Week Ending 10.19.2025
RESEARCH WATCH: 10.19.2025
OmniVinci: Enhancing Architecture and Data for Omni-Modal Understanding LLM
OmniVinci represents a significant advancement in multi-modal AI systems by enabling models to perceive and reason across vision, audio, and other modalities simultaneously. The framework introduces novel architectural innovations including OmniAlignNet for cross-modal alignment and temporal embedding techniques to synchronize audio-visual signals. With dramatically improved efficiency—using only 0.2T training tokens compared to competitors' 1.2T—OmniVinci achieves state-of-the-art performance on cross-modal understanding benchmarks. Potential applications span robotics, where multi-sensory perception is crucial for navigation and manipulation; medical AI for analyzing patient data from multiple diagnostic sources; and smart manufacturing for quality control through simultaneous visual and acoustic inspection. This open-source initiative democratizes omni-modal AI development.
Authors: Hanrong Ye, Chao-Han Huck Yang, Arushi Goel, Wei Huang, Ligeng Zhu, Yuanhang Su, Sean Lin, An-Chieh Cheng, Zhen Wan, Jinchuan Tian, Yuming Lou, Dong Yang, Zhijian Liu, Yukang Chen, Ambrish Dantrey, Ehsan Jahangiri, Sreyan Ghosh, Daguang Xu, Ehsan Hosseini-Asl, Danial Mohseni Taheri, Vidya Murali, Sifei Liu, Jason Lu, Oluwatobi Olabiyi, Frank Wang, Rafael Valle, Bryan Catanzaro, Andrew Tao, Song Han, Jan Kautz, Hongxu Yin, Pavlo Molchanov
Link: https://arxiv.org/abs/2510.15870v1
Date: 2025-10-d
Summary:
Advancing machine intelligence requires developing the ability to perceive across multiple modalities, much as humans sense the world. We introduce OmniVinci, an initiative to build a strong, open-source, omni-modal LLM. We carefully study the design choices across model architecture and data curation. For model architecture, we present three key innovations: (i) OmniAlignNet for strengthening alignment between vision and audio embeddings in a shared omni-modal latent space; (ii) Temporal Embedding Grouping for capturing relative temporal alignment between vision and audio signals; and (iii) Constrained Rotary Time Embedding for encoding absolute temporal information in omni-modal embeddings. We introduce a curation and synthesis pipeline that generates 24M single-modal and omni-modal conversations. We find that modalities reinforce one another in both perception and reasoning. Our model, OmniVinci, outperforms Qwen2.5-Omni with +19.05 on DailyOmni (cross-modal understanding), +1.7 on MMAR (audio), and +3.9 on Video-MME (vision), while using just 0.2T training tokens - a 6 times reduction compared to Qwen2.5-Omni's 1.2T. We finally demonstrate omni-modal advantages in downstream applications spanning robotics, medical AI, and smart factory.
--------------------------------------------------------------------------------------------------------
Text-to-image diffusion models face critical safety challenges when subtle, seemingly innocent prompts trigger inappropriate sexual content due to underlying model biases. NDM addresses this vulnerability through an innovative noise-driven framework that operates during the early stages of image generation. By analyzing predicted noise patterns to detect malicious intentions and employing adaptive negative guidance to suppress problematic attention regions, NDM maintains model quality while preventing harmful outputs. Applications include content moderation platforms, creative tools for educational and professional environments, and safeguarding AI-generated content in social media. This framework is particularly valuable for organizations deploying generative AI publicly, helping ensure ethical compliance without compromising creative capabilities or requiring extensive model retraining.
Authors: Yitong Sun, Yao Huang, Ruochen Zhang, Huanran Chen, Shouwei Ruan, Ranjie Duan, Xingxing Wei
Link: https://arxiv.org/abs/2510.15752v1
Date: 2025-10-d
Summary:
Despite the impressive generative capabilities of text-to-image (T2I) diffusion models, they remain vulnerable to generating inappropriate content, especially when confronted with implicit sexual prompts. Unlike explicit harmful prompts, these subtle cues, often disguised as seemingly benign terms, can unexpectedly trigger sexual content due to underlying model biases, raising significant ethical concerns. However, existing detection methods are primarily designed to identify explicit sexual content and therefore struggle to detect these implicit cues. Fine-tuning approaches, while effective to some extent, risk degrading the model's generative quality, creating an undesirable trade-off. To address this, we propose NDM, the first noise-driven detection and mitigation framework, which could detect and mitigate implicit malicious intention in T2I generation while preserving the model's original generative capabilities. Specifically, we introduce two key innovations: first, we leverage the separability of early-stage predicted noise to develop a noise-based detection method that could identify malicious content with high accuracy and efficiency; second, we propose a noise-enhanced adaptive negative guidance mechanism that could optimize the initial noise by suppressing the prominent region's attention, thereby enhancing the effectiveness of adaptive negative guidance for sexual mitigation. Experimentally, we validate NDM on both natural and adversarial datasets, demonstrating its superior performance over existing SOTA methods, including SLD, UCE, and RECE, etc. Code and resources are available at https://github.com/lorraine021/NDM.
