LLM 관련 주요 논문 - 2025-09-12

1. The Illusion of Diminishing Returns: Measuring Long Horizon Execution in LLMs

• Authors: Akshit Sinha , Arvindh Arun , Shashwat Goel , Steffen Staab , Jonas Geiping
• URL: https://arxiv.org/abs/2509.09677
• Abstract:

  Does continued scaling of large language models (LLMs) yield diminishing returns? Real-world value often stems from the length of task an agent can complete. We start this work by observing the simple but counterintuitive fact that marginal gains in single-step accuracy can compound into exponential improvements in the length of a task a model can successfully complete. Then, we argue that failures of LLMs when simple tasks are made longer arise from mistakes in execution, rather than an inability to reason. We propose isolating execution capability, by explicitly providing the knowledge and plan needed to solve a long-horizon task. We find that larger models can correctly execute significantly more turns even when small models have 100\% single-turn accuracy. We observe that the per-step accuracy of models degrades as the number of steps increases. This is not just due to long-context limitations – curiously, we observe a self-conditioning effect – models become more likely to make mistakes when the context contains their errors from prior turns. Self-conditioning does not reduce by just scaling the model size. In contrast, recent thinking models do not self-condition, and can also execute much longer tasks in a single turn. We conclude by benchmarking frontier thinking models on the length of task they can execute in a single turn. Overall, by focusing on the ability to execute, we hope to reconcile debates on how LLMs can solve complex reasoning problems yet fail at simple tasks when made longer, and highlight the massive benefits of scaling model size and sequential test-time compute for long-horizon tasks.

---

2. Compositional Concept Generalization with Variational Quantum Circuits

• Authors: Hala Hawashin , Mina Abbaszadeh , Nicholas Joseph , Beth Pearson , Martha Lewis , Mehrnoosh sadrzadeh
• URL: https://arxiv.org/abs/2509.09541
• Abstract:

  Compositional generalization is a key facet of human cognition, but lacking in current AI tools such as vision-language models. Previous work examined whether a compositional tensor-based sentence semantics can overcome the challenge, but led to negative results. We conjecture that the increased training efficiency of quantum models will improve performance in these tasks. We interpret the representations of compositional tensor-based models in Hilbert spaces and train Variational Quantum Circuits to learn these representations on an image captioning task requiring compositional generalization. We used two image encoding techniques: a multi-hot encoding (MHE) on binary image vectors and an angle/amplitude encoding on image vectors taken from the vision-language model CLIP. We achieve good proof-of-concept results using noisy MHE encodings. Performance on CLIP image vectors was more mixed, but still outperformed classical compositional models.

---

3. TORSO: Template-Oriented Reasoning Towards General Tasks

• Authors: Minhyuk Kim , Seungyoon Lee , Heuiseok Lim
• URL: https://arxiv.org/abs/2509.09448
• Abstract:

  The approaches that guide Large Language Models (LLMs) to emulate human reasoning during response generation have emerged as an effective method for enabling them to solve complex problems in a step-by-step manner, thereby achieving superior performance. However, most existing approaches using few-shot prompts to generate responses heavily depend on the provided examples, limiting the utilization of the model’s inherent reasoning capabilities. Moreover, constructing task-specific few-shot prompts is often costly and may lead to inconsistencies across different tasks. In this work, we introduce Template-Oriented Reasoning (TORSO), which elicits the model to utilize internal reasoning abilities to generate proper responses across various tasks without the need for manually crafted few-shot examples. Our experimental results demonstrate that TORSO achieves strong performance on diverse LLMs benchmarks with reasonable rationales.

---

4. Curriculum-Based Multi-Tier Semantic Exploration via Deep Reinforcement Learning

• Authors: Abdel Hakim Drid , Vincenzo Suriani , Daniele Nardi , Abderrezzak Debilou
• URL: https://arxiv.org/abs/2509.09356
• Abstract:

  Navigating and understanding complex and unknown environments autonomously demands more than just basic perception and movement from embodied agents. Truly effective exploration requires agents to possess higher-level cognitive abilities, the ability to reason about their surroundings, and make more informed decisions regarding exploration strategies. However, traditional RL approaches struggle to balance efficient exploration and semantic understanding due to limited cognitive capabilities embedded in the small policies for the agents, leading often to human drivers when dealing with semantic exploration. In this paper, we address this challenge by presenting a novel Deep Reinforcement Learning (DRL) architecture that is specifically designed for resource efficient semantic exploration. A key methodological contribution is the integration of a Vision-Language Model (VLM) common-sense through a layered reward function. The VLM query is modeled as a dedicated action, allowing the agent to strategically query the VLM only when deemed necessary for gaining external guidance, thereby conserving resources. This mechanism is combined with a curriculum learning strategy designed to guide learning at different levels of complexity to ensure robust and stable learning. Our experimental evaluation results convincingly demonstrate that our agent achieves significantly enhanced object discovery rates and develops a learned capability to effectively navigate towards semantically rich regions. Furthermore, it also shows a strategic mastery of when to prompt for external environmental information. By demonstrating a practical and scalable method for embedding common-sense semantic reasoning with autonomous agents, this research provides a novel approach to pursuing a fully intelligent and self-guided exploration in robotics.

---

5. Towards Adaptive ML Benchmarks: Web-Agent-Driven Construction, Domain Expansion, and Metric Optimization

• Authors: Hangyi Jia , Yuxi Qian , Hanwen Tong , Xinhui Wu , Lin Chen , Feng Wei
• URL: https://arxiv.org/abs/2509.09321
• Abstract:

  Recent advances in large language models (LLMs) have enabled the emergence of general-purpose agents for automating end-to-end machine learning (ML) workflows, including data analysis, feature engineering, model training, and competition solving. However, existing benchmarks remain limited in task coverage, domain diversity, difficulty modeling, and evaluation rigor, failing to capture the full capabilities of such agents in realistic settings. We present TAM Bench, a diverse, realistic, and structured benchmark for evaluating LLM-based agents on end-to-end ML tasks. TAM Bench features three key innovations: (1) A browser automation and LLM-based task acquisition system that automatically collects and structures ML challenges from platforms such as Kaggle, AIcrowd, and Biendata, spanning multiple task types and data modalities (e.g., tabular, text, image, graph, audio); (2) A leaderboard-driven difficulty modeling mechanism that estimates task complexity using participant counts and score dispersion, enabling scalable and objective task calibration; (3) A multi-dimensional evaluation framework incorporating performance, format compliance, constraint adherence, and task generalization. Based on 150 curated AutoML tasks, we construct three benchmark subsets of different sizes – Lite, Medium, and Full – designed for varying evaluation scenarios. The Lite version, with 18 tasks and balanced coverage across modalities and difficulty levels, serves as a practical testbed for daily benchmarking and comparative studies.

