论文检索

Saaduddin Mahmud, Mason Nakamura, Kyle Hollins Wray, Shlomo Zilberstein

Prompt optimization methods have demonstrated significant effectiveness in aligning black-box large language models (LLMs). In parallel, inference scaling strategies such as Best-of-N Sampling and Majority Voting have likewise been shown to improve alignment and performance by trading additional computation for better output. However, existing prompt optimization approaches are inference strategy agnostic; that is, they optimize prompts without accounting for the inference strategy. This constitutes a significant methodological gap, as our empirical and theoretical analysis reveals a strong interdependence between these two paradigms. Moreover, we find that user preferences regarding trade-offs among multiple objectives and inference budgets substantially influence the choice of prompt and inference configuration. To address this gap, we introduce a novel unified framework named IAPO (Inference-Aware Prompt Optimization) that jointly optimizes the prompt and inference scale, while being aware of the inference budget and different task objectives. We then develop a fixed-budget training algorithm for IAPO, called PSST (Prompt Scaling via Sequential Trimming), and establish finite-budget guarantees on the error probability. Finally, we evaluate the effectiveness of PSST on six tasks, including multi-objective text generation and reasoning, and demonstrate the critical role of incorporating inference-awareness in aligning black-box LLMs using prompt optimization.

Yunlong Ma, Bo Wang, Yihong Tang, Zifei Yu, Chenyun Xue, Gaoke Zhang, Yuexian Hou

The interpretative efficacy of large language models (LLMs) fundamentally hinges on the intricate alignment between user inputs and model-specific linguistic priors. Existing methodologies predominantly employ static input optimization strategies, failing to account for the empirically observed divergence in linguistic preference spaces across distinct LLM architectures, including variations in syntactic parsing heuristics, semantic grounding mechanisms, and knowledge retrieval pathways. We propose QueryAligner, an adaptive rewriting system implementing dynamic model-aware input transformation through architecture-specific preference modeling. Our framework introduces two pivotal innovations: 1) A dual-phase optimization engine integrating supervised learning on reverse-engineered cross-architectural training data with reinforcement learning driven by multi-objective reward signals, ensuring simultaneous preservation of semantic integrity and maximization of target model compatibility; 2) An architecture-informed rewriting protocol that automatically discovers latent alignment patterns encoded within distinct LLMs' parametric configurations. Experimental results demonstrate that our method achieves superior performance compared to conventional input optimization techniques.

Shichao Ma, Yunhe Guo, Jiahao Su, Qihe Huang, Zhengyang Zhou, Yang Wang

Text-to-image generation tasks have driven remarkable advances in diverse media applications, yet most focus on single-turn scenarios and struggle with iterative, multi-turn creative tasks. Recent dialogue-based systems attempt to bridge this gap, but their single-agent, sequential paradigm often causes intention drift and incoherent edits. To address these limitations, we present Talk2Image, a novel multi-agent system for interactive image generation and editing in multi-turn dialogue scenarios. Our approach integrates three key components: intention parsing from dialogue history, task decomposition and collaborative execution across specialized agents, and feedback-driven refinement based on a multi-view evaluation mechanism. Talk2Image enables step-by-step alignment with user intention and consistent image editing. Experiments demonstrate that Talk2Image outperforms existing baselines in controllability, coherence, and user satisfaction across iterative image generation and editing tasks.

Jingyuan Ma, Damai Dai, Zihang Yuan, Rui Li, Weilin Luo, Bin Wang, Qun Liu, Lei Sha, Zhifang Sui

Large Language Models (LLMs) have shown remarkable success on a wide range of math and reasoning benchmarks. However, we observe that they often struggle when faced with unreasonable math problems. Instead of recognizing these issues, models frequently proceed as if the problem is well-posed, producing incorrect answers or falling into overthinking and verbose self-correction. To systematically investigate this overlooked vulnerability, we propose the Unreasonable Math Problems (UMP) benchmark, designed to evaluate LLMs' ability to detect and respond to unreasonable math problem statements. Based on extensive experiments covering 19 LLMs, we find that even state-of-the-art general models like GPT-4o struggle on UMP. While reasoning models such as DeepSeek-R1 demonstrate a higher sensitivity to unreasonable inputs, this often comes at the cost of generating overly long and meaningless responses that fail to converge. We further find that prompting and fine-tuning enhance the detection of unreasonable inputs, with minor and acceptable trade-offs, making them practical solutions in this challenging setting.

