Evaluating semantic drift is essential for understanding dynamical discourse evolution and opinion formation in online discussions. However, sparse and uneven distributions of event-specific keywords prevent traditional models from capturing post-level semantic drift. Thus, to address this issue, we propose an LLM-embedding Semantic Adaptation Network (LLM-SAN), which is a hybrid semantic drift evaluation model with an LLM-Embedding gated recurrent unit (GRU) module, an LLM-Embedding graph convolutional network (GCN) module and a multi-expert adaptive fusion module. The GRU module is used to extract features from event related posts, and The GCN is used to extract features from temporal graphical topic posts. Then, the features are merged by the multi-expert adaptive fusion module. Finally, this module predicts the future post embedding, and the prediction error is used to evaluate and detect the semantic drift points. Extensive experiments are conducted, and the results show that LLM-SAN achieves the state-of-the-art performance on the semantic drift evaluation task, compared to the other baselines. Ablation experiments are also conducted to show the effectiveness of each module in LLM-SAN.
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
The ability of large language models (LLMs) to follow user instructions is central to their reliability, safety, and usefulness. While prior studies assess instruction adherence in the model’s main responses, we argue that it is also critical for large reasoning models (LRMs) to follow user instructions throughout their reasoning process. Reasoning instruction following makes LRMs more controllable and transparent, while reducing risks of undesirable shortcuts, hallucinations, or reward hacking within reasoning traces. To evaluate this dimension, we introduce ReasonIF, a systematic benchmark for assessing reasoning instruction following. ReasonIF includes six categories of instruction prompts, spanning multilingual reasoning, and length control. Across many open-source LRMs including GPT-OSS, Qwen3, and DeepSeek-R1, we find substantial failures in reasoning instruction adherence: the highest instruction following score (IFS) remains below 0.25, meaning that fewer than 25% of reasoning traces comply with the given instructions. Notably, as task difficulty increases, reasoning instruction following degrades further. We also explore two strategies to enhance reasoning instruction fidelity: (1) multi-turn reasoning and (2) Reasoning Instruction Finetuning (RIF) using synthetic data. RIF improves the IFS of GPT-OSS-20B from 0.11 to 0.27, indicating measurable progress but leaving ample room for improvement. We hope this work draws attention to reasoning-level instruction adherence as an underexplored but critical aspect of model alignment, and helps pave the way toward more controllable, interpretable, and trustworthy reasoning models.
We present a system that uses LLMs as a tool in the development of Constructed Languages— ConLangs, which we call IASC (Interactive Agentic System for ConLangs). The system is modular in that it creates each of the components—phonology, morphology and syntax, lexicon, orthography, and grammatical handbook, using module-specific sets of prompts. The approach is agentic in that various modules allow for refining the output given automatically-generated commentary on a previous step. Our main goals are twofold. First, we aim to provide tools that facilitate an engaging and enjoyable experience in creating artificially constructed languages. Second, the focus of this paper is on using our ConLang framework as a novel way to explore what LLMs ‘know’ about language—not what they know about any particular language or encyclopedic facts, but how much they know about and understand language and linguistic concepts. In the experiments, we particularly focus on the morphosyntax module and show that there is a fairly wide gulf in capabilities both among different LLMs and among different linguistic specifications, with it being notably easier for systems to deal with more typologically common patterns than rarer ones. All code is released: https://github.com/SakanaAI/IASC.
Large language models (LLMs) have been widely explored for embedding generation. While recent studies show that in-context learning (ICL) effectively enhances the representational capability of LLMs by prepending a few task-related demonstrations, it causes substantial token overhead due to the increased sequence length. In this work, we propose EPIC, a novel embedding-based in-context prompt training strategy that leverages ICL to generate high-quality embeddings while reducing computational burden during both training and inference. This approach replaces discrete text demonstrations with their corresponding continuous embeddings, which not only encourages the LLM to align semantically-related text pairs during contrastive learning, but also requires the model to interpret demonstration embeddings as part of the in-context prompt. Consequently, EPIC-trained models achieve excellent embedding performance both with or without in-context prompts at inference time. Comprehensive experiments demonstrate that our method establishes new state-of-the-art results on the MTEB benchmark, surpassing frontier models trained solely on publicly available retrieval data. Extensive ablation studies further validate the effectiveness and necessity of our mechanism.
