arXiv subjects

Siyu Yan

Publications and source records attributed to Siyu Yan.

16 recordsLinked to original sources

arXiv2026

AgentRM: Enhancing Agent Generalization with Reward Modeling

Existing LLM-based agents have achieved strong performance on held-in tasks, but their generalizability to unseen tasks remains poor. Hence, some recent work focus on fine-tuning the policy model with more diverse tasks to improve the generalizability. In this work, we find that finetuning a reward model to guide the policy model is more robust than directly finetuning the policy model. Based on this finding, we propose AgentRM, a generalizable reward model, to guide the policy model for effective test-time search. We comprehensively investigate three approaches to construct the reward model, including explicit reward modeling, implicit reward modeling and LLM-as-a-judge. We then use AgentRM to guide the answer generation with Best-of-N sampling and step-level beam search. On four types of nine agent tasks, AgentRM enhances the base policy model by $8.8$ points on average, surpassing the top general agent by $4.0$. Moreover, it demonstrates weak-to-strong generalization, yielding greater improvement of $12.6$ on LLaMA-3-70B policy model. As for the specializability, AgentRM can also boost a finetuned policy model and outperform the top specialized agent by $11.4$ on three held-in tasks. Further analysis verifies its effectiveness in test-time scaling. Codes will be released to facilitate the research in this area.

cs.CL↗

arXiv2026

Your LLM, Your Style: Behavioral Mode Axes for LLM Behavioral Control

Large language models (LLMs) increasingly act in interactive settings where their behavioral styles affect user experience, safety, and downstream decision making. Existing LLM personality studies largely rely on self-report questionnaires administered in first-person settings, making the resulting profiles sensitive to surface elicitation choices and poorly grounded in concrete model behavior. In this work, we introduce a situated behavioral-data (B-data) framework for studying and controlling LLM behavioral personality. We construct 3,200 contrastive behavioral scenarios spanning 20 behavioral patterns and four prompt registers, grounded in validated psychometric facets such as BFI-2, DOSPERT, and HEXACO. Using this framework, we find that LLMs exhibit stable and model-specific behavioral profiles, while also revealing register-dependent shifts across first-person decisions, advice-giving, and task execution. We then show that these behavioral patterns can be controlled through Behavioral Mode Axes (BMAs), activation-space directions derived from contrastive behavioral traces. Compared with response-derived BMAs, which are more prone to trait drift, thought-derived BMAs more faithfully capture the intended behavioral mechanism and provide cleaner control over situated behavioral styles. Our results suggest that LLM personality-like tendencies are better understood not as abstract self-report traits, but as measurable and controllable behavioral modes grounded in concrete interaction contexts. Our code and data are available at https://github.com/lhz191/LLM-Behavioral-Personality.

cs.LG↗

arXiv2026

See2Think: Do Multimodal Models Really Use Intermediate Visual States?

Multimodal large language models increasingly use sketches, annotations, tools, and intermediate images during reasoning, but it remains unclear whether they truly rely on these visual states. Existing benchmarks are limited both by task collections with narrow coverage or partially text-solvable samples and by evaluations that emphasize final answers without diagnosing how intermediate visual states are generated, rendered, and used. We introduce See2Think, a unified evaluation framework comprising See2ThinkBench and Visual Action-of-Thought (VAoT). See2ThinkBench contains 1,200 open-ended, visually dependent problems across 12 task categories spanning 2D structured, 3D scene, and real-world reasoning. VAoT records textual thoughts, visual actions, rendered states, and subsequent reasoning under four controlled inference settings. Evaluating representative proprietary and open-source multimodal models, we find that visual reasoning is strongly model- and environment-dependent, with no single setting consistently dominating across tasks. Process analysis further shows that models usually select relevant visual operations, while faithful rendering remains the clearest bottleneck and high feedback uptake does not necessarily translate into accuracy gains. Under task-relevant corrupted feedback, models exhibit behavioral dependence on visual states, with accuracy dropping by over 10 percentage points in controlled interventions.

