| Total: 51
We investigate the interaction between quantization and pruning in the compression of open-weight small language models (SLMs), as these techniques are frequently combined in practice without a clear understanding of whether their effects are antagonistic, synergistic or additive. Prior work has reported ordering effects and studied pruning or quantization in isolation, but rarely quantifies their interaction across models and evaluation metrics. We introduce an interaction coefficient that isolates non-additive effects and apply it to Falcon3-1B-Base and LLaMA-3.2-1B, with a limited study on Qwen2.5-1.5B. Our experiments cover four hybrid application orders, multiple pruning sparsities and 4/8-bit quantization. We find non-monotonic interactions: memory savings are dominated by 4-bit quantization, while pruning alone provides limited benefits. Hybrids exhibit the greatest effects at global unstructured pruning sparsity s ∈ {0.2, 0.3} while getting close to additivity at s=0.5. At low sparsity, truthfulness improves, while reasoning losses diminish at higher sparsity. Future work will extend interaction analysis to hardware-aware sparse kernels and larger models.
Subword tokenization is a standard technique for pre-trained language models, mapping text into sequences of tokens from a fixed-size vocabulary. Despite its widespread use, the impact of tokenization algorithms and vocabulary sizes on downstream performance remains underexplored, particularly for low- and medium-resource languages, where suboptimal configurations are difficult to compensate with additional data. Prior studies have mainly focused on high-resource languages, individual tokenization algorithms, or fixed vocabulary sizes, limiting their scope. In this paper, we present a systematic evaluation of subword tokenization strategies in six diverse low- and medium-resource languages: Icelandic, Estonian, Basque, Galician, Nepali, and Tajik. We pre-trained monolingual TEAMS-Small models using WordPiece, BPE, and Unigram tokenizers with vocabulary sizes of 16k, 32k, and 64k, and additionally compared byte-level and character-level tokenization for Icelandic. Evaluation on a benchmark of NLP tasks revealed that Unigram tokenization with a 64k vocabulary consistently outperformed other configurations, with vocabulary size having a greater impact on downstream performance than algorithm choice. These gains were task-dependent, with statistically significant improvements for part-of-speech tagging, named entity recognition, and question answering, but not for dependency parsing and summarization. Finally, byte-level tokenization provided no measurable advantage in a monolingual setting, suggesting that its benefits are primarily relevant for multilingual models.
Large Reasoning Models (LRMs) often suffer from overthinking due to redundant verification steps. Existing approaches for mitigating overthinking, such as fast-slow thinking switching and reasoning trajectory compression, fail to make a fine-grained distinction between beneficial and redundant steps within the LRM's reasoning process, and may thus impair reasoning capability in their pursuit of efficiency. To simultaneously improve reasoning efficiency and capability, we propose EvoThink, a framework that reduces redundant verification and encourages the exploration of new reasoning paths. EvoThink comprises two key components: Self-Pruning Training (SPT), an unsupervised method that iteratively prunes redundant reasoning steps and self-trains on the concise trajectories; and Aha-Moment Preference Optimization (AMPO), which, inspired by genetic algorithms, identifies valuable failed reasoning attempts, synthesizes from-wrong-to-right aha-moment data, and optimizes the model to internalize this reasoning pattern. Extensive evaluations across mathematical reasoning and code generation benchmarks demonstrate that EvoThink not only substantially reduces inference-time token usage but also improves the reasoning capability of LRMs.
While large-scale reasoning models have achieved remarkable performance gains by scaling test-time computation through extended Chain-of-Thought (CoT) sequences, their practical utility is severely constrained by protracted inference lengths and high computational latency. In this paper, we present Differentiable Hierarchical Fusion (DHF), a novel framework that merges reasoning models with efficient base models via differentiable optimization to produce concise, accurate outputs. We introduce a dual-factor adaptive weighting mechanism to capture intra-block (Attention vs. MLP) variance and inter-block (shallow vs. deep layers) importance hierarchies, thereby addressing key limitations of static merging heuristics. Specifically, DHF optimizes fusion coefficients using gradient descent on a loss function that jointly minimizes cross-entropy (for accuracy) and response length (for conciseness). Furthermore, we enhance an outlier-aware initialization strategy to seed coefficients based on activation density and construct multi-calibration datasets to improve the model’s generalization ability. Comprehensive evaluations on Qwen and LLaMA models across six reasoning benchmarks show that DHF reduces the average response length by 55% while boosting accuracy by 2.8%–6% compared to existing state-of-the-art merging methods.