Multimodal learning aims to learn unified multimodal representations from heterogeneous modalities and supports many natural language processing tasks. However, multimodal models often exhibit modality laziness: over-relying on a dominant modality and under-exploiting complementary signals. Existing approaches typically strengthen unimodal training or rebalance modality contributions, but they may still emphasize shared semantics and overlook modality-specific cues. To address this, we propose SCOPE, a unified framework for learning complete multimodal representations, achieving Shared-and-COmplementary cue PrEservation. Firstly, SCOPE uses a mutual information-guided disentanglement module to separate shared semantics from modality-specific cues and mitigate representation collapse. Secondly, SCOPE aligns modalities by enforcing structural consistency between modality-wise semantic graphs, avoiding brittle point-wise matching. Finally, SCOPE performs balanced fusion via structure-aware diffusion attention to integrate shared and complementary cues without feature homogenization. Experiments on four benchmark datasets show that SCOPE consistently outperforms SOTA baselines, achieving up to 27.10% accuracy improvement.
Large language models often display undesirable behaviors embedded in their internal representations, undermining fairness, inconsistency drift, amplification of harmful content, and the propagation of unwanted patterns during extended dialogue and conversations. Although training-time or data-centric methods attempt to reduce these effects, they are computationally expensive, irreversible once deployed, and slow to adapt to new conversational contexts. Pruning-based methods provide a flexible and transparent way to reduce bias by adjusting the neurons responsible for certain behaviors. However, most existing approaches are static; once a neuron is removed, the model loses the ability to adapt when the conversation or context changes. To address this, we propose a dynamic, reversible, pruning-based framework that detects context-aware neuron activations and applies adaptive masking to modulate their influence during generation. Our inference-time solution provides fine-grained, memory-aware mitigation with knowledge-preserved, more coherent behavior across multilingual single- and multi-turn dialogues, enabling dynamic fairness control in real-world conversational AI.
Towards Visually Grounded Multimodal Summarization via Cross-Modal Transformer and Gated Attention
PDF ↗Multimodal summarization requires models to jointly understand textual and visual inputs to generate concise, semantically coherent summaries. Existing methods often inject shallow visual features into deep language models, leading to representational mismatches and weak cross-modal grounding. We propose a unified framework that jointly performs text summarization and representative image selection. Our system, SPeCTrA-Sum (Sampler Perceiver with Cross-modal Transformer and gated Attention for Summarization), introduces two key innovations. First, a Deep Visual Processor (DVP) aligns the visual encoder with the language model at corresponding depths, enabling hierarchical, layer-wise fusion that preserves semantic consistency. Second, a lightweight Visual Relevance Predictor (VRP) selects salient and diverse images by distilling soft labels from a Determinantal Point Process (DPP) teacher. SPeCTrA-Sum is trained using a multi-objective loss that combines autoregressive summarization, cross-modal alignment, and DPP-based distillation. Experiments show that our system produces more accurate, visually grounded summaries and selects more representative images, demonstrating the benefits of depth-aware fusion and principled image selection for multimodal summarization.
Legal case facts are often lengthy, complex, and difficult to process, posing challenges for legal judgment prediction. Although recent advances leverage large language models (LLMs) for legal reasoning, they face high computational costs and information degradation when handling long cases. Previous approaches, such as architectural modifications and text compression methods, reduce computational complexity to some extent but still struggle to effectively capture legally salient information in complex cases. We propose a legal knowledge–adaptive compression framework for long legal judgment prediction that integrates domain-specific legal knowledge to guide adaptive context compression. Our approach selectively retains legally relevant information while reducing redundant or less informative content, enabling efficient and accurate long-context reasoning. We evaluate the proposed framework on four real-world datasets spanning multiple jurisdictions and languages. Experimental results demonstrate that our method outperforms existing approaches in both prediction performance and computational efficiency.
MonCulture-Eval: A Hierarchical Benchmark for Evaluating Mongolian Cultural Capabilities of Large Language Models across Scripts and Regions
PDF ↗While Large Language Models (LLMs) have achieved impressive linguistic fluency in low-resource languages, their capacity to process deep cultural nuances remains insufficiently quantified. This paper introduces MonCulture-Eval, a benchmark designed to assess the cultural intelligence of LLMs in the Mongolian context across two writing systems (Traditional and Cyrillic) and three regional sub-cultures (Alxa, Ordos, and Horqin). Curated entirely from primary, non-digitized archives to prevent data contamination, the benchmark employs a three-layer cognitive hierarchy—Factual, Situational, and Values—supplemented by specialized tasks including Riddles, Taboos, and Proverbs. Evaluation of frontier models reveals a severe "Script Gap" and a systematic "Etic Bias," where models sanitize spiritual rituals into secular functional norms.