Huipeng Ma, Luan Zhang, Dandan Song, Linmei Hu, Yuhang Tian, Jun Yang, Changzhi Zhou, Chenhao Li, Yizhou Jin, Xudong Li 等

In multi-hop reasoning, multi-round retrieval-augmented generation (RAG) methods typically rely on LLM-generated content as the retrieval query. However, these approaches are inherently vulnerable to knowledge overshadowing—a phenomenon where critical information is overshadowed during generation. As a result, the LLM-generated content may be incomplete or inaccurate, leading to irrelevant retrieval and causing error accumulation during the iteration process. To address this challenge, we propose ActiShade, which detects and activates overshadowed knowledge to guide large language models(LLMs) in multi-hop reasoning. Specifically, ActiShade iteratively detects the overshadowed keyphrase in the given query, retrieves documents relevant to both the query and the overshadowed keyphrase, and generates a new query based on the retrieved documents to guide the next-round iteration. By supplementing the overshadowed knowledge during the formulation of next-round queries while minimizing the introduction of irrelevant noise, ActiShade reduces the error accumulation caused by knowledge overshadowing. Extensive experiments show that ActiShade outperforms existing methods across multiple datasets and LLMs.

Fengji Ma, Chenxing Li, Li Liu

Automatic Cued Speech Recognition (ACSR) is a vital communication system designed to enhance spoken language accessibility for the hearing-impaired by combining lip movements and hand gestures to encode phonemes. Despite its effectiveness, current ACSR methods face significant challenges, including poor generalization to unseen cuers due to the limited scale of CS datasets, which restricts the ability of existing visual encoder to capture cuer-invariant CS visual features. Additionally, previous approaches relying on Connectionist Temporal Classification (CTC) decoding fail to incorporate prior linguistic sequence knowledge, further limiting their performance. To address these issues, we propose a novel Two Auxiliary Modalities guided Cross-cuer Invariant Adaptation method (TACIA), introducing pose and text modalities to help extract cuer-invariant motion and semantic features, thereby improving generalization. In addition, we introduce a Visual-guided Cued Token Prediction (VG-NTP) method, inspired by large language models. This method replaces CTC decoding by incorporating language modeling, leveraging rich linguistic knowledge, including semantics, to address the suboptimal issues present in the CTC decoding process. Extensive experiments demonstrate the superiority of our approach to the state-of-the-art (SOTA) on Chinese and British CS datasets, significantly advancing the accuracy and quality of ACSR systems.

Kailun Lyu, Zehan Li, Fu Zhang, Jingwei Cheng

Zero-Shot Relation Triplet Extraction (ZSRTE) aims to extract head-tail entity pairs and their corresponding relations from sentences, where the relations available during inference are not seen during training. Existing methods typically assume that entities are continuous; however, in practice, entities can be discontinuous, which poses challenges to these approaches. To address this issue, we are the first to discuss and study the ZSRTE task involving discontinuous entities, and propose an innovative BoG framework, which is based on our proposed Boundary Token Graph structure. This method first predicts and adds edges between boundary tokens of (dis)continuous entities to construct a token graph, and then innovatively transforms the relation triplet extraction task into a process of finding paths in the graph. Additionally, we design a Boundary Token-Aware Prompt for each relation to further enhance the interaction between boundary tokens and relation semantics. Experimental results on four ZSRTE datasets—with or without discontinuous entities—consistently demonstrate that our method outperforms previous approaches, achieving state-of-the-art results.

