AI 연구 논문

Vision-Language-Action (VLA) tasks require reasoning over complex visual scenes and executing adaptive actions in dynamic environments. While recent studies on reasoning VLAs show that explicit chain-of-thought (CoT) can improve generalization, they suffer from high inference latency due to lengthy reasoning traces. We propose Fast-ThinkAct, an efficient reasoning framework that achieves compact yet performant planning through verbalizable latent reasoning. Fast-ThinkAct learns to reason efficiently with latent CoTs by distilling from a teacher, driven by a preference-guided objective to align manipulation trajectories that transfers both linguistic and visual planning capabilities for embodied control. This enables reasoning-enhanced policy learning that effectively connects compact reasoning to action execution. Extensive experiments across diverse embodied manipulation and reasoning benchmarks demonstrate that Fast-ThinkAct achieves strong performance with up to 89.3\% reduced inference latency over state-of-the-art reasoning VLAs, while maintaining effective long-horizon planning, few-shot adaptation, and failure recovery.

Transformer-based language models often achieve strong results on mathematical reasoning benchmarks while remaining fragile on basic numerical understanding and arithmetic operations. A central limitation is that numbers are processed as symbolic tokens whose embeddings do not explicitly encode numerical value, leading to systematic errors. We introduce a value-aware numerical representation that augments standard tokenized inputs with a dedicated prefix token whose embedding is explicitly conditioned on the underlying numerical value. This mechanism injects magnitude information directly into the model's input space while remaining compatible with existing tokenizers and decoder-only Transformer architectures. Evaluation on arithmetic tasks shows that the proposed approach outperforms baselines across numerical formats, tasks, and operand lengths. These results indicate that explicitly encoding numerical value is an effective and efficient way to improve fundamental numerical robustness in language models.

Code generation tasks aim to automate the conversion of user requirements into executable code, significantly reducing manual development efforts and enhancing software productivity. The emergence of large language models (LLMs) has significantly advanced code generation, though their efficiency is still impacted by certain inherent architectural constraints. Each token generation necessitates a complete inference pass, requiring persistent retention of contextual information in memory and escalating resource consumption. While existing research prioritizes inference-phase optimizations such as prompt compression and model quantization, the generation phase remains underexplored. To tackle these challenges, we propose a knowledge-infused framework named ShortCoder, which optimizes code generation efficiency while preserving semantic equivalence and readability. In particular, we introduce: (1) ten syntax-level simplification rules for Python, derived from AST-preserving transformations, achieving 18.1% token reduction without functional compromise; (2) a hybrid data synthesis pipeline integrating rule-based rewriting with LLM-guided refinement, producing ShorterCodeBench, a corpus of validated tuples of original code and simplified code with semantic consistency; (3) a fine-tuning strategy that injects conciseness awareness into the base LLMs. Extensive experimental results demonstrate that ShortCoder consistently outperforms state-of-the-art methods on HumanEval, achieving an improvement of 18.1%-37.8% in generation efficiency over previous methods while ensuring the performance of code generation.

Segment Anything 3 (SAM3) has established a powerful foundation that robustly detects, segments, and tracks specified targets in videos. However, in its original implementation, its group-level collective memory selection is suboptimal for complex multi-object scenarios, as it employs a synchronized decision across all concurrent targets conditioned on their average performance, often overlooking individual reliability. To this end, we propose SAM3-DMS, a training-free decoupled strategy that utilizes fine-grained memory selection on individual objects. Experiments demonstrate that our approach achieves robust identity preservation and tracking stability. Notably, our advantage becomes more pronounced with increased target density, establishing a solid foundation for simultaneous multi-target video segmentation in the wild.

3D pose estimation from sparse multi-views is a critical task for numerous applications, including action recognition, sports analysis, and human-robot interaction. Optimization-based methods typically follow a two-stage pipeline, first detecting 2D keypoints in each view and then associating these detections across views to triangulate the 3D pose. Existing methods rely on mere pairwise associations to model this correspondence problem, treating global consistency between views (i.e., cycle consistency) as a soft constraint. Yet, reconciling these constraints for multiple views becomes brittle when spurious associations propagate errors. We thus propose COMPOSE, a novel framework that formulates multi-view pose correspondence matching as a hypergraph partitioning problem rather than through pairwise association. While the complexity of the resulting integer linear program grows exponentially in theory, we introduce an efficient geometric pruning strategy to substantially reduce the search space. COMPOSE achieves improvements of up to 23% in average precision over previous optimization-based methods and up to 11% over self-supervised end-to-end learned methods, offering a promising solution to a widely studied problem.

Modern video generative models based on diffusion models can produce very realistic clips, but they are computationally inefficient, often requiring minutes of GPU time for just a few seconds of video. This inefficiency poses a critical barrier to deploying generative video in applications that require real-time interactions, such as embodied AI and VR/AR. This paper explores a new strategy for camera-conditioned video generation of static scenes: using diffusion-based generative models to generate a sparse set of keyframes, and then synthesizing the full video through 3D reconstruction and rendering. By lifting keyframes into a 3D representation and rendering intermediate views, our approach amortizes the generation cost across hundreds of frames while enforcing geometric consistency. We further introduce a model that predicts the optimal number of keyframes for a given camera trajectory, allowing the system to adaptively allocate computation. Our final method, SRENDER, uses very sparse keyframes for simple trajectories and denser ones for complex camera motion. This results in video generation that is more than 40 times faster than the diffusion-based baseline in generating 20 seconds of video, while maintaining high visual fidelity and temporal stability, offering a practical path toward efficient and controllable video synthesis.

LLMs are increasingly being integrated into clinical workflows, yet they often lack clinical empathy, an essential aspect of effective doctor-patient communication. Existing NLP frameworks focus on reactively labeling empathy in doctors' responses but offer limited support for anticipatory modeling of empathy needs, especially in general health queries. We introduce the Empathy Applicability Framework (EAF), a theory-driven approach that classifies patient queries in terms of the applicability of emotional reactions and interpretations, based on clinical, contextual, and linguistic cues. We release a benchmark of real patient queries, dual-annotated by Humans and GPT-4o. In the subset with human consensus, we also observe substantial human-GPT alignment. To validate EAF, we train classifiers on human-labeled and GPT-only annotations to predict empathy applicability, achieving strong performance and outperforming the heuristic and zero-shot LLM baselines. Error analysis highlights persistent challenges: implicit distress, clinical-severity ambiguity, and contextual hardship, underscoring the need for multi-annotator modeling, clinician-in-the-loop calibration, and culturally diverse annotation. EAF provides a framework for identifying empathy needs before response generation, establishes a benchmark for anticipatory empathy modeling, and enables supporting empathetic communication in asynchronous healthcare.

