论文检索

Zhenmei Shi, Yifei Ming, Xuan-Phi Nguyen, Yingyu Liang, Shafiq Joty

Large Language Models (LLMs) have demonstrated remarkable capabilities in handling long context inputs, but this comes at the cost of increased computational resources and latency. Our research introduces a novel approach for the long context bottleneck to accelerate LLM inference and reduce GPU memory consumption. We show that LLMs can identify relevant tokens in the early layers prior to generating query responses. Leveraging this insight, we propose an algorithm that uses early layers of an LLM as filters to select and compress input tokens, significantly reducing the context length for subsequent processing. Our method, GemFilter, demonstrates substantial improvements in both speed and memory efficiency compared to existing techniques, such as standard attention and SnapKV/H2O. Notably, it achieves a 2.4X speedup and 30% reduction in GPU memory usage compared to SOTA methods. When evaluated on the Needle in a Haystack task, GemFilter significantly outperforms standard attention and SnapKV, while demonstrating comparable performance on the LongBench challenge. GemFilter is simple, training-free, and broadly applicable across different LLMs. Moreover, it provides interpretability by allowing humans to inspect the selected input sequence. Our findings provide practical benefits for deploying LLMs and deepen our understanding of their internal mechanisms, paving the way for further optimizations in LLM design and inference. Our code is available at https://github.com/SalesforceAIResearch/GemFilter.

Phung Gia Huy, Hai An Vu, Minh-Phuc Truong, Thang Duc Tran, Linh Ngo Van, Thanh Hong Nguyen, Trung Le

Representation learning is fundamental to NLP, but building embeddings that work well at different computational budgets is challenging. Matryoshka Representation Learning (MRL) offers a flexible inference paradigm through nested embeddings; however, learning such structures requires explicit coordination of how information is arranged across embedding dimensionality and model depth. In this work, we propose MIPIC (Matryoshka Representation Learning via Self-Distilled Intra-Relational Alignment and Progressive Information Chaining), a unified training framework designed to produce structurally coherent and semantically compact Matryoshka representations. MIPIC promotes cross-dimensional structural consistency through Self-Distilled Intra-Relational Alignment (SIA), which aligns token-level geometric and attention-driven relations between full and truncated representations using top-k CKA self-distillation. Complementarily, it enables depth-wise semantic consolidation via Progressive Information Chaining (PIC), a scaffolded alignment strategy that incrementally transfers mature task semantics from deeper layers into earlier layers. Extensive experiments on STS, NLI, and classification benchmarks (spanning models from TinyBERT to BGEM3, Qwen3) demonstrate that MIPIC yields Matryoshka representations that are highly competitive across all capacities, with significant performance advantages observed under extreme low-dimensional.

Jiuding Yang, Shengyao Lu, Hongxuan Liu, Shayan Shirahmad Gale Bagi, Zahra Fazel, Tomasz Czajkowski, Di Niu

Large language models (LLMs) have achieved remarkable progress in automatic code generation, yet their ability to produce high-performance code remains limited, despite its importance in real-world software systems. We argue that this limitation stems not only from data scarcity, but more fundamentally from the lack of supervision that guides interpretable and effective performance improvements. We introduce PerfCoder, a family of LLMs designed to generate performance-enhanced code through interpretable and customized optimization strategies. PerfCoder is fine-tuned on curated real-world optimization trajectories with human-readable annotations and further aligned via reinforcement fine-tuning using runtime feedback, enabling it to generate input-specific strategies and apply them directly without iterative refinement. On the PIE code performance benchmark, PerfCoder outperforms all existing models in both runtime speedup and effective optimization rate, demonstrating that code performance optimization requires strategy awareness rather than scale alone. Moreover, PerfCoder produces interpretable feedback that can guide larger LLMs in a planner–optimizer workflow, substantially improving the performance of 32B models and GPT-5 on code optimization.

