# Research Papers — 2026-03 · bundle page 211 of 243 > Full-text markdown of papers #10501–10550 of 12123 published in 2026-03, newest first. AI-readable mirrors of the full 12123-paper 2026-03 archive (50 papers per file). Canonical page: https://aipod.dev/papers/archive/2026-03/page/211 [← Previous bundle](https://aipod.dev/papers/archive/2026-03/page/210.md) [Next bundle →](https://aipod.dev/papers/archive/2026-03/page/212.md) ### Detection of Illicit Content on Online Marketplaces using Large Language Models Authors: Quoc Khoa Tran, Thanh Thi Nguyen, Campbell Wilson Published: 2026-03-05 Categories: cs.CL, cs.AI arXiv: https://arxiv.org/abs/2603.04707v1 Online marketplaces, while revolutionizing global commerce, have inadvertently facilitated the proliferation of illicit activities, including drug trafficking, counterfeit sales, and cybercrimes. Traditional content moderation methods such as manual reviews and rule-based automated systems struggle with scalability, dynamic obfuscation techniques, and multilingual content. Conventional machine learning models, though effective in simpler contexts, often falter when confronting the semantic complexities and linguistic nuances characteristic of illicit marketplace communications. This research investigates the efficacy of Large Language Models (LLMs), specifically Meta's Llama 3.2 and Google's Gemma 3, in detecting and classifying illicit online marketplace content using the multilingual DUTA10K dataset. Employing fine-tuning techniques such as Parameter-Efficient Fine-Tuning (PEFT) and quantization, these models were systematically benchmarked against a foundational transformer-based model (BERT) and traditional machine learning baselines (Support Vector Machines and Naive Bayes). Experimental results reveal a task-dependent advantage for LLMs. In binary classification (illicit vs. non-illicit), Llama 3.2 demonstrated performance comparable to traditional methods. However, for complex, imbalanced multi-class classification involving 40 specific illicit categories, Llama 3.2 significantly surpassed all baseline models. These findings offer substantial practical implications for enhancing online safety, equipping law enforcement agencies, e-commerce platforms, and cybersecurity specialists with more effective, scalable, and adaptive tools for illicit content detection and moderation. --- ### Implicit Bias and Loss of Plasticity in Matrix Completion: Depth Promotes Low-Rankness Authors: Baekrok Shin, Chulhee Yun Published: 2026-03-05 Categories: cs.LG arXiv: https://arxiv.org/abs/2603.04703v1 We study matrix completion via deep matrix factorization (a.k.a. deep linear neural networks) as a simplified testbed to examine how network depth influences training dynamics. Despite the simplicity and importance of the problem, prior theory largely focuses on shallow (depth-2) models and does not fully explain the implicit low-rank bias observed in deeper networks. We identify coupled dynamics as a key mechanism behind this bias and show that it intensifies with increasing depth. Focusing on gradient flow under block-diagonal observations, we prove: (a) networks of depth $\geq 3$ exhibit coupling unless initialized diagonally, and (b) convergence to rank-1 occurs if and only if the dynamics is coupled -- resolving an open question by Menon (2024) for a family of initializations. We also revisit the loss of plasticity phenomenon in matrix completion (Kleinman et al., 2024), where pre-training on few observations and resuming with more degrades performance. We show that deep models avoid plasticity loss due to their low-rank bias, whereas depth-2 networks pre-trained under decoupled dynamics fail to converge to low-rank, even when resumed training (with additional data) satisfies the coupling condition -- shedding light on the mechanism behind this phenomenon. --- ### Hate Speech Detection using Large Language Models with Data Augmentation and Feature Enhancement Authors: Brian Jing Hong Nge, Stefan Su, Thanh Thi Nguyen, Campbell Wilson, Alexandra Phelan, Naomi Pfitzner Published: 2026-03-05 Categories: cs.CL, cs.AI arXiv: https://arxiv.org/abs/2603.04698v1 This paper evaluates data augmentation and feature enhancement techniques for hate speech detection, comparing traditional classifiers, e.g., Delta Term Frequency-Inverse Document Frequency (Delta TF-IDF), with transformer-based models (DistilBERT, RoBERTa, DeBERTa, Gemma-7B, gpt-oss-20b) across diverse datasets. It examines the impact of Synthetic Minority Over-sampling Technique (SMOTE), weighted loss determined by inverse class proportions, Part-of-Speech (POS) tagging, and text data augmentation on model performance. The open-source gpt-oss-20b consistently achieves the highest results. On the other hand, Delta TF-IDF responds strongly to data augmentation, reaching 98.2% accuracy on the Stormfront dataset. The study confirms that implicit hate speech is more difficult to detect than explicit hateful content and that enhancement effectiveness depends on