Measuring What Matters Beyond Text: Evaluating Multimodal Summaries by Quality, Alignment, and Diversity
PDF ↗Multimodal Large Language Models (MLLMs) have facilitated Multimodal Summarization with Multimodal Output (MSMO), wherein systems generate concise textual summaries accompanied by salient visuals from multimodal sources. However, current MSMO evaluation remains fragmented: text quality, image-text alignment, and visual diversity are typically assessed in isolation using unimodal metrics, making it difficult to capture whether the modalities jointly support a faithful and useful summary. To address this gap, we introduce MM-Eval, a unified evaluation framework that integrates assessments of textual quality, cross-modal alignment, and visual diversity. MM-Eval comprises three components: (1) text quality, measured using OpenFActScore for factual consistency and G-Eval for coherence, fluency, and relevance; (2) image-text relevance, evaluated via an MLLM-as-a-judge approach; and (3) image-set diversity, quantified using Truncated CLIP Entropy. We calibrate -Eval through a learned aggregation model trained on the mLLM-EVAL news benchmark, aligning component contributions with human preferences. Our analysis reveals a text-dominant hierarchy in this setting, where factual consistency acts as a critical determinant of perceived overall quality, while visual relevance and diversity provide complementary signals. MM-Eval improves over heuristic aggregation baselines and provides an interpretable, reference-weak framework for comparative evaluation of multimodal summaries.
Key-Value (KV) caching is essential for efficient inference in multimodal large language models (MLLMs), yet its memory footprint grows linearly with context length and becomes a major bottleneck due to the large number of visual tokens. Recent prefill-stage KV selection methods estimate KV importance from prefilling statistics, implicitly assuming that prefilling-time queries are representative of those encountered during decoding. We show that this assumption breaks down in multimodal inference, where decoding-time queries exhibit substantially larger variance than prefilling-stage representations, leading to unstable KV importance estimation under tight cache budgets. As a result, small ranking errors can disproportionately discard semantically critical visual tokens and degrade grounding and reasoning performance. We propose MM-ShiftKV, a training-free, decode-aware and strictly prefill-only KV selection method. MM-ShiftKV approximates decoding-time query behavior during prefilling by constructing variance-expanded query proxies and estimates prompt KV importance based on their aggregated attention mass. Experiments on multimodal benchmarks demonstrate that MM-ShiftKV consistently outperforms existing methods under strict KV-cache budgets.
When Helpers Become Hazards: A Benchmark for Analyzing Multimodal LLM-Powered Safety in Daily Life
PDF ↗As Multimodal Large Language Models (MLLMs) become an indispensable assistant in human life, the unsafe content generated by MLLMs poses a danger to human behavior, perpetually overhanging human society like a sword of Damocles. To investigate and evaluate the safety impact of MLLMs responses on human behavior in daily life, we introduce SaLAD, a multimodal satety benchmark which contains 2,013 real-world image–text samples across 10 common categories, with a balanced design covering both unsafe scenarios and cases of oversensitivity. It emphasizes realistic risk exposure, authentic visual inputs, and fine-grained cross-modal reasoning, ensuring that safety risks cannot be inferred from text alone. We further propose a safety-warning-based evaluation framework that encourages models to provide clear and informative safety warnings, rather than generic refusals. Results on 18 MLLMs demonstrate that the top-performing models achieve a safe response rate of only 57.2% on unsafe queries. Morevoer, even popular safety alignment methods limit effectiveness of the models in our scenario, revealing the vulnerabilities of current MLLMs in identifying dangerous behaviors in daily life. Our dataset is available at https://github.com/xinyuelou/SaLAD.