Jieran Li, Xiuyuan Hu, Yang Zhao, Dongbiao Sun, Hao Zhang

Despite their strong generative capabilities, large language models frequently exhibit hallucinations, particularly due to outside-boundary confidence where incorrect assertions are produced with high statistical certainty. Existing approaches commonly use output probability as a proxy for truthfulness; however, this signal is confounded by epistemic uncertainty and cannot reliably distinguish genuine uncertainty from fabricated content. We argue that effective hallucination detection requires integrating surface-level confidence with signals that reflect the model’s underlying epistemic state. To this end, we propose Answer-level Intrinsic Cognition (AIC), a model-agnostic metric that captures epistemic boundary deviations by measuring answer-level stability across multiple stochastic forward passes. By coupling AIC with conventional output uncertainty, we derive a composite metric that disentangles within-boundary uncertainty from outside-boundary confidence. Across three public question-answering benchmarks and diverse model scales, the two-dimensional score consistently outperforms strong uncertainty-only baselines, with larger gains on adversarially constructed hallucination sets. The code is available at: https://github.com/HXYfighter/AIC-ACL2026.

Kailin Jiang, Ning Jiang, Yuntao Du, Yuchen Ren, Yuchen Li, Yifan Gao, Jinhe Bi, Yunpu Ma, Bin Li, Lei Liu 等

Large Multimodal Models (LMMs) encode rich factual knowledge via cross-modal pre-training, yet their static representations struggle to maintain an accurate understanding of time-sensitive knowledge. Existing benchmarks remain constrained by static designs, inadequately evaluating LMMs’ ability to understand time-sensitive knowledge. To address this gap, we propose MINED, a comprehensive benchmark containing 2,104 time-sensitive knowledge samples spanning six knowledge types, which evaluates temporal awareness along 6 key dimensions and 11 challenging tasks: cognition, awareness, trustworthiness, understanding, reasoning, and robustness. Evaluating 15 widely used LMMs on MINED shows that Gemini-2.5-Pro achieves the highest average CEM score of 63.07, while most open-source LMMs still lack time understanding ability. Meanwhile, LMMs perform best on organization knowledge, whereas their performance is weakest on sport. To address these challenges, we investigate the feasibility of updating time-sensitive knowledge in LMMs through knowledge editing methods and observe that LMMs can effectively update knowledge via knowledge editing methods in single editing scenarios.

Shaobo Wang, Tianle Niu, Xuan Ouyang, Xintong Li, Zhengkun Ge, Yue Min, Xiaoqian Liu, Hankun Wang, Linfeng Zhang

Dataset Pruning (DP) aims to construct a coreset that achieves performance comparable to the original, full dataset. However, few studies have explored DP in the context of Speech Classification (SC) tasks. Unlike image or text classification, SC is particularly challenging due to the difficulty in capturing the acoustic, semantic, and contextual representations. In this study, we propose a novel dataset pruning method for speech datasets, termed Meltrim, which uses a two-step coarse-to-fine framework designed to address these challenges. Specifically, in Step 1, Meltrim coarsely filters utterance-level redundant samples using DBSCAN clustering on Mel-Frequency Cepstral Coefficients (MFCC) features, which are first flattened and then reduced in dimensionality using UMAP. In Step 2, we perform frame-level redundancy pruning for each utterance via utility pruning, which aims to eliminate irrelevant frames within each utterance. To the best of our knowledge, this is the first dataset pruning approach designed for Speech Classification tasks, demonstrating outstanding performance compared to classical general DP methods. Notably, for the Speech Emotion Recognition, our method achieves up to a 49.5% improvement in WA (Weighted Accuracy) on the MEAD dataset. For the Speaker Identification tasks, it results in a 41.9% reduction in EER (Equal Error Rate) on the VoxCeleb1 dataset.