--------------------------------------------------------------------------------------------------------
RLAF: Reinforcement Learning from Automaton Feedback
RLAF introduces a paradigm shift in reinforcement learning by replacing traditional reward engineering with automaton-based preference learning. Using deterministic finite automata to encode temporal dependencies and task specifications, the framework generates trajectory preferences that guide policy optimization without manual reward design. This approach excels in environments with complex, history-dependent objectives that violate Markovian assumptions. Applications include robotic task planning where sequential constraints matter (e.g., "assemble before testing"), autonomous systems requiring temporal safety specifications, and workflow automation with procedural requirements. The framework offers both static and dynamic training approaches with convergence guarantees, making it suitable for industrial automation, process control systems, and human-robot collaboration where tasks involve multi-step procedures with specific ordering constraints.
Authors: Mahyar Alinejad, Alvaro Velasquez, Yue Wang, George Atia
Link: https://arxiv.org/abs/2510.15728v1
Date: 2025-10-d
Summary:
Reinforcement Learning (RL) in environments with complex, history-dependent reward structures poses significant challenges for traditional methods. In this work, we introduce a novel approach that leverages automaton-based feedback to guide the learning process, replacing explicit reward functions with preferences derived from a deterministic finite automaton (DFA). Unlike conventional approaches that use automata for direct reward specification, our method employs the structure of the DFA to generate preferences over trajectories that are used to learn a reward function, eliminating the need for manual reward engineering. Our framework introduces a static approach that uses the learned reward function directly for policy optimization and a dynamic approach that involves continuous refining of the reward function and policy through iterative updates until convergence. Our experiments in both discrete and continuous environments demonstrate that our approach enables the RL agent to learn effective policies for tasks with temporal dependencies, outperforming traditional reward engineering and automaton-based baselines such as reward machines and LTL-guided methods. Our results highlight the advantages of automaton-based preferences in handling non-Markovian rewards, offering a scalable, efficient, and human-independent alternative to traditional reward modeling. We also provide a convergence guarantee showing that under standard assumptions our automaton-guided preference-based framework learns a policy that is near-optimal with respect to the true non-Markovian objective.
--------------------------------------------------------------------------------------------------------
This work addresses fundamental limitations in aligning large language models with diverse human values by acknowledging that preferences vary across different user populations. The research demonstrates that binary comparisons are theoretically insufficient for capturing latent preference structures, while ternary rankings enable proper identification. The proposed Expectation-Maximization adaptation of DPO discovers distinct annotator types and trains mixture models accordingly, with an aggregation algorithm ensuring equitable performance. Applications include personalized AI assistants adapting to individual users, content recommendation systems respecting cultural differences, and educational AI tools accommodating diverse learning styles. This framework is particularly valuable for global platforms serving multicultural audiences, healthcare AI requiring demographic sensitivity, and any application where one-size-fits-all alignment proves inadequate.
Authors: Keertana Chidambaram, Karthik Vinary Seetharaman, Vasilis Syrgkanis
Link: https://arxiv.org/abs/2510.15716v1
Date: 2025-10-d
Summary:
Reinforcement Learning from Human Feedback (RLHF) has become central to aligning large language models with human values, typically by first learning a reward model from preference data which is then used to update the model with reinforcement learning. Recent alternatives such as Direct Preference Optimization (DPO) simplify this pipeline by directly optimizing on preferences. However, both approaches often assume uniform annotator preferences and rely on binary comparisons, overlooking two key limitations: the diversity of human evaluators and the limitations of pairwise feedback. In this work, we address both these issues. First, we connect preference learning in RLHF with the econometrics literature and show that binary comparisons are insufficient for identifying latent user preferences from finite user data and infinite users, while (even incomplete) rankings over three or more responses ensure identifiability. Second, we introduce methods to incorporate heterogeneous preferences into alignment algorithms. We develop an Expectation-Maximization adaptation of DPO that discovers latent annotator types and trains a mixture of LLMs accordingly. Then we propose an aggregation algorithm using a min-max regret fairness criterion to produce a single generative policy with equitable performance guarantees. Together, these contributions establish a theoretical and algorithmic framework for fairness and personalization for diverse users in generative model alignment.
--------------------------------------------------------------------------------------------------------
Adaptive Minds: Empowering Agents with LoRA-as-Tools
Adaptive Minds reimagines domain specialization in LLMs by treating LoRA adapters as dynamically selectable tools rather than static fine-tuned models. The base LLM acts as an intelligent router, analyzing queries semantically to select the most relevant domain expert on-the-fly. This architecture combines multi-agent flexibility with parameter-efficient training, enabling seamless transitions between specializations while maintaining conversational coherence. Built with LangGraph and fully open-source, applications include enterprise support systems routing between technical domains, legal AI accessing specialized practice areas, medical assistants switching between specialties, and educational platforms adapting to different subjects. The system offers scalable deployment for organizations requiring expertise across diverse domains without maintaining separate models, significantly reducing computational overhead while improving response accuracy.
Authors: Pavan C Shekar, Ashwanth Krishnan
Link: https://arxiv.org/abs/2510.15416v1
Date: 2025-10-d
Summary:
We present Adaptive Minds, an agentic system that treats LoRA adapters as domain-specific tools. Instead of relying on a single fine-tuned model or rigid rule-based routing, our approach empowers the base LLM itself to act as a semantic router analyzing each query and dynamically selecting the most relevant LoRA tool. This enables the agent to seamlessly switch between different domain experts on demand. By combining the flexibility of multi-agent orchestration with the efficiency of parameter-efficient fine-tuning, Adaptive Minds delivers accurate, specialized responses while preserving conversational ability. The system is built with LangGraph for workflow management, supports both API and web interfaces, and is fully open source, providing a scalable and extensible foundation for domain-adaptive AI assistance.
