Recent reports suggest that LLMs can handle increasingly long contexts. However, many existing benchmarks for context understanding embed substantial query-irrelevant content, which shifts evaluation toward retrieving relevant snippets rather than fully integrating all provided information. Under this setting, we view that current benchmarks can overestimate true context-understanding ability of LLMs. In particular, we demonstrate that when the context consists entirely of query-relevant text, even advanced models such as GPT-4o fail to reliably integrate inputs as short as 200 tokens. To evaluate this capability more rigorously, we introduce NeedleChain, a benchmark designed to test whether models can faithfully incorporate all given evidence. NeedleChain includes three variants that differ in the required order of comprehension, along with a parallel benchmark based on the needle-in-a-haystack(NIAH) paradigm. By comparing these variants, NeedleChain enables a more comprehensive assessment of context understanding. We further propose a training-free strategy that encourages models to reflect all available information, ROPE contraction, highlighting the importance of full-context integration and pointing to new directions for improving reliable reasoning over context.
论文检索
输入标题、作者或关键词，从 100,903 篇学术成果中精准定位
LLaTiSA: Towards Difficulty-Stratified Time Series Reasoning from Visual Perception to Semantics
PDF ↗Comprehensive understanding of time series remains a significant challenge for Large Language Models (LLMs). Current research is hindered by fragmented task definitions and benchmarks with inherent ambiguities, precluding rigorous evaluation and the development of unified Time Series Reasoning Models (TSRMs). To bridge this gap, we formalize Time Series Reasoning (TSR) via a four-level taxonomy of increasing cognitive complexity. We introduce HiTSR, a **hi**erarchical **t**ime **s**eries **r**easoning dataset comprising 83k samples with diverse task combinations and verified Chain-of-Thought (CoT) trajectories. Leveraging HiTSR, we propose LLaTiSA, a strong TSRM that integrates visualized patterns with precision-calibrated numerical tables to enhance the temporal perception of Vision-Language Models (VLMs). Through a multi-stage curriculum fine-tuning strategy, LLaTiSA achieves superior performance and exhibits robust out-of-distribution generalization across diverse TSR tasks and real-world scenarios. We will publicly release the code, dataset, and model weights.
Retrieval Augmented Generation faces a trade-off: concatenating documents in a long prompt enables multi-document reasoning but creates prefill bottlenecks, while encoding document KV caches separately offers speed but breaks cross-document interaction. We propose Parallel Context-of-Experts Decoding (PCED), a training-free framework that shifts evidence aggregation from the attention mechanism to the decoding. PCED treats retrieved documents as isolated "experts", synchronizing their predictions via a retrieval-aware extension of context-aware decoding. This approach recovers cross-document reasoning capabilities without constructing a shared attention across documents.
Despite significant advances in Large Reasoning Models (LRMs) driven by reinforcement learning with verifiable rewards (RLVR), this paradigm is fundamentally limited in specialized or novel domains where such supervision is prohibitively expensive or unavailable, posing a key challenge for test-time adaptation. While existing test-time methods offer a potential solution, they are constrained by learning from static query sets, risking overfitting to textual patterns. To address this gap, we introduce Test-Time Variational Synthesis (TTVS), a novel framework that enables LRMs to self-evolve by dynamically augmenting the training stream from unlabeled test queries. TTVS comprises two synergistic modules: (1) Online Variational Synthesis, which transforms static test queries into a dynamic stream of diverse, semantically-equivalent variations, enforcing the model to learn underlying problem logic rather than superficial patterns; (2) Test-time Hybrid Exploration, which balances accuracy-driven exploitation with consistency-driven exploration across synthetic variants. Extensive experiments show TTVS yields superior performance across eight model architectures. Notably, using only unlabeled test-time data, TTVS not only surpasses other test-time adaptation methods but also outperforms state-of-the-art supervised RL-based techniques trained on vast, high-quality labeled data.