Lei Jiang, Desheng Wu, Xiaolong Zheng, Cuicui Luo

In recent years, the rapid advancement of large language models (LLMs) has enabled generated texts to closely mimic human writing, posing significant challenges to the detection of AI-generated content. Current mainstream zero-shot detection methods largely adopt a machine-centric perspective, relying on proxy models to compute token-level AI-likelihood scores and treating all tokens equally during overall detection. However, such approaches overlook the prediction discrepancies that arise when humans and large language models interpret the same text. We argue that tokens exhibiting greater divergence between human and machine predictions can provide stronger clues for determining the authorship of a text. To address this limitation, we propose HAPDA—a human-machine prediction discrepancy adapter for AI-generated text detection (AGTD). The framework consists of two core components: (1) a joint fine-tuning strategy for training paired human-preference and machine-preference models, and (2) a discrepancy-aware reweighting mechanism designed to calibrate token-level detection scores in downstream detectors. Extensive experiments demonstrate that HAPDA consistently and significantly enhances the detection performance of five representative baseline models under various evaluation scenarios.

Jian Zhang, Zhiyuan Wang, Zhangqi Wang, Yu He, Haoran Luo, li Yuan, Lingling Zhang, Rui Mao, Qika Lin, Jun Liu

Large Language Model (LLM) Agents exhibit inherent reasoning abilities through the collaboration of multiple tools.However, during agent inference, existing methods often suffer from (i) locally myopic generation, due to the absence of lookahead, and (ii) trajectory instability, where minor early errors can escalate into divergent reasoning paths. These issues make it difficult to balance global effectiveness and computational efficiency. To address these two issues, we propose meta-adaptive exploration with LLM agents (MAXS)[<https://github.com/exoskeletonzj/MAXS>], a meta-adaptive reasoning framework based on LLM Agents that flexibly integrates tool execution and reasoning planning. MAXS employs a lookahead strategy to extend reasoning paths a few steps ahead, estimating the advantage value of tool usage, and combines step consistency variance and inter-step trend slopes to jointly select stable, consistent, and high-value reasoning steps. Additionally, we introduce a trajectory convergence mechanism that controls computational cost by halting further rollouts once path consistency is achieved, enabling a balance between resource efficiency and global effectiveness in multi-tool reasoning. We conduct extensive empirical studies across three base models (MiMo-VL-7B, Qwen2.5-VL-7B, Qwen2.5-VL-32B) and five datasets, demonstrating that MAXS consistently outperforms existing methods in both performance and inference efficiency. Further analysis confirms the effectiveness of our lookahead strategy and tool usage.

Xiaoyu Chen, Fengge Wu, Zhao Junsuo, Yun Fan

Full-state latent communication in LLM-based multi-agent systems offers richer semantics than text but suffers from memory overhead scaling linearly with collaboration rounds. We propose CondenseFlow, which introduces the Latent Thought Condenser (LTC)—a lightweight module using learnable semantic probes to compress KV caches into fixed-size representations, achieving \mathcal{O}(1) communication complexity regardless of context length. We theoretically prove that compression error is bounded by attention concentration and accumulates controllably across rounds. On seven benchmarks spanning six models, CondenseFlow reduces KV cache memory by over 99% and inference latency by approximately 20% compared to dense transfer with negligible accuracy degradation, while outperforming text-based methods by 1.7 percentage points on average across all configurations. Code is available at https://github.com/xxy33/condenseflow.

Qinzhuo Wu, Zhizhuo Yang, Hanhao Li, Pengzhi Gao, Wei Liu, Jian Luan

Recent advances in mobile Graphical User Interface (GUI) agents highlight the growing need for comprehensive evaluation benchmarks. While new online benchmarks offer more realistic testing than offline ones, they tend to focus on the agents’ task instruction-following ability while neglecting their reasoning and exploration ability. Moreover, these benchmarks do not consider the random noise in real-world mobile environments. This leads to a gap between benchmarks and real-world environments. To addressing these limitations, we propose MobileBench-OL, an online benchmark with 1080 tasks from 80 Chinese apps. It measures task execution, complex reasoning, and noise robustness of agents by including 5 subsets, which set multiple evaluation dimensions. We also provide an auto-eval framework with a reset mechanism, enabling stable and repeatable real-world benchmarking. Evaluating 13 leading GUI agents on MobileBench-OL shows significant room for improvement to meet real-world requirements. Human evaluation further confirms that MobileBench-OL can reliably measure the performance of leading GUI agents in real environments.