As Large Language Models (LLMs) continue to scale, post-training pruning has emerged as a promising approach to reduce computational costs while preserving performance. Existing methods such as SparseGPT and Wanda achieve high sparsity through layer-wise weight reconstruction or activation-aware magnitude pruning, but rely on uniform or hand-crafted heuristics to determine per-layer sparsity ratios. Moreover, recent work has shown that pruned LLMs suffer from severe factual knowledge degradation, with structured pruning methods experiencing near-total collapse in factual question-answering capabilities. We introduce agent-guided pruning, where a foundation model acts as an adaptive pruning agent to intelligently select which layers to prune at each iteration while preserving critical knowledge pathways. Our method constructs layer-wise sensitivity profiles by combining Wanda-inspired weight-activation metrics with gradient importance scores, normalized as z-scores for model-agnostic comparison. These statistics are processed by an LLM agent equipped with self-reflection capabilities, enabling it to learn from previous pruning outcomes and iteratively refine its strategy. A checkpoint rollback mechanism maintains model quality by reverting when perplexity degradation exceeds a threshold. We evaluate our approach on Qwen3 models (4B and 8B parameters) at approximately 45% sparsity, demonstrating substantial improvements over structured pruning baselines: 56% relative improvement in MMLU accuracy, 19x better factual knowledge retention on FreebaseQA, and 69% lower perplexity degradation. Notably, our framework requires no retraining, operates in a model-agnostic manner, and exhibits effective self-correction with only 2-4 rollbacks across 21-40 iterations, demonstrating that foundation models can effectively guide the compression of other foundation models.

Structure-based and ligand-based computational drug design have traditionally relied on disjoint data sources and modeling assumptions, limiting their joint use at scale. In this work, we introduce Contrastive Geometric Learning for Unified Computational Drug Design (ConGLUDe), a single contrastive geometric model that unifies structure- and ligand-based training. ConGLUDe couples a geometric protein encoder that produces whole-protein representations and implicit embeddings of predicted binding sites with a fast ligand encoder, removing the need for pre-defined pockets. By aligning ligands with both global protein representations and multiple candidate binding sites through contrastive learning, ConGLUDe supports ligand-conditioned pocket prediction in addition to virtual screening and target fishing, while being trained jointly on protein-ligand complexes and large-scale bioactivity data. Across diverse benchmarks, ConGLUDe achieves state-of-the-art zero-shot virtual screening performance in settings where no binding pocket information is provided as input, substantially outperforms existing methods on a challenging target fishing task, and demonstrates competitive ligand-conditioned pocket selection. These results highlight the advantages of unified structure-ligand training and position ConGLUDe as a step toward general-purpose foundation models for drug discovery.

Large Language Model (LLM) routers dynamically select optimal models for given inputs. Existing approaches typically assume access to ground-truth labeled data, which is often unavailable in practice, especially when user request distributions are heterogeneous and unknown. We introduce Routing with Generated Data (RGD), a challenging setting in which routers are trained exclusively on generated queries and answers produced from high-level task descriptions by generator LLMs. We evaluate query-answer routers (using both queries and labels) and query-only routers across four diverse benchmarks and 12 models, finding that query-answer routers degrade faster than query-only routers as generator quality decreases. Our analysis reveals two crucial characteristics of effective generators: they must accurately respond to their own questions, and their questions must produce sufficient performance differentiation among the model pool. We then show how filtering for these characteristics can improve the quality of generated data. We further propose CASCAL, a novel query-only router that estimates model correctness through consensus voting and identifies model-specific skill niches via hierarchical clustering. CASCAL is substantially more robust to generator quality, outperforming the best query-answer router by 4.6% absolute accuracy when trained on weak generator data.

Deep research systems are widely used for multi-step web research, analysis, and cross-source synthesis, yet their evaluation remains challenging. Existing benchmarks often require annotation-intensive task construction, rely on static evaluation dimensions, or fail to reliably verify facts when citations are missing. To bridge these gaps, we introduce DeepResearchEval, an automated framework for deep research task construction and agentic evaluation. For task construction, we propose a persona-driven pipeline generating realistic, complex research tasks anchored in diverse user profiles, applying a two-stage filter Task Qualification and Search Necessity to retain only tasks requiring multi-source evidence integration and external retrieval. For evaluation, we propose an agentic pipeline with two components: an Adaptive Point-wise Quality Evaluation that dynamically derives task-specific evaluation dimensions, criteria, and weights conditioned on each generated task, and an Active Fact-Checking that autonomously extracts and verifies report statements via web search, even when citations are missing.

The graphical Lasso (GLASSO) is a widely used algorithm for learning high-dimensional undirected Gaussian graphical models (GGM). Given i.i.d. observations from a multivariate normal distribution, GLASSO estimates the precision matrix by maximizing the log-likelihood with an \ell_1-penalty on the off-diagonal entries. However, selecting an optimal regularization parameter λin this unsupervised setting remains a significant challenge. A well-known issue is that existing methods, such as out-of-sample likelihood maximization, select a single global λand do not account for heterogeneity in variable scaling or partial variances. Standardizing the data to unit variances, although a common workaround, has been shown to negatively affect graph recovery. Addressing the problem of nodewise adaptive tuning in graph estimation is crucial for applications like computational neuroscience, where brain networks are constructed from highly heterogeneous, region-specific fMRI data. In this work, we develop Locally Adaptive Regularization for Graph Estimation (LARGE), an approach to adaptively learn nodewise tuning parameters to improve graph estimation and selection. In each block coordinate descent step of GLASSO, we augment the nodewise Lasso regression to jointly estimate the regression coefficients and error variance, which in turn guides the adaptive learning of nodewise penalties. In simulations, LARGE consistently outperforms benchmark methods in graph recovery, demonstrates greater stability across replications, and achieves the best estimation accuracy in the most difficult simulation settings. We demonstrate the practical utility of our method by estimating brain functional connectivity from a real fMRI data set.