--------------------------------------------------------------------------------------------------------
This research extends kernel ridge regression theory beyond traditional independent and identically distributed assumptions to handle real-world data with structured dependencies. Focusing on signal-noise causal structures where multiple observations derive from shared underlying signals, the work develops novel blockwise decomposition methods for precise concentration analysis. The derived bounds explicitly account for kernel spectrum, causal structure parameters, and sampling mechanisms. Primary applications center on denoising score learning for generative models, providing principled guidance for sampling strategies in diffusion models and score-based generative systems. Additional applications include computer vision tasks with multiple noisy measurements (medical imaging, astronomical observations), financial modeling with correlated time-series data, and sensor networks where measurements share common underlying phenomena but contain independent noise.
Authors: Dechen Zhang, Zhenmei Shi, Yi Zhang, Yingyu Liang, Difan Zou
Link: https://arxiv.org/abs/2510.15363v1
Date: 2025-10-d
Summary:
Kernel ridge regression (KRR) is a foundational tool in machine learning, with recent work emphasizing its connections to neural networks. However, existing theory primarily addresses the i.i.d. setting, while real-world data often exhibits structured dependencies - particularly in applications like denoising score learning where multiple noisy observations derive from shared underlying signals. We present the first systematic study of KRR generalization for non-i.i.d. data with signal-noise causal structure, where observations represent different noisy views of common signals. By developing a novel blockwise decomposition method that enables precise concentration analysis for dependent data, we derive excess risk bounds for KRR that explicitly depend on: (1) the kernel spectrum, (2) causal structure parameters, and (3) sampling mechanisms (including relative sample sizes for signals and noises). We further apply our results to denoising score learning, establishing generalization guarantees and providing principled guidance for sampling noisy data points. This work advances KRR theory while providing practical tools for analyzing dependent data in modern machine learning applications.
--------------------------------------------------------------------------------------------------------
Identifying internal patterns in (1+1)-dimensional directed percolation using neural networks
This paper demonstrates deep learning's capability to automatically discover phase transitions and classify hidden patterns in complex physical systems without manual feature engineering. Combining CNN, TCN, and GRU architectures, the neural network learns directly from raw percolation configurations to reproduce phase diagrams and assign phase labels. This approach reveals that hierarchical deep architectures can extract meaningful structures from numerical experiment data automatically. Applications extend beyond physics to materials science for identifying structural phase transitions, epidemiological modeling for detecting critical thresholds in disease spread, network analysis for discovering percolation patterns in information diffusion, and climate science for identifying regime shifts. The methodology provides a template for applying deep learning to complex systems where theoretical understanding is incomplete or traditional analytical methods prove insufficient.
Authors: Danil Parkhomenko, Pavel Ovchinnikov, Konstantin Soldatov, Vitalii Kapitan, Gennady Y. Chitov
Link: https://arxiv.org/abs/2510.15294v1
Date: 2025-10-d
Summary:
In this paper we present a neural network-based method for the automatic detection of phase transitions and classification of hidden percolation patterns in a (1+1)-dimensional replication process. The proposed network model is based on the combination of CNN, TCN and GRU networks, which are trained directly on raw configurations without any manual feature extraction. The network reproduces the phase diagram and assigns phase labels to configurations. It shows that deep architectures are capable of extracting hierarchical structures from the raw data of numerical experiments.
--------------------------------------------------------------------------------------------------------
Structure-R1 addresses limitations in retrieval-augmented generation by transforming retrieved content into task-specific structured representations rather than using fixed schemas or unstructured text. Through reinforcement learning, the framework learns a content representation policy that dynamically generates optimal structural formats tailored to multi-step reasoning demands. A self-reward structural verification mechanism ensures correctness and completeness. With a 7B-parameter backbone matching larger models' performance, applications include knowledge-intensive question answering, technical documentation navigation, scientific literature review, legal research requiring complex reasoning chains, and enterprise knowledge management. The approach particularly benefits domains where information relationships and dependencies critically impact reasoning quality, offering improved information density and contextual clarity compared to traditional flat-text RAG systems.
Authors: Junlin Wu, Xianrui Zhong, Jiashuo Sun, Bolian Li, Bowen Jin, Jiawei Han, Qingkai Zeng
Link: https://arxiv.org/abs/2510.15191v1
Date: 2025-10-d
Summary:
Large language models (LLMs) have demonstrated remarkable advances in reasoning capabilities. However, their performance remains constrained by limited access to explicit and structured domain knowledge. Retrieval-Augmented Generation (RAG) addresses this by incorporating external information as context to augment reasoning. Nevertheless, traditional RAG systems typically operate over unstructured and fragmented text, resulting in low information density and suboptimal reasoning. To overcome these limitations, we propose \textsc{Structure-R1}, a novel framework that transforms retrieved content into structured representations optimized for reasoning. Leveraging reinforcement learning, \textsc{Structure-R1} learns a content representation policy that dynamically generates and adapts structural formats based on the demands of multi-step reasoning. Unlike prior methods that rely on fixed schemas, our approach adopts a generative paradigm capable of producing task-specific structures tailored to individual queries. To ensure the quality and reliability of these representations, we introduce a self-reward structural verification mechanism that checks whether the generated structures are both correct and self-contained. Extensive experiments on seven knowledge-intensive benchmarks show that \textsc{Structure-R1} consistently achieves competitive performance with a 7B-scale backbone model and matches the performance of much larger models. Additionally, our theoretical analysis demonstrates how structured representations enhance reasoning by improving information density and contextual clarity. Our code and data are available at: https://github.com/jlwu002/sr1.