The rapid growth of scientific literature calls for automated methods to assess and predict research impact.Prior work has largely focused on citation-based metrics, leaving limited evaluation of models’ capability to reason about other impact dimensions.To this end, we introduce SciImpact, a large-scale, multi-dimensional benchmark for scientific impact prediction spanning 19 fields.SciImpact captures various forms of scientific influence, ranging from citation counts to award recognition, media attention, patent reference, and artifact adoption, by integrating heterogeneous data sources and targeted web crawling.It comprises 215,928 contrastive paper pairs reflecting meaningful impact differences in both short- (e.g., Best Paper Award) and long-term settings (e.g., Nobel Prize).We evaluate 11 widely used large language models (LLMs) on SciImpact.Results show that off-the-shelf models show substantial variability across dimensions and fields, while multi-task supervised fine-tuning consistently enables smaller LLMs (e.g., 4B) to markedly outperform much larger models (e.g., 30B) and surpass powerful closed-source LLMs (e.g., o4-mini).These results establish SciImpact as a challenging benchmark and demonstrate its value for multi-dimensional, multi-field scientific impact prediction.Our project homepage is https://flypig23.github.io/sciimpact-homepage/.
GFT: From Imitation to Reward Fine-Tuning with Unbiased Group Advantages and Dynamic Coefficient Rectification
PDF ↗Large language models are typically post-trained using supervised fine-tuning (SFT) and reinforcement learning (RL), yet effectively unifying efficient knowledge injection with robust generalization remains challenging. In this work, we provide a training-dynamics analysis showing that SFT can be interpreted as a special case of policy gradient optimization with an extremely sparse implicit reward and unstable inverse-probability weighting, which together lead to single-path dependency, entropy collapse, and gradient explosion. Motivated by this diagnosis, we propose Group Fine-Tuning (GFT), a unified post-training framework that addresses these intrinsic limitations through two mechanisms: Group Advantage Learning, which constructs diverse response groups and derives normalized contrastive supervision to alleviate reward sparsity, and Dynamic Coefficient Rectification, which adaptively bounds inverse-probability weights to stabilize optimization while preserving efficient knowledge injection. Experiments demonstrate that GFT consistently surpasses SFT-based methods and yields policies that integrate more smoothly with subsequent RL training.Our code is publicly available athttps://github.com/ZJU-OmniAI/GFT.
The success of Large Language Models in mathematical reasoning relies heavily on the generation of diverse and valid solution paths during the rollout phase. However, current rollout techniques face a fundamental trade-off: token-level sampling often yields redundant trajectories that differ only in rephrasing, while embedding-level methods utilizing random noise frequently disrupt semantic consistency. To resolve this, we introduce **N-GRPO**, a novel exploration strategy integrated into the Group Relative Policy Optimization (GRPO) framework. Rather than relying on token-level sampling or native embedding-level noise, our approach leverages Semantic Neighbor Mixing. This mechanism dynamically constructs input representations by mixing the embeddings of an anchor token and its nearest semantic neighbors, thereby injecting diversity while strictly adhering to the local semantic manifold. Experimental evaluations on the DeepSeek-R1-Distill-Qwen models across different sizes show that not only achieves consistent improvements over strong baselines on math reasoning benchmarks but also exhibits robust generalization capabilities on out-of-distribution tasks.
Current large language models (LLMs), even those explicitly trained for reasoning, often struggle with ambiguous content moderation cases due to misleading "decision shortcuts" embedded in context. Inspired by cognitive psychology insights into expert moderation, we introduce CᴀʀO (Chain-of-Analogy Reasoning Optimization), a novel two-stage training framework to induce robust analogical reasoning in LLMs. First, CᴀʀO bootstraps analogical reasoning chains via retrieval-augmented generation (RAG) on moderation data and performs supervised fine-tuning (SFT). Second, we propose a customized direct preference optimization (DPO) approach to reinforce analogical reasoning behaviors explicitly. Unlike static retrieval methods, CᴀʀO dynamically generates tailored analogical references during inference, effectively mitigating harmful decision shortcuts. Extensive experiments demonstrate that CᴀʀO substantially outperforms state-of-the-art reasoning models (DeepSeek R1, QwQ), specialized moderation models (LLaMA Guard), and advanced fine-tuning and retrieval-augmented methods, achieving an average F1 score improvement of 24.9% on challenging ambiguous moderation benchmarks.