cs.CV↗

arXiv2026

DataEvolver: Self-Evolving Multi-Agent Data Construction for Text-Rich Image Generation

Text-rich image generation is one of the most challenging settings in image generation, since models must simultaneously produce visually realistic images and render legible, semantically aligned, and layout-consistent text. Existing data pipelines usually follow a static crawl-filter-freeze paradigm. They collect candidate samples, filter them once, and freeze the accepted data for training. However, rejected samples are usually discarded, although they often contain useful failure signals such as OCR errors and semantic mismatches. As a result, later construction rounds may repeat the same failure modes. To address these limitations, we propose DataEvolver, a self-evolving multi-agent framework for text-rich image data construction. DataEvolver treats data construction as feedback-driven construction policy evolution. A Retriever collects candidate samples, a Verifier assigns quality scores and rejection causes, a Critic summarizes round-level feedback into semantic feedback, and a Generator completes under-covered regions through targeted synthesis. The updated feedback memory then guides the next construction round. Experiments on text-rich image generation benchmarks show that DataEvolver produces more useful training data than fixed-dataset baselines under matched data budgets. At the 0.75M scale on PixArt-alpha, DataEvolver improves OCR-F1 over the strongest baseline by 85.3 percent on TextScenesHQ and 35.3 percent on LongTextBench. The improvements are consistent across both evaluated benchmarks and also transfer to Show-o2, indicating that the benefit of DataEvolver is not tied to a single downstream generator. These results suggest that rejected samples can provide actionable feedback for improving text-rich image data construction.

cs.CV↗

arXiv2026

Differentiable Mixture-of-Agents Incentivizes Swarm Intelligence of Large Language Models

Recent advances in Large Language Models (LLMs) have catalyzed the development of multi-agent systems (MAS) for complex reasoning tasks. However, existing MAS typically rely on pre-defined or pre-compiled communication topologies, which limits their flexibility and adaptability to dynamic task requirements. In this work, we propose Differentiable Mixture-of-Agents (DMoA), a self-evolving multi-agent framework that enables elastic and adaptive agent collaboration during inference. Instead of statically constructing workflows, DMoA dynamically routes and activates agents at each reasoning step, allowing the system to implicitly simulate diverse communication topologies and adapt to evolving demands. To achieve this, we design a differentiable, context-aware routing mechanism that leverages recurrent structures to incorporate historical and contextual information, producing sparse agent activations in a step-wise manner. Furthermore, we introduce predictive entropy as self-supervised signals to optimize the routing process, enabling efficient test-time adaptation without external annotations. Extensive experiments across 9 benchmarks demonstrate that DMoA achieves state-of-the-art performance while exhibiting strong efficiency, robustness, and ensembling capabilities.

cs.LG↗

arXiv2026

A Survey of Inductive Reasoning for Large Language Models

Reasoning is an important task for large language models (LLMs). Among all the reasoning paradigms, inductive reasoning is one of the fundamental types, which is characterized by its particular-to-general thinking process and the non-uniqueness of its answers. The inductive mode is crucial for knowledge generalization and aligns better with human cognition, so it is a fundamental mode of learning, hence attracting increasing interest. Despite the importance of inductive reasoning, there is no systematic summary of it. Therefore, this paper presents the first comprehensive survey of inductive reasoning for LLMs. First, methods for improving inductive reasoning are categorized into three main areas: post-training, test-time scaling, and data augmentation. Then, current benchmarks of inductive reasoning are summarized, and a unified sandbox-based evaluation approach with the observation coverage metric is derived. Finally, we offer some analyses regarding the source of inductive ability and how simple model architectures and data help with inductive tasks, providing a solid foundation for future research.

cs.CL↗

arXiv2026

Atomic Structure of Grain Boundaries, Dislocations and Associated Strain in Templated Co-evaporated Photoactive Halide Perovskites