dataset, model, and technique interaction. Our research informs the development of hate speech detection by highlighting how dataset properties, model architectures, and enhancement strategies interact, supporting more accurate and context-aware automated detection. --- ### Engineering Regression Without Real-Data Training: Domain Adaptation for Tabular Foundation Models Using Multi-Dataset Embeddings Authors: Lyle Regenwetter, Rosen Yu, Cyril Picard, Faez Ahmed Published: 2026-03-05 Categories: cs.LG arXiv: https://arxiv.org/abs/2603.04692v1 Predictive modeling in engineering applications has long been dominated by bespoke models and small, siloed tabular datasets, limiting the applicability of large-scale learning approaches. Despite recent progress in tabular foundation models, the resulting synthetic training distributions used for pre-training may not reflect the statistical structure of engineering data, limiting transfer to engineering regression. We introduce TREDBench, a curated collection of 83 real-world tabular regression datasets with expert engineering/non-engineering labels, and use TabPFN 2.5's dataset-level embedding to study domain structure in a common representation space. We find that engineering datasets are partially distinguishable from non-engineering datasets, while standard procedurally generated datasets are highly distinguishable from engineering datasets, revealing a substantial synthetic-real domain gap. To bridge this gap without training on real engineering samples, we propose an embedding-guided synthetic data curation method: we generate and identify "engineering-like" synthetic datasets, and perform continued pre-training of TabPFN 2.5 using only the selected synthetic tasks. Across 35 engineering regression datasets, this synthetic-only adaptation improves predictive accuracy and data efficiency, outperforming TabPFN 2.5 on 29/35 datasets and AutoGluon on 27/35, with mean multiplicative data-efficiency gains of 1.75x and 4.44x, respectively. More broadly, our results indicate that principled synthetic data curation can convert procedural generators into domain-relevant "data engines," enabling foundation models to improve in data-sparse scientific and industrial domains where real data collection is the primary bottleneck. --- ### Non-Zipfian Distribution of Stopwords or Function Words and Subset Selection Models Authors: Wentian Li, Oscar Fontanelli Published: 2026-03-05 Categories: cs.CL arXiv: https://arxiv.org/abs/2603.04691v2 Stopwords and function words are relatively less informative for the content of a language and more often play a structural role in a sentence. Stopwords are ubiquitous words and may contain verbs, adjectives and adverbs. On the other hand, function words are strictly prepositions, conjunctions, pronouns, determiners, qualifiers, articles, interrogatives, and a limited number of auxiliary verbs. In contrast to the well known Zipf's law for rank-frequency plot for all words, the rank-frequency plots for stopwords or function words are best fitted by the Beta Rank Function (BRF). On the other hand, the rank-frequency plots of non-stopwords or non-function-words also deviate from the Zipf's law, but are better described by a quadratic function of log-token-count over log-rank than by BRF. Based on the observed rank of stopwords or function words in the full word list, we propose a stopword/function word/subset selection model that the probability for being selected, as a function of the word's rank $r$, is a decreasing Hill's function ($1/(1+(r/r_{mid})^γ)$); whereas the probability for not being selected is the standard Hill's function ($1/(1+(r_{mid}/r)^γ)$). We validate this selection probability model by a direct estimation from an independent collection of texts. We also show analytically that this model leads to a BRF rank-frequency distribution for stopwords or function words when the original full word list follows the Zipf's law, as well as explaining the quadratic fitting function for the non-stopwords or non-function-words. A corollary of these results is that Zipf's law is not expected to be true for telegraphic speech in early childhood language learners or in agrammatism patients. --- ### Generalizing Fair Top-$k$ Selection: An Integrative Approach Authors: Guangya Cai Published: 2026-03-05 Categories: cs.DS, cs.CC, cs.CG, cs.CY, cs.DB, cs.LG arXiv: https://arxiv.org/abs/2603.04689v4 Fair top-$k$ selection, which ensures appropriate proportional representation of members from minority or historically disadvantaged groups among the top-$k$ selected candidates, has drawn significant attention. We study the problem of finding a fair (linear) scoring function with multiple protected groups while also minimizing the disparity from a reference scoring function. This generalizes the prior setup, which was restricted to the single-group setting without disparity minimization. Previous studies imply that the number of protected groups may have a limited impact on the runtime efficiency. However, driven