--------------------------------------------------------------------------------------------------------
This large-scale study extends the Unified Theory of Acceptance and Use of Technology framework to understand AI adoption patterns among 2,257 professionals across global regions and organizational hierarchies. By reintroducing affective dimensions like attitude, self-efficacy, and anxiety, the research reveals that organizational level significantly predicts adoption, while emotional and cognitive responses vary modestly across contexts. Applications include organizational change management for AI deployment, designing targeted training programs addressing specific demographic concerns, developing equitable AI access policies, and informing human-centered AI design. The findings guide HR professionals implementing workplace AI, technology vendors designing adoption strategies, and policymakers ensuring inclusive digital transformation. Understanding these nuanced patterns helps organizations maximize AI benefits while addressing barriers to confident, sustainable engagement.
Authors: Diana Wolfe, Matt Price, Alice Choe, Fergus Kidd, Hannah Wagner
Link: https://arxiv.org/abs/2510.15142v1
Date: 2025-10-d
Summary:
This study investigates whether demographic factors shape adoption and attitudes among employees toward artificial intelligence (AI) technologies at work. Building on an extended Unified Theory of Acceptance and Use of Technology (UTAUT), which reintroduces affective dimensions such as attitude, self-efficacy, and anxiety, we surveyed 2,257 professionals across global regions and organizational levels within a multinational consulting firm. Non-parametric tests examined whether three demographic factors (i.e., years of experience, hierarchical level in the organization, and geographic region) were associated with AI adoption, usage intensity, and eight UTAUT constructs. Organizational level significantly predicted AI adoption, with senior employees showing higher usage rates, while experience and region were unrelated to adoption. Among AI users (n = 1,256), frequency and duration of use showed minimal demographic variation. However, omnibus tests revealed small but consistent group differences across several UTAUT constructs, particularly anxiety, performance expectancy, and behavioral intention, suggesting that emotional and cognitive responses to AI vary modestly across contexts. These findings highlight that demographic factors explain limited variance in AI acceptance but remain relevant for understanding contextual nuances in technology-related attitudes. The results underscore the need to integrate affective and organizational factors into models of technology acceptance to support equitable, confident, and sustainable engagement with AI in modern workplaces.
--------------------------------------------------------------------------------------------------------
DLER tackles the critical challenge of maximizing intelligence per token in reasoning language models that generate unnecessarily verbose outputs. Through refined reinforcement learning addressing advantage estimation bias, entropy collapse, and sparse rewards, DLER achieves dramatic efficiency gains—reducing output length by 70% while exceeding baseline accuracy. The framework introduces batch-wise reward normalization, dynamic sampling, and difficulty-aware adaptive truncation. Applications include cost-sensitive production deployments where inference costs scale with token generation, real-time systems requiring low-latency responses, mobile and edge AI where computational resources are constrained, and interactive applications where concise responses improve user experience. The update-selective merging method enables preserving baseline capabilities while gaining efficiency, making it practical for scenarios with limited retraining data or strict backward compatibility requirements.
Authors: Shih-Yang Liu, Xin Dong, Ximing Lu, Shizhe Diao, Mingjie Liu, Min-Hung Chen, Hongxu Yin, Yu-Chiang Frank Wang, Kwang-Ting Cheng, Yejin Choi, Jan Kautz, Pavlo Molchanov
Link: https://arxiv.org/abs/2510.15110v1
Date: 2025-10-d
Summary:
Reasoning language models such as OpenAI-o1, DeepSeek-R1, and Qwen achieve strong performance via extended chains of thought but often generate unnecessarily long outputs. Maximizing intelligence per token--accuracy relative to response length--remains an open problem. We revisit reinforcement learning (RL) with the simplest length penalty--truncation--and show that accuracy degradation arises not from the lack of sophisticated penalties but from inadequate RL optimization. We identify three key challenges: (i) large bias in advantage estimation, (ii) entropy collapse, and (iii) sparse reward signal. We address them with Doing Length pEnalty Right (DLER), a training recipe combining batch-wise reward normalization, higher clipping, dynamic sampling, and a simple truncation length penalty. DLER achieves state-of-the-art accuracy--efficiency trade-offs, cutting output length by over 70 percent while surpassing all previous baseline accuracy. It also improves test-time scaling: compared to DeepSeek-R1-7B, DLER-7B generates multiple concise responses in parallel with 28 percent higher accuracy and lower latency. We further introduce Difficulty-Aware DLER, which adaptively tightens truncation on easier questions for additional efficiency gains. We also propose an update-selective merging method that preserves baseline accuracy while retaining the concise reasoning ability of the DLER model, which is useful for scenarios where RL training data is scarce.
