LLM 관련 주요 논문 - 2025-12-12

1. SCOPE: Language Models as One-Time Teacher for Hierarchical Planning in Text Environments

Authors: Haoye Lu , Pavan Seshadri , Kaheer Suleman
URL: https://arxiv.org/abs/2512.09897
Abstract:

Long-term planning in complex, text-based environments presents significant challenges due to open-ended action spaces, ambiguous observations, and sparse feedback. Recent research suggests that large language models (LLMs) encode rich semantic knowledge about the world, which can be valuable for guiding agents in high-level reasoning and planning across both embodied and purely textual settings. However, existing approaches often depend heavily on querying LLMs during training and inference, making them computationally expensive and difficult to deploy efficiently. In addition, these methods typically employ a pretrained, unaltered LLM whose parameters remain fixed throughout training, providing no opportunity for adaptation to the target task. To address these limitations, we introduce SCOPE (Subgoal-COnditioned Pretraining for Efficient planning), a one-shot hierarchical planner that leverages LLM-generated subgoals only at initialization to pretrain a lightweight student model. Unlike prior approaches that distill LLM knowledge by repeatedly prompting the model to adaptively generate subgoals during training, our method derives subgoals directly from example trajectories. This design removes the need for repeated LLM queries, significantly improving efficiency, though at the cost of reduced explainability and potentially suboptimal subgoals. Despite their suboptimality, our results on the TextCraft environment show that LLM-generated subgoals can still serve as a strong starting point for hierarchical goal decomposition in text-based planning tasks. Compared to the LLM-based hierarchical agent ADaPT (Prasad et al., 2024), which achieves a 0.52 success rate, our method reaches 0.56 and reduces inference time from 164.4 seconds to just 3.0 seconds.

2. RIFT: A Scalable Methodology for LLM Accelerator Fault Assessment using Reinforcement Learning

Authors: Khurram Khalil , Muhammad Mahad Khaliq , Khaza Anuarul Hoque
URL: https://arxiv.org/abs/2512.09829
Abstract:

The massive scale of modern AI accelerators presents critical challenges to traditional fault assessment methodologies, which face prohibitive computational costs and provide poor coverage of critical failure modes. This paper introduces RIFT (Reinforcement Learning-guided Intelligent Fault Targeting), a scalable framework that automates the discovery of minimal, high-impact fault scenarios for efficient design-time fault assessment. RIFT transforms the complex search for worst-case faults into a sequential decision-making problem, combining hybrid sensitivity analysis for search space pruning with reinforcement learning to intelligently generate minimal, high-impact test suites. Evaluated on billion-parameter Large Language Model (LLM) workloads using NVIDIA A100 GPUs, RIFT achieves a \textbf{2.2$\times$} fault assessment speedup over evolutionary methods and reduces the required test vector volume by over \textbf{99\%} compared to random fault injection, all while achieving \textbf{superior fault coverage}. The proposed framework also provides actionable data to enable intelligent hardware protection strategies, demonstrating that RIFT-guided selective error correction code provides a \textbf{12.8$\times$} improvement in \textbf{cost-effectiveness} (coverage per unit area) compared to uniform triple modular redundancy protection. RIFT automatically generates UVM-compliant verification artifacts, ensuring its findings are directly actionable and integrable into commercial RTL verification workflows.

3. An End-to-end Planning Framework with Agentic LLMs and PDDL

Authors: Emanuele La Malfa , Ping Zhu , Samuele Marro , Sara Bernardini , Michael Wooldridge
URL: https://arxiv.org/abs/2512.09629
Abstract:

We present an end-to-end framework for planning supported by verifiers. An orchestrator receives a human specification written in natural language and converts it into a PDDL (Planning Domain Definition Language) model, where the domain and problem are iteratively refined by sub-modules (agents) to address common planning requirements, such as time constraints and optimality, as well as ambiguities and contradictions that may exist in the human specification. The validated domain and problem are then passed to an external planning engine to generate a plan. The orchestrator and agents are powered by Large Language Models (LLMs) and require no human intervention at any stage of the process. Finally, a module translates the final plan back into natural language to improve human readability while maintaining the correctness of each step. We demonstrate the flexibility and effectiveness of our framework across various domains and tasks, including the Google NaturalPlan benchmark and PlanBench, as well as planning problems like Blocksworld and the Tower of Hanoi (where LLMs are known to struggle even with small instances). Our framework can be integrated with any PDDL planning engine and validator (such as Fast Downward, LPG, POPF, VAL, and uVAL, which we have tested) and represents a significant step toward end-to-end planning aided by LLMs.

4. Toward Closed-loop Molecular Discovery via Language Model, Property Alignment and Strategic Search

Authors: Junkai Ji , Zhangfan Yang , Dong Xu , Ruibin Bai , Jianqiang Li , Tingjun Hou , Zexuan Zhu
URL: https://arxiv.org/abs/2512.09566
Abstract:

Drug discovery is a time-consuming and expensive process, with traditional high-throughput and docking-based virtual screening hampered by low success rates and limited scalability. Recent advances in generative modelling, including autoregressive, diffusion, and flow-based approaches, have enabled de novo ligand design beyond the limits of enumerative screening. Yet these models often suffer from inadequate generalization, limited interpretability, and an overemphasis on binding affinity at the expense of key pharmacological properties, thereby restricting their translational utility. Here we present Trio, a molecular generation framework integrating fragment-based molecular language modeling, reinforcement learning, and Monte Carlo tree search, for effective and interpretable closed-loop targeted molecular design. Through the three key components, Trio enables context-aware fragment assembly, enforces physicochemical and synthetic feasibility, and guides a balanced search between the exploration of novel chemotypes and the exploitation of promising intermediates within protein binding pockets. Experimental results show that Trio reliably achieves chemically valid and pharmacologically enhanced ligands, outperforming state-of-the-art approaches with improved binding affinity (+7.85%), drug-likeness (+11.10%) and synthetic accessibility (+12.05%), while expanding molecular diversity more than fourfold.

5. SDialog: A Python Toolkit for End-to-End Agent Building, User Simulation, Dialog Generation, and Evaluation

Authors: Sergio Burdisso , Séverin Baroudi , Yanis Labrak , David Grunert , Pawel Cyrta , Yiyang Chen , Srikanth Madikeri , Esaú Villatoro-Tello , Thomas Schaaf , Ricard Marxer , Petr Motlicek
URL: https://arxiv.org/abs/2512.09142
Abstract:

We present SDialog, an MIT-licensed open-source Python toolkit that unifies dialog generation, evaluation and mechanistic interpretability into a single end-to-end framework for building and analyzing LLM-based conversational agents. Built around a standardized \texttt{Dialog} representation, SDialog provides: (1) persona-driven multi-agent simulation with composable orchestration for controlled, synthetic dialog generation, (2) comprehensive evaluation combining linguistic metrics, LLM-as-a-judge and functional correctness validators, (3) mechanistic interpretability tools for activation inspection and steering via feature ablation and induction, and (4) audio generation with full acoustic simulation including 3D room modeling and microphone effects. The toolkit integrates with all major LLM backends, enabling mixed-backend experiments under a unified API. By coupling generation, evaluation, and interpretability in a dialog-centric architecture, SDialog enables researchers to build, benchmark and understand conversational systems more systematically.

6. A Categorical Analysis of Large Language Models and Why LLMs Circumvent the Symbol Grounding Problem

Authors: Luciano Floridi , Yiyang Jia , Fernando Tohmé
URL: https://arxiv.org/abs/2512.09117
Abstract:

This paper presents a formal, categorical framework for analysing how humans and large language models (LLMs) transform content into truth-evaluated propositions about a state space of possible worlds W , in order to argue that LLMs do not solve but circumvent the symbol grounding problem.

7. Calibrated Trust in Dealing with LLM Hallucinations: A Qualitative Study

Authors: Adrian Ryser , Florian Allwein , Tim Schlippe
URL: https://arxiv.org/abs/2512.09088
Abstract:

Hallucinations are outputs by Large Language Models (LLMs) that are factually incorrect yet appear plausible [1]. This paper investigates how such hallucinations influence users’ trust in LLMs and users’ interaction with LLMs. To explore this in everyday use, we conducted a qualitative study with 192 participants. Our findings show that hallucinations do not result in blanket mistrust but instead lead to context-sensitive trust calibration. Building on the calibrated trust model by Lee & See [2] and Afroogh et al.’s trust-related factors [3], we confirm expectancy [3], [4], prior experience [3], [4], [5], and user expertise & domain knowledge [3], [4] as userrelated (human) trust factors, and identify intuition as an additional factor relevant for hallucination detection. Additionally, we found that trust dynamics are further influenced by contextual factors, particularly perceived risk [3] and decision stakes [6]. Consequently, we validate the recursive trust calibration process proposed by Blöbaum [7] and extend it by including intuition as a user-related trust factor. Based on these insights, we propose practical recommendations for responsible and reflective LLM use.