Structural defects, particularly grain boundaries, play a crucial role in governing charge transport and the optoelectronic properties of metal halide perovskites, thereby limiting the performance of devices. Solar cells incorporating templated FA0.9Cs0.1PbI3-xClx show significant improvements in grain orientation and steady-state power conversion efficiency; however, the underlying mechanisms remain unclear. In this study, we address this gap by employing a suite of tailored low-dose electron microscopy techniques to investigate the templated FA0.9Cs0.1PbI3-xClx film, revealing that it exhibits a preferred crystallographic orientation along the <001> zone axis, with arbitrary grain rotations about that axis, indicative of a Volmer-Weber growth mechanism. We determine the atomic structure of the resulting high-angle and low-angle grain boundaries. We also reveal the presence of edge dislocations and their associated strain fields, demonstrating the compressive strain on one side of the dislocation core and tensile strain on the opposite side. Furthermore, we find dislocations associated with stacking faults. These atomic-level insights uncover which grain boundaries and intra-grain defects are likely to act as recombination centres or modify band gaps, crucial for understanding which defects influence the performance of perovskite solar cell devices.

cond-mat.mtrl-sci↗

arXiv2026

Eliciting Trustworthiness Priors of Large Language Models via Economic Games

One critical aspect of building human-centered, trustworthy artificial intelligence (AI) systems is maintaining calibrated trust: appropriate reliance on AI systems outperforms both overtrust (e.g., automation bias) and undertrust (e.g., disuse). A fundamental challenge, however, is how to characterize the level of trust exhibited by an AI system itself. Here, we propose a novel elicitation method based on iterated in-context learning (Zhu and Griffiths, 2024a) and apply it to elicit trustworthiness priors using the Trust Game from behavioral game theory. The Trust Game is particularly well suited for this purpose because it operationalizes trust as voluntary exposure to risk based on beliefs about another agent, rather than self-reported attitudes. Using our method, we elicit trustworthiness priors from several leading large language models (LLMs) and find that GPT-4.1's trustworthiness priors closely track those observed in humans. Building on this result, we further examine how GPT-4.1 responds to different player personas in the Trust Game, providing an initial characterization of how such models differentiate trust across agent characteristics. Finally, we show that variation in elicited trustworthiness can be well predicted by a stereotype-based model grounded in perceived warmth and competence.

cs.CL↗

arXiv2025

MUSE: MCTS-Driven Red Teaming Framework for Enhanced Multi-Turn Dialogue Safety in Large Language Models

As large language models~(LLMs) become widely adopted, ensuring their alignment with human values is crucial to prevent jailbreaks where adversaries manipulate models to produce harmful content. While most defenses target single-turn attacks, real-world usage often involves multi-turn dialogues, exposing models to attacks that exploit conversational context to bypass safety measures. We introduce MUSE, a comprehensive framework tackling multi-turn jailbreaks from both attack and defense angles. For attacks, we propose MUSE-A, a method that uses frame semantics and heuristic tree search to explore diverse semantic trajectories. For defense, we present MUSE-D, a fine-grained safety alignment approach that intervenes early in dialogues to reduce vulnerabilities. Extensive experiments on various models show that MUSE effectively identifies and mitigates multi-turn vulnerabilities. Code is available at \href{https://github.com/yansiyu02/MUSE}{https://github.com/yansiyu02/MUSE}.

cs.CL↗

arXiv2025

DeepVIS: Bridging Natural Language and Data Visualization Through Step-wise Reasoning