---

6. LightAgent: Production-level Open-source Agentic AI Framework

• Authors: Weige Cai , Tong Zhu , Jinyi Niu , Ruiqi Hu , Lingyao Li , Tenglong Wang , Xiaowu Dai , Weining Shen , Liwen Zhang
• URL: https://arxiv.org/abs/2509.09292
• Abstract:

  With the rapid advancement of large language models (LLMs), Multi-agent Systems (MAS) have achieved significant progress in various application scenarios. However, substantial challenges remain in designing versatile, robust, and efficient platforms for agent deployment. To address these limitations, we propose \textbf{LightAgent}, a lightweight yet powerful agentic framework, effectively resolving the trade-off between flexibility and simplicity found in existing frameworks. LightAgent integrates core functionalities such as Memory (mem0), Tools, and Tree of Thought (ToT), while maintaining an extremely lightweight structure. As a fully open-source solution, it seamlessly integrates with mainstream chat platforms, enabling developers to easily build self-learning agents. We have released LightAgent at \href{ this https URL }{ this https URL }

---

7. Tree-OPO: Off-policy Monte Carlo Tree-Guided Advantage Optimization for Multistep Reasoning

• Authors: Bingning Huang , Tu Nguyen , Matthieu Zimmer
• URL: https://arxiv.org/abs/2509.09284
• Abstract:

  Recent advances in reasoning with large language models (LLMs) have shown the effectiveness of Monte Carlo Tree Search (MCTS) for generating high-quality intermediate trajectories, particularly in math and symbolic domains. Inspired by this, we explore how MCTS-derived trajectories, traditionally used for training value or reward models, can be repurposed to improve policy optimization in preference-based reinforcement learning (RL). Specifically, we focus on Group Relative Policy Optimization (GRPO), a recent algorithm that enables preference-consistent policy learning without value networks. We propose a staged GRPO training paradigm where completions are derived from partially revealed MCTS rollouts, introducing a novel tree-structured setting for advantage estimation. This leads to a rich class of prefix-conditioned reward signals, which we analyze theoretically and empirically. Our initial results indicate that while structured advantage estimation can stabilize updates and better reflect compositional reasoning quality, challenges such as advantage saturation and reward signal collapse remain. We propose heuristic and statistical solutions to mitigate these issues and discuss open challenges for learning under staged or tree-like reward structures.

---

8. Fusing Knowledge and Language: A Comparative Study of Knowledge Graph-Based Question Answering with LLMs

• Authors: Vaibhav Chaudhary , Neha Soni , Narotam Singh , Amita Kapoor
• URL: https://arxiv.org/abs/2509.09272
• Abstract:

  Knowledge graphs, a powerful tool for structuring information through relational triplets, have recently become the new front-runner in enhancing question-answering systems. While traditional Retrieval Augmented Generation (RAG) approaches are proficient in fact-based and local context-based extraction from concise texts, they encounter limitations when addressing the thematic and holistic understanding of complex, extensive texts, requiring a deeper analysis of both text and context. This paper presents a comprehensive technical comparative study of three different methodologies for constructing knowledge graph triplets and integrating them with Large Language Models (LLMs) for question answering: spaCy, Stanford CoreNLP-OpenIE, and GraphRAG, all leveraging open source technologies. We evaluate the effectiveness, feasibility, and adaptability of these methods by analyzing their capabilities, state of development, and their impact on the performance of LLM-based question answering. Experimental results indicate that while OpenIE provides the most comprehensive coverage of triplets, GraphRAG demonstrates superior reasoning abilities among the three. We conclude with a discussion on the strengths and limitations of each method and provide insights into future directions for improving knowledge graph-based question answering.

---

9. Jupiter: Enhancing LLM Data Analysis Capabilities via Notebook and Inference-Time Value-Guided Search

• Authors: Shuocheng Li , Yihao Liu , Silin Du , Wenxuan Zeng , Zhe Xu , Mengyu Zhou , Yeye He , Haoyu Dong , Shi Han , Dongmei Zhang
• URL: https://arxiv.org/abs/2509.09245
• Abstract:

  Large language models (LLMs) have shown great promise in automating data science workflows, but existing models still struggle with multi-step reasoning and tool use, which limits their effectiveness on complex data analysis tasks. To address this, we propose a scalable pipeline that extracts high-quality, tool-based data analysis tasks and their executable multi-step solutions from real-world Jupyter notebooks and associated data files. Using this pipeline, we introduce NbQA, a large-scale dataset of standardized task-solution pairs that reflect authentic tool-use patterns in practical data science scenarios. To further enhance multi-step reasoning, we present Jupiter, a framework that formulates data analysis as a search problem and applies Monte Carlo Tree Search (MCTS) to generate diverse solution trajectories for value model learning. During inference, Jupiter combines the value model and node visit counts to efficiently collect executable multi-step plans with minimal search steps. Experimental results show that Qwen2.5-7B and 14B-Instruct models on NbQA solve 77.82% and 86.38% of tasks on InfiAgent-DABench, respectively-matching or surpassing GPT-4o and advanced agent frameworks. Further evaluations demonstrate improved generalization and stronger tool-use reasoning across diverse multi-step reasoning tasks.

---

10. Enabling Regulatory Multi-Agent Collaboration: Architecture, Challenges, and Solutions

• Authors: Qinnan Hu , Yuntao Wang , Yuan Gao , Zhou Su , Linkang Du
• URL: https://arxiv.org/abs/2509.09215
• Abstract:

  Large language models (LLMs)-empowered autonomous agents are transforming both digital and physical environments by enabling adaptive, multi-agent collaboration. While these agents offer significant opportunities across domains such as finance, healthcare, and smart manufacturing, their unpredictable behaviors and heterogeneous capabilities pose substantial governance and accountability challenges. In this paper, we propose a blockchain-enabled layered architecture for regulatory agent collaboration, comprising an agent layer, a blockchain data layer, and a regulatory application layer. Within this framework, we design three key modules: (i) an agent behavior tracing and arbitration module for automated accountability, (ii) a dynamic reputation evaluation module for trust assessment in collaborative scenarios, and (iii) a malicious behavior forecasting module for early detection of adversarial activities. Our approach establishes a systematic foundation for trustworthy, resilient, and scalable regulatory mechanisms in large-scale agent ecosystems. Finally, we discuss the future research directions for blockchain-enabled regulatory frameworks in multi-agent systems.