8. Provably Learning from Modern Language Models via Low Logit Rank

Authors: Noah Golowich , Allen Liu , Abhishek Shetty
URL: https://arxiv.org/abs/2512.09892
Abstract:

While modern language models and their inner workings are incredibly complex, recent work (Golowich, Liu & Shetty; 2025) has proposed a simple and potentially tractable abstraction for them through the observation that empirically, these language models all seem to have approximately low logit rank. Roughly, this means that a matrix formed by the model’s log probabilities of various tokens conditioned on certain sequences of tokens is well approximated by a low rank matrix. In this paper, our focus is on understanding how this structure can be exploited algorithmically for obtaining provable learning guarantees. Since low logit rank models can encode hard-to-learn distributions such as noisy parities, we study a query learning model with logit queries that reflects the access model for common APIs. Our main result is an efficient algorithm for learning any approximately low logit rank model from queries. We emphasize that our structural assumption closely reflects the behavior that is empirically observed in modern language models. Thus, our result gives what we believe is the first end-to-end learning guarantee for a generative model that plausibly captures modern language models.

9. FlipLLM: Efficient Bit-Flip Attacks on Multimodal LLMs using Reinforcement Learning

Authors: Khurram Khalil , Khaza Anuarul Hoque
URL: https://arxiv.org/abs/2512.09872
Abstract:

Generative Artificial Intelligence models, such as Large Language Models (LLMs) and Large Vision Models (VLMs), exhibit state-of-the-art performance but remain vulnerable to hardware-based threats, specifically bit-flip attacks (BFAs). Existing BFA discovery methods lack generalizability and struggle to scale, often failing to analyze the vast parameter space and complex interdependencies of modern foundation models in a reasonable time. This paper proposes FlipLLM, a reinforcement learning (RL) architecture-agnostic framework that formulates BFA discovery as a sequential decision-making problem. FlipLLM combines sensitivity-guided layer pruning with Q-learning to efficiently identify minimal, high-impact bit sets that can induce catastrophic failure. We demonstrate the effectiveness and generalizability of FlipLLM by applying it to a diverse set of models, including prominent text-only LLMs (GPT-2 Large, LLaMA 3.1 8B, and DeepSeek-V2 7B), VLMs such as LLaVA 1.6, and datasets, such as MMLU, MMLU-Pro, VQAv2, and TextVQA. Our results show that FlipLLM can identify critical bits that are vulnerable to BFAs up to 2.5x faster than SOTA methods. We demonstrate that flipping the FlipLLM-identified bits plummets the accuracy of LLaMA 3.1 8B from 69.9% to ~0.2%, and for LLaVA’s VQA score from 78% to almost 0%, by flipping as few as 5 and 7 bits, respectively. Further analysis reveals that applying standard hardware protection mechanisms, such as ECC SECDED, to the FlipLLM-identified bit locations completely mitigates the BFA impact, demonstrating the practical value of our framework in guiding hardware-level defenses. FlipLLM offers the first scalable and adaptive methodology for exploring the BFA vulnerability of both language and multimodal foundation models, paving the way for comprehensive hardware-security evaluation.

10. MedForget: Hierarchy-Aware Multimodal Unlearning Testbed for Medical AI

Authors: Fengli Wu , Vaidehi Patil , Jaehong Yoon , Yue Zhang , Mohit Bansal
URL: https://arxiv.org/abs/2512.09867
Abstract:

Pretrained Multimodal Large Language Models (MLLMs) are increasingly deployed in medical AI systems for clinical reasoning, diagnosis support, and report generation. However, their training on sensitive patient data raises critical privacy and compliance challenges under regulations such as HIPAA and GDPR, which enforce the “right to be forgotten”. Unlearning, the process of tuning models to selectively remove the influence of specific training data points, offers a potential solution, yet its effectiveness in complex medical settings remains underexplored. To systematically study this, we introduce MedForget, a Hierarchy-Aware Multimodal Unlearning Testbed with explicit retain and forget splits and evaluation sets containing rephrased variants. MedForget models hospital data as a nested hierarchy (Institution -> Patient -> Study -> Section), enabling fine-grained assessment across eight organizational levels. The benchmark contains 3840 multimodal (image, question, answer) instances, each hierarchy level having a dedicated unlearning target, reflecting distinct unlearning challenges. Experiments with four SOTA unlearning methods on three tasks (generation, classification, cloze) show that existing methods struggle to achieve complete, hierarchy-aware forgetting without reducing diagnostic performance. To test whether unlearning truly deletes hierarchical pathways, we introduce a reconstruction attack that progressively adds hierarchical level context to prompts. Models unlearned at a coarse granularity show strong resistance, while fine-grained unlearning leaves models vulnerable to such reconstruction. MedForget provides a practical, HIPAA-aligned testbed for building compliant medical AI systems.

11. LLMs in Interpreting Legal Documents

Authors: Simone Corbo
URL: https://arxiv.org/abs/2512.09830
Abstract:

This chapter explores the application of Large Language Models in the legal domain, showcasing their potential to optimise and augment traditional legal tasks by analysing possible use cases, such as assisting in interpreting statutes, contracts, and case law, enhancing clarity in legal summarisation, contract negotiation, and information retrieval. There are several challenges that can arise from the application of such technologies, such as algorithmic monoculture, hallucinations, and compliance with existing regulations, including the EU’s AI Act and recent U.S. initiatives, alongside the emerging approaches in China. Furthermore, two different benchmarks are presented.

12. The Ky Fan Norms and Beyond: Dual Norms and Combinations for Matrix Optimization

Authors: Alexey Kravatskiy , Ivan Kozyrev , Nikolai Kozlov , Alexander Vinogradov , Daniil Merkulov , Ivan Oseledets
URL: https://arxiv.org/abs/2512.09678
Abstract:

In this article, we explore the use of various matrix norms for optimizing functions of weight matrices, a crucial problem in training large language models. Moving beyond the spectral norm underlying the Muon update, we leverage duals of the Ky Fan $k$-norms to introduce a family of Muon-like algorithms we name Fanions, which are closely related to Dion. By working with duals of convex combinations of the Ky Fan $k$-norms with either the Frobenius norm or the $l_\infty$ norm, we construct the families of F-Fanions and S-Fanions, respectively. Their most prominent members are F-Muon and S-Muon. We complement our theoretical analysis with an extensive empirical study of these algorithms across a wide range of tasks and settings, demonstrating that F-Muon and S-Muon consistently match Muon’s performance, while outperforming vanilla Muon on a synthetic linear least squares problem.

13. Can LLMs Evaluate What They Cannot Annotate? Revisiting LLM Reliability in Hate Speech Detection

Authors: Paloma Piot , David Otero , Patricia Martín-Rodilla , Javier Parapar
URL: https://arxiv.org/abs/2512.09662
Abstract:

Hate speech spreads widely online, harming individuals and communities, making automatic detection essential for large-scale moderation, yet detecting it remains difficult. Part of the challenge lies in subjectivity: what one person flags as hate speech, another may see as benign. Traditional annotation agreement metrics, such as Cohen’s $\kappa$, oversimplify this disagreement, treating it as an error rather than meaningful diversity. Meanwhile, Large Language Models (LLMs) promise scalable annotation, but prior studies demonstrate that they cannot fully replace human judgement, especially in subjective tasks. In this work, we reexamine LLM reliability using a subjectivity-aware framework, cross-Rater Reliability (xRR), revealing that even under fairer lens, LLMs still diverge from humans. Yet this limitation opens an opportunity: we find that LLM-generated annotations can reliably reflect performance trends across classification models, correlating with human evaluations. We test this by examining whether LLM-generated annotations preserve the relative ordering of model performance derived from human evaluation (i.e. whether models ranked as more reliable by human annotators preserve the same order when evaluated with LLM-generated labels). Our results show that, although LLMs differ from humans at the instance level, they reproduce similar ranking and classification patterns, suggesting their potential as proxy evaluators. While not a substitute for human annotators, they might serve as a scalable proxy for evaluation in subjective NLP tasks.

