dynamic sparsity often introduces severe computational load imbalance and synchronization long-tails, which hinder hardware acceleration during autoregressive decoding. To bridge this gap, we introduce Flux Attention, a context-aware framework that dynamically optimizes attention computation at the layer level. By integrating a lightweight Layer Router into frozen pretrained LLMs, the proposed method adaptively routes each layer to FA or SA based on the input context. This layer-wise routing preserves high-fidelity information retrieval while ensuring contiguous memory access, translating theoretical computational reductions into practical wall-clock speedups. As a parameter-efficient approach, our framework requires only 12 hours of training on 8$\times$A800 GPUs. Extensive experiments across multiple long-context and mathematical reasoning benchmarks demonstrate that Flux Attention achieves a superior trade-off between performance and inference speed compared with baseline models, with speed improvements of up to $2.8\times$ and $2.0\times$ in the prefill and decode stages.

Zecheng Tang, Haitian Wang, Quantong Qiu et al.

Long-context processing has become a fundamental capability for large language models~(LLMs). To assess model's long-context performance, numerous long-context evaluation benchmarks have been proposed. However, variations in evaluation settings across these benchmarks lead to inconsistent results, making it difficult to draw reliable comparisons. Besides, the high computational cost of long-context evaluation poses a significant barrier for the community to conduct comprehensive assessments of long-context models. In this paper, we propose LOOM-Scope, a comprehensive and efficient framework for long-context evaluation. LOOM-Scope standardizes evaluation settings across diverse benchmarks, supports deployment of efficient long-context inference acceleration methods, and introduces a holistic yet lightweight benchmark suite to evaluate models comprehensively. Homepage: https://loomscope.github.io

Zecheng Tang, Baibei Ji, Quantong Qiu et al.

Reward model (RM) plays a pivotal role in aligning large language model (LLM) with human preferences. As real-world applications increasingly involve long history trajectories, e.g., LLM agent, it becomes indispensable to evaluate whether a model's responses are not only high-quality but also grounded in and consistent with the provided context. Yet, current RMs remain confined to short-context settings and primarily focus on response-level attributes (e.g., safety or helpfulness), while largely neglecting the critical dimension of long context-response consistency. In this work, we introduce Long-RewardBench, a benchmark specifically designed for long-context RM evaluation, featuring both Pairwise Comparison and Best-of-N tasks. Our preliminary study reveals that even state-of-the-art generative RMs exhibit significant fragility in long-context scenarios, failing to maintain context-aware preference judgments. Motivated by the analysis of failure patterns observed in model outputs, we propose a general multi-stage training strategy that effectively scales arbitrary models into robust Long-context RMs (LongRMs). Experiments show that our approach not only substantially improves performance on long-context evaluation but also preserves strong short-context capability. Notably, our 8B LongRM outperforms much larger 70B-scale baselines and matches the performance of the proprietary Gemini 2.5 Pro model.

Yi Su, Quantong Qiu, Yuechi Zhou et al.

Large Language Models (LLMs) require substantial computational resources during generation. While the Key-Value (KV) cache significantly accelerates this process by storing attention intermediates, its memory footprint grows linearly with sequence length, batch size, and model size, creating a bottleneck in long-context scenarios. Various KV cache compression techniques, including token eviction, quantization, and low-rank projection, have been proposed to mitigate this bottleneck, often complementing each other. This paper focuses on enhancing token eviction strategies. Token eviction leverages the observation that the attention patterns are often sparse, allowing for the removal of less critical KV entries to save memory. However, this reduction usually comes at the cost of notable accuracy degradation, particularly under high compression ratios. To address this issue, we propose \textbf{CaliDrop}, a novel strategy that enhances token eviction through calibration. Our preliminary experiments show that queries at nearby positions exhibit high similarity. Building on this observation, CaliDrop performs speculative calibration on the discarded tokens to mitigate the accuracy loss caused by token eviction. Extensive experiments demonstrate that CaliDrop significantly improves the accuracy of existing token eviction methods.