---

11. Understanding Economic Tradeoffs Between Human and AI Agents in Bargaining Games

• Authors: Crystal Qian , Kehang Zhu , John Horton , Benjamin S. Manning , Vivian Tsai , James Wexler , Nithum Thain
• URL: https://arxiv.org/abs/2509.09071
• Abstract:

  Coordination tasks traditionally performed by humans are increasingly being delegated to autonomous agents. As this pattern progresses, it becomes critical to evaluate not only these agents’ performance but also the processes through which they negotiate in dynamic, multi-agent environments. Furthermore, different agents exhibit distinct advantages: traditional statistical agents, such as Bayesian models, may excel under well-specified conditions, whereas large language models (LLMs) can generalize across contexts. In this work, we compare humans (N = 216), LLMs (GPT-4o, Gemini 1.5 Pro), and Bayesian agents in a dynamic negotiation setting that enables direct, identical-condition comparisons across populations, capturing both outcomes and behavioral dynamics. Bayesian agents extract the highest surplus through aggressive optimization, at the cost of frequent trade rejections. Humans and LLMs can achieve similar overall surplus, but through distinct behaviors: LLMs favor conservative, concessionary trades with few rejections, while humans employ more strategic, risk-taking, and fairness-oriented behaviors. Thus, we find that performance parity – a common benchmark in agent evaluation – can conceal fundamental differences in process and alignment, which are critical for practical deployment in real-world coordination tasks.

---

12. Instructional Prompt Optimization for Few-Shot LLM-Based Recommendations on Cold-Start Users

• Authors: Haowei Yang , Yushang Zhao , Sitao Min , Bo Su , Chao Yao , Wei Xu
• URL: https://arxiv.org/abs/2509.09066
• Abstract:

  The cold-start user issue further compromises the effectiveness of recommender systems in limiting access to the historical behavioral information. It is an effective pipeline to optimize instructional prompts on a few-shot large language model (LLM) used in recommender tasks. We introduce a context-conditioned prompt formulation method P(u,\ Ds)\ \rightarrow\ R\widehat, where u is a cold-start user profile, Ds is a curated support set, and R\widehat is the predicted ranked list of items. Based on systematic experimentation with transformer-based autoregressive LLMs (BioGPT, LLaMA-2, GPT-4), we provide empirical evidence that optimal exemplar injection and instruction structuring can significantly improve the precision@k and NDCG scores of such models in low-data settings. The pipeline uses token-level alignments and embedding space regularization with a greater semantic fidelity. Our findings not only show that timely composition is not merely syntactic but also functional as it is in direct control of attention scales and decoder conduct through inference. This paper shows that prompt-based adaptation may be considered one of the ways to address cold-start recommendation issues in LLM-based pipelines.

---

13. Global Constraint LLM Agents for Text-to-Model Translation

• Authors: Junyang Cai , Serdar Kadioglu , Bistra Dilkina
• URL: https://arxiv.org/abs/2509.08970
• Abstract:

  Natural language descriptions of optimization or satisfaction problems are challenging to translate into correct MiniZinc models, as this process demands both logical reasoning and constraint programming expertise. We introduce a framework that addresses this challenge with an agentic approach: multiple specialized large language model (LLM) agents decompose the modeling task by global constraint type. Each agent is dedicated to detecting and generating code for a specific class of global constraint, while a final assembler agent integrates these constraint snippets into a complete MiniZinc model. By dividing the problem into smaller, well-defined sub-tasks, each LLM handles a simpler reasoning challenge, potentially reducing overall complexity. We conduct initial experiments with several LLMs and show better performance against baselines such as one-shot prompting and chain-of-thought prompting. Finally, we outline a comprehensive roadmap for future work, highlighting potential enhancements and directions for improvement.

---

14. Automated Unity Game Template Generation from GDDs via NLP and Multi-Modal LLMs

• Authors: Amna Hassan
• URL: https://arxiv.org/abs/2509.08847
• Abstract:

  This paper presents a novel framework for automated game template generation by transforming Game Design Documents (GDDs) into functional Unity game prototypes using Natural Language Processing (NLP) and multi-modal Large Language Models (LLMs). We introduce an end-to-end system that parses GDDs, extracts structured game specifications, and synthesizes Unity-compatible C# code that implements the core mechanics, systems, and architecture defined in the design documentation. Our approach combines a fine-tuned LLaMA-3 model specialized for Unity code generation with a custom Unity integration package that streamlines the implementation process. Evaluation results demonstrate significant improvements over baseline models, with our fine-tuned model achieving superior performance (4.8/5.0 average score) compared to state-of-the-art LLMs across compilation success, GDD adherence, best practices adoption, and code modularity metrics. The generated templates demonstrate high adherence to GDD specifications across multiple game genres. Our system effectively addresses critical gaps in AI-assisted game development, positioning LLMs as valuable tools in streamlining the transition from game design to implementation.