14. Rethinking Chain-of-Thought Reasoning for Videos

Authors: Yiwu Zhong , Zi-Yuan Hu , Yin Li , Liwei Wang
URL: https://arxiv.org/abs/2512.09616
Abstract:

Chain-of-thought (CoT) reasoning has been highly successful in solving complex tasks in natural language processing, and recent multimodal large language models (MLLMs) have extended this paradigm to video reasoning. However, these models typically build on lengthy reasoning chains and large numbers of input visual tokens. Motivated by empirical observations from our benchmark study, we hypothesize that concise reasoning combined with a reduced set of visual tokens can be sufficient for effective video reasoning. To evaluate this hypothesis, we design and validate an efficient post-training and inference framework that enhances a video MLLM’s reasoning capability. Our framework enables models to operate on compressed visual tokens and generate brief reasoning traces prior to answering. The resulting models achieve substantially improved inference efficiency, deliver competitive performance across diverse benchmarks, and avoid reliance on manual CoT annotations or supervised fine-tuning. Collectively, our results suggest that long, human-like CoT reasoning may not be necessary for general video reasoning, and that concise reasoning can be both effective and efficient. Our code will be released at this https URL .

15. Auto-BenchmarkCard: Automated Synthesis of Benchmark Documentation

Authors: Aris Hofmann , Inge Vejsbjerg , Dhaval Salwala , Elizabeth M. Daly
URL: https://arxiv.org/abs/2512.09577
Abstract:

We present Auto-BenchmarkCard, a workflow for generating validated descriptions of AI benchmarks. Benchmark documentation is often incomplete or inconsistent, making it difficult to interpret and compare benchmarks across tasks or domains. Auto-BenchmarkCard addresses this gap by combining multi-agent data extraction from heterogeneous sources (e.g., Hugging Face, Unitxt, academic papers) with LLM-driven synthesis. A validation phase evaluates factual accuracy through atomic entailment scoring using the FactReasoner tool. This workflow has the potential to promote transparency, comparability, and reusability in AI benchmark reporting, enabling researchers and practitioners to better navigate and evaluate benchmark choices.

16. System Report for CCL25-Eval Task 10: Prompt-Driven Large Language Model Merge for Fine-Grained Chinese Hate Speech Detection

Authors: Binglin Wu , Jiaxiu Zou , Xianneng Li
URL: https://arxiv.org/abs/2512.09563
Abstract:

The proliferation of hate speech on Chinese social media poses urgent societal risks, yet traditional systems struggle to decode context-dependent rhetorical strategies and evolving slang. To bridge this gap, we propose a novel three-stage LLM-based framework: Prompt Engineering, Supervised Fine-tuning, and LLM Merging. First, context-aware prompts are designed to guide LLMs in extracting implicit hate patterns. Next, task-specific features are integrated during supervised fine-tuning to enhance domain adaptation. Finally, merging fine-tuned LLMs improves robustness against out-of-distribution cases. Evaluations on the STATE-ToxiCN benchmark validate the framework’s effectiveness, demonstrating superior performance over baseline methods in detecting fine-grained hate speech.

17. SWEnergy: An Empirical Study on Energy Efficiency in Agentic Issue Resolution Frameworks with SLMs

Authors: Arihant Tripathy , Ch Pavan Harshit , Karthik Vaidhyanathan
URL: https://arxiv.org/abs/2512.09543
Abstract:

Context. LLM-based autonomous agents in software engineering rely on large, proprietary models, limiting local deployment. This has spurred interest in Small Language Models (SLMs), but their practical effectiveness and efficiency within complex agentic frameworks for automated issue resolution remain poorly understood. Goal. We investigate the performance, energy efficiency, and resource consumption of four leading agentic issue resolution frameworks when deliberately constrained to using SLMs. We aim to assess the viability of these systems for this task in resource-limited settings and characterize the resulting trade-offs. Method. We conduct a controlled evaluation of four leading agentic frameworks (SWE-Agent, OpenHands, Mini SWE Agent, AutoCodeRover) using two SLMs (Gemma-3 4B, Qwen-3 1.7B) on the SWE-bench Verified Mini benchmark. On fixed hardware, we measure energy, duration, token usage, and memory over 150 runs per configuration. Results. We find that framework architecture is the primary driver of energy consumption. The most energy-intensive framework, AutoCodeRover (Gemma), consumed 9.4x more energy on average than the least energy-intensive, OpenHands (Gemma). However, this energy is largely wasted. Task resolution rates were near-zero, demonstrating that current frameworks, when paired with SLMs, consume significant energy on unproductive reasoning loops. The SLM’s limited reasoning was the bottleneck for success, but the framework’s design was the bottleneck for efficiency. Conclusions. Current agentic frameworks, designed for powerful LLMs, fail to operate efficiently with SLMs. We find that framework architecture is the primary driver of energy consumption, but this energy is largely wasted due to the SLMs’ limited reasoning. Viable low-energy solutions require shifting from passive orchestration to architectures that actively manage SLM weaknesses.

18. Representation Invariance and Allocation: When Subgroup Balance Matters

Authors: Anissa Alloula , Charles Jones , Zuzanna Wakefield-Skorniewska , Francesco Quinzan , Bartłomiej Papież
URL: https://arxiv.org/abs/2512.09496
Abstract:

Unequal representation of demographic groups in training data poses challenges to model generalisation across populations. Standard practice assumes that balancing subgroup representation optimises performance. However, recent empirical results contradict this assumption: in some cases, imbalanced data distributions actually improve subgroup performance, while in others, subgroup performance remains unaffected by the absence of an entire subgroup during training. We conduct a systematic study of subgroup allocation across four vision and language models, varying training data composition to characterise the sensitivity of subgroup performance to data balance. We propose the latent separation hypothesis, which states that a partially fine-tuned model’s dependence on subgroup representation is determined by the degree of separation between subgroups in the latent space of the pre-trained model. We formalise this hypothesis, provide theoretical analysis, and validate it empirically. Finally, we present a practical application to foundation model fine-tuning, demonstrating that quantitative analysis of latent subgroup separation can inform data collection and balancing decisions.

19. RouteRAG: Efficient Retrieval-Augmented Generation from Text and Graph via Reinforcement Learning

Authors: Yucan Guo , Miao Su , Saiping Guan , Zihao Sun , Xiaolong Jin , Jiafeng Guo , Xueqi Cheng
URL: https://arxiv.org/abs/2512.09487
Abstract:

Retrieval-Augmented Generation (RAG) integrates non-parametric knowledge into Large Language Models (LLMs), typically from unstructured texts and structured graphs. While recent progress has advanced text-based RAG to multi-turn reasoning through Reinforcement Learning (RL), extending these advances to hybrid retrieval introduces additional challenges. Existing graph-based or hybrid systems typically depend on fixed or handcrafted retrieval pipelines, lacking the ability to integrate supplementary evidence as reasoning unfolds. Besides, while graph evidence provides relational structures crucial for multi-hop reasoning, it is substantially more expensive to retrieve. To address these limitations, we introduce \model{}, an RL-based framework that enables LLMs to perform multi-turn and adaptive graph-text hybrid RAG. \model{} jointly optimizes the entire generation process via RL, allowing the model to learn when to reason, what to retrieve from either texts or graphs, and when to produce final answers, all within a unified generation policy. To guide this learning process, we design a two-stage training framework that accounts for both task outcome and retrieval efficiency, enabling the model to exploit hybrid evidence while avoiding unnecessary retrieval overhead. Experimental results across five question answering benchmarks demonstrate that \model{} significantly outperforms existing RAG baselines, highlighting the benefits of end-to-end RL in supporting adaptive and efficient retrieval for complex reasoning.

20. Advancing LLM-Based Security Automation with Customized Group Relative Policy Optimization for Zero-Touch Networks

Authors: Xinye Cao , Yihan Lin , Guoshun Nan , Qinchuan Zhou , Yuhang Luo , Yurui Gao , Zeliang Zhang , Haolang Lu , Qimei Cui , Yanzhao Hou , Xiaofeng Tao , Tony Q.S. Quek
URL: https://arxiv.org/abs/2512.09485
Abstract:

Zero-Touch Networks (ZTNs) represent a transformative paradigm toward fully automated and intelligent network management, providing the scalability and adaptability required for the complexity of sixth-generation (6G) networks. However, the distributed architecture, high openness, and deep heterogeneity of 6G networks expand the attack surface and pose unprecedented security challenges. To address this, security automation aims to enable intelligent security management across dynamic and complex environments, serving as a key capability for securing 6G ZTNs. Despite its promise, implementing security automation in 6G ZTNs presents two primary challenges: 1) automating the lifecycle from security strategy generation to validation and update under real-world, parallel, and adversarial conditions, and 2) adapting security strategies to evolving threats and dynamic environments. This motivates us to propose SecLoop and SA-GRPO. SecLoop constitutes the first fully automated framework that integrates large language models (LLMs) across the entire lifecycle of security strategy generation, orchestration, response, and feedback, enabling intelligent and adaptive defenses in dynamic network environments, thus tackling the first challenge. Furthermore, we propose SA-GRPO, a novel security-aware group relative policy optimization algorithm that iteratively refines security strategies by contrasting group feedback collected from parallel SecLoop executions, thereby addressing the second challenge. Extensive real-world experiments on five benchmarks, including 11 MITRE ATT&CK processes and over 20 types of attacks, demonstrate the superiority of the proposed SecLoop and SA-GRPO. We will release our platform to the community, facilitating the advancement of security automation towards next generation communications.