Code-switching (CS) speech translation (ST) aims to translate speech that alternates between multiple languages into a target language text, posing significant challenges due to the complexity of semantic modeling and the scarcity of CS data. Previous studies mainly rely on the models themselves to implicitly learn semantic representations and resort to costly manual annotations. To mitigate these limitations, we propose enhancing Large Language Models (LLMs) with a Mixture-of-Experts (MoE) speech projector composed of language expert groups, where each group specializes in the semantic space of a specific language for fine-grained speech feature modeling. A language-specific loss and an intra-group load balancing loss are jointly introduced to guide efficient token routing across and within expert groups. Furthermore, we introduce a multi-stage training paradigm that utilizes readily available automatic speech recognition (ASR) and monolingual ST data, facilitating speech-text alignment and improving translation performance. To bridge the data gap for smooth domain transfer, a transition loss is employed to improve adaptation to CS scenarios. Extensive experiments on widely used datasets demonstrate the effectiveness and generality of our approach, achieving average improvements of 0.86 BLEU and 0.93 COMET over SeamlessM4T, with maximum improvements of 1.49 BLEU and 1.41 COMET across different test sets. Our code and supplementary appendices are available at https://github.com/XMUDeepLIT/CSST-SSA.
Affective computing has achieved notable success in recognizing explicit emotions from short, isolated dialogue segments. However, human emotions are often implicitly expressed, internally regulated, and dynamically evolve over extended interactions. Existing models struggle to disentangle internal emotional states from external expressions, and fail to capture the emotional inconsistency that emerges across long-horizon dialogues. To address this limitation, we introduce Emotional Inconsistency Analysis (EIA), a novel task that aims to identify and reason about discrepancies between implicit and explicit emotions over long-term conversational contexts. To support this task, we construct the MaskDialog dataset carefully curated from television drama and large language models (LLMs). We further propose two LLM-based baseline approaches, i.e., One-shot Self-consistent Inference and Cascaded Multi-step Inference, and conduct comprehensive analyses on dialogue construction strategies and inference behaviors. Extensive experiments across multiple mainstream LLMs reveal that EIA remains highly challenging, particularly in modeling implicit emotional trajectories and cross-turn inconsistency. Overall, EIA reframes emotion understanding from short-term recognition to longitudinal, implicit emotion tracking, with implications for dialogue systems and human–computer interaction.
Cross-task generalization (CTG) enables large language models (LLMs) to handle unseen tasks proficiently, enhancing their adaptability in real-world scenarios. However, existing methods relying on per-token dynamic routing to multiple trained LoRA adapters face high computational and GPU memory costs. Recent Representation Fine-Tuning (ReFT) enhances efficiency for single-task adaptation by editing only prefix and suffix token representations. However, the semantic ambiguity of tokens and absence of a self-guided mechanism for parameter selection in unseen tasks limits their application to CTG. To this end, we propose RaMod, a Representation-Aware Modularity framework to extend the ReFT paradigm to CTG through two novel components: (i) Dual-Modular Representation & Parameter Fine-tuning, which manipulates only a strategically chosen subset of hidden representations with modular interventions to guide the model toward solving unseen tasks; and (ii) Asynchronous Orchestrator, which proactively allocates and releases GPU memory for selected interventions, thereby minimizing storage overhead. Extensive experiments demonstrate that RaMod not only achieves superior CTG performance but also substantially reduces the overhead of the latest CTG baseline, achieving 83%, 100%, and 79% reduction in its additional prefill time, generation delays, and memory consumption relative to original LLMs.
In many large language model (LLM) applications, generating structurally diverse candidate outputs is crucial for downstream decision-making and reasoning. However, conventional beam search allocates search budget at the token-prefix level, so multiple beam slots may be occupied by lexically different but structurally similar hypotheses. To address these limitations, we propose Grammar-State Aware Beam Search (GSA-Beam). Our method first organizes search at the constraint state level rather than individual prefix level, enabling explicit control over structural diversity. It further incorporates a dynamic state-level beam regulation mechanism that adaptively adjusts beam allocation based on the branching of constraint states, efficiently exploring the feasible structured solution space. Experiments on JSON-schema constrained generation show that GSA-Beam improves explicit state coverage and final structural diversity while preserving schema validity; across several LLM backbones, it reduces reference latency by 22.5% on average compared with standard constrained beam search. Code is available at https://github.com/Paradozile/GSABeam.