Large Language Models (LLMs) have demonstrated remarkable capabilities in agentic collaborative problem-solving, albeit a gap exists. Existing frameworks predominantly rely on natural language as a primary representation (format) for agentic communication. However natural language could be ambiguous and verbose. Furthermore, recent works have shown that alternative representations can enhance performance in LLMs on certain tasks. But current approaches lack the intelligence necessary to understand, learn or apply optimal communication representations adaptively. In this paper, we propose to dynamically learn the optimal message representations to enhance agentic performance. We model the optimization problem as an Expanding Markov Decision Process (EMDP) and propose our method named OPTiMACS. We evaluate our system across benchmark datasets of collaborative problem-solving. The results show significant performance improvements while maintaining efficiency. Our work bridges the gap between rigid communication protocols and open-ended natural language by providing an adaptive framework that learns task-aware structural representations.
Mixture of Experts (MoE) dynamically routes inputs to specialized expert networks, enabling large language models to scale capacity with low inference overhead. To further improve MoE’s parameter efficiency in resource-constrained scenarios, LoRA–MoE integrates LoRA for lightweight adaptation while preserving MoE’s specialization. Despite these benefits, the effectiveness of LoRA–MoE still hinges on balanced expert utilization, where certain experts dominate activations while most remain underutilized. Existing balancing strategies focus on constraining the final distribution of expert usage, but overlook the routing decisions made at each layer. As a result, imbalances gradually accumulate across the routing hierarchy. To address this challenge, we propose LayerMoE, a novel three-stage framework that leverages process-level rewards to guide balanced expert routing. Specifically, to overcome the limitation of focusing only on final losses and ignoring intermediate routing, we introduce Monte Carlo Tree Search (MCTS)-based sampling that decomposes outcome-level supervision into layer-wise reward signals, guiding expert choices throughout the routing process. For efficiency, we organize Transformer layers into groups, which constrain the search space of MCTS and keep exploration overhead tractable while retaining the hierarchical structure. Extensive experiments on representative datasets (e.g., ARC, RACE, OBQA) show that applying LayerMoE consistently improves the performance of state-of-the-art LoRA-MoE baselines, yielding an average accuracy gain of 1.39%. Notably, the maximum improvement reaches 2.50%.
Mobile Phone Agents (MPAs) have emerged as a promising research direction due to their broad applicability across diverse scenarios. While Multimodal Large Language Models (MLLMs) serve as the foundation for MPAs, their effectiveness in handling multiple mobile phone tasks simultaneously remains limited. Although multitask supervised fine-tuning (SFT) is widely adopted for multitask learning, existing approaches struggle to determine optimal training data compositions for peak performance. To address this challenge, we propose DaMo (Data Mixture Optimizer) – a novel solution employing a trainable network that predicts optimal data mixtures by forecasting downstream task performance for any given dataset ratio. To support comprehensive evaluation, we introduce PhoneAgentBench, the first specialized benchmark to evaluate MLLMs on multimodal mobile phone tasks, comprising 1,235 QA pairs spanning diverse real-world industrial mobile application scenarios. Demonstrating strong predictive capability (R²=0.81) in small-scale pilot experiments, DaMo efficiently extrapolates optimal data mixing configurations. Our results show DaMo achieves 3.06% average score improvement on PhoneAgentBench and open-source benchmarks, including BFCL-v3, MME-Reasoning, MME-Perception, and OCRBench, compared to alternative methods. Through predicting optimal data mixture only on open-source benchmarks, DaMo outperforms other approaches by 6.70% in terms of average score. Moreover, DaMo improves the metrics by 12.74% than other methods when used solely for MLLM optimization on the BFCL-v3 task. Notably, DaMo maintains robust scalability, preserving its effectiveness when applied to other model architectures.
Benchmarking Vision-Language Models on Chinese Ancient Documents: From OCR to Knowledge Reasoning
PDF ↗Chinese ancient documents, invaluable carriers of millennia of Chinese history and culture, hold rich knowledge across diverse fields but face challenges in digitization and understanding—traditional methods only scan images, while current Vision-Language Models (VLMs) struggle with their visual/linguistic complexity. Existing document benchmarks focus on English printed texts or simplified Chinese, leaving a gap for evaluating VLMs on ancient Chinese documents. To address this, we present AncientDoc, the first benchmark for Chinese ancient documents, designed to assess VLMs from OCR to knowledge reasoning. AncientDoc includes five tasks (page-level OCR, vernacular translation, reasoning-based QA, knowledge-based QA, linguistic variant QA) and covers 14 document types, over 100 books, and about 3,000 pages. Based on AncientDoc, we evaluate mainstream VLMs using multiple metrics, supplemented by a human-aligned large language model for scoring.