Chuan Qin, Xi Chen, Jinpeng Li, Hengshu Zhu

Text classification has long been a cornerstone of NLP, yet most prior work and benchmarks have been limited to closed-world settings, where all classes are assumed to be known in advance. In contrast, open-world learning has recently emerged as a critical paradigm for building more robust and realistic systems. However, existing benchmarks largely focus on out-of-distribution (OOD) detection, while overlooking broader challenges such as the discovery of novel categories. To address this gap, we introduce BOLT, a unified Benchmark and evaluation toolkit supporting Open-world Learning for Text classification. BOLT encompasses two representative tasks: Open-set Text Classification (OSTC), which requires models to classify in-distribution (ID) samples while rejecting OOD inputs, and Generalized Category Discovery (GCD), which aims to identify both known and novel categories from partially labeled corpora. We carefully curate 12 publicly available datasets spanning diverse domains and benchmark 22 methods, including 15 for OSTC and 7 for GCD, under a standardized protocol that explicitly accounts for varying labeled ratios and known class ratios. Our results reveal key challenges: most current methods tend to overfit training distributions and struggle to generalize to unseen classes. Moreover, by comparing our lightweight LLM-based variants with prior open-set baselines, we demonstrate the promise of leveraging LLMs for open-world text classification. BOLT provides standardized evaluation protocols that enable fair comparison and support future research in this emerging area. All datasets, baselines, and tools are available at https://github.com/CNIC-DSL/BOLT.

Zhibo Zhang, Yang Xu, Kai Ming Ting, Cam-Tu Nguyen

Large language models (LLMs) have recently enabled remarkable progress in text representation. However, their embeddings are typically high-dimensional, leading to substantial storage and retrieval overhead. Although recent approaches such as Matryoshka Representation Learning (MRL) and Contrastive Sparse Representation (CSR) alleviate these issues to some extent, they still suffer from retrieval accuracy degradation. This paper proposes Isolation Kernel Embedding or IKE, a learning-free method that transforms an LLM embedding into a binary embedding using Isolation Kernel (IK). Lightweight and based on binary encoding, IKE offers a low memory footprint and fast bitwise computation, lowering retrieval latency. Experiments on multiple text retrieval datasets demonstrate that IKE offers up to 16.7\times faster retrieval and 16\times lower memory usage than the original LLM embeddings, while maintaining comparable accuracy. Theoretically, we show that IKE works because it satisfies four essential criteria for effective binary hashing that other methods do not possess. Compared to CSR, IKE consistently achieves better retrieval efficiency and effectiveness. IKE also works effectively with graph-based indexing, demonstrating its superiority in balancing accuracy and latency compared to alternative compression techniques in the approximate nearest neighbor (ANN) search setting.

Pengyu Li, Lingling Zhang, Zhitao Gao, Yanrui Wu, Yuxuan Dong, Huan Liu, Bifan Wei, Jun Liu

While Large Language Models (LLMs) have achieved remarkable capabilities, they unintentionally memorize sensitive data, posing critical privacy and security risks.Machine unlearning is pivotal for mitigating these risks, yet existing paradigms face a fundamental dilemma: aggressive unlearning often induces catastrophic forgetting that degrades model utility, whereas conservative strategies risk superficial forgetting, leaving models vulnerable to adversarial recovery. To address this trade-off, we propose AGTAO (Adversarial Gating Training with Adaptive Orthogonality), a unified framework designed to reconcile robust erasure with utility preservation. Specifically, our approach introduces Adaptive Orthogonality (AO) to dynamically mitigate geometric gradient conflicts between forgetting and retention objectives, thereby minimizing unintended knowledge degradation. Concurrently, Adversarial Gating Training (AGT) formulates unlearning as a latent-space min-max game, employing a curriculum-based gating mechanism to simulate and counter internal recovery attempts. Extensive experiments demonstrate that AGTAO achieves a superior trade-off between unlearning efficacy (KUR \approx 0.01) and model utility (MMLU 58.30).[Code is available at <https://anonymous.4open.science/r/AGT-unlearning>.].