Effective document-level machine translation (DocMT) requires capturing long-range discourse dependencies. Recent work has explored retrieval-based and discourse-aware context selection. However, these approaches often lack an explicit mechanism for modeling structured discourse dependencies between distant paragraphs in a document. In this paper, we propose G²C-MT (Graph-Guided Context for Machine Translation), which views DocMT context selection as a structured path discovery problem on a lightweight discourse graph, rather than retrieving unstructured context sets or relying on expensive LLM-based discourse modeling. In detail, we represent each paragraph as a node and model the relationship between each pair of nodes, considering their semantic similarity, adjacency, and keyword overlap. Furthermore, we propose a depth-biased random walk over the graph to sample a backward context path for each target paragraph. The context path will be used to prompt a large language model (LLM) for translation. This framework naturally supports multi-path context sampling, which can improve robustness by aggregating diverse translation candidates for discourse-ambiguous inputs. Experiments conducted across various domains show that G²C-MT outperforms strong baselines on multiple LLMs, including DeepSeek-V3, Gemini-2.5-Flash-lite, and the Qwen-2.5/3 series.
Improving the memory efficiency, throughput, and serving cost of large language models (LLMs) is critical for edge deployment, interactive applications, and sustainable inference. Pruning is a promising approach, but existing methods have limitations: width pruning disrupts the standard transformer architecture and requires custom inference code, while depth pruning causes abrupt accuracy drops. Moreover, many approaches that work well for LLMs fail to preserve performance on small language models (SLMs). We propose COMPACT, which jointly prunes (i) rare vocabulary to shrink embedding layers and (ii) FFN intermediate channels using common-token–weighted activations aligned with the post-pruning token distribution. COMPACT inherits strengths of both depth and width pruning, such as deployment-friendliness, scale-adaptivity, competitive pruning speed, and strong inference performance. Experiments on several LLM families (0.5B–70B) show state-of-the-art downstream performance, with substantial improvements in inference throughput and GPU memory. Project code is at https://github.com/ekwek1/COMPACT.
Preference alignment is a critical step in making Large Language Models (LLMs) useful and aligned with (human) preferences. Existing approaches such as Reinforcement Learning from Human Feedback or Direct Preference Optimization typically require curated data and expensive optimization over billions of parameters, and eventually lead to persistent task-specific models. In this work, we introduce Preference alignment of Large Language Models via Residual Steering (PaLRS), a training-free method that exploits preference signals encoded in the residual streams of LLMs. From as few as one hundred preference pairs, PaLRS extracts lightweight, plug-and-play steering vectors that can be applied at inference time to push models toward preferred behaviors. We evaluate PaLRS on various small-to-medium-scale open-source LLMs, showing that PaLRS-aligned models achieve consistent gains on mathematical reasoning and code generation benchmarks while preserving baseline general-purpose performance. Moreover, when compared to models aligned with DPO and SimPO, they perform better with great time savings. Our findings highlight that PaLRS offers an effective, much more efficient and flexible alternative to standard preference optimization pipelines, offering a training-free, plug-and-play mechanism for alignment with minimal data. Extended version with Suppl. Mat. is available at https://doi.org/10.48550/arXiv.2509.23982.
Retrieval‑augmented generation (RAG) improves knowledge‑intensive QA tasks by incorporating external evidence, yet retrieval remains imperfect and often returns irrelevant or misleading passages. We refer to this as Noisy RAG, where LLMs suffer from inconsistent relevance signals due to the fact that relevance identification is carried out by only a small subset of internal modules (e.g., retrieval heads) and can easily be confounded by other modules. Recent work attempts to enhance the denoising ability of the LLM through external emphasis, prompting the LLM to highlight helpful passages as the emphasis before answering. However, these textual identifications themselves depend on the same inconsistent signals and thus become less reliable. Our key insight is that, because relevance signals reside in only a few modules, emphasis should be applied directly within these modules rather than only at the text level. We therefore propose \textit{selective internal emphasis} to amplify the relevance signals, implemented via a lightweight, plug‑and‑play signal amplifier that operates inside the LLM. The amplifier performs token‑ and channel‑level selective emphasis during a standard RAG fine‑tuning pipeline, with the base LLM frozen. Across four QA benchmarks and three LLM scales, our method consistently improves accuracy and robustness. Furthermore, generalization and interpretability analyses show that the amplifier captures retrieval‑related patterns rather than only dataset‑specific patterns, enabling more reliable passage-denoising.