Although data visualization is powerful for revealing patterns and communicating insights, creating effective visualizations requires familiarity with authoring tools and often disrupts the analysis flow. While large language models show promise for automatically converting analysis intent into visualizations, existing methods function as black boxes without transparent reasoning processes, which prevents users from understanding design rationales and refining suboptimal outputs. To bridge this gap, we propose integrating Chain-of-Thought (CoT) reasoning into the Natural Language to Visualization (NL2VIS) pipeline. First, we design a comprehensive CoT reasoning process for NL2VIS and develop an automatic pipeline to equip existing datasets with structured reasoning steps. Second, we introduce nvBench-CoT, a specialized dataset capturing detailed step-by-step reasoning from ambiguous natural language descriptions to finalized visualizations, which enables state-of-the-art performance when used for model fine-tuning. Third, we develop DeepVIS, an interactive visual interface that tightly integrates with the CoT reasoning process, allowing users to inspect reasoning steps, identify errors, and make targeted adjustments to improve visualization outcomes. Quantitative benchmark evaluations, two use cases, and a user study collectively demonstrate that our CoT framework effectively enhances NL2VIS quality while providing insightful reasoning steps to users.

cs.AI↗

arXiv2025

Beyond Quality: Unlocking Diversity in Ad Headline Generation with Large Language Models

The generation of ad headlines plays a vital role in modern advertising, where both quality and diversity are essential to engage a broad range of audience segments. Current approaches primarily optimize language models for headline quality or click-through rates (CTR), often overlooking the need for diversity and resulting in homogeneous outputs. To address this limitation, we propose DIVER, a novel framework based on large language models (LLMs) that are jointly optimized for both diversity and quality. We first design a semantic- and stylistic-aware data generation pipeline that automatically produces high-quality training pairs with ad content and multiple diverse headlines. To achieve the goal of generating high-quality and diversified ad headlines within a single forward pass, we propose a multi-stage multi-objective optimization framework with supervised fine-tuning (SFT) and reinforcement learning (RL). Experiments on real-world industrial datasets demonstrate that DIVER effectively balances quality and diversity. Deployed on a large-scale content-sharing platform serving hundreds of millions of users, our framework improves advertiser value (ADVV) and CTR by 4.0% and 1.4%.

cs.CL↗

arXiv2025

DrugMCTS: a drug repurposing framework combining multi-agent, RAG and Monte Carlo Tree Search

Recent advances in large language models have demonstrated considerable potential in scientific domains such as drug repositioning. However, their effectiveness remains constrained when reasoning extends beyond the knowledge acquired during pretraining. Conventional approaches, such as fine-tuning or retrieval-augmented generation, face limitations in either imposing high computational overhead or failing to fully exploit structured scientific data. To overcome these challenges, we propose DrugMCTS, a novel framework that synergistically integrates RAG, multi-agent collaboration, and Monte Carlo Tree Search for drug repositioning. The framework employs five specialized agents tasked with retrieving and analyzing molecular and protein information, thereby enabling structured and iterative reasoning. Extensive experiments on the DrugBank and KIBA datasets demonstrate that DrugMCTS achieves substantially higher recall and robustness compared to both general-purpose LLMs and deep learning baselines. Our results highlight the importance of structured reasoning, agent-based collaboration, and feedback-driven search mechanisms in advancing LLM applications for drug repositioning.

cs.AI↗

arXiv2025

HKD4VLM: A Progressive Hybrid Knowledge Distillation Framework for Robust Multimodal Hallucination and Factuality Detection in VLMs

Driven by the rapid progress in vision-language models (VLMs), the responsible behavior of large-scale multimodal models has become a prominent research area, particularly focusing on hallucination detection and factuality checking. In this paper, we present the solution for the two tracks of Responsible AI challenge. Inspirations from the general domain demonstrate that a smaller distilled VLM can often outperform a larger VLM that is directly tuned on downstream tasks, while achieving higher efficiency. We thus jointly tackle two tasks from the perspective of knowledge distillation and propose a progressive hybrid knowledge distillation framework termed HKD4VLM. Specifically, the overall framework can be decomposed into Pyramid-like Progressive Online Distillation and Ternary-Coupled Refinement Distillation, hierarchically moving from coarse-grained knowledge alignment to fine-grained refinement. Besides, we further introduce the mapping shift-enhanced inference and diverse augmentation strategies to enhance model performance and robustness. Extensive experimental results demonstrate the effectiveness of our HKD4VLM. Ablation studies provide insights into the critical design choices driving performance gains.