--------------------------------------------------------------------------------------------------------
C4D: 4D Made from 3D through Dual Correspondences
C4D advances dynamic scene reconstruction by extending static 3D methods to 4D, jointly estimating geometry, motion, and camera poses from monocular video. The framework captures both short-term optical flow and long-term point tracking, using a dynamic-aware tracker to separate moving objects from static backgrounds. By lifting 2D trajectories into smooth 3D paths through correspondence-based optimization, C4D achieves complete 4D recovery. Applications include autonomous driving for understanding dynamic environments, augmented reality requiring real-time scene reconstruction, robotics for manipulation in changing environments, visual effects and animation, sports analysis tracking player movements, and surveillance systems. The method's ability to handle both static and dynamic elements simultaneously makes it valuable for any computer vision application requiring temporal geometry understanding from standard video input.
Authors: Shizun Wang, Zhenxiang Jiang, Xingyi Yang, Xinchao Wang
Link: https://arxiv.org/abs/2510.14960v1
Date: 2025-10-d
Summary:
Recovering 4D from monocular video, which jointly estimates dynamic geometry and camera poses, is an inevitably challenging problem. While recent pointmap-based 3D reconstruction methods (e.g., DUSt3R) have made great progress in reconstructing static scenes, directly applying them to dynamic scenes leads to inaccurate results. This discrepancy arises because moving objects violate multi-view geometric constraints, disrupting the reconstruction. To address this, we introduce C4D, a framework that leverages temporal Correspondences to extend existing 3D reconstruction formulation to 4D. Specifically, apart from predicting pointmaps, C4D captures two types of correspondences: short-term optical flow and long-term point tracking. We train a dynamic-aware point tracker that provides additional mobility information, facilitating the estimation of motion masks to separate moving elements from the static background, thus offering more reliable guidance for dynamic scenes. Furthermore, we introduce a set of dynamic scene optimization objectives to recover per-frame 3D geometry and camera parameters. Simultaneously, the correspondences lift 2D trajectories into smooth 3D trajectories, enabling fully integrated 4D reconstruction. Experiments show that our framework achieves complete 4D recovery and demonstrates strong performance across multiple downstream tasks, including depth estimation, camera pose estimation, and point tracking. Project Page: https://littlepure2333.github.io/C4D
--------------------------------------------------------------------------------------------------------
Circuit Insights: Towards Interpretability Beyond Activations
This work advances mechanistic interpretability by moving beyond activation-based analysis to understand neural network circuits through learned weights and component interactions. WeightLens interprets features directly from weights without requiring datasets or external models, while CircuitLens captures how feature activations emerge from inter-component interactions. These complementary methods reveal circuit-level dynamics invisible to activation-only approaches. Applications include AI safety through understanding model decision-making mechanisms, debugging neural networks by identifying problematic circuits, model auditing for bias detection, automated red-teaming discovering adversarial vulnerabilities, and efficient model compression by identifying critical versus redundant circuits. The framework is particularly valuable for high-stakes domains like healthcare and finance where understanding model reasoning is essential, and for AI alignment research requiring mechanistic understanding of model behavior.
Authors: Elena Golimblevskaia, Aakriti Jain, Bruno Puri, Ammar Ibrahim, Wojciech Samek, Sebastian Lapuschkin
Link: https://arxiv.org/abs/2510.14936v1
Date: 2025-10-d
Summary:
The fields of explainable AI and mechanistic interpretability aim to uncover the internal structure of neural networks, with circuit discovery as a central tool for understanding model computations. Existing approaches, however, rely on manual inspection and remain limited to toy tasks. Automated interpretability offers scalability by analyzing isolated features and their activations, but it often misses interactions between features and depends strongly on external LLMs and dataset quality. Transcoders have recently made it possible to separate feature attributions into input-dependent and input-invariant components, providing a foundation for more systematic circuit analysis. Building on this, we propose WeightLens and CircuitLens, two complementary methods that go beyond activation-based analysis. WeightLens interprets features directly from their learned weights, removing the need for explainer models or datasets while matching or exceeding the performance of existing methods on context-independent features. CircuitLens captures how feature activations arise from interactions between components, revealing circuit-level dynamics that activation-only approaches cannot identify. Together, these methods increase interpretability robustness and enhance scalable mechanistic analysis of circuits while maintaining efficiency and quality.
--------------------------------------------------------------------------------------------------------
The Bidding Games: Reinforcement Learning for MEV Extraction on Polygon Blockchain
This research addresses strategic challenges in blockchain Maximal Extractable Value extraction on Polygon's Atlas auction system, where searchers must make optimal bidding decisions within sub-second windows without competitor information. The PPO-based framework includes a simulation environment modeling stochastic arbitrage opportunities and probabilistic competition, with agents achieving production-ready inference speeds. The history-conditioned agent captures 49% of available profits alongside existing searchers and 81% when replacing market leaders. Applications extend beyond MEV to high-frequency trading requiring rapid strategic decisions, algorithmic market making under uncertainty, real-time bidding in computational advertising, spectrum auctions with incomplete information, and any sealed-bid auction mechanism with extreme time constraints. The framework demonstrates reinforcement learning's advantages over traditional game theory in partially observable, high-frequency strategic environments.