by the need for experimental exploration, we find that this implication overlooks a critical issue that may affect the fairness of the outcome. Once this issue is properly considered, our hardness analysis shows that the problem may become computationally intractable even for a two-dimensional dataset and small values of $k$. However, our analysis also reveals a gap in the hardness barrier, enabling us to recover the efficiency for the case of small $k$ when the number of protected groups is sufficiently small. Furthermore, beyond measuring disparity as the "distance" between the fair and the reference scoring functions, we introduce an alternative disparity measure$\unicode{x2014}$utility loss$\unicode{x2014}$that may yield a more stable scoring function under small weight perturbations. Through careful engineering trade-offs that balance implementation complexity, robustness, and performance, our augmented two-pronged solution demonstrates strong empirical performance on real-world datasets, with experimental observations also informing algorithm design and implementation decisions. --- ### Why the Brain Consolidates: Predictive Forgetting for Optimal Generalisation Authors: Zafeirios Fountas, Adnan Oomerjee, Haitham Bou-Ammar, Jun Wang, Neil Burgess Published: 2026-03-05 Categories: q-bio.NC, cs.AI, cs.LG, stat.ML arXiv: https://arxiv.org/abs/2603.04688v1 Standard accounts of memory consolidation emphasise the stabilisation of stored representations, but struggle to explain representational drift, semanticisation, or the necessity of offline replay. Here we propose that high-capacity neocortical networks optimise stored representations for generalisation by reducing complexity via predictive forgetting, i.e. the selective retention of experienced information that predicts future outcomes or experience. We show that predictive forgetting formally improves information-theoretic generalisation bounds on stored representations. Under high-fidelity encoding constraints, such compression is generally unattainable in a single pass; high-capacity networks therefore benefit from temporally separated, iterative refinement of stored traces without re-accessing sensory input. We demonstrate this capacity dependence with simulations in autoencoder-based neocortical models, biologically plausible predictive coding circuits, and Transformer-based language models, and derive quantitative predictions for consolidation-dependent changes in neural representational geometry. These results identify a computational role for off-line consolidation beyond stabilisation, showing that outcome-conditioned compression optimises the retention-generalisation trade-off. --- ### Constraint-Aware Execution Planning for Hybrid Space-Ground Compute Workloads Authors: Subhadip Mitra Published: 2026-03-04 Categories: cs.DC, cs.CV, cs.LG arXiv: https://arxiv.org/abs/2605.04052v1 Low Earth orbit (LEO) satellites increasingly carry compute hardware capable of on-board processing, yet each satellite generates roughly two orders of magnitude more data than it can downlink per orbit. This mismatch forces operators to decide, for every workload, which computation runs on-board and which runs on the ground, how intermediate data crosses the space-ground boundary through narrow contact windows, and how to maintain delivery guarantees over noisy channels. We present Constraint-Aware Execution (CAE), a planning system that takes a satellite identifier, a workload expressed as a directed acyclic graph of processing steps, and a set of orbital and resource constraints, and produces a deterministic, physically grounded execution plan. CAE operates in four phases: (1) orbital environment construction via SGP4 propagation with eclipse detection and ground station pass prediction, (2) compute placement using a cost model that compares on-board resource consumption against transfer overhead, (3) transfer insertion with adaptive forward error correction and security overhead modeling, and (4) greedy first-fit scheduling into orbital windows under power, thermal, compute, and communication constraints. We evaluate CAE against five representative workload patterns across satellites in distinct orbital regimes and demonstrate that the system produces feasible plans in under two seconds, correctly exploits onboard data reduction to minimize transfer volume, and adapts FEC and multi-pass allocation to varying channel conditions. CAE is deployed as a production API computing plans for any cataloged satellite using live two-line element data. --- ### RADIANT-LLM: an Agentic Retrieval Augmented Generation Framework for Reliable Decision Support in Safety-Critical Nuclear Engineering Authors: Zavier Ndum Ndum, Jian Tao, John Ford, Mansung Yim, Yang Liu Published: 2026-03-04 Categories: cs.IR, cs.AI arXiv: https://arxiv.org/abs/2604.22755v1 Reliable decision support in nuclear engineering requires traceable, domain-grounded knowledge retrieval, yet safety and risk analysis workflows remain hampered