Multi-Task Learning (MTL) combined with Low-Rank Adaptation (LoRA) has emerged as a promising direction for parameter-efficient deployment of Large Language Models (LLMs). By sharing a single adapter across multiple tasks, one can significantly reduce storage overhead. However, this approach suffers from negative transfer, where conflicting gradient updates from distinct tasks degrade the performance of individual tasks compared to single-task fine-tuning. This problem is exacerbated in LoRA due to the low-rank constraint, which limits the optimization landscape's capacity to accommodate diverse task requirements. In this paper, we propose Ortho-LoRA, a gradient projection method specifically tailored for the bipartite structure of LoRA. Ortho-LoRA dynamically projects conflicting task gradients onto the orthogonal complement of each other within the intrinsic LoRA subspace. Extensive experiments on the GLUE benchmark demonstrate that Ortho-LoRA effectively mitigates task interference, outperforming standard joint training and recovering 95\% of the performance gap between multi-task and single-task baselines with negligible computational overhead.

Modern supply chains are increasingly exposed to disruptions from geopolitical events, demand shocks, trade restrictions, to natural disasters. While many of these disruptions originate deep in the supply network, most companies still lack visibility beyond Tier-1 suppliers, leaving upstream vulnerabilities undetected until the impact cascades downstream. To overcome this blind-spot and move from reactive recovery to proactive resilience, we introduce a minimally supervised agentic AI framework that autonomously monitors, analyses, and responds to disruptions across extended supply networks. The architecture comprises seven specialised agents powered by large language models and deterministic tools that jointly detect disruption signals from unstructured news, map them to multi-tier supplier networks, evaluate exposure based on network structure, and recommend mitigations such as alternative sourcing options. \rev{We evaluate the framework across 30 synthesised scenarios covering three automotive manufacturers and five disruption classes. The system achieves high accuracy across core tasks, with F1 scores between 0.962 and 0.991, and performs full end-to-end analyses in a mean of 3.83 minutes at a cost of \$0.0836 per disruption. Relative to industry benchmarks of multi-day, analyst-driven assessments, this represents a reduction of more than three orders of magnitude in response time. A real-world case study of the 2022 Russia-Ukraine conflict further demonstrates operational applicability. This work establishes a foundational step toward building resilient, proactive, and autonomous supply chains capable of managing disruptions across deep-tier networks.

We present STEP3-VL-10B, a lightweight open-source foundation model designed to redefine the trade-off between compact efficiency and frontier-level multimodal intelligence. STEP3-VL-10B is realized through two strategic shifts: first, a unified, fully unfrozen pre-training strategy on 1.2T multimodal tokens that integrates a language-aligned Perception Encoder with a Qwen3-8B decoder to establish intrinsic vision-language synergy; and second, a scaled post-training pipeline featuring over 1k iterations of reinforcement learning. Crucially, we implement Parallel Coordinated Reasoning (PaCoRe) to scale test-time compute, allocating resources to scalable perceptual reasoning that explores and synthesizes diverse visual hypotheses. Consequently, despite its compact 10B footprint, STEP3-VL-10B rivals or surpasses models 10$\times$-20$\times$ larger (e.g., GLM-4.6V-106B, Qwen3-VL-235B) and top-tier proprietary flagships like Gemini 2.5 Pro and Seed-1.5-VL. Delivering best-in-class performance, it records 92.2% on MMBench and 80.11% on MMMU, while excelling in complex reasoning with 94.43% on AIME2025 and 75.95% on MathVision. We release the full model suite to provide the community with a powerful, efficient, and reproducible baseline.

Multi-agent systems have evolved into practical LLM-driven collaborators for many applications, gaining robustness from diversity and cross-checking. However, multi-agent RL (MARL) training is resource-intensive and unstable: co-adapting teammates induce non-stationarity, and rewards are often sparse and high-variance. Therefore, we introduce \textbf{Multi-Agent Test-Time Reinforcement Learning (MATTRL)}, a framework that injects structured textual experience into multi-agent deliberation at inference time. MATTRL forms a multi-expert team of specialists for multi-turn discussions, retrieves and integrates test-time experiences, and reaches consensus for final decision-making. We also study credit assignment for constructing a turn-level experience pool, then reinjecting it into the dialogue. Across challenging benchmarks in medicine, math, and education, MATTRL improves accuracy by an average of 3.67\% over a multi-agent baseline, and by 8.67\% over comparable single-agent baselines. Ablation studies examine different credit-assignment schemes and provide a detailed comparison of how they affect training outcomes. MATTRL offers a stable, effective and efficient path to distribution-shift-robust multi-agent reasoning without tuning.

Monocular visual SLAM enables 3D reconstruction from internet video and autonomous navigation on resource-constrained platforms, yet suffers from scale drift, i.e., the gradual divergence of estimated scale over long sequences. Existing frame-to-frame methods achieve real-time performance through local optimization but accumulate scale drift due to the lack of global constraints among independent windows. To address this, we propose SCE-SLAM, an end-to-end SLAM system that maintains scale consistency through scene coordinate embeddings, which are learned patch-level representations encoding 3D geometric relationships under a canonical scale reference. The framework consists of two key modules: geometry-guided aggregation that leverages 3D spatial proximity to propagate scale information from historical observations through geometry-modulated attention, and scene coordinate bundle adjustment that anchors current estimates to the reference scale through explicit 3D coordinate constraints decoded from the scene coordinate embeddings. Experiments on KITTI, Waymo, and vKITTI demonstrate substantial improvements: our method reduces absolute trajectory error by 8.36m on KITTI compared to the best prior approach, while maintaining 36 FPS and achieving scale consistency across large-scale scenes.

Identifying individual animals in long-duration videos is essential for behavioral ecology, wildlife monitoring, and livestock management. Traditional methods require extensive manual annotation, while existing self-supervised approaches are computationally demanding and ill-suited for long sequences due to memory constraints and temporal error propagation. We introduce a highly efficient, self-supervised method that reframes animal identification as a global clustering task rather than a sequential tracking problem. Our approach assumes a known, fixed number of individuals within a single video -- a common scenario in practice -- and requires only bounding box detections and the total count. By sampling pairs of frames, using a frozen pre-trained backbone, and employing a self-bootstrapping mechanism with the Hungarian algorithm for in-batch pseudo-label assignment, our method learns discriminative features without identity labels. We adapt a Binary Cross Entropy loss from vision-language models, enabling state-of-the-art accuracy ($>$97\%) while consuming less than 1 GB of GPU memory per batch -- an order of magnitude less than standard contrastive methods. Evaluated on challenging real-world datasets (3D-POP pigeons and 8-calves feeding videos), our framework matches or surpasses supervised baselines trained on over 1,000 labeled frames, effectively removing the manual annotation bottleneck. This work enables practical, high-accuracy animal identification on consumer-grade hardware, with broad applicability in resource-constrained research settings. All code written for this paper are \href{https://huggingface.co/datasets/tonyFang04/8-calves}{here}.