Zhao Lv, Haoran Zhou, Ying Chen, Youdian Gao, Xinhui Li, Ruibo Fu, Cunhang Fan

Brain-assisted target speaker extraction (TSE) isolates a target speaker's voice from a mixture by leveraging task-specific representations in Electroencephalogram (EEG) signals. However, existing methods rely on fixed interpolation for EEG-audio alignment, introducing redundant computations. They also employ single-path encoders that extract only target-relevant features while neglecting complementary, irrelevant ones, limiting discriminability. To address these limitations, this paper proposes a Trainable EEG Interpolation and Structure-sharing Dual-path Encoders network (TIDENet). The proposed Trainable EEG Interpolation (TEI) uses a neural network module to leverage cross-sample EEG information during resampling by parameters updating, thereby overcoming the limitations of fixed interpolation. The Structure-sharing Dual-path Encoders (SSDPE) extend existing speech and EEG encoders by introducing dual paths that separately process features relevant and irrelevant to the target speaker and incorporates interactive fusion between them, which enhances the encoder's ability to capture task-relevant information. Experimental results on public datasets demonstrate that TIDENet achieves relative improvements of up to 20.47%, 22.22%, 2.91%, 6.20%, and 15.84% in signal-to-distortion ratio (SDR), scale-invariant SDR (SI-SDR), short-time objective intelligibility (STOI), extended STOI (ESTOI), and perceptual evaluation of speech quality (PESQ), respectively, compared to the state-of-the-art. These significant gains validate the effectiveness of the proposed TEI method and SSDPE architecture.

Fangrui Lv, Lei Wang, Ruixin Hong, Yong Du, Xiangyu Wu, Tingting Gao, Guorui Zhou, Changshui Zhang

Chain-of-Thought prompting has remarkably advanced LLM reasoning by generating explicit step-by-step tokens, yet its discrete nature inherently limits expressiveness and efficiency, struggling with abstract, ambiguous, or semantically divergent cognition beyond linguistic tokens. Latent reasoning offers a promising alternative by operating in the model’s internal continuous space for richer cognitive representations. However, existing methods typically rely on finetuning or token interpolation to bridge latent and input spaces, introducing training difficulty or semantic degradation. To this end, we propose Dynamic Latent Reasoning (DyLaR), a training-free framework that preserves semantic fidelity to latent space. DyLaR introduces a Semantic Residual Refinement module that progressively refines latent inputs by integrating semantic residuals from prior hidden states, thus capturing expressive semantic hierarchies that closely approximate continuous latent representations. To enhance flexibility, DyLaR further incorporates a dynamic switching policy that allows LLMs to alternate between discrete and latent reasoning based on model uncertainty, favoring explicit reasoning when confident and latent exploration under ambiguity. Empirical experiments across knowledge- and reasoning-intensive tasks demonstrate that DyLaR consistently outperforms strong baselines in both effectiveness and token efficiency. Qualitative analyses further illustrate its interpretability and flexibility in navigating complex reasoning scenarios.

Xiaonan Luo, Yue Huang, Ping He, Xiangliang Zhang

High‑quality Question–Answer (QA) datasets are foundational for reliable Large Language Model (LLM) evaluation, yet even expert‑crafted datasets exhibit persistent gaps in domain coverage, misaligned difficulty distributions, and factual inconsistencies. The recent surge in generative model-powered datasets has compounded these quality challenges. In this work, we introduce RefineLab, the first LLM‑driven framework that automatically refines raw QA textual data into high-quality datasets under a controllable token‑budget constraint. RefineLab takes a set of target quality attributes as refinement objectives and performs selective edits within a predefined token budget to ensure practicality and efficiency. In essence, RefineLab addresses a constrained optimization problem: improving the quality of QA samples as much as possible while respecting resource limitations. With a set of available refinement operations, RefineLab takes as input the original dataset, a specified set of target quality dimensions, and a token budget, and determines which refinement operations should be applied to each QA sample. This process is guided by an assignment module that selects optimal refinement strategies to maximize overall dataset quality while adhering to the budget constraint. Experiments demonstrate that RefineLab consistently narrows divergence from expert datasets across coverage, difficulty alignment, factual fidelity, and distractor quality. RefineLab pioneers a scalable, customizable path to reproducible dataset design, with broad implications for LLM evaluation.