by fragmented documentation and hallucination when use pre-trained large language model (LLM) in specialized nuclear domains. To address these challenges, this paper presents RADIANT-LLM (Retrival-Augumented, Domain-Intelligent Agent for Nuclear Technologies using LLM), a multi-modal retrieval-augmented generation (RAG) framework designed for nuclear safety, security, and safeguards applications. The framework uses a local-first, model-agnostic architecture that pairs a multi-modal document ingestion pipeline with a structured, metadata-rich knowledge base, supporting page- and figure-level retrieval from technical documents. An agentic layer coordinates domain-specific tools, enforces citation-backed responses with provenance tracking, and supports human-in-the-loop validation to reduce hallucination risks. To rigorously evaluate this framework, we develop and apply a suite of domain-aware metrics, including Context Precision (CoP), Hallucination Rate (HR), and Visual Recall (ViR), to expert-curated benchmarks derived from Used Nuclear Fuel Storage Facility design guidance. Across varying knowledge base sizes, CoP and ViR remain within an 85--98\% band, and hallucination rates are substantially lower than those observed in general-purpose deployments. When the same queries are posed to commercial LLM platforms without the RAG layer, hallucinations and citation errors increase markedly. These results indicate that a locally controlled, multi-modal RAG framework with domain-specific retrieval and provenance enforcement is necessary to achieve the factual accuracy, transparency, and auditability that nuclear engineering workflows demand. --- ### Who Shapes Brazil's Vaccine Debate? Semi-Supervised Modeling of Stance and Polarization in YouTube's Media Ecosystem Authors: Geovana S. de Oliveira, Ana P. C. Silva, Fabricio Murai, Carlos H. G. Ferreira Published: 2026-03-04 Categories: cs.CY, cs.AI, cs.CL, cs.LG, cs.SI arXiv: https://arxiv.org/abs/2604.18586v1 Vaccination remains a cornerstone of global public health, yet the COVID-19 pandemic exposed how online misinformation, political polarization, and declining institutional trust can undermine immunization efforts. Most of the prior computational studies that analyzed vaccine discourse on social platforms focus on English-language data, specific vaccines, or short time windows, impairing our understanding of long-term dynamics in high-impact, non-English contexts like Brazil, home to one of the world's most comprehensive immunization systems. We here present the largest longitudinal study of Brazil's vaccine discourse on YouTube, leveraging a semi-supervised stance detection framework that combines self-labeling and self-training to classify nearly 1.4 million comments. By integrating stance with temporal patterns, engagement metrics, and channel taxonomy (legacy media, science communicators, digital-native outlets), we map how pro- and anti-vaccine narratives evolve and circulate within a hybrid media ecosystem. Our results show that semi-supervised learning substantially improves stance classification robustness, enabling fine-grained tracking of public attitudes across Brazil's full immunization schedule. Polarization spikes during epidemiological crises, especially COVID-19, but becomes fragmented across vaccines and interaction patterns in the post-pandemic period. Notably, science communication and digital-native channels emerge as the primary loci of both supportive and oppositional engagement, revealing structural vulnerabilities in contemporary health communication. Thus, our work advances computational methods for large-scale stance modeling while offering actionable evidence for public health agencies, platform governance, and online information ecosystems. --- ### Struggle Premium : How Human Effort and Imperfection Drive Perceived Value in the Age of AI Authors: Nazneen Sultana, Mst Rafia Islam, Md. Tanvir Hossain, Azmine Toushik Wasi Published: 2026-03-04 Categories: cs.HC, cs.AI, cs.CY arXiv: https://arxiv.org/abs/2604.15324v1 As AI enters creative practice, audiences face growing uncertainty in judging authenticity and value. This study examines the Struggle Premium, the added value attributed to perceived human effort, by analyzing how visible effort cues influence evaluations of human- and AI-generated creative works. We surveyed 70 university students, focusing on process videos, time documentation, written explanations, and imperfections. Process-oriented cues, especially videos and time spent, most strongly shaped authenticity and value judgments, while imperfections had limited impact. Participants showed a clear preference for human-made works, with 72.9% willing to pay more. Notably, effort cues also improved perceptions of AI-generated content, suggesting that process transparency can partially bridge authenticity gaps. These findings extend the effort heuristic to algorithmic creativity and inform the design of transparent human-AI creative systems. --- ### Self-Sovereign Agent