21. Advancing Research via Human-AI Interactive Theorem Proving

Authors: Chenyi Li , Zhijian Lai , Dong An , Jiang Hu , Zaiwen Wen
URL: https://arxiv.org/abs/2512.09443
Abstract:

We investigate how large language models can be used as research tools in scientific computing while preserving mathematical rigor. We propose a human-in-the-loop workflow for interactive theorem proving and discovery with LLMs. Human experts retain control over problem formulation and admissible assumptions, while the model searches for proofs or contradictions, proposes candidate properties and theorems, and helps construct structures and parameters that satisfy explicit constraints, supported by numerical experiments and simple verification checks. Experts treat these outputs as raw material, further refine them, and organize the results into precise statements and rigorous proofs. We instantiate this workflow in a case study on the connection between manifold optimization and Grover’s quantum search algorithm, where the pipeline helps identify invariant subspaces, explore Grover-compatible retractions, and obtain convergence guarantees for the retraction-based gradient method. The framework provides a practical template for integrating large language models into frontier mathematical research, enabling faster exploration of proof space and algorithm design while maintaining transparent reasoning responsibilities. Although illustrated on manifold optimization problems in quantum computing, the principles extend to other core areas of scientific computing.

22. Representation Calibration and Uncertainty Guidance for Class-Incremental Learning based on Vision Language Model

Authors: Jiantao Tan , Peixian Ma , Tong Yu , Wentao Zhang , Ruixuan Wang
URL: https://arxiv.org/abs/2512.09441
Abstract:

Class-incremental learning requires a learning system to continually learn knowledge of new classes and meanwhile try to preserve previously learned knowledge of old classes. As current state-of-the-art methods based on Vision-Language Models (VLMs) still suffer from the issue of differentiating classes across learning tasks. Here a novel VLM-based continual learning framework for image classification is proposed. In this framework, task-specific adapters are added to the pre-trained and frozen image encoder to learn new knowledge, and a novel cross-task representation calibration strategy based on a mixture of light-weight projectors is used to help better separate all learned classes in a unified feature space, alleviating class confusion across tasks. In addition, a novel inference strategy guided by prediction uncertainty is developed to more accurately select the most appropriate image feature for class prediction. Extensive experiments on multiple datasets under various settings demonstrate the superior performance of our method compared to existing ones.

23. CourtPressGER: A German Court Decision to Press Release Summarization Dataset

Authors: Sebastian Nagl , Mohamed Elganayni , Melanie Pospisil , Matthias Grabmair
URL: https://arxiv.org/abs/2512.09434
Abstract:

Official court press releases from Germany’s highest courts present and explain judicial rulings to the public, as well as to expert audiences. Prior NLP efforts emphasize technical headnotes, ignoring citizen-oriented communication needs. We introduce CourtPressGER, a 6.4k dataset of triples: rulings, human-drafted press releases, and synthetic prompts for LLMs to generate comparable releases. This benchmark trains and evaluates LLMs in generating accurate, readable summaries from long judicial texts. We benchmark small and large LLMs using reference-based metrics, factual-consistency checks, LLM-as-judge, and expert ranking. Large LLMs produce high-quality drafts with minimal hierarchical performance loss; smaller models require hierarchical setups for long judgments. Initial benchmarks show varying model performance, with human-drafted releases ranking highest.

24. ODMA: On-Demand Memory Allocation Framework for LLM Serving on LPDDR-Class Accelerators

Authors: Guoqiang Zou , Wanyu Wang , Hao Zheng , Longxiang Yin , Yinhe Han
URL: https://arxiv.org/abs/2512.09427
Abstract:

Serving large language models (LLMs) on accelerators with poor random-access bandwidth (e.g., LPDDR5-based) is limited by current memory managers. Static pre-allocation wastes memory, while fine-grained paging (e.g., PagedAttention) is ill-suited due to high random-access costs. Existing HBM-centric solutions do not exploit the characteristics of random-access-constrained memory (RACM) accelerators like Cambricon MLU370. We present ODMA, an on-demand memory allocation framework for RACM. ODMA addresses distribution drift and heavy-tailed requests by coupling a lightweight length predictor with dynamic bucket partitioning and a large-bucket safeguard. Boundaries are periodically updated from live traces to maximize utilization. On Alpaca and Google-NQ, ODMA improves prediction accuracy of prior work significantly (e.g., from 82.68% to 93.36%). Serving DeepSeek-R1-Distill-Qwen-7B on Cambricon MLU370-X4, ODMA raises memory utilization from 55.05% to 72.45% and improves RPS and TPS by 29% and 27% over static baselines. This demonstrates that hardware-aware allocation unlocks efficient LLM serving on RACM platforms.

25. GLACIA: Instance-Aware Positional Reasoning for Glacial Lake Segmentation via Multimodal Large Language Model

Authors: Lalit Maurya , Saurabh Kaushik , Beth Tellman
URL: https://arxiv.org/abs/2512.09251
Abstract:

Glacial lake monitoring bears great significance in mitigating the anticipated risk of Glacial Lake Outburst Floods. However, existing segmentation methods based on convolutional neural networks (CNNs) and Vision Transformers (ViTs), remain constrained to pixel-level predictions, lacking high-level global scene semantics and human-interpretable reasoning. To address this, we introduce GLACIA (\textbf{G}lacial \textbf{LA}ke segmentation with \textbf{C}ontextual \textbf{I}nstance \textbf{A}wareness), the first framework that integrates large language models with segmentation capabilities to produce both accurate segmentation masks and corresponding spatial reasoning outputs. We construct the Glacial Lake Position Reasoning (GLake-Pos) dataset pipeline, which provides diverse, spatially grounded question-answer pairs designed to overcome the lack of instance-aware positional reasoning data in remote sensing. Comparative evaluation demonstrate that GLACIA (mIoU: 87.30) surpasses state-of-the-art method based on CNNs (mIoU: 78.55 - 79.01), ViTs (mIoU: 69.27 - 81.75), Geo-foundation models (mIoU: 76.37 - 87.10), and reasoning based segmentation methods (mIoU: 60.12 - 75.66). Our approach enables intuitive disaster preparedness and informed policy-making in the context of rapidly changing glacial environments by facilitating natural language interaction, thereby supporting more efficient and interpretable decision-making. The code is released on this https URL

26. CORE: A Conceptual Reasoning Layer for Large Language Models

Authors: Vishwas Hegde , Vindhya Shigehalli
URL: https://arxiv.org/abs/2512.09222
Abstract:

Large language models handle single-turn generation well, but multi-turn interactions still require the model to reconstruct user intent and task state from an expanding token history because internal representations do not persist across turns. This token-first paradigm leads to drift, inconsistent reasoning modes, and growing prompts as conversations deepen. We propose CORE, a concept-first interaction layer that improves multi-turn stability without modifying model weights. CORE combines a small library of universal cognitive operators with a persistent Local Concept - a compact semantic state capturing the task, constraints, preferences, and intermediate results. Each model call receives only this concept state, the user’s latest instruction, and the selected operator, eliminating the need to replay full history. A preliminary prototype simulating CORE’s behavior shows about 42% reduction in cumulative prompt tokens, though this number reflects prototype conditions and should not be interpreted as a real-world performance estimate. CORE offers a model-agnostic mechanism that separates conceptual reasoning from language generation, suggesting a scalable direction for more stable multi-turn systems.

27. LLMs for Analog Circuit Design Continuum (ACDC)

Authors: Yasaman Esfandiari , Jocelyn Rego , Austin Meyer , Jonathan Gallagher , Mia Levy
URL: https://arxiv.org/abs/2512.09199
Abstract:

Large Language Models (LLMs) and transformer architectures have shown impressive reasoning and generation capabilities across diverse natural language tasks. However, their reliability and robustness in real-world engineering domains remain largely unexplored, limiting their practical utility in human-centric workflows. In this work, we investigate the applicability and consistency of LLMs for analog circuit design – a task requiring domain-specific reasoning, adherence to physical constraints, and structured representations – focusing on AI-assisted design where humans remain in the loop. We study how different data representations influence model behavior and compare smaller models (e.g., T5, GPT-2) with larger foundation models (e.g., Mistral-7B, GPT-oss-20B) under varying training conditions. Our results highlight key reliability challenges, including sensitivity to data format, instability in generated designs, and limited generalization to unseen circuit configurations. These findings provide early evidence on the limits and potential of LLMs as tools to enhance human capabilities in complex engineering tasks, offering insights into designing reliable, deployable foundation models for structured, real-world applications.