Kun Luo, Zheng Liu, Shitao Xiao, Jiabei Chen, Hongjin Qian, Peitian Zhang, Shanshan Jiang, Bin Dong, Jun Zhao, Kang Liu

Long-context processing remains a significant challenge for large language models (LLMs). Retrieval-augmented generation (RAG) has recently emerged as a promising approach, enabling LLMs to selectively access relevant information from extended contexts to improve efficiency. However, existing RAG approaches often lag behind other efficient long-context processing methods primarily due to inherent limitations on inaccurate retrieval and fragmented contexts. To address these limitations, we propose RetroLM, a novel RAG framework designed for effective long-context processing. Unlike traditional approaches, RetroLM introduces KV-level retrieval augmentation, which partitions the LLM's KV cache into contiguous pages and performs encoding and decoding operations based on the retrieved KV pages. Built upon this framework, we further develop a specialized retriever for precise retrieval of critical pages and conduct unsupervised post-training to optimize the model’s ability to leverage retrieved information. Compared with traditional RAG, the new approach enhances robustness to retrieval inaccuracy, facilitates effective utilization of fragmented contexts, and saves the cost from repeated context-encoding operations. We conduct extensive evaluations across several popular benchmarks, including LongBench, InfiniteBench, and RULER. RetroLM consistently outperforms existing long-LLMs and RAG-based methods, especially in tasks requiring deep reasoning or extreme context lengths.

Haitong Luo, Weiyao Zhang, Suhang Wang, Wenji Zou, Chungang Lin, Xuying Meng, Yujun Zhang

The proliferation of high-quality text from Large Language Models (LLMs) demands reliable and efficient detection methods. While existing training-free approaches show promise, they often rely on surface-level statistics and overlook fundamental signal properties of the text generation process. In this work, we reframe detection as a signal processing problem, introducing a novel paradigm that analyzes the sequence of token log-probabilities in the frequency domain. By systematically analyzing the signal's spectral properties using the global Discrete Fourier Transform (DFT) and the local Short-Time Fourier Transform (STFT), we find that human-written text consistently exhibits significantly higher spectral energy. This higher energy reflects the larger-amplitude fluctuations inherent in human writing compared to the suppressed dynamics of LLM-generated text. Based on this key insight, we construct SpecDetect, a detector built on a single, robust feature from the global DFT: DFT total energy. We also propose an enhanced version, SpecDetect++, which incorporates a sampling discrepancy mechanism to further boost robustness. Extensive experiments show that our approach outperforms the state-of-the-art model while running in nearly half the time. Our work introduces a new, efficient, and interpretable pathway for LLM-generated text detection, showing that classical signal processing techniques offer a surprisingly powerful solution to this modern challenge.

Yujing Lu, Ling Zhong, Jing Yang, Weiming Li, Peng Wei, Yongheng Wang, Manni Duan, Qing Zhang

Chart Question Answering (CQA) evaluates Multimodal Large Language Models (MLLMs) on visual understanding and reasoning over chart data. However, existing benchmarks mostly test surface-level parsing, such as reading labels and legends, while overlooking deeper scientific reasoning. We propose DomainCQA, a framework for constructing domain-specific CQA benchmarks that emphasize both visual comprehension and knowledge-intensive reasoning. It integrates complexity-aware chart selection, multitier QA generation, and expert validation. Applied to astronomy, DomainCQA yields AstroChart, a benchmark of 1,690 QA pairs over 482 charts, exposing persistent weaknesses in fine-grained perception, numerical reasoning, and domain knowledge integration across 21 MLLMs. Fine-tuning on AstroChart improves performance across fundamental and advanced tasks. Pilot QA sets in biochemistry, economics, medicine, and social science further demonstrate DomainCQA’s generality. Together, our results establish DomainCQA as a unified pipeline for constructing and augmenting domain-specific chart reasoning benchmarks.