Large Multimodal Models (LMMs) have achieved significant success across various tasks. These models usually encode visual inputs into dense token sequences, which are then concatenated with textual tokens and jointly processed by a language model. However, the increased token count substantially raises computational and memory costs during inference. Token pruning has emerged as a promising approach to address this issue. Existing token pruning methods often rely on costly calibration or suboptimal importance metrics, leading to redundant retained tokens. In this paper, we analyze the redundancy differences between visual and textual tokens and propose pruning exclusively on visual tokens. Based on this, we propose a visual token pruning strategy that explicitly preserves both cross-modal alignment and intra-modal informational diversity. We introduce a mutual information-based token pruning strategy that removes visual tokens semantically misaligned with textual tokens, effectively preserving the alignment between the visual and textual modalities. We further refine the retained tokens by maximizing their expected pairwise distances in the latent space to enhance representational quality and reduce redundancy. which is solved efficiently with a greedy algorithm. Extensive experiments demonstrate that our method maintains strong performance while reducing tokens by 88.9% on models such as LLaVA-1.5-7B and LLaVA-NEXT-7B, resulting in a 56.7% improvement in inference speed.
Tokenization plays a pivotal role in NLP and is fundamental to training language models. However, existing tokenizers are often skewed towards high-resource languages, limiting their effectiveness for linguistically diverse and morphologically rich languages such as those in the Indian subcontinent. In this work, we present a comprehensive empirical study of multilingual tokenization across 17 Indic languages spanning 11 scripts and two language families. We systematically evaluate the effects of (i) widely used subword algorithms: BPE (CITATION) and Unigram LM (CITATION), (ii) script and orthography-aware normalization, (iii) vocabulary size, and (iv) multilingual vocabulary construction strategies. We use a combination of intrinsic and extrinsic evaluations to obtain the following observations: (i) script-specific normalization improves tokenization quality, (ii) Unigram LM better preserves morphological boundaries than BPE, (iii) cluster-based vocabulary construction shows improvement in downstream tasks compared to the joint method. Our findings highlight the importance of linguistically informed design choices in multilingual tokenization and offer practical guidance for building effective tokenizers for low-resource and morphologically complex languages.
Distilling Examples into Task Instructions: Enhanced In-Context Learning for Real-World B2B Conversations
PDF ↗In-context learning (ICL) is the standard method for low-resource classification, yet its efficacy in specialized domains remains largely unexplored. We address the challenge of classifying semantically complex, multi-party B2B conversations, where traditional ICL encounters significant limitations, especially as context length increases due to the concatenation of multiple few-shot examples. We introduce the Call Playbook dataset, featuring five classification tasks derived from real-world B2B conversations targeting core sales concepts. To bridge the gap between performance and practical utility, we propose novel knowledge extraction methods that distill verbose examples into compact, interpretable representations of structured classification criteria and precise task descriptions. Our approach achieves a 99% reduction in token usage and improves macro-averaged AUC by up to 7% over traditional ICL. Notably, it remains robust as context grows, unlike advanced token compression baselines which degrade by over 9 F1 points. Importantly, our framework enables direct refinement of classification logic, addressing critical needs for transparency, efficiency, and user interaction in real-world NLP applications.
Accurate and interpretable predictions of depression severity are essential for clinical decision support, yet existing models often lack uncertainty estimates and temporal interpretability. We propose PTTSD, a Probabilistic framework for Depression Detection from clinical interview utterance sequences that predicts PHQ-8 scores while modeling calibrated uncertainty. PTTSD includes sequence-to-sequence and sequence-to-one variants, both combining LSTMs, self-attention, and residual connections with Gaussian or Student’s-t output heads trained via negative log-likelihood. The sequence-to-sequence variant enables temporal analysis of how predictive confidence evolves over an interview, despite the target being a single session-level score. Evaluated on E-DAIC and DAIC-WOZ, PTTSD achieves competitive performance among text-only systems (e.g., MAE = 3.85 on E-DAIC, 3.55 on DAIC) and produces well-calibrated prediction intervals. Ablations confirm the value of attention and probabilistic modeling, while a three-part calibration analysis and qualitative case studies highlight the clinical relevance of uncertainty-aware prediction.
Speculative decoding (SD) accelerates large language model (LLM) inference by using a smaller model to draft future tokens, which are then verified by the target LLM. This preserves generation quality by accepting only aligned tokens. However, individual drafters, often trained for specific tasks or domains, exhibit limited effectiveness across diverse applications. To address this, we introduce MetaSD, a unified framework that integrates multiple drafters into the SD process. MetaSD dynamically allocates computational resources to heterogeneous drafters by leveraging alignment feedback and framing drafter selection as a multi-armed bandit problem. Extensive experiments show MetaSD consistently outperforms single-drafter approaches.