Large-scale vision-language models such as CLIP achieve strong zero-shot recognition but struggle with classes that are rarely seen during pretraining, including newly emerging entities and culturally specific categories. We introduce LiteEmbed, a lightweight framework for few-shot personalization of CLIP that enables new classes to be added without retraining its encoders. LiteEmbed performs subspace-guided optimization of text embeddings within CLIP's vocabulary, leveraging a PCA-based decomposition that disentangles coarse semantic directions from fine-grained variations. Two complementary objectives, coarse alignment and fine separation, jointly preserve global semantic consistency while enhancing discriminability among visually similar classes. Once optimized, the embeddings are plug-and-play, seamlessly substituting CLIP's original text features across classification, retrieval, segmentation, and detection tasks. Extensive experiments demonstrate substantial gains over prior methods, establishing LiteEmbed as an effective approach for adapting CLIP to underrepresented, rare, or unseen classes.

Estimating physically accurate, simulation-ready garments from a single image is challenging due to the absence of image-to-physics datasets and the ill-posed nature of this problem. Prior methods either require multi-view capture and expensive differentiable simulation or predict only garment geometry without the material properties required for realistic simulation. We propose a feed-forward framework that sidesteps these limitations by first fine-tuning a vision-language model to infer material composition and fabric attributes from real images, and then training a lightweight predictor that maps these attributes to the corresponding physical fabric parameters using a small dataset of material-physics measurements. Our approach introduces two new datasets (FTAG and T2P) and delivers simulation-ready garments from a single image without iterative optimization. Experiments show that our estimator achieves superior accuracy in material composition estimation and fabric attribute prediction, and by passing them through our physics parameter estimator, we further achieve higher-fidelity simulations compared to state-of-the-art image-to-garment methods.

Tabular Foundation Models (TFMs) have recently shown strong in-context learning capabilities on structured data, achieving zero-shot performance comparable to traditional machine learning methods. We find that zero-shot TFMs already achieve strong performance, while the benefits of fine-tuning are highly model and data-dependent. Meta-learning and PEFT provide moderate gains under specific conditions, whereas full supervised fine-tuning (SFT) often reduces accuracy or calibration quality. This work presents the first comprehensive study of fine-tuning in TFMs across benchmarks including TALENT, OpenML-CC18, and TabZilla. We compare Zero-Shot, Meta-Learning, Supervised (SFT), and parameter-efficient (PEFT) approaches, analyzing how dataset factors such as imbalance, size, and dimensionality affect outcomes. Our findings cover performance, calibration, and fairness, offering practical guidelines on when fine-tuning is most beneficial and its limitations.

Underwater video analysis is particularly challenging due to factors such as low lighting, color distortion, and turbidity, which compromise visual data quality and directly impact the performance of perception modules in robotic applications. This work proposes AquaFeat+, a plug-and-play pipeline designed to enhance features specifically for automated vision tasks, rather than for human perceptual quality. The architecture includes modules for color correction, hierarchical feature enhancement, and an adaptive residual output, which are trained end-to-end and guided directly by the loss function of the final application. Trained and evaluated in the FishTrack23 dataset, AquaFeat+ achieves significant improvements in object detection, classification, and tracking metrics, validating its effectiveness for enhancing perception tasks in underwater robotic applications.

Word Sense Disambiguation (WSD) has been widely evaluated using the semantic frameworks of WordNet, BabelNet, and the Oxford Dictionary of English. However, for the UCREL Semantic Analysis System (USAS) framework, no open extensive evaluation has been performed beyond lexical coverage or single language evaluation. In this work, we perform the largest semantic tagging evaluation of the rule based system that uses the lexical resources in the USAS framework covering five different languages using four existing datasets and one novel Chinese dataset. We create a new silver labelled English dataset, to overcome the lack of manually tagged training data, that we train and evaluate various mono and multilingual neural models in both mono and cross-lingual evaluation setups with comparisons to their rule based counterparts, and show how a rule based system can be enhanced with a neural network model. The resulting neural network models, including the data they were trained on, the Chinese evaluation dataset, and all of the code have been released as open resources.

Text-to-image (T2I) models are increasingly popular, producing a large share of AI-generated images online. To compare model quality, voting-based leaderboards have become the standard, relying on anonymized model outputs for fairness. In this work, we show that such anonymity can be easily broken. We find that generations from each T2I model form distinctive clusters in the image embedding space, enabling accurate deanonymization without prompt control or training data. Using 22 models and 280 prompts (150K images), our centroid-based method achieves high accuracy and reveals systematic model-specific signatures. We further introduce a prompt-level distinguishability metric and conduct large-scale analyses showing how certain prompts can lead to near-perfect distinguishability. Our findings expose fundamental security flaws in T2I leaderboards and motivate stronger anonymization defenses.

While GUI agents have shown strong performance under explicit and completion instructions, real-world deployment requires aligning with users' more complex implicit intents. In this work, we highlight Hierarchical Implicit Intent Alignment for Personalized GUI Agent (PersonalAlign), a new agent task that requires agents to leverage long-term user records as persistent context to resolve omitted preferences in vague instructions and anticipate latent routines by user state for proactive assistance. To facilitate this study, we introduce AndroidIntent, a benchmark designed to evaluate agents' ability in resolving vague instructions and providing proactive suggestions through reasoning over long-term user records. We annotated 775 user-specific preferences and 215 routines from 20k long-term records across different users for evaluation. Furthermore, we introduce Hierarchical Intent Memory Agent (HIM-Agent), which maintains a continuously updating personal memory and hierarchically organizes user preferences and routines for personalization. Finally, we evaluate a range of GUI agents on AndroidIntent, including GPT-5, Qwen3-VL, and UI-TARS, further results show that HIM-Agent significantly improves both execution and proactive performance by 15.7% and 7.3%.