Kuofei Fang, Siyan Wu, Yu Guo, Bin Wu

Node Importance Estimation (NIE) in Knowledge Graphs (KGs) aims to quantify the significance of entities, serving as a pivotal instrument for deciphering the latent mechanisms of social dynamics. However, existing methods are often confined to supervised paradigms and rely heavily on topological aggregation, resulting in limited generalization capability. To address these challenges, we propose GenNIE, the first end-to-end generative reasoning framework for NIE. Specifically, GenNIE leverages Large Language Models (LLMs) integrated with topological information to generate precise importance scores for entities in KGs. Furthermore, GenNIE introduces a Global-Structural Graph Perception mechanism to empower the LLMs with holistic graph cognition. Extensive experiments demonstrate the performance superiority of GenNIE and its robust generalization across diverse domains. Our code is available at https://github.com/CoffeyF/GenNIE.git.

Ziniu Liu, Shuheng Zhou, Mingqing Liu, Hao Deng, Huijia Zhu

Visual token pruning has emerged as a pivotal strategy to alleviate the computational bottleneck in Multimodal Large Language Models (MLLMs), yet it frequently compromises the integrity of visual understanding in pursuit of efficiency. Existing methods face a fundamental tension: vision-centric approaches are susceptible to the attention sink phenomenon and operate in a query-agnostic manner, whereas text-guided methods often create an overly narrow focus, discarding essential background context and failing on ambiguous queries. In this paper, we propose CrisPrune, a training-free and model-agnostic method that reconciles efficiency with understanding by integrating visual saliency and text relevance. Specifically, we introduce intrinsic visual saliency with robust normalization to identify information-rich regions characterized by significant visual features. Simultaneously, we design dual-source text relevance to synergize explicit instruction alignment with implicit scene priors. Finally, we reformulate the selection process using a Determinantal Point Process (DPP) to balance token quality and spatial diversity. Extensive experiments demonstrate that CrisPrune significantly outperforms state-of-the-art methods. On LLaVA-NeXT, it achieves a 13 \times decrease in FLOPs while maintaining 97% of the original performance with 94.4% of visual tokens pruned, effectively bridging the gap between efficiency and holistic understanding.

Wenqian Yu, Shuo Chen, Zhijiang Li, Zhipeng Wang, Jindong Gu

Large language models (LLMs) face a critical alignment challenge: balancing safety with helpfulness. Excessive safety can lead to over-refusal, where models reject harmful-looking yet benign queries, severely limiting utility.Existing training-free interventions offer an efficient way to mitigate over-refusal without re-training, but suffer from high inference overhead and architecture dependency. Our work explores a complementary direction: rather than applying post-hoc corrections to model outputs, our goal is to intrinsically reshape the distributions of harmful and benign samples within the model’s decision space. In this paper, we argue that a lightweight training-based approach can more effectively distinguish between harmful and benign samples. We propose Single Token Alignment (STA), which optimizes only a single-token prefix (e.g., 4,096 parameters) while keeping the base model frozen. To address the inherent challenge of achieving robust refinement through such a minimal parameter interface, STA employs a mixed weighting mechanism integrated with its optimization objective. This mechanism incorporates hard weighting via stringent data filtering to provide clear, unbiased learning signals, and soft weighting through a focal mechanism to prioritize challenging cases.Extensive experiments across 9 models and 10 datasets demonstrate that STA achieves a superior safety-helpfulness balance for LLMs, MLLMs, and reasoning models, offering a highly efficient and generalizable solution for refining safety alignment.