Data marketplaces analyze strategic data exchanges among users, platforms, and buyers. However, most existing studies model static equilibria and complete information, which limits their realism. In this work, we study whether large language model (LLM)-driven agents can make equilibrium-consistent decisions in analytically tractable data marketplaces with evolving and incomplete-information. Specifically, we introduce EvoDM, an agent-based modeling framework that extends the static data marketplace to dynamic and incomplete-information settings while providing tractable equilibrium benchmarks for evaluating agent decisions. Building upon EvoDM, we propose Datamart-Agent, an LLM-driven game-theoretic agent that improves equilibrium-consistent decision execution through dynamic game tree memory and mechanism-guided reflection, without requiring parameter updates. Experiments demonstrate that Datamart-Agent closely matches equilibrium-consistent decision-making, achieving the lowest utility gap and over 20% higher Pass@\epsilon than strong baselines. After validating its effectiveness, we employ EvoDM with Datamart-Agent to analyze competition and regulation in assumption-relaxed settings where closed-form ground truth is unavailable, providing exploratory simulation-based insights into market dynamics and regulatory effects.
Sounding vs. Being an Expert: Disentangling Authority, Register and Cultural Impact in Sycophantic LLMs
PDF ↗Large Language Models (LLMs) have been shown to exhibit sycophancy, a tendency to align with user assertions even when they conflict with facts. We frame sycophancy as a sociolinguistic phenomenon, disentangling two distinct drivers of credibility: explicit authority (credentials) and implicit authority (linguistic register). We introduce the Sycophancy Matrix, an adversarial evaluation framework that isolates these variables. Using a controlled subset of TruthfulQA, we evaluate open-weight models across English, Spanish, and Portuguese variants. Our findings reveal that models often conflate high register with truthfulness: for some architectures, sophisticated tone triggers deference more effectively than explicit expertise. Furthermore, we observe statistically significant variability across cultural variants of Spanish and Portuguese, supporting the hypothesis that LLMs internalise language-specific sociolinguistic norms and that sycophancy is not a purely technical deficit but an emergent property of multilingual training and alignment. Finally, we identify stable sycophancy fingerprints–domain-specific vulnerability profiles that persist across languages–suggesting that alignment artefacts are intrinsic to model families rather than linguistic context.
Causal Evidence Extraction and Triangulation in Crisis Reports using Large Language Models: A ReliefWeb-based Study
PDF ↗Humanitarian reports are long, noisy, and multi-topic, making it difficult to consolidate decision-relevant causal evidence. We present a ReliefWeb study (2000–2024) and a two-stage Large Language Model (LLM) pipeline that extracts structured intervention-outcome records with direction and strength attributes. Query-conditioned extraction restricts output to a specified intervention class, reducing retrieval-induced over-extraction, while snippet grounding links each relation to supporting text for auditability and classification. In an expert-annotated dataset of 100 reports, the best closed-source LLM achieved a weighted F1 score of 90.73% with strong cost-efficiency, while Llama-3.1-8B with supervised fine-tuning reached 94.15% weighted F1 score. We further propose context-preserving triangulation that aggregates strength-weighted evidence within disaster\timessource cells, applies Laplace smoothing and equally weights cells to quantify cross-context convergence via a Level-of-Evidence score. Applied to cash assistance, food-related outcomes show strong positive convergence (LoE=0.865) and stable long-horizon trajectories.