Wayne Lu, Xiaoxi Cui

Real-world text classification datasets frequently exhibit long-tail distributions, where numerous classes have sparse data, significantly degrading model performance on these underrepresented categories. While Large Language Models (LLMs) offer promise for data augmentation, existing methods often produce semantically limited samples, neglect "implicit long-tails" (sparse sub-patterns within classes), and lack cost-effective optimization. To address these challenges, we propose \textbf{DEALT (LLM-driven Diversity-Enhanced Data Augmentation for Long-Tail Text Classification)}, a novel cognitive-inspired framework emulating the human learning process of "recognize, explore, generate, and optimize." DEALT systematically enhances augmented data diversity by first detecting both explicit and implicit long-tails. It then employs an LLM for diversity-aware planning of augmentation strategies, followed by conditional generation. A low-overhead quality and diversity validator filters the synthetic data, and an adaptive incremental sampler refines future augmentation efforts based on proxy model feedback, ensuring efficient and budget-aware optimization. Extensive experiments on multiple public text classification datasets demonstrate DEALT's superiority over state-of-the-art methods in improving tail-class performance and overall model robustness by generating more diverse and high-fidelity augmented data.

Songshuo Lu, Hua Wang, Zhi Chen, Yaohua Tang

Large-scale alignment pipelines typically pair a policy model with a separately trained reward model whose parameters remain frozen during reinforcement learning (RL). This separation creates a complex, resource-intensive pipeline and leads to a performance ceiling. We propose a novel framework, Unified Reward & Policy Optimization (URPO), that unifies instruction-following (“player”) and reward modeling (“referee”) into a single model and a single training phase. Our method recasts all alignment data-including preference pairs, verifiable reasoning, and open-ended instructions-into a unified generative format optimized by a single Group-Relative Policy Optimization (GRPO) loop. This enables the model to learn from ground-truth preferences and verifiable logic while simultaneously generating its own rewards for open-ended tasks. Experiments on the Qwen2.5-7B model demonstrate that URPO significantly outperforms a strong baseline using a separate generative reward model, boosting the instructionfollowing score on AlpacaEval to 44.84 and achieving a 36% relative improvement on the challenging AIME reasoning benchmark. Furthermore, URPO cultivates a superior internal evaluator as a byproduct of training, achieving a RewardBench score of 85.15 and surpassing the dedicated reward model it replaces (83.55). By eliminating the need for a separate reward model and fostering a co-evolutionary dynamic, URPO presents a simpler, more efficient, and more effective path towards robustly aligned language models.

Hao Lu, Yanchi Gu, Haoyuan Huang, Yulin Zhou, Ningxin Zhu, Chen Li

The integration of Monte Carlo Tree Search (MCTS) with Large Language Models (LLMs) has demonstrated significant success in structured, problem-oriented tasks. However, applying these methods to open-ended dialogues, such as those in psychological counseling, presents unique challenges. Unlike tasks with objective correctness, success in therapeutic conversations depends on subjective factors like empathetic engagement, ethical adherence, and alignment with human preferences, for which strict correctness criteria are ill-defined. Existing result-oriented MCTS approaches can therefore produce misaligned responses. To address this, we introduce MCTSr-Zero, an MCTS framework designed for open-ended, human-centric dialogues. Its core innovation is domain alignment, which shifts the MCTS search objective from predefined end-states towards conversational trajectories that conform to target domain principles (e.g., empathy in counseling). Furthermore, MCTSr-Zero incorporates Regeneration and Meta-Prompt Adaptation mechanisms to substantially broaden exploration by allowing the MCTS to consider fundamentally different initial dialogue strategies. We evaluate MCTSr-Zero in psychological counseling by generating multi-turn dialogue data, which is used to fine-tune an LLM, PsyLLM. We also introduce PsyEval, a benchmark for assessing multi-turn psychological counseling dialogues. Experiments demonstrate that PsyLLM achieves state-of-the-art performance on PsyEval and other relevant metrics, validating MCTSr-Zero's effectiveness in generating high-quality, principle-aligned conversational data for human-centric domains and addressing the LLM challenge of consistently adhering to complex psychological standards.