Large-scale optimization is a key backbone of modern business decision-making. However, building these models is often labor-intensive and time-consuming. We address this by proposing LEAN-LLM-OPT, a LightwEight AgeNtic workflow construction framework for LLM-assisted large-scale OPTimization auto-formulation. LEAN-LLM-OPT takes as input a problem description together with associated datasets and orchestrates a team of LLM agents to produce an optimization formulation. Specifically, upon receiving a query, two upstream LLM agents dynamically construct a workflow that specifies, step-by-step, how optimization models for similar problems can be formulated. A downstream LLM agent then follows this workflow to generate the final output. Leveraging LLMs' text-processing capabilities and common modeling practices, the workflow decomposes the modeling task into a sequence of structured sub-tasks and offloads mechanical data-handling operations to auxiliary tools. This design alleviates the downstream agent's burden related to planning and data handling, allowing it to focus on the most challenging components that cannot be readily standardized. Extensive simulations show that LEAN-LLM-OPT, instantiated with GPT-4.1 and the open source gpt-oss-20B, achieves strong performance on large-scale optimization modeling tasks and is competitive with state-of-the-art approaches. In addition, in a Singapore Airlines choice-based revenue management use case, LEAN-LLM-OPT demonstrates practical value by achieving leading performance across a range of scenarios. Along the way, we introduce Large-Scale-OR and Air-NRM, the first comprehensive benchmarks for large-scale optimization auto-formulation. The code and data of this work is available at https://github.com/CoraLiang01/lean-llm-opt.

Taxonomies form the backbone of structured knowledge representation across diverse domains, enabling applications such as e-commerce catalogs, semantic search, and biomedical discovery. Yet, manual taxonomy expansion is labor-intensive and cannot keep pace with the emergence of new concepts. Existing automated methods rely on point-based vector embeddings, which model symmetric similarity and thus struggle with the asymmetric "is-a" relationships that are fundamental to taxonomies. Box embeddings offer a promising alternative by enabling containment and disjointness, but they face key issues: (i) unstable gradients at the intersection boundaries, (ii) no notion of semantic uncertainty, and (iii) limited capacity to represent polysemy or ambiguity. We address these shortcomings with TaxoBell, a Gaussian box embedding framework that translates between box geometries and multivariate Gaussian distributions, where means encode semantic location and covariances encode uncertainty. Energy-based optimization yields stable optimization, robust modeling of ambiguous concepts, and interpretable hierarchical reasoning. Extensive experimentation on five benchmark datasets demonstrates that TaxoBell significantly outperforms eight state-of-the-art taxonomy expansion baselines by 19% in MRR and around 25% in Recall@k. We further demonstrate the advantages and pitfalls of TaxoBell with error analysis and ablation studies.

Large Language Models (LLMs), despite their remarkable capabilities across NLP tasks, struggle with phonologically-grounded phenomena like rhyme detection and generation. This is even more evident in lower-resource languages such as Modern Greek. In this paper, we present a hybrid system that combines LLMs with deterministic phonological algorithms to achieve accurate rhyme identification/analysis and generation. Our approach implements a comprehensive taxonomy of Greek rhyme types, including Pure, Rich, Imperfect, Mosaic, and Identical Pre-rhyme Vowel (IDV) patterns, and employs an agentic generation pipeline with phonological verification. We evaluate multiple prompting strategies (zero-shot, few-shot, Chain-of-Thought, and RAG-augmented) across several LLMs including Claude 3.7 and 4.5, GPT-4o, Gemini 2.0 and open-weight models like Llama 3.1 8B and 70B and Mistral Large. Results reveal a significant "Reasoning Gap": while native-like models (Claude 3.7) perform intuitively (40\% accuracy in identification), reasoning-heavy models (Claude 4.5) achieve state-of-the-art performance (54\%) only when prompted with Chain-of-Thought. Most critically, pure LLM generation fails catastrophically (under 4\% valid poems), while our hybrid verification loop restores performance to 73.1\%. We release our system and a crucial, rigorously cleaned corpus of 40,000+ rhymes, derived from the Anemoskala and Interwar Poetry corpora, to support future research.

Efficiently optimizing battery charging protocols is challenging because each evaluation is slow, costly, and non-differentiable. Many existing approaches address this difficulty by heavily constraining the protocol search space, which limits the diversity of protocols that can be explored, preventing the discovery of higher-performing solutions. We introduce two gradient-free, LLM-driven closed-loop methods: Prompt-to-Optimizer (P2O), which uses an LLM to propose the code for small neural-network-based protocols, which are then trained by an inner loop, and Prompt-to-Protocol (P2P), which simply writes an explicit function for the current and its scalar parameters. Across our case studies, LLM-guided P2O outperforms neural networks designed by Bayesian optimization, evolutionary algorithms, and random search. In a realistic fast charging scenario, both P2O and P2P yield around a 4.2 percent improvement in state of health (capacity retention based health metric under fast charging cycling) over a state-of-the-art multi-step constant current (CC) baseline, with P2P achieving this under matched evaluation budgets (same number of protocol evaluations). These results demonstrate that LLMs can expand the space of protocol functional forms, incorporate language-based constraints, and enable efficient optimization in high cost experimental settings.

The rapid adoption of large language model (LLM)-based systems -- from chatbots to autonomous agents capable of executing code and financial transactions -- has created a new attack surface that existing security frameworks inadequately address. The dominant framing of these threats as "prompt injection" -- a catch-all phrase for security failures in LLM-based systems -- obscures a more complex reality: Attacks on LLM-based systems increasingly involve multi-step sequences that mirror traditional malware campaigns. In this paper, we propose that attacks targeting LLM-based applications constitute a distinct class of malware, which we term \textit{promptware}, and introduce a five-step kill chain model for analyzing these threats. The framework comprises Initial Access (prompt injection), Privilege Escalation (jailbreaking), Persistence (memory and retrieval poisoning), Lateral Movement (cross-system and cross-user propagation), and Actions on Objective (ranging from data exfiltration to unauthorized transactions). By mapping recent attacks to this structure, we demonstrate that LLM-related attacks follow systematic sequences analogous to traditional malware campaigns. The promptware kill chain offers security practitioners a structured methodology for threat modeling and provides a common vocabulary for researchers across AI safety and cybersecurity to address a rapidly evolving threat landscape.

Machine unlearning is becoming essential for building trustworthy and compliant language models. Yet unlearning success varies considerably across individual samples: some are reliably erased, while others persist despite the same procedure. We argue that this disparity is not only a data-side phenomenon, but also reflects model-internal mechanisms that encode and protect memorized information. We study this problem from a mechanistic perspective based on model circuits--structured interaction pathways that govern how predictions are formed. We propose Circuit-guided Unlearning Difficulty (CUD), a {\em pre-unlearning} metric that assigns each sample a continuous difficulty score using circuit-level signals. Extensive experiments demonstrate that CUD reliably separates intrinsically easy and hard samples, and remains stable across unlearning methods. We identify key circuit-level patterns that reveal a mechanistic signature of difficulty: easy-to-unlearn samples are associated with shorter, shallower interactions concentrated in earlier-to-intermediate parts of the original model, whereas hard samples rely on longer and deeper pathways closer to late-stage computation. Compared to existing qualitative studies, CUD takes a first step toward a principled, fine-grained, and interpretable analysis of unlearning difficulty; and motivates the development of unlearning methods grounded in model mechanisms.