Rejection-to-Acceptance Transition: Model Editing-Based Jailbreak Backdoor Injection Not Limited to Few Output Tokens
PDF ↗Model editing-based jailbreak backdoor attacks against LLMs have gained attention for being lightweight, enabling vulnerability discovery in LLMs. Existing methods are implemented by binding backdoors to predefined phrases as first few output tokens, inducing the LLM’s next-token prediction to produce continuous responses. However, their effectiveness is heavily dependent on the number of bound phrases, with attack costs rising as this number increases. In this work, we propose JEST, which achieves jailbreak backdoor attacks by hijacking LLM representations into a acceptance domain rather than binding to a few output tokens. Specifically, we propose a representation transition-guided model editing to inject jailbreak backdoors into LLMs. The activated backdoor transitions the LLM from rejection domain to acceptance domain, causing it to accept and generate jailbreak behavior. To clearly distinguish between rejection and acceptance domains within LLMs, we also design a domain modeling strategy for JEST that models these two opposing domains within the representation space. Additionally, JEST-hijacked LLMs exhibit greater vulnerability to direct prompt attacks. Experimental results show that JEST outperforms existing model editing methods, demonstrating stronger jailbreak capabilities across various LLMs and datasets. We also provide analysis to explore the safety boundary of LLM.
Recent research has extensively explored the graph-reasoning capabilities of Large Language Models (LLMs) through textual descriptions. However, benchmarks specifically designed for Graph-Language Models (GLMs), which integrate Graph Neural Networks (GNNs) with LLMs, remain significantly underdeveloped. In this work, we first demonstrate that existing GLM evaluations, largely repurposed from unimodal node and edge level tasks, fail to assess true multimodal integration. Our analysis reveals that strong performance on these benchmarks is achievable using textual or structural features in isolation, bypassing the need for joint reasoning. To bridge this gap, we introduce CLEGR (Compositional Language-Graph Reasoning), a benchmark explicitly designed to evaluate multimodal reasoning over graph topology and textual semantics. Evaluation of representative GLMs on CLEGR shows that they exhibit significant performance degradation on CLEGR tasks and unimodal soft-prompted LLMs perform on par with complex multimodal GLMs. These findings collectively highlight limitations in the graph reasoning capabilities of existing GLMs and provide a foundation for advancing the community toward explicit multimodal reasoning involving graph structure and language.
Embodied question answering (EQA) in 3D environments often requires collecting context that is distributed across multiple viewpoints and partially occluded. However, most recent vision–language models (VLMs) are constrained to a fixed and finite set of input views, which limits their ability to acquire question-relevant context at inference time and hinders complex spatial reasoning. We propose Chain-of-View (CoV) prompting, a training-free, test-time reasoning framework that transforms a VLM into an active viewpoint reasoner through a coarse-to-fine exploration process. CoV first employs a View Selection agent to filter redundant frames and identify question-aligned anchor views. It then performs fine-grained view adjustment by interleaving iterative reasoning with discrete camera actions, obtaining new observations from the underlying 3D scene representation until sufficient context is gathered or a step budget is reached. We evaluate CoV on OpenEQA across four mainstream VLMs and obtain an average 11.98% improvement in LLM-Match, with a maximum gain of 13.62% on Qwen3-VL-Flash. CoV further exhibits test-time scaling: increasing the minimum action budget yields an additional 2.54% average improvement, peaking at 3.73% on Gemini-2.5-Flash. On ScanQA and SQA3D, CoV delivers strong performance (e.g., 116 CIDEr 31.9 EM@1 on ScanQA and 51.1 EM@1 on SQA3D). Overall, these results suggest that question-aligned view selection coupled with open-view search is an effective, model-agnostic strategy for improving spatial reasoning in 3D EQA without additional training.
Hierarchical Representation Alignment Learning of Diffusion Transformers for Neural Audio Codec
PDF ↗Despite recent progress in diffusion and conditional flow matching (CFM) models for low-resolution domains such as latent representations, their application to high-resolution data like raw waveform signals remains underexplored. Generative adversarial networks (GANs) have been the dominant approach in neural vocoder and neural audio codecs for realistic waveform generation. However, under low-bitrate conditions, these models suffer from degraded performance due to information loss caused by heavy compression and quantization, often resulting in mispronunciations. To address the aforementioned problem, we first leverage CFM to iteratively generate raw waveform in an extremely low-bitrate scenario. We then introduce hierarchical representation alignment learning (REPA-H) to enable efficient and robust CFM training. Furthermore, we propose dense vector quantization (DVQ), a novel factorized quantization method using a single quantizer. Our model, FlowTokenizer, outperforms state-of-the-art neural audio codecs in audio quality and semantic intelligibility under low-bitrate conditions, using only 25 tokens per second for 24 kHz waveform generation.