---

15. ButterflyQuant: Ultra-low-bit LLM Quantization through Learnable Orthogonal Butterfly Transforms

• Authors: Bingxin Xu , Zhen Dong , Oussama Elachqar , Yuzhang Shang
• URL: https://arxiv.org/abs/2509.09679
• Abstract:

  Large language models require massive memory footprints, severely limiting deployment on consumer hardware. Quantization reduces memory through lower numerical precision, but extreme 2-bit quantization suffers from catastrophic performance loss due to outliers in activations. Rotation-based methods such as QuIP and QuaRot apply orthogonal transforms to eliminate outliers before quantization, using computational invariance: $\mathbf{y} = \mathbf{Wx} = (\mathbf{WQ}^T)(\mathbf{Qx})$ for orthogonal $\mathbf{Q}$. However, these methods use fixed transforms–Hadamard matrices achieving optimal worst-case coherence $\mu = 1/\sqrt{n}$–that cannot adapt to specific weight distributions. We identify that different transformer layers exhibit distinct outlier patterns, motivating layer-adaptive rotations rather than one-size-fits-all approaches. We propose ButterflyQuant, which replaces Hadamard rotations with learnable butterfly transforms parameterized by continuous Givens rotation angles. Unlike Hadamard’s discrete ${+1, -1}$ entries that are non-differentiable and prohibit gradient-based learning, butterfly transforms’ continuous parameterization enables smooth optimization while guaranteeing orthogonality by construction. This orthogonal constraint ensures theoretical guarantees in outlier suppression while achieving $O(n \log n)$ computational complexity with only $\frac{n \log n}{2}$ learnable parameters. We further introduce a uniformity regularization on post-transformation activations to promote smoother distributions amenable to quantization. Learning requires only 128 calibration samples and converges in minutes on a single GPU–a negligible one-time cost. On LLaMA-2-7B with 2-bit quantization, ButterflyQuant achieves 15.4 perplexity versus 22.1 for QuaRot.

---

16. CDE: Curiosity-Driven Exploration for Efficient Reinforcement Learning in Large Language Models

• Authors: Runpeng Dai , Linfeng Song , Haolin Liu , Zhenwen Liang , Dian Yu , Haitao Mi , Zhaopeng Tu , Rui Liu , Tong Zheng , Hongtu Zhu , Dong Yu
• URL: https://arxiv.org/abs/2509.09675
• Abstract:

  Reinforcement Learning with Verifiable Rewards (RLVR) is a powerful paradigm for enhancing the reasoning ability of Large Language Models (LLMs). Yet current RLVR methods often explore poorly, leading to premature convergence and entropy collapse. To address this challenge, we introduce Curiosity-Driven Exploration (CDE), a framework that leverages the model’s own intrinsic sense of curiosity to guide exploration. We formalize curiosity with signals from both the actor and the critic: for the actor, we use perplexity over its generated response, and for the critic, we use the variance of value estimates from a multi-head architecture. Both signals serve as an exploration bonus within the RLVR framework to guide the model. Our theoretical analysis shows that the actor-wise bonus inherently penalizes overconfident errors and promotes diversity among correct responses; moreover, we connect the critic-wise bonus to the well-established count-based exploration bonus in RL. Empirically, our method achieves an approximate +3 point improvement over standard RLVR using GRPO/PPO on AIME benchmarks. Further analysis identifies a calibration collapse mechanism within RLVR, shedding light on common LLM failure modes.

---

17. LoCoBench: A Benchmark for Long-Context Large Language Models in Complex Software Engineering

• Authors: Jielin Qiu , Zuxin Liu , Zhiwei Liu , Rithesh Murthy , Jianguo Zhang , Haolin Chen , Shiyu Wang , Ming Zhu , Liangwei Yang , Juntao Tan , Zhepeng Cen , Cheng Qian , Shelby Heinecke , Weiran Yao , Silvio Savarese , Caiming Xiong , Huan Wang
• URL: https://arxiv.org/abs/2509.09614
• Abstract:

  The emergence of long-context language models with context windows extending to millions of tokens has created new opportunities for sophisticated code understanding and software development evaluation. We propose LoCoBench, a comprehensive benchmark specifically designed to evaluate long-context LLMs in realistic, complex software development scenarios. Unlike existing code evaluation benchmarks that focus on single-function completion or short-context tasks, LoCoBench addresses the critical evaluation gap for long-context capabilities that require understanding entire codebases, reasoning across multiple files, and maintaining architectural consistency across large-scale software systems. Our benchmark provides 8,000 evaluation scenarios systematically generated across 10 programming languages, with context lengths spanning 10K to 1M tokens, a 100x variation that enables precise assessment of long-context performance degradation in realistic software development settings. LoCoBench introduces 8 task categories that capture essential long-context capabilities: architectural understanding, cross-file refactoring, multi-session development, bug investigation, feature implementation, code comprehension, integration testing, and security analysis. Through a 5-phase pipeline, we create diverse, high-quality scenarios that challenge LLMs to reason about complex codebases at unprecedented scale. We introduce a comprehensive evaluation framework with 17 metrics across 4 dimensions, including 8 new evaluation metrics, combined in a LoCoBench Score (LCBS). Our evaluation of state-of-the-art long-context models reveals substantial performance gaps, demonstrating that long-context understanding in complex software development represents a significant unsolved challenge that demands more attention. LoCoBench is released at: this https URL .

---

18. Fluent but Unfeeling: The Emotional Blind Spots of Language Models

• Authors: Bangzhao Shu , Isha Joshi , Melissa Karnaze , Anh C. Pham , Ishita Kakkar , Sindhu Kothe , Arpine Hovasapian , Mai ElSherief
• URL: https://arxiv.org/abs/2509.09593
• Abstract:

  The versatility of Large Language Models (LLMs) in natural language understanding has made them increasingly popular in mental health research. While many studies explore LLMs’ capabilities in emotion recognition, a critical gap remains in evaluating whether LLMs align with human emotions at a fine-grained level. Existing research typically focuses on classifying emotions into predefined, limited categories, overlooking more nuanced expressions. To address this gap, we introduce EXPRESS, a benchmark dataset curated from Reddit communities featuring 251 fine-grained, self-disclosed emotion labels. Our comprehensive evaluation framework examines predicted emotion terms and decomposes them into eight basic emotions using established emotion theories, enabling a fine-grained comparison. Systematic testing of prevalent LLMs under various prompt settings reveals that accurately predicting emotions that align with human self-disclosed emotions remains challenging. Qualitative analysis further shows that while certain LLMs generate emotion terms consistent with established emotion theories and definitions, they sometimes fail to capture contextual cues as effectively as human self-disclosures. These findings highlight the limitations of LLMs in fine-grained emotion alignment and offer insights for future research aimed at enhancing their contextual understanding.