Authors: Andrei Seoev, Leonid Gremyachikh, Anastasiia Smirnova, Yash Madhwal, Alisa Kalacheva, Dmitry Belousov, Ilia Zubov, Aleksei Smirnov, Denis Fedyanin, Vladimir Gorgadze, Yury Yanovich
Link: https://arxiv.org/abs/2510.14642v1
Date: 2025-10-d
Summary:
In blockchain networks, the strategic ordering of transactions within blocks has emerged as a significant source of profit extraction, known as Maximal Extractable Value (MEV). The transition from spam-based Priority Gas Auctions to structured auction mechanisms like Polygon Atlas has transformed MEV extraction from public bidding wars into sealed-bid competitions under extreme time constraints. While this shift reduces network congestion, it introduces complex strategic challenges where searchers must make optimal bidding decisions within a sub-second window without knowledge of competitor behavior or presence. Traditional game-theoretic approaches struggle in this high-frequency, partially observable environment due to their reliance on complete information and static equilibrium assumptions. We present a reinforcement learning framework for MEV extraction on Polygon Atlas and make three contributions: (1) A novel simulation environment that accurately models the stochastic arrival of arbitrage opportunities and probabilistic competition in Atlas auctions; (2) A PPO-based bidding agent optimized for real-time constraints, capable of adaptive strategy formulation in continuous action spaces while maintaining production-ready inference speeds; (3) Empirical validation demonstrating our history-conditioned agent captures 49\% of available profits when deployed alongside existing searchers and 81\% when replacing the market leader, significantly outperforming static bidding strategies. Our work establishes that reinforcement learning provides a critical advantage in high-frequency MEV environments where traditional optimization methods fail, offering immediate value for industrial participants and protocol designers alike.
--------------------------------------------------------------------------------------------------------
LLM Agents Beyond Utility: An Open-Ended Perspective
This exploratory work examines whether LLM agents can transcend tool-like utility to become entities capable of autonomous goal pursuit and self-directed exploration. By augmenting pretrained models with task generation, knowledge accumulation, and extensive environmental interaction capabilities, the research reveals both promise and limitations. Agents demonstrate reliable multi-step instruction following, cross-run knowledge reuse, and autonomous task proposal, but struggle with prompt sensitivity, repetitive task generation, and self-representation. Applications include autonomous research assistants continuously exploring knowledge domains, creative AI systems generating novel artistic directions, open-ended game NPCs with emergent behaviors, and robotic systems adapting to unpredictable environments. The findings inform future development of agents pursuing abstract long-term goals, managing memory effectively, and exploring productively beyond immediate objectives.
Authors: Asen Nachkov, Xi Wang, Luc Van Gool
Link: https://arxiv.org/abs/2510.14548v1
Date: 2025-10-d
Summary:
Recent LLM agents have made great use of chain of thought reasoning and function calling. As their capabilities grow, an important question arises: can this software represent not only a smart problem-solving tool, but an entity in its own right, that can plan, design immediate tasks, and reason toward broader, more ambiguous goals? To study this question, we adopt an open-ended experimental setting where we augment a pretrained LLM agent with the ability to generate its own tasks, accumulate knowledge, and interact extensively with its environment. We study the resulting open-ended agent qualitatively. It can reliably follow complex multi-step instructions, store and reuse information across runs, and propose and solve its own tasks, though it remains sensitive to prompt design, prone to repetitive task generation, and unable to form self-representations. These findings illustrate both the promise and current limits of adapting pretrained LLMs toward open-endedness, and point to future directions for training agents to manage memory, explore productively, and pursue abstract long-term goals.
--------------------------------------------------------------------------------------------------------
IMAGINE: Integrating Multi-Agent System into One Model for Complex Reasoning and Planning
IMAGINE addresses the inefficiency of multi-agent systems—high reasoning costs, long latency, and training difficulties—by distilling their collective reasoning and planning capabilities into single compact models. Through end-to-end training, an 8B-parameter model achieves 82.7% accuracy on TravelPlanner, far exceeding the 40% of much larger models like DeepSeek-R1-671B. This demonstrates that structured reasoning patterns from multi-agent collaboration can be internalized efficiently. Applications include complex task planning and scheduling, multi-hop question answering, strategic game playing, automated workflow design, and resource allocation optimization. The framework is particularly valuable for edge deployment where computational resources are limited, real-time applications requiring low latency, and scenarios where multi-agent communication overhead is prohibitive, democratizing sophisticated reasoning capabilities through accessible model sizes.
Authors: Xikai Zhang, Bo Wang, Likang Xiao, Yongzhi Li, Quan Chen, Wenju Wu, Liu Liu
Link: https://arxiv.org/abs/2510.14406v1
Date: 2025-10-d
Summary:
Although large language models (LLMs) have made significant strides across various tasks, they still face significant challenges in complex reasoning and planning. For example, even with carefully designed prompts and prior information explicitly provided, GPT-4o achieves only a 7% Final Pass Rate on the TravelPlanner dataset in the sole-planning mode. Similarly, even in the thinking mode, Qwen3-8B-Instruct and DeepSeek-R1-671B, only achieve Final Pass Rates of 5.9% and 40%, respectively. Although well-organized Multi-Agent Systems (MAS) can offer improved collective reasoning, they often suffer from high reasoning costs due to multi-round internal interactions, long per-response latency, and difficulties in end-to-end training. To address these challenges, we propose a general and scalable framework called IMAGINE, short for Integrating Multi-Agent System into One Model. This framework not only integrates the reasoning and planning capabilities of MAS into a single, compact model, but also significantly surpass the capabilities of the MAS through a simple end-to-end training. Through this pipeline, a single small-scale model is not only able to acquire the structured reasoning and planning capabilities of a well-organized MAS but can also significantly outperform it. Experimental results demonstrate that, when using Qwen3-8B-Instruct as the base model and training it with our method, the model achieves an 82.7% Final Pass Rate on the TravelPlanner benchmark, far exceeding the 40% of DeepSeek-R1-671B, while maintaining a much smaller model size.