Large language models have demonstrated strong reasoning capabilities in general knowledge question answering. However, their ability to handle temporal information remains limited. To address this limitation, existing approaches often design time-sensitive reasoning pipelines that rely on external tools or manual verification and are tailored to specific scenarios, leading to poor generalizability. Moreover, these methods apply a fixed pipeline to all questions, overlooking the fact that different types of temporal questions often require distinct reasoning strategies, which leads to unnecessary processing for simple cases and inadequate reasoning for more complex ones. To this end, we propose AdapTime, an adaptive temporal reasoning method that dynamically executes reasoning steps based on the input context and task requirements. Specifically, it involves three temporal reasoning actions: reformulate, rewrite and review, with an LLM planner guiding the reasoning process. AdapTime integrates seamlessly with state-of-the-art LLMs and significantly enhances their temporal reasoning capabilities without relying on external support. Extensive experiments on two temporal QA benchmarks demonstrate the effectiveness of our approach.
Let the Comments Speak: A Multi-Agent Framework based on Large Language Model for Comment-Guided Code Refactoring
PDF ↗Code refactoring is essential for software maintainability, yet current Large Language Model (LLM) based frameworks primarily focus on syntax and neglect the vital semantic signals in code comments. As pointed out in Fowler’s refactoring theory, explanatory comments function as semantic anchors that provide necessary guidance for method extraction. Moreover, research reports show that more than 84 percent of codebases lack appropriate code comments, which hinders the leverage of such guidance. To bridge this gap, we propose MACOR, a Multi-Agent framework for COmment-guided code Refactoring. In specific, MACOR populates original code with precise comments to provide necessary semantic guidance for the subsequent refactoring process. These generated signals are employed to retrieve expert examples. An iterative feedback is incorporated loop for validation. Experiments are conducted on three benchmarks using three base LLMs. Experimental results show that MACOR significantly optimizes code quality and achieves higher developer acceptance compared to the representative baselines.
RecMem: Recurrence-based Memory Consolidation for Efficient and Effective Long-Running LLM Agents
PDF ↗Memory systems often organize user-agent interactions as retrievable external memory and are crucial for long-running agents by overcoming the limited context windows of LLMs. However, existing memory systems invoke LLMs to process every incoming interaction for memory extraction, and such an eager memory consolidation scheme leads to substantial token consumption. To tackle this problem, we propose RecMem by rethinking when memory consolidation should be conducted. RecMem stores incoming interactions in a subconscious memory layer and encode them using lightweight embedding models for retrieval. LLMs are only invoked to extract episodic and semantic memory when sustained recurrence are observed for semantically similar interactions. Such recurrence-based consolidation works because these interactions correspond to a semantic cluster with rich information and thus are worth extraction and summarization. To improve accuracy, RecMem also incorporates a semantic refinement mechanism that recovers the fine-grained facts omitted by memory extraction. Experiments show that RecMem reduces the memory construction token cost of three SOTA memory systems by up to 87% while exceeding their accuracy.
Large language models (LLMs) often encode word-form variation (e.g., *walk* vs. *walk**ed***) as linear directions in the embedding space. However, standard tokenization algorithms treat such variants as distinct words with different vocabulary entries—quickly filling the size-capped token vocabulary with surface-form variation (e.g., *walk*, *walk**ing***, ***W**alk*), at the expense of diversity and multilingual coverage. We show that many of these variations can be captured by *transformation* vectors—additive offsets that yield the appropriate word representation when applied to a *base form* embedding, in both the input and output spaces. Building on this, we propose a compact reshaping of the vocabulary: instead of assigning unique tokens to each surface form, we compose them from shared *base form* and *transformation* vectors (e.g., *walked* is *walk*+*past tense*). Our approach is lightweight—keeping the pretrained backbone frozen and only training small adaptation modules. We apply it across five languages and multiple LLMs in both pretraining and post-hoc adaptation, freeing 10-40% of vocabulary slots to be reallocated where tokenization is inefficient. Importantly, we do so while also expanding vocabulary coverage to out-of-vocabulary words, and with minimal impact on downstream performance. Our findings motivate a rethinking of vocabulary design, towards a representation that better matches the underlying structure of language and the practical needs of multilingual coverage.