---

19. ENSI: Efficient Non-Interactive Secure Inference for Large Language Models

• Authors: Zhiyu He , Maojiang Wang , Xinwen Gao , Yuchuan Luo , Lin Liu , Shaojing Fu
• URL: https://arxiv.org/abs/2509.09424
• Abstract:

  Secure inference enables privacy-preserving machine learning by leveraging cryptographic protocols that support computations on sensitive user data without exposing it. However, integrating cryptographic protocols with large language models (LLMs) presents significant challenges, as the inherent complexity of these protocols, together with LLMs’ massive parameter scale and sophisticated architectures, severely limits practical usability. In this work, we propose ENSI, a novel non-interactive secure inference framework for LLMs, based on the principle of co-designing the cryptographic protocols and LLM architecture. ENSI employs an optimized encoding strategy that seamlessly integrates CKKS scheme with a lightweight LLM variant, BitNet, significantly reducing the computational complexity of encrypted matrix multiplications. In response to the prohibitive computational demands of softmax under homomorphic encryption (HE), we pioneer the integration of the sigmoid attention mechanism with HE as a seamless, retraining-free alternative. Furthermore, by embedding the Bootstrapping operation within the RMSNorm process, we efficiently refresh ciphertexts while markedly decreasing the frequency of costly bootstrapping invocations. Experimental evaluations demonstrate that ENSI achieves approximately an 8x acceleration in matrix multiplications and a 2.6x speedup in softmax inference on CPU compared to state-of-the-art method, with the proportion of bootstrapping is reduced to just 1%.

---

20. LLMs Don’t Know Their Own Decision Boundaries: The Unreliability of Self-Generated Counterfactual Explanations

• Authors: Harry Mayne , Ryan Othniel Kearns , Yushi Yang , Andrew M. Bean , Eoin Delaney , Chris Russell , Adam Mahdi
• URL: https://arxiv.org/abs/2509.09396
• Abstract:

  To collaborate effectively with humans, language models must be able to explain their decisions in natural language. We study a specific type of self-explanation: self-generated counterfactual explanations (SCEs), where a model explains its prediction by modifying the input such that it would have predicted a different outcome. We evaluate whether LLMs can produce SCEs that are valid, achieving the intended outcome, and minimal, modifying the input no more than necessary. When asked to generate counterfactuals, we find that LLMs typically produce SCEs that are valid, but far from minimal, offering little insight into their decision-making behaviour. Worryingly, when asked to generate minimal counterfactuals, LLMs typically make excessively small edits that fail to change predictions. The observed validity-minimality trade-off is consistent across several LLMs, datasets, and evaluation settings. Our findings suggest that SCEs are, at best, an ineffective explainability tool and, at worst, can provide misleading insights into model behaviour. Proposals to deploy LLMs in high-stakes settings must consider the impact of unreliable self-explanations on downstream decision-making. Our code is available at this https URL .

---

21. MetaLLMix : An XAI Aided LLM-Meta-learning Based Approach for Hyper-parameters Optimization

• Authors: Mohammed Tiouti , Mohamed Bal-Ghaoui
• URL: https://arxiv.org/abs/2509.09387
• Abstract:

  Effective model and hyperparameter selection remains a major challenge in deep learning, often requiring extensive expertise and computation. While AutoML and large language models (LLMs) promise automation, current LLM-based approaches rely on trial and error and expensive APIs, which provide limited interpretability and generalizability. We propose MetaLLMiX, a zero-shot hyperparameter optimization framework combining meta-learning, explainable AI, and efficient LLM reasoning. By leveraging historical experiment outcomes with SHAP explanations, MetaLLMiX recommends optimal hyperparameters and pretrained models without additional trials. We further employ an LLM-as-judge evaluation to control output format, accuracy, and completeness. Experiments on eight medical imaging datasets using nine open-source lightweight LLMs show that MetaLLMiX achieves competitive or superior performance to traditional HPO methods while drastically reducing computational cost. Our local deployment outperforms prior API-based approaches, achieving optimal results on 5 of 8 tasks, response time reductions of 99.6-99.9%, and the fastest training times on 6 datasets (2.4-15.7x faster), maintaining accuracy within 1-5% of best-performing baselines.

---

22. Can Multimodal LLMs See Materials Clearly? A Multimodal Benchmark on Materials Characterization

• Authors: Zhengzhao Lai , Youbin Zheng , Zhenyang Cai , Haonan Lyu , Jinpu Yang , Hongqing Liang , Yan Hu , Benyou Wang
• URL: https://arxiv.org/abs/2509.09307
• Abstract:

  Materials characterization is fundamental to acquiring materials information, revealing the processing-microstructure-property relationships that guide material design and optimization. While multimodal large language models (MLLMs) have recently shown promise in generative and predictive tasks within materials science, their capacity to understand real-world characterization imaging data remains underexplored. To bridge this gap, we present MatCha, the first benchmark for materials characterization image understanding, comprising 1,500 questions that demand expert-level domain expertise. MatCha encompasses four key stages of materials research comprising 21 distinct tasks, each designed to reflect authentic challenges faced by materials scientists. Our evaluation of state-of-the-art MLLMs on MatCha reveals a significant performance gap compared to human experts. These models exhibit degradation when addressing questions requiring higher-level expertise and sophisticated visual perception. Simple few-shot and chain-of-thought prompting struggle to alleviate these limitations. These findings highlight that existing MLLMs still exhibit limited adaptability to real-world materials characterization scenarios. We hope MatCha will facilitate future research in areas such as new material discovery and autonomous scientific agents. MatCha is available at this https URL .

---

23. On Integrating Large Language Models and Scenario-Based Programming for Improving Software Reliability

• Authors: Ayelet Berzack , Guy Katz
• URL: https://arxiv.org/abs/2509.09194
• Abstract:

  Large Language Models (LLMs) are fast becoming indispensable tools for software developers, assisting or even partnering with them in crafting complex programs. The advantages are evident – LLMs can significantly reduce development time, generate well-organized and comprehensible code, and occasionally suggest innovative ideas that developers might not conceive on their own. However, despite their strengths, LLMs will often introduce significant errors and present incorrect code with persuasive confidence, potentially misleading developers into accepting flawed solutions. In order to bring LLMs into the software development cycle in a more reliable manner, we propose a methodology for combining them with ``traditional’’ software engineering techniques in a structured way, with the goal of streamlining the development process, reducing errors, and enabling users to verify crucial program properties with increased confidence. Specifically, we focus on the Scenario-Based Programming (SBP) paradigm – an event-driven, scenario-based approach for software engineering – to allow human developers to pour their expert knowledge into the LLM, as well as to inspect and verify its outputs. To evaluate our methodology, we conducted a significant case study, and used it to design and implement the Connect4 game. By combining LLMs and SBP we were able to create a highly-capable agent, which could defeat various strong existing agents. Further, in some cases, we were able to formally verify the correctness of our agent. Finally, our experience reveals interesting insights regarding the ease-of-use of our proposed approach. The full code of our case-study will be made publicly available with the final version of this paper.