Target speaker extraction (TSE) aims to extract the speech of a target speaker from mixtures containing multiple competing speakers. Conventional TSE systems predominantly rely on speaker cues, such as pre-enrolled speech, to identify and isolate the target speaker. However, in many practical scenarios, clean enrollment utterances are unavailable, limiting the applicability of existing approaches. In this work, we propose DAE-TSE, a keyword-guided TSE framework that specifies the target speaker through distinct keywords they utter. By leveraging keywords (i.e., partial transcriptions) as cues, our approach provides a flexible and practical alternative to enrollment-based TSE. DAE-TSE follows the Detect-Attend-Extract (DAE) paradigm: it first detects the presence of the given keywords, then attends to the corresponding speaker based on the keyword content, and finally extracts the target speech. Experimental results demonstrate that DAE-TSE outperforms standard TSE systems that rely on clean enrollment speech. To the best of our knowledge, this is the first study to utilize partial transcription as a cue for specifying the target speaker in TSE, offering a flexible and practical solution for real-world scenarios. Our code (https://github.com/GnafiY/DAE-TSE) and demo page (https://gnafiy.github.io/DAE-TSE_demo) are now publicly available.
Recent reasoning-enhanced Large Language Models (LLMs) have achieved promising results in solving complex competitive programming problems. However, it remains unclear whether these reasoning abilities generalize to relevant tasks, like identifying algorithmically similar problems (ASPs). We introduce AlgoSimBench, a benchmark of 402 multiple-choice questions curated in an adversarial setting: each given reference problem is paired with one algorithmically similar problem and three distractors that are semantically close but algorithmically dissimilar. This design forces models to rely on algorithmic reasoning rather than superficial textual cues. Our evaluation shows that LLMs consistently struggle under this setting. To address this gap, we propose Attempted Solution Matching (ASM), which leverages LLM-generated solution attempts to assess similarity, yielding an average accuracy improvement of 9% across models. Beyond LLM evaluation, AlgoSimBench also probes code retrieval methods; when combined with BM25, ASM achieves an additional 11.8% gain over state-of-the-art embedding models. AlgoSimBench offers a challenging testbed that facilitates future studies on LLMs and retrieval methods.
Large language models often face challenges in low-resource agglutinative language text summarization tasks due to poorly designed prompts, leading to core information dilution, reduced fidelity, and critical information loss caused by the complex grammatical structures of agglutinative languages. For traditional Mongolian, a typical low-resource agglutinative language, this paper proposes a Nucleus-to-Periphery Retrieval-Iterative Prompting (NPRIP). This method first guides the model to extract highly condensed semantic core information (events, persons, time, etc.) from the original text. Subsequently, through multiple rounds of self-refinement iteration, it progressively expands peripheral details (background, causes, consequences, secondary facts, etc.). The model performs fact consistency checks, redundancy removal, and fidelity correction on the current draft, achieving gradual improvements in information completeness and fidelity. To enhance Mongolian language representation, we perform parameter-efficient fine-tuning on LLaMA3-8B using the CCMT2019 Mongolian-Chinese parallel corpus, and construct a larger abstractive Mongolian news summarization dataset MoSum along with its augmented version. Experiments on traditional Mongolian text summarization tasks demonstrate that our proposed method significantly outperforms multiple baseline models on automatic evaluation metrics including ROUGE-1, ROUGE-2, and ROUGE-L. This validates the effectiveness of core-priority structured iterative prompting in low-resource agglutinative language summarization scenarios.
Conversational Aspect-based Sentiment Quadruple Analysis (DiaASQ) needs to capture the complex interrelationships in multiple rounds of dialogues. Existing methods usually employ simple Graph Convolutional Networks (GCN), which introduce structural noise and fail to consider the temporal sequence of the dialogues, or use standard RoPE, which implicitly captures relative distances in a flat sequence but cannot clearly separate the token-level syntactic order from the utterance-level progression, and may suffer from the Distance Dilution problem. To address these issues, we propose a new framework that combines Thread-Constrained Directed Acyclic Graph (TC-DAG) and Discourse-Aware Rotary Position Embedding (D-RoPE). Specifically, TC-DAG filters out cross-thread noise based on thread constraints, maintains global connectivity through root anchoring, and incorporates the temporal sequence of the dialogues. D-RoPE aligns multi-layer semantics using dual-stream projection and multi-scale frequency signals, captures thread dependencies using tree-like distances, and alleviates the token-level Distance Dilution problem by incorporating utterance-level progressions. Experimental results on two benchmark datasets demonstrate that our framework achieves state-of-the-art performance.