cs.CV↗

arXiv2025

The Joint Entity-Relation Extraction Model Based on Span and Interactive Fusion Representation for Chinese Medical Texts with Complex Semantics

Joint entity-relation extraction is a critical task in transforming unstructured or semi-structured text into triplets, facilitating the construction of large-scale knowledge graphs, and supporting various downstream applications. Despite its importance, research on Chinese text, particularly with complex semantics in specialized domains like medicine, remains limited. To address this gap, we introduce the CH-DDI, a Chinese drug-drug interactions dataset designed to capture the intricacies of medical text. Leveraging the strengths of attention mechanisms in capturing long-range dependencies, we propose the SEA module, which enhances the extraction of complex contextual semantic information, thereby improving entity recognition and relation extraction. Additionally, to address the inefficiencies of existing methods in facilitating information exchange between entity recognition and relation extraction, we present an interactive fusion representation module. This module employs Cross Attention for bidirectional information exchange between the tasks and further refines feature extraction through BiLSTM. Experimental results on both our CH-DDI dataset and public CoNLL04 dataset demonstrate that our model exhibits strong generalization capabilities. On the CH-DDI dataset, our model achieves an F1-score of 96.73% for entity recognition and 78.43% for relation extraction. On the CoNLL04 dataset, it attains an entity recognition precision of 89.54% and a relation extraction accuracy of 71.64%.

cs.CL↗

arXiv2025

ChartEditor: A Human-AI Paired Tool for Authoring Pictorial Charts

Pictorial charts are favored for their memorability and visual appeal, offering a more engaging alternative to basic charts. However, their creation can be complex and time-consuming due to the lack of native support in popular visualization tools like Tableau. While AI-generated content (AIGC) tools have lowered the barrier to creating pictorial charts, they often lack precise design control. To address this issue, we introduce ChartEditor, a human-AI paired tool that transforms basic charts into pictorial versions based on user intent. ChartEditor decomposes chart images into visual components and organizes them within a hierarchical tree. Based on this tree, users can express their intent in natural language, which is then translated into modifications to the hierarchy. In addition, users can directly interact with and modify specific chart components via an intuitive interface to achieve fine-grained design control. A user study demonstrates the effectiveness and usability of ChartEditor in simplifying the creation of pictorial charts.

cs.HC↗

arXiv2024

Static and multivariate-temporal attentive fusion transformer for readmission risk prediction

Background: Accurate short-term readmission prediction of ICU patients is significant in improving the efficiency of resource assignment by assisting physicians in making discharge decisions. Clinically, both individual static static and multivariate temporal data collected from ICU monitors play critical roles in short-term readmission prediction. Informative static and multivariate temporal feature representation capturing and fusion present challenges for accurate readmission prediction. Methods:We propose a novel static and multivariate-temporal attentive fusion transformer (SMTAFormer) to predict short-term readmission of ICU patients by fully leveraging the potential of demographic and dynamic temporal data. In SMTAFormer, we first apply an MLP network and a temporal transformer network to learn useful static and temporal feature representations, respectively. Then, the well-designed static and multivariate temporal feature fusion module is applied to fuse static and temporal feature representations by modeling intra-correlation among multivariate temporal features and constructing inter-correlation between static and multivariate temporal features. Results: We construct a readmission risk assessment (RRA) dataset based on the MIMIC-III dataset. The extensive experiments show that SMTAFormer outperforms advanced methods, in which the accuracy of our proposed method is up to 86.6%, and the area under the receiver operating characteristic curve (AUC) is up to 0.717. Conclusion: Our proposed SMTAFormer can efficiently capture and fuse static and multivariate temporal feature representations. The results show that SMTAFormer significantly improves the short-term readmission prediction performance of ICU patients through comparisons to strong baselines.

cs.LG↗