28. WOLF: Werewolf-based Observations for LLM Deception and Falsehoods

Authors: Mrinal Agarwal , Saad Rana , Theo Sundoro , Hermela Berhe , Spencer Kim , Vasu Sharma , Sean O’Brien , Kevin Zhu
URL: https://arxiv.org/abs/2512.09187
Abstract:

Deception is a fundamental challenge for multi-agent reasoning: effective systems must strategically conceal information while detecting misleading behavior in others. Yet most evaluations reduce deception to static classification, ignoring the interactive, adversarial, and longitudinal nature of real deceptive dynamics. Large language models (LLMs) can deceive convincingly but remain weak at detecting deception in peers. We present WOLF, a multi-agent social deduction benchmark based on Werewolf that enables separable measurement of deception production and detection. WOLF embeds role-grounded agents (Villager, Werewolf, Seer, Doctor) in a programmable LangGraph state machine with strict night-day cycles, debate turns, and majority voting. Every statement is a distinct analysis unit, with self-assessed honesty from speakers and peer-rated deceptiveness from others. Deception is categorized via a standardized taxonomy (omission, distortion, fabrication, misdirection), while suspicion scores are longitudinally smoothed to capture both immediate judgments and evolving trust dynamics. Structured logs preserve prompts, outputs, and state transitions for full reproducibility. Across 7,320 statements and 100 runs, Werewolves produce deceptive statements in 31% of turns, while peer detection achieves 71-73% precision with ~52% overall accuracy. Precision is higher for identifying Werewolves, though false positives occur against Villagers. Suspicion toward Werewolves rises from ~52% to over 60% across rounds, while suspicion toward Villagers and the Doctor stabilizes near 44-46%. This divergence shows that extended interaction improves recall against liars without compounding errors against truthful roles. WOLF moves deception evaluation beyond static datasets, offering a dynamic, controlled testbed for measuring deceptive and detective capacity in adversarial multi-agent interaction.

29. Prompt-Based Continual Compositional Zero-Shot Learning

Authors: Sauda Maryam , Sara Nadeem , Faisal Qureshi , Mohsen Ali
URL: https://arxiv.org/abs/2512.09172
Abstract:

We tackle continual adaptation of vision-language models to new attributes, objects, and their compositions in Compositional Zero-Shot Learning (CZSL), while preventing forgetting of prior knowledge. Unlike classical continual learning where classes are disjoint, CCZSL is more complex as attributes and objects may reoccur across sessions while compositions remain unique. Built on a frozen VLM backbone, we propose the first Prompt-based Continual Compositional Zero-Shot Learning (PromptCCZSL) framework that retains prior knowledge through recency-weighted multi-teacher distillation. It employs session-aware compositional prompts to fuse multimodal features for new compositions, while attribute and object prompts are learned through session-agnostic fusion to maintain global semantic consistency, which is further stabilized by a Cosine Anchor Loss (CAL) to preserve prior knowledge. To enhance adaptation in the current session, an Orthogonal Projection Loss (OPL) ensures that new attribute and object embeddings remain distinct from previous ones, preventing overlap, while an Intra-Session Diversity Loss (IDL) promotes variation among current-session embeddings for richer, more discriminative representations. We also introduce a comprehensive protocol that jointly measures catastrophic forgetting and compositional generalization. Extensive experiments on UT-Zappos and C-GQA benchmarks demonstrate that PromptCCZSL achieves substantial improvements over prior VLM-based and non-VLM baselines, setting a new benchmark for CCZSL in closed-world settings.

30. MindShift: Analyzing Language Models’ Reactions to Psychological Prompts

Authors: Anton Vasiliuk , Irina Abdullaeva , Polina Druzhinina , Anton Razzhigaev , Andrey Kuznetsov
URL: https://arxiv.org/abs/2512.09149
Abstract:

Large language models (LLMs) hold the potential to absorb and reflect personality traits and attitudes specified by users. In our study, we investigated this potential using robust psychometric measures. We adapted the most studied test in psychological literature, namely Minnesota Multiphasic Personality Inventory (MMPI) and examined LLMs’ behavior to identify traits. To asses the sensitivity of LLMs’ prompts and psychological biases we created personality-oriented prompts, crafting a detailed set of personas that vary in trait intensity. This enables us to measure how well LLMs follow these roles. Our study introduces MindShift, a benchmark for evaluating LLMs’ psychological adaptability. The results highlight a consistent improvement in LLMs’ role perception, attributed to advancements in training datasets and alignment techniques. Additionally, we observe significant differences in responses to psychometric assessments across different model types and families, suggesting variability in their ability to emulate human-like personality traits. MindShift prompts and code for LLM evaluation will be publicly available.

31. Detecting Hallucinations in Graph Retrieval-Augmented Generation via Attention Patterns and Semantic Alignment

Authors: Shanghao Li , Jinda Han , Yibo Wang , Yuanjie Zhu , Zihe Song , Langzhou He , Kenan Kamel A Alghythee , Philip S. Yu
URL: https://arxiv.org/abs/2512.09148
Abstract:

Graph-based Retrieval-Augmented Generation (GraphRAG) enhances Large Language Models (LLMs) by incorporating external knowledge from linearized subgraphs retrieved from knowledge graphs. However, LLMs struggle to interpret the relational and topological information in these inputs, resulting in hallucinations that are inconsistent with the retrieved knowledge. To analyze how LLMs attend to and retain structured knowledge during generation, we propose two lightweight interpretability metrics: Path Reliance Degree (PRD), which measures over-reliance on shortest-path triples, and Semantic Alignment Score (SAS), which assesses how well the model’s internal representations align with the retrieved knowledge. Through empirical analysis on a knowledge-based QA task, we identify failure patterns associated with over-reliance on salient paths and weak semantic grounding, as indicated by high PRD and low SAS scores. We further develop a lightweight post-hoc hallucination detector, Graph Grounding and Alignment (GGA), which outperforms strong semantic and confidence-based baselines across AUC and F1. By grounding hallucination analysis in mechanistic interpretability, our work offers insights into how structural limitations in LLMs contribute to hallucinations, informing the design of more reliable GraphRAG systems in the future.

32. Knowledge-Guided Large Language Model for Automatic Pediatric Dental Record Understanding and Safe Antibiotic Recommendation

Authors: Zihan Han , Junyan Ge , Caifeng Li
URL: https://arxiv.org/abs/2512.09127
Abstract:

Accurate interpretation of pediatric dental clinical records and safe antibiotic prescribing remain persistent challenges in dental informatics. Traditional rule-based clinical decision support systems struggle with unstructured dental narratives, incomplete radiographic descriptions, and complex safety constraints. To address these limitations, this study proposes a Knowledge-Guided Large Language Model (KG-LLM) that integrates a pediatric dental knowledge graph, retrieval-augmented generation (RAG), and a multi-stage safety validation pipeline for evidence-grounded antibiotic recommendation. The framework first employs a clinical NER/RE module to extract structured entities and relations from dental notes and radiology reports. Relevant guidelines, drug-safety rules, and analogous historical cases are subsequently retrieved from the knowledge graph and supplied to the LLM for diagnostic summarization and dose-drug-duration prediction. Safety assurance is achieved through a dual-layer validation mechanism combining deterministic rule checking with a learned classifier for detecting allergies, contraindications, and dosing errors. Experiments on 32,000 de-identified pediatric dental visit records demonstrate the effectiveness of the proposed approach. Compared with a domain-adapted Llama-2 clinical baseline, KG-LLM improves record-understanding performance (F1: 0.914 vs. 0.867), drug-dose-duration accuracy (Top-1: 0.782 vs. 0.716), and reduces unsafe antibiotic suggestions by 50%. Additional evaluation across summary quality, recommendation accuracy, and global safety scores further confirms the robustness of the system. Ablation analyses indicate that the knowledge graph, RAG, and safety modules each contribute substantially to clinical reliability and interpretability.