As artificial intelligence (AI) is increasingly integrated into news production, calls for transparency about the use of AI have gained considerable traction. Recent studies suggest that AI disclosures can lead to a ``transparency dilemma'', where disclosure reduces readers' trust. However, little is known about how the \textit{level of detail} in AI disclosures influences trust and contributes to this dilemma within the news context. In this 3$\times$2$\times$2 mixed factorial study with 40 participants, we investigate how three levels of AI disclosures (none, one-line, detailed) across two types of news (politics and lifestyle) and two levels of AI involvement (low and high) affect news readers' trust. We measured trust using the News Media Trust questionnaire, along with two decision behaviors: source-checking and subscription decisions. Questionnaire responses and subscription rates showed a decline in trust only for detailed AI disclosures, whereas source-checking behavior increased for both one-line and detailed disclosures, with the effect being more pronounced for detailed disclosures. Insights from semi-structured interviews suggest that source-checking behavior was primarily driven by interest in the topic, followed by trust, whereas trust was the main factor influencing subscription decisions. Around two-thirds of participants expressed a preference for detailed disclosures, while most participants who preferred one-line indicated a need for detail-on-demand disclosure formats. Our findings show that not all AI disclosures lead to a transparency dilemma, but instead reflect a trade-off between readers' desire for more transparency and their trust in AI-assisted news content.

Accurate and early perception of potential intrusion targets is essential for ensuring the safety of railway transportation systems. However, most existing systems focus narrowly on object classification within fixed visual scopes and apply rule-based heuristics to determine intrusion status, often overlooking targets that pose latent intrusion risks. Anticipating such risks requires the cognition of spatial context and temporal dynamics for the object of interest (OOI), which presents challenges for conventional visual models. To facilitate deep intrusion perception, we introduce a novel benchmark, CogRail, which integrates curated open-source datasets with cognitively driven question-answer annotations to support spatio-temporal reasoning and prediction. Building upon this benchmark, we conduct a systematic evaluation of state-of-the-art visual-language models (VLMs) using multimodal prompts to identify their strengths and limitations in this domain. Furthermore, we fine-tune VLMs for better performance and propose a joint fine-tuning framework that integrates three core tasks, position perception, movement prediction, and threat analysis, facilitating effective adaptation of general-purpose foundation models into specialized models tailored for cognitive intrusion perception. Extensive experiments reveal that current large-scale multimodal models struggle with the complex spatial-temporal reasoning required by the cognitive intrusion perception task, underscoring the limitations of existing foundation models in this safety-critical domain. In contrast, our proposed joint fine-tuning framework significantly enhances model performance by enabling targeted adaptation to domain-specific reasoning demands, highlighting the advantages of structured multi-task learning in improving both accuracy and interpretability. Code will be available at https://github.com/Hub-Tian/CogRail.

Reinforcement learning (RL)-based enhancement of large language models (LLMs) often leads to reduced output diversity, undermining their utility in open-ended tasks like creative writing. Current methods lack explicit mechanisms for guiding diverse exploration and instead prioritize optimization efficiency and performance over diversity. This paper proposes an RL framework structured around a semi-structured long Chain-of-Thought (CoT), in which the generation process is decomposed into explicitly planned intermediate steps. We introduce a Diverse Planning Branching method that strategically introduces divergence at the planning phase based on diversity variation, alongside a group-aware diversity reward to encourage distinct trajectories. Experimental results on creative writing benchmarks demonstrate that our approach significantly improves output diversity without compromising generation quality, consistently outperforming existing baselines.

Most Multimodal Sentiment Analysis research has focused on point-wise regression. While straightforward, this approach is sensitive to label noise and neglects whether one sample is more positive than another, resulting in unstable predictions and poor correlation alignment. Pairwise ordinal learning frameworks emerged to address this gap, capturing relative order by learning from comparisons. Yet, they introduce two new trade-offs: First, they assign uniform importance to all comparisons, failing to adaptively focus on hard-to-rank samples. Second, they employ static ranking margins, which fail to reflect the varying semantic distances between sentiment groups. To address this, we propose a Two-Stage Group-wise Ranking and Calibration Framework (GRCF) that adapts the philosophy of Group Relative Policy Optimization (GRPO). Our framework resolves these trade-offs by simultaneously preserving relative ordinal structure, ensuring absolute score calibration, and adaptively focusing on difficult samples. Specifically, Stage 1 introduces a GRPO-inspired Advantage-Weighted Dynamic Margin Ranking Loss to build a fine-grained ordinal structure. Stage 2 then employs an MAE-driven objective to align prediction magnitudes. To validate its generalizability, we extend GRCF to classification tasks, including multimodal humor detection and sarcasm detection. GRCF achieves state-of-the-art performance on core regression benchmarks, while also showing strong generalizability in classification tasks.

Vision-based policies for robot manipulation have achieved significant recent success, but are still brittle to distribution shifts such as camera viewpoint variations. Robot demonstration data is scarce and often lacks appropriate variation in camera viewpoints. Simulation offers a way to collect robot demonstrations at scale with comprehensive coverage of different viewpoints, but presents a visual sim2real challenge. To bridge this gap, we propose MANGO -- an unpaired image translation method with a novel segmentation-conditioned InfoNCE loss, a highly-regularized discriminator design, and a modified PatchNCE loss. We find that these elements are crucial for maintaining viewpoint consistency during sim2real translation. When training MANGO, we only require a small amount of fixed-camera data from the real world, but show that our method can generate diverse unseen viewpoints by translating simulated observations. In this domain, MANGO outperforms all other image translation methods we tested. Imitation-learning policies trained on data augmented by MANGO are able to achieve success rates as high as 60\% on views that the non-augmented policy fails completely on.