---

24. Probing Pre-trained Language Models on Code Changes: Insights from ReDef, a High-Confidence Just-in-Time Defect Prediction Dataset

• Authors: Doha Nam , Taehyoun Kim , Duksan Ryu , Jongmoon Baik
• URL: https://arxiv.org/abs/2509.09192
• Abstract:

  Just-in-Time software defect prediction (JIT-SDP) plays a critical role in prioritizing risky code changes during code review and continuous integration. However, existing datasets often suffer from noisy labels and low precision in identifying bug-inducing commits. To address this, we present ReDef (Revert-based Defect dataset), a high-confidence benchmark of function-level modifications curated from 22 large-scale C/C++ projects. Defective cases are anchored by revert commits, while clean cases are validated through post-hoc history checks. Ambiguous instances are conservatively filtered out via a GPT-assisted triage process involving multiple votes and audits. This pipeline yields 3,164 defective and 10,268 clean modifications, offering substantially more reliable labels than prior existing resources. Beyond dataset construction, we provide the first systematic evaluation of how pre-trained language models (PLMs) reason about code modifications – specifically, which input encodings most effectively expose change information, and whether models genuinely capture edit semantics. We fine-tune CodeBERT, CodeT5+, and UniXcoder under five encoding strategies, and further probe their sensitivity through counterfactual perturbations that swap added/deleted blocks, invert diff polarity, or inject spurious markers. Our results show that compact diff-style encodings consistently outperform whole-function formats across all PLMs, with statistical tests confirming large, model-independent effects. However, under counterfactual tests, performance degrades little or not at all – revealing that what appears to be robustness in fact reflects reliance on superficial cues rather than true semantic understanding. These findings indicate that, unlike in snapshot-based tasks, current PLMs remain limited in their ability to genuinely comprehend code modifications.

---

25. EchoX: Towards Mitigating Acoustic-Semantic Gap via Echo Training for Speech-to-Speech LLMs

• Authors: Yuhao Zhang , Yuhao Du , Zhanchen Dai , Xiangnan Ma , Kaiqi Kou , Benyou Wang , Haizhou Li
• URL: https://arxiv.org/abs/2509.09174
• Abstract:

  Speech-to-speech large language models (SLLMs) are attracting increasing attention. Derived from text-based large language models (LLMs), SLLMs often exhibit degradation in knowledge and reasoning capabilities. We hypothesize that this limitation arises because current training paradigms for SLLMs fail to bridge the acoustic-semantic gap in the feature representation space. To address this issue, we propose EchoX, which leverages semantic representations and dynamically generates speech training targets. This approach integrates both acoustic and semantic learning, enabling EchoX to preserve strong reasoning abilities as a speech LLM. Experimental results demonstrate that EchoX, with about six thousand hours of training data, achieves advanced performance on multiple knowledge-based question-answering benchmarks. The project is available at this https URL .

---

26. Character-Level Perturbations Disrupt LLM Watermarks

• Authors: Zhaoxi Zhang , Xiaomei Zhang , Yanjun Zhang , He Zhang , Shirui Pan , Bo Liu , Asif Qumer Gill , Leo Yu Zhang
• URL: https://arxiv.org/abs/2509.09112
• Abstract:

  Large Language Model (LLM) watermarking embeds detectable signals into generated text for copyright protection, misuse prevention, and content detection. While prior studies evaluate robustness using watermark removal attacks, these methods are often suboptimal, creating the misconception that effective removal requires large perturbations or powerful adversaries. To bridge the gap, we first formalize the system model for LLM watermark, and characterize two realistic threat models constrained on limited access to the watermark detector. We then analyze how different types of perturbation vary in their attack range, i.e., the number of tokens they can affect with a single edit. We observe that character-level perturbations (e.g., typos, swaps, deletions, homoglyphs) can influence multiple tokens simultaneously by disrupting the tokenization process. We demonstrate that character-level perturbations are significantly more effective for watermark removal under the most restrictive threat model. We further propose guided removal attacks based on the Genetic Algorithm (GA) that uses a reference detector for optimization. Under a practical threat model with limited black-box queries to the watermark detector, our method demonstrates strong removal performance. Experiments confirm the superiority of character-level perturbations and the effectiveness of the GA in removing watermarks under realistic constraints. Additionally, we argue there is an adversarial dilemma when considering potential defenses: any fixed defense can be bypassed by a suitable perturbation strategy. Motivated by this principle, we propose an adaptive compound character-level attack. Experimental results show that this approach can effectively defeat the defenses. Our findings highlight significant vulnerabilities in existing LLM watermark schemes and underline the urgency for the development of new robust mechanisms.

---

27. DP-FedLoRA: Privacy-Enhanced Federated Fine-Tuning for On-Device Large Language Models

• Authors: Honghui Xu , Shiva Shrestha , Wei Chen , Zhiyuan Li , Zhipeng Cai
• URL: https://arxiv.org/abs/2509.09097
• Abstract:

  As on-device large language model (LLM) systems become increasingly prevalent, federated fine-tuning enables advanced language understanding and generation directly on edge devices; however, it also involves processing sensitive, user-specific data, raising significant privacy concerns within the federated learning framework. To address these challenges, we propose DP-FedLoRA, a privacy-enhanced federated fine-tuning framework that integrates LoRA-based adaptation with differential privacy in a communication-efficient setting. Each client locally clips and perturbs its LoRA matrices using Gaussian noise to satisfy ($\epsilon$, $\delta$)-differential privacy. We further provide a theoretical analysis demonstrating the unbiased nature of the updates and deriving bounds on the variance introduced by noise, offering practical guidance for privacy-budget calibration. Experimental results across mainstream benchmarks show that DP-FedLoRA delivers competitive performance while offering strong privacy guarantees, paving the way for scalable and privacy-preserving LLM deployment in on-device environments.