33. Evolving Excellence: Automated Optimization of LLM-based Agents

Authors: Paul Brookes , Vardan Voskanyan , Rafail Giavrimis , Matthew Truscott , Mina Ilieva , Chrystalla Pavlou , Alexandru Staicu , Manal Adham , Will Evers- Hood , Jingzhi Gong , Kejia Zhang , Matvey Fedoseev , Vishal Sharma , Roman Bauer , Zheng Wang , Hema Nair , Wei Jie , Tianhua Xu , Aurora Constantin , Leslie Kanthan , Michail Basios
URL: https://arxiv.org/abs/2512.09108
Abstract:

Agentic AI systems built on large language models (LLMs) offer significant potential for automating complex workflows, from software development to customer support. However, LLM agents often underperform due to suboptimal configurations; poorly tuned prompts, tool descriptions, and parameters that typically require weeks of manual refinement. Existing optimization methods either are too complex for general use or treat components in isolation, missing critical interdependencies. We present ARTEMIS, a no-code evolutionary optimization platform that jointly optimizes agent configurations through semantically-aware genetic operators. Given only a benchmark script and natural language goals, ARTEMIS automatically discovers configurable components, extracts performance signals from execution logs, and evolves configurations without requiring architectural modifications. We evaluate ARTEMIS on four representative agent systems: the \emph{ALE Agent} for competitive programming on AtCoder Heuristic Contest, achieving a \textbf{$13.6\%$ improvement} in acceptance rate; the \emph{Mini-SWE Agent} for code optimization on SWE-Perf, with a statistically significant \textbf{10.1\% performance gain}; and the \emph{CrewAI Agent} for cost and mathematical reasoning on Math Odyssey, achieving a statistically significant \textbf{$36.9\%$ reduction} in the number of tokens required for evaluation. We also evaluate the \emph{MathTales-Teacher Agent} powered by a smaller open-source model (Qwen2.5-7B) on GSM8K primary-level mathematics problems, achieving a \textbf{22\% accuracy improvement} and demonstrating that ARTEMIS can optimize agents based on both commercial and local models.

34. ORCA: Open-ended Response Correctness Assessment for Audio Question Answering

Authors: Šimon Sedláček , Sara Barahona , Bolaji Yusuf , Laura Herrera-Alarcón , Santosh Kesiraju , Cecilia Bolaños , Alicia Lozano-Diez , Sathvik Udupa , Fernando López , Allison Ferner , Ramani Duraiswami , Jan Černocký
URL: https://arxiv.org/abs/2512.09066
Abstract:

Evaluating open-ended responses from large audio language models (LALMs) is challenging because human annotators often genuinely disagree on answer correctness due to multiple valid interpretations, partial correctness, and subjective judgment. Traditional metrics reporting only mean scores fail to capture this uncertainty. We present ORCA (Open-ended Response Correctness Assessment), a framework that models the variability in human judgments using Beta distributions to predict both expected correctness and uncertainty. Our three-stage annotation framework combines human judgment with structured feedback and iterative refinement to simultaneously curate training data and improve benchmark quality. We collected 11,721 annotations across 3,580 question-answer pairs from 15 LALMs on two audio QA benchmarks, achieving inter-annotator agreement of 0.82 (Krippendorff’s alpha). ORCA achieves 0.91 Spearman correlation with mean human judgments, matching or outperforming LLM-judge baselines while providing uncertainty estimates and requiring significantly less compute. We release our models, code, and curated dataset.

35. Towards Lossless Ultimate Vision Token Compression for VLMs

Authors: Dehua Zheng , Mouxiao Huang , Borui Jiang , Hailin Hu , Xinghao Chen
URL: https://arxiv.org/abs/2512.09010
Abstract:

Visual language models encounter challenges in computational efficiency and latency, primarily due to the substantial redundancy in the token representations of high-resolution images and videos. Current attention/similarity-based compression algorithms suffer from either position bias or class imbalance, leading to significant accuracy degradation. They also fail to generalize to shallow LLM layers, which exhibit weaker cross-modal interactions. To address this, we extend token compression to the visual encoder through an effective iterative merging scheme that is orthogonal in spatial axes to accelerate the computation across the entire VLM. Furthermoer, we integrate a spectrum pruning unit into LLM through an attention/similarity-free low-pass filter, which gradually prunes redundant visual tokens and is fully compatible to modern FlashAttention. On this basis, we propose Lossless Ultimate Vision tokens Compression (LUVC) framework. LUVC systematically compresses visual tokens until complete elimination at the final layer of LLM, so that the high-dimensional visual features are gradually fused into the multimodal queries. The experiments show that LUVC achieves a 2 speedup inference in language model with negligible accuracy degradation, and the training-free characteristic enables immediate deployment across multiple VLMs.

36. Llama-based source code vulnerability detection: Prompt engineering vs Fine tuning

Authors: Dyna Soumhane Ouchebara , Stéphane Dupont
URL: https://arxiv.org/abs/2512.09006
Abstract:

The significant increase in software production, driven by the acceleration of development cycles over the past two decades, has led to a steady rise in software vulnerabilities, as shown by statistics published yearly by the CVE program. The automation of the source code vulnerability detection (CVD) process has thus become essential, and several methods have been proposed ranging from the well established program analysis techniques to the more recent AI-based methods. Our research investigates Large Language Models (LLMs), which are considered among the most performant AI models to date, for the CVD task. The objective is to study their performance and apply different state-of-the-art techniques to enhance their effectiveness for this task. We explore various fine-tuning and prompt engineering settings. We particularly suggest one novel approach for fine-tuning LLMs which we call Double Fine-tuning, and also test the understudied Test-Time fine-tuning approach. We leverage the recent open-source Llama-3.1 8B, with source code samples extracted from BigVul and PrimeVul datasets. Our conclusions highlight the importance of fine-tuning to resolve the task, the performance of Double tuning, as well as the potential of Llama models for CVD. Though prompting proved ineffective, Retrieval augmented generation (RAG) performed relatively well as an example selection technique. Overall, some of our research questions have been answered, and many are still on hold, which leaves us many future work perspectives. Code repository is available here: this https URL .

37. DermETAS-SNA LLM: A Dermatology Focused Evolutionary Transformer Architecture Search with StackNet Augmented LLM Assistant

Authors: Nitya Phani Santosh Oruganty , Keerthi Vemula Murali , Chun-Kit Ngan , Paulo Bandeira Pinho
URL: https://arxiv.org/abs/2512.08998
Abstract:

Our work introduces the DermETAS-SNA LLM Assistant that integrates Dermatology-focused Evolutionary Transformer Architecture Search with StackNet Augmented LLM. The assistant dynamically learns skin-disease classifiers and provides medically informed descriptions to facilitate clinician-patient interpretation. Contributions include: (1) Developed an ETAS framework on the SKINCON dataset to optimize a Vision Transformer (ViT) tailored for dermatological feature representation and then fine-tuned binary classifiers for each of the 23 skin disease categories in the DermNet dataset to enhance classification performance; (2) Designed a StackNet architecture that integrates multiple fine-tuned binary ViT classifiers to enhance predictive robustness and mitigate class imbalance issues; (3) Implemented a RAG pipeline, termed Diagnostic Explanation and Retrieval Model for Dermatology, which harnesses the capabilities of the Google Gemini 2.5 Pro LLM architecture to generate personalized, contextually informed diagnostic descriptions and explanations for patients, leveraging a repository of verified dermatological materials; (4) Performed extensive experimental evaluations on 23 skin disease categories to demonstrate performance increase, achieving an overall F1-score of 56.30% that surpasses SkinGPT-4 (48.51%) by a considerable margin, representing a performance increase of 16.06%; (5) Conducted a domain-expert evaluation, with eight licensed medical doctors, of the clinical responses generated by our AI assistant for seven dermatological conditions. Our results show a 92% agreement rate with the assessments provided by our AI assistant (6) Created a proof-of-concept prototype that fully integrates our DermETAS-SNA LLM into our AI assistant to demonstrate its practical feasibility for real-world clinical and educational applications.

38. RAG-HAR: Retrieval Augmented Generation-based Human Activity Recognition

Authors: Nirhoshan Sivaroopan , Hansi Karunarathna , Chamara Madarasingha , Anura Jayasumana , Kanchana Thilakarathna
URL: https://arxiv.org/abs/2512.08984
Abstract:

Human Activity Recognition (HAR) underpins applications in healthcare, rehabilitation, fitness tracking, and smart environments, yet existing deep learning approaches demand dataset-specific training, large labeled corpora, and significant computational this http URL introduce RAG-HAR, a training-free retrieval-augmented framework that leverages large language models (LLMs) for HAR. RAG-HAR computes lightweight statistical descriptors, retrieves semantically similar samples from a vector database, and uses this contextual evidence to make LLM-based activity identification. We further enhance RAG-HAR by first applying prompt optimization and introducing an LLM-based activity descriptor that generates context-enriched vector databases for delivering accurate and highly relevant contextual information. Along with these mechanisms, RAG-HAR achieves state-of-the-art performance across six diverse HAR benchmarks. Most importantly, RAG-HAR attains these improvements without requiring model training or fine-tuning, emphasizing its robustness and practical applicability. RAG-HAR moves beyond known behaviors, enabling the recognition and meaningful labelling of multiple unseen human activities.