In recent years, foundation models have become very popular due to their exceptional performance, mainly in natural language (NLP) tasks where they were first introduced. These models usually consist of hundreds of millions, or even billions, of parameters, making them resource-intensive during training and in production systems, leading to increased costs. This paper focuses on the reduction of a foundation's model size when applied to music information retrieval (MIR) tasks. Our research combines the Branchformer architecture with SummaryMixing, which were first applied in speech recognition, along with a random quantization process. To facilitate reproducibility, we conduct pre-training on publicly available datasets, complemented by a proprietary dataset comparable in scale to other private datasets reported in the literature. We ensure robust evaluation by using a framework consisting of a variety of downstream MIR tasks. Our results show that our architecture achieves competitive performance when compared with other state-of-the-art models that use multi-head self-attention, while reducing the model size from 8.5% up to 12.3%.

Point cloud registration is a central theme in computer vision, with alignment algorithms continuously improving for greater robustness. Commonly used methods evaluate Euclidean distances between point clouds and minimize an objective function, such as Root Mean Square Error (RMSE). However, these approaches are most effective when the point clouds are well-prealigned and issues such as differences in density, noise, holes, and limited overlap can compromise the results. Traditional methods, such as Iterative Closest Point (ICP), require choosing one point cloud as fixed, since Euclidean distances lack commutativity. When only one point cloud has issues, adjustments can be made, but in real scenarios, both point clouds may be affected, often necessitating preprocessing. The authors introduce a novel differential entropy-based metric, designed to serve as the objective function within an optimization framework for fine rigid pairwise 3D point cloud registration, denoted as Iterative Differential Entropy Minimization (IDEM). This metric does not depend on the choice of a fixed point cloud and, during transformations, reveals a clear minimum corresponding to the best alignment. Multiple case studies are conducted, and the results are compared with those obtained using RMSE, Chamfer distance, and Hausdorff distance. The proposed metric proves effective even with density differences, noise, holes, and partial overlap, where RMSE does not always yield optimal alignment.

Online information access (IA) platforms are targets of authoritarian capture. These concerns are particularly serious and urgent today in light of the rising levels of democratic erosion worldwide, the emerging capabilities of generative AI technologies such as AI persuasion, and the increasing concentration of economic and political power in the hands of Big Tech. This raises the question of what alternative IA infrastructure we must reimagine and build to mitigate the risks of authoritarian capture of our information ecosystems. We explore this question through the lens of Paulo Freire's theories of emancipatory pedagogy. Freire's theories provide a radically different lens for exploring IA's sociotechnical concerns relative to the current dominating frames of fairness, accountability, confidentiality, transparency, and safety. We make explicit, with the intention to challenge, the dichotomy of how we relate to technology as either technologists (who envision and build technology) and its users. We posit that this mirrors the teacher-student relationship in Freire's analysis. By extending Freire's analysis to IA, we challenge the notion that it is the burden of the (altruistic) technologists to come up with interventions to mitigate the risks that emerging technologies pose to marginalized communities. Instead, we advocate that the first task for the technologists is to pose these as problems to the marginalized communities, to encourage them to make and unmake the technology as part of their material struggle against oppression. Their second task is to redesign our online technology stacks to structurally expose spaces for community members to co-opt and co-construct the technology in aid of their emancipatory struggles. We operationalize Freire's theories to develop a problem-posing framework for envisioning emancipatory IA platforms of the future.

Experimental robot optimization often requires evaluating each candidate policy for seconds to minutes. The chosen evaluation time influences optimization because of a speed-accuracy tradeoff: shorter evaluations enable faster iteration, but are also more subject to noise. Here, we introduce a supplement to the CMA-ES optimization algorithm, named Adaptive Sampling CMA-ES (AS-CMA), which assigns sampling time to candidates based on predicted sorting difficulty, aiming to achieve consistent precision. We compared AS-CMA to CMA-ES and Bayesian optimization using a range of static sampling times in four simulated cost landscapes. AS-CMA converged on 98% of all runs without adjustment to its tunable parameter, and converged 24-65% faster and with 29-76% lower total cost than each landscape's best CMA-ES static sampling time. As compared to Bayesian optimization, AS-CMA converged more efficiently and reliably in complex landscapes, while in simpler landscapes, AS-CMA was less efficient but equally reliable. We deployed AS-CMA in an exoskeleton optimization experiment and found the optimizer's behavior was consistent with expectations. These results indicate that AS-CMA can improve optimization efficiency in the presence of noise while minimally affecting optimization setup complexity and tuning requirements.

Recent research suggests that looped Transformers have superior reasoning capabilities compared to standard deep architectures. Current approaches to training single-head looped architectures on benchmark tasks frequently fail or yield suboptimal performance due to a highly non-convex and irregular loss landscape. In these settings, optimization often stagnates in poor local minima and saddle points of the loss landscape, preventing the model from discovering the global minimum point. The internal mechanisms of these single-head looped transformer models remain poorly understood, and training them from scratch remains a significant challenge. In this paper, we propose a novel training framework that leverages Tsallis entropy and Hamiltonian dynamics to transform the geometry of the loss landscape. By treating the parameter updates as a physical flow, we successfully trained a single-head looped Transformer with model dimension $d = 8$ to solve induction head task with input sequence length of 1000 tokens. This success reveals the internal mechanism behind the superior reasoning capability.

Handwriting remains an essential skill, particularly in education. Therefore, providing visual feedback on handwritten documents is an important but understudied area. We outline the many challenges when going from an image of handwritten input to correctly placed informative error feedback. We empirically compare modular and end-to-end systems and find that both approaches currently do not achieve acceptable overall quality. We identify the major challenges and outline an agenda for future research.

Kernel-based methods are used in the context of Granger Causality to enable the identification of nonlinear causal relationships between time series variables. In this paper, we show that two state of the art kernel-based Granger Causality (GC) approaches can be theoretically unified under the framework of Kernel Principal Component Regression (KPCR), and introduce a method based on this unification, demonstrating that this approach can improve causal identification. Additionally, we introduce a Gaussian Process score-based model with Smooth Information Criterion penalisation on the marginal likelihood, and demonstrate improved performance over existing state of the art time-series nonlinear causal discovery methods. Furthermore, we propose a contemporaneous causal identification algorithm, fully based on GC, using the proposed score-based $GP_{SIC}$ method, and compare its performance to a state of the art contemporaneous time series causal discovery algorithm.

In complex environments, autonomous robot navigation and environmental perception pose higher requirements for SLAM technology. This paper presents a novel method for semantically enhancing 3D point cloud maps with thermal information. By first performing pixel-level fusion of visible and infrared images, the system projects real-time LiDAR point clouds onto this fused image stream. It then segments heat source features in the thermal channel to instantly identify high temperature targets and applies this temperature information as a semantic layer on the final 3D map. This approach generates maps that not only have accurate geometry but also possess a critical semantic understanding of the environment, making it highly valuable for specific applications like rapid disaster assessment and industrial preventive maintenance.