Detecting text generated by large language models~(LLMs) is crucial but challenging. Existing detectors depend on impractical assumptions, such as white-box settings, or solely rely on text-level features, leading to imprecise detection ability. In this paper, we propose a simple but effective and practical LLM-generated text detection method, VaryBalance. The core of VaryBalance is that, compared to LLM-generated texts, there is a greater difference between human texts and their rewritten version via LLMs. Leveraging this observation, VaryBalance quantifies this through mean standard deviation and distinguishes human texts and LLM-generated texts. Comprehensive experiments demonstrated that VaryBalance outperforms the state-of-the-art detectors, i.e., Binoculars, by up to 34.3% in terms of AUROC, and maintains robustness against multiple generating models and languages.
Generating executable tool plans requires selecting appropriate subsets from tool libraries, a combinatorial search problem with an exponentially large solution space. However, we identify a critical misalignment in predominant approaches: standard autoregressive (AR) decoding suffers from early commitment, where initial token choices rigidly constrain the search trajectory. A controlled study shows that masked denoising raises Pass@10 solution coverage from 0.320 to 0.943 over AR sampling under matched compute. Motivated by this, we propose DiG-Plan, a framework that decouples combinatorial exploration from structural refinement. DiG-Plan employs a diffusion-based proposer to generate diverse tool sets via iterative refinement, followed by an AR refiner for dependency prediction. On TaskBench, DiG-Plan improves over AR baselines by a 10% relative margin, with the largest gains on complex compositional tasks; API-Bank results show that the propose-refine-select design remains effective across domains. Code is available at https://github.com/puddingyeah/DiG-Plan.
Evaluating LLM forecasting capabilities is constrained by a fundamental tension: prospective evaluation offers methodological rigor but prohibitive latency, while retrospective forecasting (RF)—evaluating on already-resolved events—faces rapidly shrinking clean evaluation data as SOTA models possess increasingly recent knowledge cutoffs. Simulated Ignorance (SI), prompting models to suppress pre-cutoff knowledge, has emerged as a potential solution. We provide the first systematic test of whether SI can approximate True Ignorance (TI). Across 470 competition-level questions and 9 models, we find that SI fails systematically: (1) cutoff instructions leave a 52% performance gap between SI and TI; (2) chain-of-thought reasoning fails to suppress prior knowledge, even when reasoning traces contain no explicit post-cutoff references; (3) reasoning-optimized models exhibit worse SI fidelity despite superior reasoning trace quality. These findings demonstrate that prompts cannot reliably "rewind" model knowledge. We conclude that RF on pre-cutoff events is methodologically flawed; we recommend against using SI-based retrospective setups to benchmark forecasting capabilities.
Zero-Shot Relation Triplet Extraction (ZSRTE) aims to extract relation triplets for unseen relation types without any annotated training data. Recent advancements in Large Language Models (LLMs) have significantly enhanced ZSRTE performance, enabling the direct generation of relational triplets from unstructured text. However, LLMs often introduce biases such as entity shift, relation confusion, and over-prediction, which limit the reliability of the extracted triplets. In this paper, we introduce DETRI, a novel debiasing framework for ZSRTE that addresses these biases by leveraging a discriminative structural expert model. DETRI treats LLMs as inspiration generators and refines their outputs via a three-stage debiasing pipeline, consisting of inspiration-based re-prediction, confidence-based filtering, and entity shift correction through span perturbation. The framework improves LLM-generated triplets without requiring fine-tuning the LLM, thus offering a transferable and scalable solution. Extensive experiments on two datasets show that DETRI outperforms state-of-the-art methods by reducing bias and improving extraction accuracy, achieving 2.89% F1 improvement over fine-tuned LLM methods while keeping the LLM parameters frozen. Our approach shows strong generalization across different LLMs, offering a solution to mitigate biases.
Large language model (LLM)-based text-to-speech (TTS) systems enable prompt-conditioned emotional control but struggle with fine-grained emotion intensity due to the semantic--acoustic gap between text and speech. To address this challenge, we formulate emotion intensity control in LLM-based TTS as a learning-to-rank problem and propose Emo-LiPO, a listwise preference optimization framework that aligns prompt-conditioned speech generation with relative emotion intensity expressed in text. Emo-LiPO explicitly models global intensity ordering within each emotion under fixed transcripts, enabling more faithful and continuous emotional expression. We further construct ESD-plus, a multi-speaker dataset with explicit emotion intensity variations, to support fine-grained emotion modeling and evaluation. Experiments on ESD-plus demonstrate that Emo-LiPO significantly improves emotion accuracy and intensity controllability over both supervised- and DPO-based LLM TTS baselines, with particularly pronounced gains at high intensity levels.