39. Mitigating Bias with Words: Inducing Demographic Ambiguity in Face Recognition Templates by Text Encoding

Authors: Tahar Chettaoui , Naser Damer , Fadi Boutros
URL: https://arxiv.org/abs/2512.08981
Abstract:

Face recognition (FR) systems are often prone to demographic biases, partially due to the entanglement of demographic-specific information with identity-relevant features in facial embeddings. This bias is extremely critical in large multicultural cities, especially where biometrics play a major role in smart city infrastructure. The entanglement can cause demographic attributes to overshadow identity cues in the embedding space, resulting in disparities in verification performance across different demographic groups. To address this issue, we propose a novel strategy, Unified Text-Image Embedding (UTIE), which aims to induce demographic ambiguity in face embeddings by enriching them with information related to other demographic groups. This encourages face embeddings to emphasize identity-relevant features and thus promotes fairer verification performance across groups. UTIE leverages the zero-shot capabilities and cross-modal semantic alignment of Vision-Language Models (VLMs). Given that VLMs are naturally trained to align visual and textual representations, we enrich the facial embeddings of each demographic group with text-derived demographic features extracted from other demographic groups. This encourages a more neutral representation in terms of demographic attributes. We evaluate UTIE using three VLMs, CLIP, OpenCLIP, and SigLIP, on two widely used benchmarks, RFW and BFW, designed to assess bias in FR. Experimental results show that UTIE consistently reduces bias metrics while maintaining, or even improving in several cases, the face verification accuracy.

40. Peek-a-Boo Reasoning: Contrastive Region Masking in MLLMs

Authors: Isha Chaturvedi , Anjana Nair , Yushen Li , Adhitya Rajendra Kumar , Kevin Zhu , Sunishchal Dev , Ashwinee Panda , Vasu Sharma
URL: https://arxiv.org/abs/2512.08976
Abstract:

We introduce Contrastive Region Masking (CRM), a training free diagnostic that reveals how multimodal large language models (MLLMs) depend on specific visual regions at each step of chain-of-thought (CoT) reasoning. Unlike prior approaches limited to final answers or attention maps, CRM provides causal, step-level attri- bution by systematically masking annotated regions and contrasting the resulting reasoning traces with unmasked baselines. Applied to datasets such as VisArgs, CRM reveals distinct failure modes: some models preserve reasoning structure, but hallucinate when evidence is missing, while others ground tightly to visual cues yet collapse under perturbations. By shifting the evaluation from correctness of an- swers to faithfulness of reasoning, CRM reframes visual benchmarks as diagnostic tools, highlighting the need for multimodal evaluation frameworks that measure not just performance, but also robustness and fidelity of reasoning.

41. CluCERT: Certifying LLM Robustness via Clustering-Guided Denoising Smoothing

Authors: Zixia Wang , Gaojie Jin , Jia Hu , Ronghui Mu
URL: https://arxiv.org/abs/2512.08967
Abstract:

Recent advancements in Large Language Models (LLMs) have led to their widespread adoption in daily applications. Despite their impressive capabilities, they remain vulnerable to adversarial attacks, as even minor meaning-preserving changes such as synonym substitutions can lead to incorrect predictions. As a result, certifying the robustness of LLMs against such adversarial prompts is of vital importance. Existing approaches focused on word deletion or simple denoising strategies to achieve robustness certification. However, these methods face two critical limitations: (1) they yield loose robustness bounds due to the lack of semantic validation for perturbed outputs and (2) they suffer from high computational costs due to repeated sampling. To address these limitations, we propose CluCERT, a novel framework for certifying LLM robustness via clustering-guided denoising smoothing. Specifically, to achieve tighter certified bounds, we introduce a semantic clustering filter that reduces noisy samples and retains meaningful perturbations, supported by theoretical analysis. Furthermore, we enhance computational efficiency through two mechanisms: a refine module that extracts core semantics, and a fast synonym substitution strategy that accelerates the denoising process. Finally, we conduct extensive experiments on various downstream tasks and jailbreak defense scenarios. Experimental results demonstrate that our method outperforms existing certified approaches in both robustness bounds and computational efficiency.

42. Financial Instruction Following Evaluation (FIFE)

Authors: Glenn Matlin , Siddharth , Anirudh JM , Aditya Shukla , Yahya Hassan , Sudheer Chava
URL: https://arxiv.org/abs/2512.08965
Abstract:

Language Models (LMs) struggle with complex, interdependent instructions, particularly in high-stakes domains like finance where precision is critical. We introduce FIFE, a novel, high-difficulty benchmark designed to assess LM instruction-following capabilities for financial analysis tasks. FIFE comprises 88 human-authored prompts and employs a verification system with chainable, verifiable constraints for fine-grained reward signals. We evaluate 53 models (proprietary, open-weight, open-source) in a zero-shot setting. Our key findings reveal a clear performance hierarchy: the top open-weight model (76.1 strict / 79.5 loose) surpasses the leading proprietary system (65.9 strict / 70.5 loose), while the best open-source models lag significantly (45.5 strict / 48.9 loose). However, even top-performing models struggle with FIFE’s complex requirements, failing to achieve perfect compliance. We release our dataset and code as an open-source resource to promote research in Reinforcement Learning for the financial domain.

43. LLM4XCE: Large Language Models for Extremely Large-Scale Massive MIMO Channel Estimation

Authors: Renbin Li , Shuangshuang Li , Peihao Dong
URL: https://arxiv.org/abs/2512.08955
Abstract:

Extremely large-scale massive multiple-input multiple-output (XL-MIMO) is a key enabler for sixth-generation (6G) networks, offering massive spatial degrees of freedom. Despite these advantages, the coexistence of near-field and far-field effects in hybrid-field channels presents significant challenges for accurate estimation, where traditional methods often struggle to generalize effectively. In recent years, large language models (LLMs) have achieved impressive performance on downstream tasks via fine-tuning, aligning with the semantic communication shift toward task-oriented understanding over bit-level accuracy. Motivated by this, we propose Large Language Models for XL-MIMO Channel Estimation (LLM4XCE), a novel channel estimation framework that leverages the semantic modeling capabilities of large language models to recover essential spatial-channel representations for downstream tasks. The model integrates a carefully designed embedding module with Parallel Feature-Spatial Attention, enabling deep fusion of pilot features and spatial structures to construct a semantically rich representation for LLM input. By fine-tuning only the top two Transformer layers, our method effectively captures latent dependencies in the pilot data while ensuring high training efficiency. Extensive simulations demonstrate that LLM4XCE significantly outperforms existing state-of-the-art methods under hybrid-field conditions, achieving superior estimation accuracy and generalization performance.

44. AI Co-Artist: A LLM-Powered Framework for Interactive GLSL Shader Animation Evolution

Authors: Kamer Ali Yuksel , Hassan Sawaf
URL: https://arxiv.org/abs/2512.08951
Abstract:

Creative coding and real-time shader programming are at the forefront of interactive digital art, enabling artists, designers, and enthusiasts to produce mesmerizing, complex visual effects that respond to real-time stimuli such as sound or user interaction. However, despite the rich potential of tools like GLSL, the steep learning curve and requirement for programming fluency pose substantial barriers for newcomers and even experienced artists who may not have a technical background. In this paper, we present AI Co-Artist, a novel interactive system that harnesses the capabilities of large language models (LLMs), specifically GPT-4, to support the iterative evolution and refinement of GLSL shaders through a user-friendly, visually-driven interface. Drawing inspiration from the user-guided evolutionary principles pioneered by the Picbreeder platform, our system empowers users to evolve shader art using intuitive interactions, without needing to write or understand code. AI Co-Artist serves as both a creative companion and a technical assistant, allowing users to explore a vast generative design space of real-time visual art. Through comprehensive evaluations, including structured user studies and qualitative feedback, we demonstrate that AI Co-Artist significantly reduces the technical threshold for shader creation, enhances creative outcomes, and supports a wide range of users in producing professional-quality visual effects. Furthermore, we argue that this paradigm is broadly generalizable. By leveraging the dual strengths of LLMs-semantic understanding and program synthesis, our method can be applied to diverse creative domains, including website layout generation, architectural visualizations, product prototyping, and infographics.

45. The Linguistic Architecture of Reflective Thought: Evaluation of a Large Language Model as a Tool to Isolate the Formal Structure of Mentalization

Authors: Stefano Epifani (1 and 2), Giuliano Castigliego (2 and 3), Laura Kecskemeti (4), Giuliano Razzicchia (2), Elisabeth Seiwald-Sonderegger (4) ((1) University of Pavia Italy, (2) Digital Transformation Institute Italy, (3) Psychoanalytic Academy of Italian-Speaking Switzerland, (4) Psychiatric Services of the Canton of Grisons Switzerland)
URL: https://arxiv.org/abs/2512.08945
Abstract:

Background: Mentalization integrates cognitive, affective, and intersubjective components. Large Language Models (LLMs) display an increasing ability to generate reflective texts, raising questions regarding the relationship between linguistic form and mental representation. This study assesses the extent to which a single LLM can reproduce the linguistic structure of mentalization according to the parameters of Mentalization-Based Treatment (MBT). Methods: Fifty dialogues were generated between human participants and an LLM configured in standard mode. Five psychiatrists trained in MBT, working under blinded conditions, evaluated the mentalization profiles produced by the model along the four MBT axes, assigning Likert-scale scores for evaluative coherence, argumentative coherence, and global quality. Inter-rater agreement was estimated using ICC(3,1). Results: Mean scores (3.63-3.98) and moderate standard deviations indicate a high level of structural coherence in the generated profiles. ICC values (0.60-0.84) show substantial-to-high agreement among raters. The model proved more stable in the Implicit-Explicit and Self-Other dimensions, while presenting limitations in the integration of internal states and external contexts. The profiles were coherent and clinically interpretable yet characterized by affective neutrality.