--------------------------------------------------------------------------------------------------------
Rethinking Toxicity Evaluation in Large Language Models: A Multi-Label Perspective
This work addresses fundamental limitations in current toxicity detection by introducing multi-label benchmarks (Q-A-MLL, R-A-MLL, H-X-MLL) capturing the inherently ambiguous and multi-dimensional nature of toxic content across 15 fine-grained categories. The research provides theoretical proof that pseudo-label training outperforms single-label supervision, with the proposed method significantly surpassing GPT-4o and DeepSeek. Applications include content moderation systems requiring nuanced categorization, automated comment filtering for social platforms, AI safety evaluation tools, training data curation for responsible AI, and compliance monitoring for regulatory requirements. The multi-label approach better reflects real-world toxicity complexity, enabling more accurate detection while reducing both false positives and missed detections. This is essential for platforms balancing safety with free expression and organizations deploying LLMs in sensitive contexts.
Authors: Zhiqiang Kou, Junyang Chen, Xin-Qiang Cai, Ming-Kun Xie, Biao Liu, Changwei Wang, Lei Feng, Yuheng Jia, Gang Niu, Masashi Sugiyama, Xin Geng
Link: https://arxiv.org/abs/2510.15007v1
Date: 2025-10-d
Summary:
Large language models (LLMs) have achieved impressive results across a range of natural language processing tasks, but their potential to generate harmful content has raised serious safety concerns. Current toxicity detectors primarily rely on single-label benchmarks, which cannot adequately capture the inherently ambiguous and multi-dimensional nature of real-world toxic prompts. This limitation results in biased evaluations, including missed toxic detections and false positives, undermining the reliability of existing detectors. Additionally, gathering comprehensive multi-label annotations across fine-grained toxicity categories is prohibitively costly, further hindering effective evaluation and development. To tackle these issues, we introduce three novel multi-label benchmarks for toxicity detection: \textbf{Q-A-MLL}, \textbf{R-A-MLL}, and \textbf{H-X-MLL}, derived from public toxicity datasets and annotated according to a detailed 15-category taxonomy. We further provide a theoretical proof that, on our released datasets, training with pseudo-labels yields better performance than directly learning from single-label supervision. In addition, we develop a pseudo-label-based toxicity detection method. Extensive experimental results show that our approach significantly surpasses advanced baselines, including GPT-4o and DeepSeek, thus enabling more accurate and reliable evaluation of multi-label toxicity in LLM-generated content.
--------------------------------------------------------------------------------------------------------
This research develops a highly accurate machine learning pipeline for early Alzheimer's detection through language analysis, achieving 91% accuracy and 97% AUC by combining hybrid word embeddings from Doc2Vec and ELMo with linguistic features. The method captures sentence fluency through perplexity scores while analyzing syntax and semantics, with stability validated across repeated experiments. Applications include large-scale population screening for cognitive decline, complementary diagnostic tools supporting clinical assessment, monitoring disease progression through longitudinal language analysis, early intervention enabling symptom management, and reducing healthcare costs through timely detection. The non-invasive nature makes it suitable for regular monitoring in elderly populations, telemedicine platforms, and community health programs, potentially transforming Alzheimer's detection from clinical examination to accessible computational screening.
Authors: Yangyang Li
Link: https://arxiv.org/abs/2510.14332v1
Date: 2025-10-d
Summary:
Early detection of Alzheimer's Disease (AD) is greatly beneficial to AD patients, leading to early treatments that lessen symptoms and alleviating financial burden of health care. As one of the leading signs of AD, language capability changes can be used for early diagnosis of AD. In this paper, I develop a robust classification method using hybrid word embedding and fine-tuned hyperparameters to achieve state-of-the-art accuracy in the early detection of AD. Specifically, we create a hybrid word embedding based on word vectors from Doc2Vec and ELMo to obtain perplexity scores of the sentences. The scores identify whether a sentence is fluent or not and capture semantic context of the sentences. I enrich the word embedding by adding linguistic features to analyze syntax and semantics. Further, we input an embedded feature vector into logistic regression and fine tune hyperparameters throughout the pipeline. By tuning hyperparameters of the machine learning pipeline (e.g., model regularization parameter, learning rate and vector size of Doc2Vec, and vector size of ELMo), I achieve 91% classification accuracy and an Area Under the Curve (AUC) of 97% in distinguishing early AD from healthy subjects. Based on my knowledge, my model with 91% accuracy and 97% AUC outperforms the best existing NLP model for AD diagnosis with an accuracy of 88% [32]. I study the model stability through repeated experiments and find that the model is stable even though the training data is split randomly (standard deviation of accuracy = 0.0403; standard deviation of AUC = 0.0174). This affirms our proposed method is accurate and stable. This model can be used as a large-scale screening method for AD, as well as a complementary examination for doctors to detect AD.