Reasoning chains in Large Language Models (LLMs) often exhibit heavy-tailed length distributions, yet existing efficiency methods rely on suboptimal linear penalties that suppress complex reasoning, limiting both accuracy and generalization. To address this, we first empirically observe that reasoning lengths are well approximated by a log-normal distribution, and provide an intuitive explanation for this phenomenon. Based on this insight, we propose the Powered Length Penalty (PLP), an adaptive regularizer that penalizes redundancy in short sequences while gradually reducing penalties for longer sequences, preserving deep reasoning. Trained solely on the elementary GSM8K dataset, PLP significantly improves reasoning efficiency by reducing inference costs on GSM8K and MATH500, while simultaneously enhancing accuracy on the challenging AIME2024 benchmark. Furthermore, PLP transfers effectively to diverse domains, including MMLU and GPQA. These results suggest that modeling reasoning length distributions and adapting penalties accordingly can mitigate the typical trade-off between efficiency and performance, enabling more reliable and cost-effective reasoning across tasks.
Movie dubbing is the task of synthesizing speech from scripts conditioned on video scenes, requiring accurate lip sync, faithful timbre transfer, and proper modeling of character identity and emotion. However, existing methods face two major limitations: (1) high-quality multimodal dubbing datasets are limited in scale, suffer from high word error rates, contain sparse annotations, rely on costly manual labeling, and are restricted to monologue scenes, all of which hinder effective model training; (2) existing dubbing models rely solely on the lip region to learn audio-visual alignment, which limits their applicability to complex live-action cinematic scenes, and exhibit suboptimal performance in lip sync, speech quality, and emotional expressiveness. To address these issues, we propose FunCineForge, which comprises an end-to-end production pipeline for large-scale dubbing datasets and an MLLM-based dubbing model designed for diverse cinematic scenes. The pipeline enables the construction of the first television dubbing dataset, CineDub, which serves as a high-quality foundation for training and evaluation. Building on this, our dubbing model effectively captures multimodal cues and supports complex dubbing scenarios, including monologue, narration, dialogue, and multi-speaker settings. Experiments demonstrate that our approach consistently outperforms state-of-the-art methods in audio quality, word error rate, lip sync, temporal alignment, timbre transfer, and instruction following. Code and demos are available at https://funcineforge.github.io/.
Knowledge editing is pivotal for efficiently updating the parametric memory of Large Language Models (LLMs), enabling them to function as evolving agents in dynamic environments. However, mainstream in-parameter knowledge editing approaches suffer from Subject-Dominant Memory Interference: modifying a specific fact inadvertently corrupts the broader structural knowledge associated with the same subject within LLMs. We diagnose the root cause as a shortcut learning pathology, where the optimization objective overfits subject representations while bypassing the essential relational context. To rectify this, we propose Causal Path Alignment (CPA), a principled framework designed to anchor the optimization trajectory to valid causal pathways. CPA enforces parameter updates to route through relation-aware intermediate states, thereby preventing the erasure of contextual dependencies. Experimental results across diverse LLM backbones demonstrate that CPA consistently eliminates the shortcut, significantly improving relation specificity while exhibiting minimal side-effects. Moreover, CPA serves as a model-agnostic plug-in for existing editors, paving the way for reliable and trustworthy in-parameter knowledge editing.
AI-driven survey generation has advanced rapidly, yet related work generation (RWG) remains relatively underexplored. Unlike surveys that provide broad literature overviews, RWG synthesizes prior studies for a single focal paper, requiring contextual fit, cross-paper comparison, and accurate attribution. To address this gap, we propose MA-RWG, a fully automated multi-agent framework that generates polished related work sections from only a title and abstract. MA-RWG first retrieves high-quality candidate papers through semantic retrieval, optionally enhanced with a diversity-aware term. It then coordinates four specialized agents for summarization, organization, integration, and fact checking, enabling DAG-based taxonomy construction, feedback-guided refinement, and dual-model verification. For evaluation, we introduce a dedicated benchmark for paper-specific related work generation, covering generation quality, citation quality, and claim-level semantic similarity. Experimental results show that MA-RWG outperforms RAG-based baselines and survey-oriented agentic methods on the RWG task. Further ablation and cross-domain experiments demonstrate the soundness and robustness of the proposed framework.