46. Noise-Robust Abstractive Compression in Retrieval-Augmented Language Models

Authors: Singon Kim
URL: https://arxiv.org/abs/2512.08943
Abstract:

Abstractive compression utilizes smaller langauge models to condense query-relevant context, reducing computational costs in retrieval-augmented generation (RAG). However, retrieved documents often include information that is either irrelevant to answering the query or misleading due to factual incorrect content, despite having high relevance scores. This behavior indicates that abstractive compressors are more likely to omit important information essential for the correct answer, especially in long contexts where attention dispersion occurs. To address this issue, we categorize retrieved documents in a more fine-grained manner and propose Abstractive Compression Robust against Noise (ACoRN), which introduces two novel training steps. First, we use offline data augmentation on the training dataset to enhance compressor robustness against two distinct types of retrieval noise. Second, since the language model based compressor cannot fully utilize information from multiple retrieved documents and exhibits positional bias, we perform finetuning to generate summaries centered around key information that directly supports the correct answer. Our experiments demonstrate that T5-large, trained with ACoRN as a compressor, improves EM and F1 scores while preserving the answer string, which could serve as direct evidence. ACoRN excels on datasets with many accuracy reducing documents, making it highly useful in real-world scenarios.

47. Assessing the Human-Likeness of LLM-Driven Digital Twins in Simulating Health Care System Trust

Authors: Yuzhou Wu , Mingyang Wu , Di Liu , Rong Yin , Kang Li
URL: https://arxiv.org/abs/2512.08939
Abstract:

Serving as an emerging and powerful tool, Large Language Model (LLM)-driven Human Digital Twins are showing great potential in healthcare system research. However, its actual simulation ability for complex human psychological traits, such as distrust in the healthcare system, remains unclear. This research gap particularly impacts health professionals’ trust and usage of LLM-based Artificial Intelligence (AI) systems in assisting their routine work. In this study, based on the Twin-2K-500 dataset, we systematically evaluated the simulation results of the LLM-driven human digital twin using the Health Care System Distrust Scale (HCSDS) with an established human-subject sample, analyzing item-level distributions, summary statistics, and demographic subgroup patterns. Results showed that the simulated responses by the digital twin were significantly more centralized with lower variance and had fewer selections of extreme options (all p<0.001). While the digital twin broadly reproduces human results in major demographic patterns, such as age and gender, it exhibits relatively low sensitivity in capturing minor differences in education levels. The LLM-based digital twin simulation has the potential to simulate population trends, but it also presents challenges in making detailed, specific distinctions in subgroups of human beings. This study suggests that the current LLM-driven Digital Twins have limitations in modeling complex human attitudes, which require careful calibration and validation before applying them in inferential analyses or policy simulations in health systems engineering. Future studies are necessary to examine the emotional reasoning mechanism of LLMs before their use, particularly for studies that involve simulations sensitive to social topics, such as human-automation trust.

48. When AI Gives Advice: Evaluating AI and Human Responses to Online Advice-Seeking for Well-Being

Authors: Harsh Kumar , Jasmine Chahal , Yinuo Zhao , Zeling Zhang , Annika Wei , Louis Tay , Ashton Anderson
URL: https://arxiv.org/abs/2512.08937
Abstract:

Seeking advice is a core human behavior that the Internet has reinvented twice: first through forums and Q\&A communities that crowdsource public guidance, and now through large language models (LLMs) that deliver private, on-demand counsel at scale. Yet the quality of this synthesized LLM advice remains unclear. How does it compare, not only against arbitrary human comments, but against the wisdom of the online crowd? We conducted two studies (N = 210) in which experts compared top-voted Reddit advice with LLM-generated advice. LLMs ranked significantly higher overall and on effectiveness, warmth, and willingness to seek advice again. GPT-4o beat GPT-5 on all metrics except sycophancy, suggesting that benchmark gains need not improve advice-giving. In our second study, we examined how human and algorithmic advice could be combined, and found that human advice can be unobtrusively polished to compete with AI-generated comments. Finally, to surface user expectations, we ran an exploratory survey with undergraduates (N=148) that revealed heterogeneous, persona-dependent preferences for agent qualities (e.g., coach-like: goal-focused structure; friend-like: warmth and humor). We conclude with design implications for advice-giving agents and ecosystems blending AI, crowd input, and expert oversight.

49. A Principle-based Framework for the Development and Evaluation of Large Language Models for Health and Wellness

Authors: Brent Winslow , Jacqueline Shreibati , Javier Perez , Hao-Wei Su , Nichole Young-Lin , Nova Hammerquist , Daniel McDuff , Jason Guss , Jenny Vafeiadou , Nick Cain , Alex Lin , Erik Schenck , Shiva Rajagopal , Jia-Ru Chung , Anusha Venkatakrishnan , Amy Armento Lee , Maryam Karimzadehgan , Qingyou Meng , Rythm Agarwal , Aravind Natarajan , Tracy Giest
URL: https://arxiv.org/abs/2512.08936
Abstract:

The incorporation of generative artificial intelligence into personal health applications presents a transformative opportunity for personalized, data-driven health and fitness guidance, yet also poses challenges related to user safety, model accuracy, and personal privacy. To address these challenges, a novel, principle-based framework was developed and validated for the systematic evaluation of LLMs applied to personal health and wellness. First, the development of the Fitbit Insights explorer, a large language model (LLM)-powered system designed to help users interpret their personal health data, is described. Subsequently, the safety, helpfulness, accuracy, relevance, and personalization (SHARP) principle-based framework is introduced as an end-to-end operational methodology that integrates comprehensive evaluation techniques including human evaluation by generalists and clinical specialists, autorater assessments, and adversarial testing, into an iterative development lifecycle. Through the application of this framework to the Fitbit Insights explorer in a staged deployment involving over 13,000 consented users, challenges not apparent during initial testing were systematically identified. This process guided targeted improvements to the system and demonstrated the necessity of combining isolated technical evaluations with real-world user feedback. Finally, a comprehensive, actionable approach is established for the responsible development and deployment of LLM-powered health applications, providing a standardized methodology to foster innovation while ensuring emerging technologies are safe, effective, and trustworthy for users.

50. Motion2Meaning: A Clinician-Centered Framework for Contestable LLM in Parkinson’s Disease Gait Interpretation

Authors: Loc Phuc Truong Nguyen , Hung Thanh Do , Hung Truong Thanh Nguyen , Hung Cao
URL: https://arxiv.org/abs/2512.08934
Abstract:

AI-assisted gait analysis holds promise for improving Parkinson’s Disease (PD) care, but current clinical dashboards lack transparency and offer no meaningful way for clinicians to interrogate or contest AI decisions. To address this issue, we present Motion2Meaning, a clinician-centered framework that advances Contestable AI through a tightly integrated interface designed for interpretability, oversight, and procedural recourse. Our approach leverages vertical Ground Reaction Force (vGRF) time-series data from wearable sensors as an objective biomarker of PD motor states. The system comprises three key components: a Gait Data Visualization Interface (GDVI), a one-dimensional Convolutional Neural Network (1D-CNN) that predicts Hoehn & Yahr severity stages, and a Contestable Interpretation Interface (CII) that combines our novel Cross-Modal Explanation Discrepancy (XMED) safeguard with a contestable Large Language Model (LLM). Our 1D-CNN achieves 89.0% F1-score on the public PhysioNet gait dataset. XMED successfully identifies model unreliability by detecting a five-fold increase in explanation discrepancies in incorrect predictions (7.45%) compared to correct ones (1.56%), while our LLM-powered interface enables clinicians to validate correct predictions and successfully contest a portion of the model’s errors. A human-centered evaluation of this contestable interface reveals a crucial trade-off between the LLM’s factual grounding and its readability and responsiveness to clinical feedback. This work demonstrates the feasibility of combining wearable sensor analysis with Explainable AI (XAI) and contestable LLMs to create a transparent, auditable system for PD gait interpretation that maintains clinical oversight while leveraging advanced AI capabilities. Our implementation is publicly available at: this https URL .