--------------------------------------------------------------------------------------------------------
This training-free approach tackles hallucination in Large Vision-Language Models through tri-layer contrastive decoding with watermarking. The method selects mature and amateur decoding layers, identifies a pivot layer using watermark-related questions to assess visual grounding, and applies contrastive decoding for final output generation. Achieving state-of-the-art hallucination reduction on POPE, MME, and AMBER benchmarks, the technique ensures more visually grounded responses. Applications include medical image analysis requiring accurate visual interpretation, autonomous vehicle perception systems, visual question answering in educational contexts, content moderation verifying image-text correspondence, and assistive technologies for visually impaired users. The training-free nature enables immediate deployment to existing models without retraining, making it practical for production systems requiring improved factuality and visual grounding.
Authors: Kyungryul Back, Seongbeom Park, Milim Kim, Mincheol Kwon, SangHyeok Lee, Hyunyoung Lee, Junhee Cho, Seunghyun Park, Jinkyu Kim
Link: https://arxiv.org/abs/2510.14304v1
Date: 2025-10-d
Summary:
Large Vision-Language Models (LVLMs) have recently shown promising results on various multimodal tasks, even achieving human-comparable performance in certain cases. Nevertheless, LVLMs remain prone to hallucinations -- they often rely heavily on a single modality or memorize training data without properly grounding their outputs. To address this, we propose a training-free, tri-layer contrastive decoding with watermarking, which proceeds in three steps: (1) select a mature layer and an amateur layer among the decoding layers, (2) identify a pivot layer using a watermark-related question to assess whether the layer is visually well-grounded, and (3) apply tri-layer contrastive decoding to generate the final output. Experiments on public benchmarks such as POPE, MME and AMBER demonstrate that our method achieves state-of-the-art performance in reducing hallucinations in LVLMs and generates more visually grounded responses.
--------------------------------------------------------------------------------------------------------
DPRF optimizes behavioral alignment between LLM role-playing agents and target individuals by iteratively identifying cognitive divergences through free-form or theory-grounded analysis and refining persona profiles accordingly. Evaluated across formal debates, social media, interviews, and reviews with five LLMs, DPRF consistently improves behavioral fidelity across models and scenarios. Applications include user simulation for product testing and user experience research, social science studies modeling individual or demographic behaviors, personalized AI assistants adapting to individual communication styles, synthetic data generation for training recommendation systems, and entertainment applications creating believable character interactions. The framework is particularly valuable for studying human behavior computationally, testing systems under diverse user conditions, and developing highly personalized AI experiences while maintaining validity and methodological rigor.
Authors: Bingsheng Yao, Bo Sun, Yuanzhe Dong, Yuxuan Lu, Dakuo Wang
Link: https://arxiv.org/abs/2510.14205v1
Date: 2025-10-d
Summary:
The emerging large language model role-playing agents (LLM RPAs) aim to simulate individual human behaviors, but the persona fidelity is often undermined by manually-created profiles (e.g., cherry-picked information and personality characteristics) without validating the alignment with the target individuals. To address this limitation, our work introduces the Dynamic Persona Refinement Framework (DPRF).DPRF aims to optimize the alignment of LLM RPAs' behaviors with those of target individuals by iteratively identifying the cognitive divergence, either through free-form or theory-grounded, structured analysis, between generated behaviors and human ground truth, and refining the persona profile to mitigate these divergences.We evaluate DPRF with five LLMs on four diverse behavior-prediction scenarios: formal debates, social media posts with mental health issues, public interviews, and movie reviews.DPRF can consistently improve behavioral alignment considerably over baseline personas and generalizes across models and scenarios.Our work provides a robust methodology for creating high-fidelity persona profiles and enhancing the validity of downstream applications, such as user simulation, social studies, and personalized AI.
--------------------------------------------------------------------------------------------------------
Implementation of AI in Precision Medicine
This scoping review synthesizes 2019-2024 literature on implementing AI in precision medicine, identifying critical barriers and enablers across data quality, clinical reliability, workflow integration, and governance dimensions. While AI shows immense potential for integrating and interpreting multimodal patient data, real-world clinical adoption remains limited. The ecosystem-based framework reveals interdependent relationships shaping successful translation. Applications include personalized treatment planning based on genomic and clinical data, drug discovery targeting specific patient subpopulations, disease risk prediction from multi-source health records, clinical decision support systems, and population health management. The review informs healthcare administrators implementing AI systems, researchers developing clinically viable algorithms, and policymakers establishing governance frameworks, ultimately supporting trustworthy and sustainable AI integration that transforms precision medicine from promise to practice.
Authors: Göktuğ Bender, Samer Faraj, Anand Bhardwaj
Link: https://arxiv.org/abs/2510.14194v1
Date: 2025-10-d
Summary:
Artificial intelligence (AI) has become increasingly central to precision medicine by enabling the integration and interpretation of multimodal data, yet implementation in clinical settings remains limited. This paper provides a scoping review of literature from 2019-2024 on the implementation of AI in precision medicine, identifying key barriers and enablers across data quality, clinical reliability, workflow integration, and governance. Through an ecosystem-based framework, we highlight the interdependent relationships shaping real-world translation and propose future directions to support trustworthy and sustainable implementation.
--------------------------------------------------------------------------------------------------------