Haowei Lou, Hye-young Paik, Wen Hu, Lina Yao

Learning representative embeddings for different types of speaking styles, such as emotion, age, and gender, is critical for both recognition tasks (e.g., cognitive computing and human-computer interaction) and generative tasks (e.g., style-controllable speech generation). In this work, we introduce ParaMETA, a unified and flexible framework for learning and controlling speaking styles directly from speech. Unlike existing methods that rely on single-task models or cross-modal alignment, ParaMETA learns disentangled, task-specific embeddings by projecting speech into dedicated subspaces for each style type. This design reduces inter-task interference, mitigates negative transfer, and allows a single model to handle multiple paralinguistic tasks such as emotion, gender, age, and nationality classification. Beyond recognition, ParaMETA enables fine-grained style control in Text-To-Speech (TTS) generative models. It supports both speech- and text-based prompting and allows users to modify one speaking style while preserving others. Extensive experiments demonstrate that ParaMETA outperforms strong baselines in classification accuracy and generates more natural and expressive speech, while maintaining a lightweight and efficient model suitable for real-world applications.

Fanghao Lou, Qiqi Wang, Guanyu Chen, Kaiqi Zhao, Huijia Li

Large Language Models (LLMs) are widely used in legal judgment prediction tasks, which aim to enhance judicial efficiency. However, the length of legal fact descriptions poses a significant challenge to the application of LLMs. Long inputs not only introduce noise, affecting output quality, but also increase processing time. While existing text compression methods, such as generating summaries or training models to implicitly reduce text dimensionality, can shorten input length, they often face the slow generation speeds and limited interpretability issues. To address these issues and inspired by information bottleneck-based text compression, we propose the Zipped Information Processor for Legal Judgment Prediction method, ZipLJP. By effectively integrating legal knowledge into the compression process, ZipLJP not only reduces input length but also improves processing efficiency and prediction quality. Experiments show that our approach achieves better performance compared to the previous methods on two widely used open-source and real-world datasets.

Lintao Long, Ruizhang Huang, Ruina Bai, Yongbin Qin, Qihang Fu

Chinese Grammar Error Correction (CGEC) aims to identify and correct grammatical errors in Chinese sentences. Fine-tuning Large Language Models (LLMs) is a popular current method. However, we have observed a significant flaw: LLMs learn grammatical knowledge but often fail to explicitly use specific grammatical concepts to correct erroneous sentences, leading to multiple corrections without a clear indication of which is the most reliable. Humans possess an "intuitive thinking" mode, which allows them to quickly decide which correction is more reliable based on experience and intuition. To address this deficiency in LLMs, we propose the Expanding Intuitive Thinking Model (ExIT). ExIT extends the thinking process of LLMs for CGEC, providing them with a human-like rapid decision-making process. This enables LLMs to quickly select a more reliable correction from multiple alternatives based on experience and intuition. Unlike the LLM decoding process, which focuses only on the trustworthiness of local tokens, this is a global thinking process concerning the erroneous sentence and its correction. ExIT is a lightweight model that performs rapid computations without significantly increasing overhead. Our experimental results on CGEC datasets demonstrate that the proposed ExIT can substantially unleash the error correction potential of LLMs.

Lingkun Long, Rubing Yang, Yushi Huang, Desheng Hui, Ao Zhou, Jianlei Yang

Long-context inference for Large Language Models (LLMs) is heavily limited by high computational demands. While several existing methods optimize attention computation, they still process the full set of hidden states at each layer, limiting overall efficiency. In this work, we propose SlimInfer, an innovative framework that aims to accelerate inference by directly pruning less critical prompt tokens during the forward pass. Our key insight is an information diffusion phenomenon: As information from critical tokens propagates through layers, it becomes distributed across the entire sequence. This diffusion process suggests that LLMs can maintain their semantic integrity when excessive tokens, even including these critical ones, are pruned in hidden states. Motivated by this, SlimInfer introduces a dynamic fine-grained pruning mechanism that accurately removes redundant tokens of hidden state at intermediate layers. This layer-wise pruning naturally enables an asynchronous KV cache manager that prefetches required token blocks without complex predictors, reducing both memory usage and I/O costs. Extensive experiments show that SlimInfer can achieve up to 2.53× time-to-first-token (TTFT) speedup and 1.88× end-to-end latency reduction for LLaMA3.1-8B-Instruct on a single RTX 4090, without sacrificing performance on LongBench.