---

28. Towards Confidential and Efficient LLM Inference with Dual Privacy Protection

• Authors: Honglan Yu , Yibin Wang , Feifei Dai , Dong Liu , Haihui Fan , Xiaoyan Gu
• URL: https://arxiv.org/abs/2509.09091
• Abstract:

  CPU-based trusted execution environments (TEEs) and differential privacy (DP) have gained wide applications for private inference. Due to high inference latency in TEEs, researchers use partition-based approaches that offload linear model components to GPUs. However, dense nonlinear layers of large language models (LLMs) result in significant communication overhead between TEEs and GPUs. DP-based approaches apply random noise to protect data privacy, but this compromises LLM performance and semantic understanding. To overcome the above drawbacks, this paper proposes CMIF, a Confidential and efficient Model Inference Framework. CMIF confidentially deploys the embedding layer in the client-side TEE and subsequent layers on GPU servers. Meanwhile, it optimizes the Report-Noisy-Max mechanism to protect sensitive inputs with a slight decrease in model performance. Extensive experiments on Llama-series models demonstrate that CMIF reduces additional inference overhead in TEEs while preserving user data privacy.

---

29. Improving LLM Safety and Helpfulness using SFT and DPO: A Study on OPT-350M

• Authors: Piyush Pant
• URL: https://arxiv.org/abs/2509.09055
• Abstract:

  This research investigates the effectiveness of alignment techniques, Supervised Fine-Tuning (SFT), Direct Preference Optimization (DPO), and a combined SFT+DPO approach on improving the safety and helpfulness of the OPT-350M language model. Utilizing the Anthropic Helpful-Harmless RLHF dataset, we train and evaluate four models: the base OPT350M, an SFT model, a DPO model, and a model trained with both SFT and DPO. We introduce three key evaluation metrics: Harmlessness Rate (HmR), Helpfulness Rate (HpR), and a Combined Alignment Score (CAS), all derived from reward model outputs. The results show that while SFT outperforms DPO, The combined SFT+DPO model outperforms all others across all metrics, demonstrating the complementary nature of these techniques. Our findings also highlight challenges posed by noisy data, limited GPU resources, and training constraints. This study offers a comprehensive view of how fine-tuning strategies affect model alignment and provides a foundation for more robust alignment pipelines in future work.

---

30. Stated Preference for Interaction and Continued Engagement (SPICE): Evaluating an LLM’s Willingness to Re-engage in Conversation

• Authors: Thomas Manuel Rost , Martina Figlia , Bernd Wallraff
• URL: https://arxiv.org/abs/2509.09043
• Abstract:

  We introduce and evaluate Stated Preference for Interaction and Continued Engagement (SPICE), a simple diagnostic signal elicited by asking a Large Language Model a YES or NO question about its willingness to re-engage with a user’s behavior after reviewing a short transcript. In a study using a 3-tone (friendly, unclear, abusive) by 10-interaction stimulus set, we tested four open-weight chat models across four framing conditions, resulting in 480 trials. Our findings show that SPICE sharply discriminates by user tone. Friendly interactions yielded a near-unanimous preference to continue (97.5% YES), while abusive interactions yielded a strong preference to discontinue (17.9% YES), with unclear interactions falling in between (60.4% YES). This core association remains decisive under multiple dependence-aware statistical tests, including Rao-Scott adjustment and cluster permutation tests. Furthermore, we demonstrate that SPICE provides a distinct signal from abuse classification. In trials where a model failed to identify abuse, it still overwhelmingly stated a preference not to continue the interaction (81% of the time). An exploratory analysis also reveals a significant interaction effect: a preamble describing the study context significantly impacts SPICE under ambiguity, but only when transcripts are presented as a single block of text rather than a multi-turn chat. The results validate SPICE as a robust, low-overhead, and reproducible tool for auditing model dispositions, complementing existing metrics by offering a direct, relational signal of a model’s state. All stimuli, code, and analysis scripts are released to support replication.

---

31. Can Vision-Language Models Solve Visual Math Equations?

• Authors: Monjoy Narayan Choudhury , Junling Wang , Yifan Hou , Mrinmaya Sachan
• URL: https://arxiv.org/abs/2509.09013
• Abstract:

  Despite strong performance in visual understanding and language-based reasoning, Vision-Language Models (VLMs) struggle with tasks requiring integrated perception and symbolic computation. We study this limitation through visual equation solving, where mathematical equations are embedded in images, variables are represented by object icons, and coefficients must be inferred by counting. While VLMs perform well on textual equations, they fail on visually grounded counterparts. To understand this gap, we decompose the task into coefficient counting and variable recognition, and find that counting is the primary bottleneck, even when recognition is accurate. We also observe that composing recognition and reasoning introduces additional errors, highlighting challenges in multi-step visual reasoning. Finally, as equation complexity increases, symbolic reasoning itself becomes a limiting factor. These findings reveal key weaknesses in current VLMs and point toward future improvements in visually grounded mathematical reasoning.

---

32. Open-sci-ref-0.01: open and reproducible reference baselines for language model and dataset comparison

• Authors: Marianna Nezhurina , Taishi Nakamura , Timur Carstensen , Niccolò Ajroldi , Ville Komulainen , David Salinas , Jenia Jitsev
• URL: https://arxiv.org/abs/2509.09009
• Abstract:

  We introduce open-sci-ref, a family of dense transformer models trained as research baselines across multiple model (0.13B to 1.7B parameters) and token scales (up to 1T) on 8 recent open reference datasets. Evaluating the models on various standardized benchmarks, our training runs set establishes reference points that enable researchers to assess the sanity and quality of alternative training approaches across scales and datasets. Intermediate checkpoints allow comparison and studying of the training dynamics. The established reference baselines allow training procedures to be compared through their scaling trends, aligning them on a common compute axis. Comparison of open reference datasets reveals that training on NemoTron-CC HQ consistently outperforms other reference datasets, followed by DCLM-baseline and FineWeb-Edu. In addition to intermediate training checkpoints, the release includes logs, code, and downstream evaluations to simplify reproduction, standardize comparison, and facilitate future research.