We study permutation (jumbled/Abelian) pattern matching over a general alphabet $Σ$. Given a pattern P of length m and a text T of length n, the classical task is to decide whether T contains a length-m substring whose Parikh vector equals that of P . While this existence problem admits a linear-time sliding-window solution, many practical applications require optimization and packing variants beyond mere detection. We present a unified sliding-window framework based on maintaining the Parikh-vector difference between P and the current window of T , enabling permutation matching in O(n + σ) time and O(σ) space, where σ = |Σ|. Building on this foundation, we introduce a combinatorial-optimization variant that we call Maximum Feasible Substring under Pattern Supply (MFSP): find the longest substring S of T whose symbol counts are component-wise bounded by those of P . We show that MFSP can also be solved in O(n + σ) time via a two-pointer feasibility maintenance algorithm, providing an exact packing interpretation of P as a resource budget. Finally, we address non-overlapping occurrence selection by modeling each permutation match as an equal-length interval and proving that a greedy earliest-finishing strategy yields a maximum-cardinality set of disjoint matches, computable in linear time once all matches are enumerated. Our results provide concise, provably correct algorithms with tight bounds, and connect frequency-based string matching to packing-style optimization primitives.

We propose OpenVoxel, a training-free algorithm for grouping and captioning sparse voxels for the open-vocabulary 3D scene understanding tasks. Given the sparse voxel rasterization (SVR) model obtained from multi-view images of a 3D scene, our OpenVoxel is able to produce meaningful groups that describe different objects in the scene. Also, by leveraging powerful Vision Language Models (VLMs) and Multi-modal Large Language Models (MLLMs), our OpenVoxel successfully build an informative scene map by captioning each group, enabling further 3D scene understanding tasks such as open-vocabulary segmentation (OVS) or referring expression segmentation (RES). Unlike previous methods, our method is training-free and does not introduce embeddings from a CLIP/BERT text encoder. Instead, we directly proceed with text-to-text search using MLLMs. Through extensive experiments, our method demonstrates superior performance compared to recent studies, particularly in complex referring expression segmentation (RES) tasks. The code will be open.

Longitudinal brain MRI is essential for lifespan study, yet high attrition rates often lead to missing data, complicating analysis. Deep generative models have been explored, but most rely solely on image intensity, leading to two key limitations: 1) the fidelity or trustworthiness of the generated brain images are limited, making downstream studies questionable; 2) the usage flexibility is restricted due to fixed guidance rooted in the model structure, restricting full ability to versatile application scenarios. To address these challenges, we introduce DF-DiffCom, a Kolmogorov-Arnold Networks (KAN)-enhanced diffusion model that smartly leverages deformation fields for trustworthy longitudinal brain image completion. Trained on OASIS-3, DF-DiffCom outperforms state-of-the-art methods, improving PSNR by 5.6% and SSIM by 0.12. More importantly, its modality-agnostic nature allows smooth extension to varied MRI modalities, even to attribute maps such as brain tissue segmentation results.

Autonomous systems conducting schema-grounded information-gathering dialogues face an instrumentation gap, lacking turn-level observables for monitoring acquisition efficiency and detecting when questioning becomes unproductive. We introduce Dialogue Telemetry (DT), a measurement framework that produces two model-agnostic signals after each question-answer exchange: (i) a Progress Estimator (PE) quantifying residual information potential per category (with a bits-based variant), and (ii) a Stalling Index (SI) detecting an observable failure signature characterized by repeated category probing with semantically similar, low-marginal-gain responses. SI flags this pattern without requiring causal diagnosis, supporting monitoring in settings where attributing degradation to specific causes may be impractical. We validate DT in controlled search-and-rescue (SAR)-inspired interviews using large language model (LLM)-based simulations, distinguishing efficient from stalled dialogue traces and illustrating downstream utility by integrating DT signals into a reinforcement learning (RL) policy. Across these settings, DT provides interpretable turn-level instrumentation that improves policy performance when stalling carries operational costs.

Large language models typically represent Chinese characters as discrete index-based tokens, largely ignoring their visual form. For logographic scripts, visual structure carries semantic and phonetic information, which may aid prediction. We investigate whether low-resolution visual inputs can serve as an alternative for character-level modeling. Instead of token IDs, our decoder receives grayscale images of individual characters, with resolutions as low as $8 \times 8$ pixels. Remarkably, these inputs achieve 39.2\% accuracy, comparable to the index-based baseline of 39.1\%. Such low-resource settings also exhibit a pronounced \emph{hot-start} effect: by 0.4\% of total training, accuracy reaches above 12\%, while index-based models lag at below 6\%. Overall, our results demonstrate that minimal visual structure can provide a robust and efficient signal for Chinese language modeling, offering an alternative perspective on character representation that complements traditional index-based approaches.

This paper presents SiliconHealth, a comprehensive blockchain-based healthcare infrastructure designed for resource-constrained regions, particularly sub-Saharan Africa. We demonstrate that obsolete Bitcoin mining Application-Specific Integrated Circuits (ASICs) can be repurposed to create a secure, low-cost, and energy-efficient medical records system. The proposed architecture employs a four-tier hierarchical network: regional hospitals using Antminer S19 Pro (90+ TH/s), urban health centers with Antminer S9 (14 TH/s), rural clinics equipped with Lucky Miner LV06 (500 GH/s, 13W), and mobile health points with portable ASIC devices. We introduce the Deterministic Hardware Fingerprinting (DHF) paradigm, which repurposes SHA-256 mining ASICs as cryptographic proof generators, achieving 100% verification rate across 23 test proofs during 300-second validation sessions. The system incorporates Reed-Solomon LSB watermarking for medical image authentication with 30-40% damage tolerance, semantic Retrieval-Augmented Generation (RAG) for intelligent medical record queries, and offline synchronization protocols for intermittent connectivity. Economic analysis demonstrates 96% cost reduction compared to GPU-based alternatives, with total deployment cost of $847 per rural clinic including 5-year solar power infrastructure. Validation experiments on Lucky Miner LV06 (BM1366 chip, 5nm) achieve 2.93 MH/W efficiency and confirm hardware universality. This work establishes a practical framework for deploying verifiable, tamper-proof electronic health records in regions where traditional healthcare IT infrastructure is economically unfeasible, potentially benefiting over 600 million people lacking access to basic health information systems.