Dong She, Xianrong Yao, Liqun Chen, Jinghe Yu, Yang Gao, Zhanpeng Jin

Vision-Language Models (VLMs) have demonstrated strong capabilities in perception, yet holistic Affective Image Content Analysis (AICA)—which integrates perception, reasoning, and generation into a unified framework—remains underexplored. To address this, we introduce AICA-Bench, a comprehensive benchmark comprising three core tasks: Emotion Understanding (EU), Reasoning (ER), and Generation (EGCG). We evaluate 23 VLMs, revealing critical gaps: models struggle with intensity calibration and suffer from descriptive shallowness in open-ended tasks. To bridge these gaps, we propose Grounded Affective Tree (GAT) Prompting, a training-free framework that integrates visual scaffolding with hierarchical reasoning. Experiments show that GAT effectively corrects intensity errors and significantly enhances descriptive depth, establishing a robust baseline for future affective multimodal research.

Zijian Zheng, Yonghe Lu

Short text clustering has gained significant prominence due to its ubiquity in real-world applications. Despite the recent success of contrastive clustering, existing paradigms still suffer from two critical bottlenecks: (1) conventional data augmentation provides limited semantic granularity and may introduce unintended noise; and (2) the absence of global optimization for cluster assignments often precipitates the accumulation of pseudo-label noise, thereby compromising semantic consistency. To bridge these gaps, we propose MAST, a Multi-view Alignment Strategy with Transport-based clustering. MAST constructs complementary structural views to capture multi-granularity semantic features and introduces a multi-view contrastive objective that jointly aligns original, augmented, and structure-enhanced embeddings. To mitigate representation over-smoothing, we incorporate structure-aware negative reweighting and intermediate-layer negative sampling. Furthermore, MAST employs high-confidence guided refinement and an optimal transport-based pseudo-label alignment mechanism to enforce global semantic consistency across multiple views. Extensive experiments on several benchmark datasets demonstrate that MAST consistently outperforms state-of-the-art methods, establishing a new competitive baseline for short text clustering.

Yiju Guo, Tianyi Hu, Zexu Sun, Yankai Lin

Reinforcement Learning with Verifiable Rewards (RLVR) has advanced LLM reasoning, but remains constrained by inefficient exploration under limited rollout budgets, leading to low sampling success and unstable training in complex tasks. We find that many exploration failures arise not from problem difficulty, but from a small number of prompt tokens that introduce interference. Building on this insight, we propose the Less Noise Sampling Framework (LENS), which first purifies prompts by identifying and removing interference tokens. then transfers successful rollouts from the purified setting to supervise policy optimization on the original noisy prompts, enabling the model to learn to ignore interference in the real-world, noisy prompting settings. Experimental results show that LENS significantly outperforms GRPO, delivering higher performance and faster convergence, with a 3.88% average gain and over 1.6 \times speedup. Our work highlights the critical role of pruning interference tokens in improving rollout efficiency, offering a new perspective for RLVR research.

Yueqin Yin, Yaxi Li, Xin Liu, Xun Wang, Kaiqiang Song, Simin Ma, Shujian Liu, Sathish Reddy Indurthi, Haoyun Deng, Pengcheng He 等

Large language models often hallucinate, producing content that is factually incorrect or not grounded in the sources. Reliable faithfulness verification is critical for trustworthy deployment. In the provided-source (closed-world) setting, existing verifiers either classify whole passages in one step or check sentences independently, overlooking cross-sentence context. We present ContextCheck, a framework for sentence-level faithfulness verification with context-aware disambiguation. Each sentence is verified against the grounding document while conditioning on preceding sentences, enabling pronouns and references to be resolved directly in context. This design avoids the separate decontextualization step of rewriting claims into self-contained forms, casting verification as a context-conditioned task. Fine-tuned from Llama-3.1-8B-Instruct, ContextCheck sets a new state of the art on three context-dependent datasets; it improves Macro F1 by over 10 points compared to the strongest baselines, and matches or slightly surpasses the strongest baselines on 14 standard single-sentence datasets compared to prior 8B-scale verifiers (average Macro F1 73.5 vs. 72.8). These results show that ContextCheck offers a practical and effective approach for sentence-level hallucination detection.