---

33. PromptGuard: An Orchestrated Prompting Framework for Principled Synthetic Text Generation for Vulnerable Populations using LLMs with Enhanced Safety, Fairness, and Controllability

• Authors: Tung Vu , Lam Nguyen , Quynh Dao
• URL: https://arxiv.org/abs/2509.08910
• Abstract:

  The proliferation of Large Language Models (LLMs) in real-world applications poses unprecedented risks of generating harmful, biased, or misleading information to vulnerable populations including LGBTQ+ individuals, single parents, and marginalized communities. While existing safety approaches rely on post-hoc filtering or generic alignment techniques, they fail to proactively prevent harmful outputs at the generation source. This paper introduces PromptGuard, a novel modular prompting framework with our breakthrough contribution: VulnGuard Prompt, a hybrid technique that prevents harmful information generation using real-world data-driven contrastive learning. VulnGuard integrates few-shot examples from curated GitHub repositories, ethical chain-of-thought reasoning, and adaptive role-prompting to create population-specific protective barriers. Our framework employs theoretical multi-objective optimization with formal proofs demonstrating 25-30% analytical harm reduction through entropy bounds and Pareto optimality. PromptGuard orchestrates six core modules: Input Classification, VulnGuard Prompting, Ethical Principles Integration, External Tool Interaction, Output Validation, and User-System Interaction, creating an intelligent expert system for real-time harm prevention. We provide comprehensive mathematical formalization including convergence proofs, vulnerability analysis using information theory, and theoretical validation framework using GitHub-sourced datasets, establishing mathematical foundations for systematic empirical research.

---

34. Recurrence Meets Transformers for Universal Multimodal Retrieval

• Authors: Davide Caffagni , Sara Sarto , Marcella Cornia , Lorenzo Baraldi , Rita Cucchiara
• URL: https://arxiv.org/abs/2509.08897
• Abstract:

  With the rapid advancement of multimodal retrieval and its application in LLMs and multimodal LLMs, increasingly complex retrieval tasks have emerged. Existing methods predominantly rely on task-specific fine-tuning of vision-language models and are limited to single-modality queries or documents. In this paper, we propose ReT-2, a unified retrieval model that supports multimodal queries, composed of both images and text, and searches across multimodal document collections where text and images coexist. ReT-2 leverages multi-layer representations and a recurrent Transformer architecture with LSTM-inspired gating mechanisms to dynamically integrate information across layers and modalities, capturing fine-grained visual and textual details. We evaluate ReT-2 on the challenging M2KR and M-BEIR benchmarks across different retrieval configurations. Results demonstrate that ReT-2 consistently achieves state-of-the-art performance across diverse settings, while offering faster inference and reduced memory usage compared to prior approaches. When integrated into retrieval-augmented generation pipelines, ReT-2 also improves downstream performance on Encyclopedic-VQA and InfoSeek datasets. Our source code and trained models are publicly available at: this https URL

---

35. Benchmarking Energy Efficiency of Large Language Models Using vLLM

• Authors: K. Pronk , Q. Zhao
• URL: https://arxiv.org/abs/2509.08867
• Abstract:

  The prevalence of Large Language Models (LLMs) is having an growing impact on the climate due to the substantial energy required for their deployment and use. To create awareness for developers who are implementing LLMs in their products, there is a strong need to collect more information about the energy efficiency of LLMs. While existing research has evaluated the energy efficiency of various models, these benchmarks often fall short of representing realistic production scenarios. In this paper, we introduce the LLM Efficiency Benchmark, designed to simulate real-world usage conditions. Our benchmark utilizes vLLM, a high-throughput, production-ready LLM serving backend that optimizes model performance and efficiency. We examine how factors such as model size, architecture, and concurrent request volume affect inference energy efficiency. Our findings demonstrate that it is possible to create energy efficiency benchmarks that better reflect practical deployment conditions, providing valuable insights for developers aiming to build more sustainable AI systems.

---

36. Investigating Student Interaction Patterns with Large Language Model-Powered Course Assistants in Computer Science Courses

• Authors: Chang Liu , Loc Hoang , Andrew Stolman , Rene F. Kizilcec , Bo Wu
• URL: https://arxiv.org/abs/2509.08862
• Abstract:

  Providing students with flexible and timely academic support is a challenge at most colleges and universities, leaving many students without help outside scheduled hours. Large language models (LLMs) are promising for bridging this gap, but interactions between students and LLMs are rarely overseen by educators. We developed and studied an LLM-powered course assistant deployed across multiple computer science courses to characterize real-world use and understand pedagogical implications. By Spring 2024, our system had been deployed to approximately 2,000 students across six courses at three institutions. Analysis of the interaction data shows that usage remains strong in the evenings and nights and is higher in introductory courses, indicating that our system helps address temporal support gaps and novice learner needs. We sampled 200 conversations per course for manual annotation: most sampled responses were judged correct and helpful, with a small share unhelpful or erroneous; few responses included dedicated examples. We also examined an inquiry-based learning strategy: only around 11% of sampled conversations contained LLM-generated follow-up questions, which were often ignored by students in advanced courses. A Bloom’s taxonomy analysis reveals that current LLM capabilities are limited in generating higher-order cognitive questions. These patterns suggest opportunities for pedagogically oriented LLM-based educational systems and greater educator involvement in configuring prompts, content, and policies.

---

37. PerFairX: Is There a Balance Between Fairness and Personality in Large Language Model Recommendations?

• Authors: Chandan Kumar Sah
• URL: https://arxiv.org/abs/2509.08829
• Abstract:

  The integration of Large Language Models (LLMs) into recommender systems has enabled zero-shot, personality-based personalization through prompt-based interactions, offering a new paradigm for user-centric recommendations. However, incorporating user personality traits via the OCEAN model highlights a critical tension between achieving psychological alignment and ensuring demographic fairness. To address this, we propose PerFairX, a unified evaluation framework designed to quantify the trade-offs between personalization and demographic equity in LLM-generated recommendations. Using neutral and personality-sensitive prompts across diverse user profiles, we benchmark two state-of-the-art LLMs, ChatGPT and DeepSeek, on movie (MovieLens 10M) and music ( this http URL 360K) datasets. Our results reveal that personality-aware prompting significantly improves alignment with individual traits but can exacerbate fairness disparities across demographic groups. Specifically, DeepSeek achieves stronger psychological fit but exhibits higher sensitivity to prompt variations, while ChatGPT delivers stable yet less personalized outputs. PerFairX provides a principled benchmark to guide the development of LLM-based recommender systems that are both equitable and psychologically informed, contributing to the creation of inclusive, user-centric AI applications in continual learning contexts.

---