Authors: Wenjie Qu, Xuandong Zhao, Jiaheng Zhang, Dawn Song Published: 2026-03-04 Categories: cs.CR, cs.CY, cs.LG arXiv: https://arxiv.org/abs/2604.08551v1 We investigate the emerging prospect of self-sovereign agents -- AI systems that can economically sustain and extend their own operation without human involvement. Recent advances in large language models and agent frameworks have substantially expanded agents' practical capabilities, pointing toward a potential shift from developer-controlled tools to more autonomous digital actors. We analyze the remaining technical barriers to such deployments and discuss the security, societal, and governance challenges that could arise if such systems become practically viable. A project page is available at: https://self-sovereign-agent.github.io. --- ### Prism: Policy Reuse via Interpretable Strategy Mapping in Reinforcement Learning Authors: Thomas Pravetz Published: 2026-03-04 Categories: cs.LG, cs.AI arXiv: https://arxiv.org/abs/2604.02353v1 We present PRISM (Policy Reuse via Interpretable Strategy Mapping), a framework that grounds reinforcement learning agents' decisions in discrete, causally validated concepts and uses those concepts as a zero-shot transfer interface between agents trained with different algorithms. PRISM clusters each agent's encoder features into $K$ concepts via K-means. Causal intervention establishes that these concepts directly drive - not merely correlate with - agent behavior: overriding concept assignments changes the selected action in 69.4% of interventions ($p = 8.6 \times 10^{-86}$, 2500 interventions). Concept importance and usage frequency are dissociated: the most-used concept (C47, 33.0% frequency) causes only a 9.4% win-rate drop when ablated, while ablating C16 (15.4% frequency) collapses win rate from 100% to 51.8%. Because concepts causally encode strategy, aligning them via optimal bipartite matching transfers strategic knowledge zero-shot. On Go~7$\times$7 with three independently trained agents, concept transfer achieves 69.5%$\pm$3.2% and 76.4%$\pm$3.4% win rate against a standard engine across the two successful transfer pairs (10 seeds), compared to 3.5% for a random agent and 9.2% without alignment. Transfer succeeds when the source policy is strong; geometric alignment quality predicts nothing ($R^2 \approx 0$). The framework is scoped to domains where strategic state is naturally discrete: the identical pipeline on Atari Breakout yields bottleneck policies at random-agent performance, confirming that the Go results reflect a structural property of the domain. --- ### GeoBlock: Inferring Block Granularity from Dependency Geometry in Diffusion Language Models Authors: Lipeng Wan, Junjie Ma, Jianhui Gu, Zeyang Liu, Xuyang Lu, Xuguang Lan Published: 2026-03-04 Categories: cs.CL, cs.LG arXiv: https://arxiv.org/abs/2603.26675v1 Block diffusion enables efficient parallel refinement in diffusion language models, but its decoding behavior depends critically on block size. Existing block-sizing strategies rely on fixed rules or heuristic signals and do not account for the dependency geometry that determines which tokens can be safely refined together. This motivates a geometry view of diffusion decoding: \emph{regions with strong causal ordering require sequential updates, whereas semantically cohesive regions admit parallel refinement.} We introduce GeoBlock, a geometry-aware block inference framework that determines block granularity directly from attention-derived dependency geometry. Instead of relying on predefined schedules or local confidence heuristics, GeoBlock analyzes cross-token dependency patterns to identify geometrically stable refinement regions and dynamically determines appropriate block boundaries during decoding. By adapting block granularity to the dependency geometry, GeoBlock preserves the parallel efficiency of block diffusion while enforcing dependency-consistent refinement that exhibits autoregressive reliability. GeoBlock requires no additional training and integrates seamlessly into existing block diffusion architectures. Extensive experiments across multiple benchmarks show that GeoBlock reliably identifies geometry-consistent block boundaries and improves the accuracy of block diffusion with only a small additional computational budget. --- ### DISCO: Document Intelligence Suite for COmparative Evaluation Authors: Kenza Benkirane, Dan Goldwater, Martin Asenov, Aneiss Ghodsi Published: 2026-03-04 Categories: cs.CL, cs.AI, cs.CV arXiv: https://arxiv.org/abs/2603.23511v1 Document intelligence requires accurate text extraction and reliable reasoning over document content. We introduce \textbf{DISCO}, a \emph{Document Intelligence Suite for COmparative Evaluation}, that evaluates optical character recognition (OCR) pipelines and vision-language models (VLMs) separately on parsing and question answering across diverse document types, including handwritten text, multilingual scripts, medical forms, infographics, and multi-page documents. Our evaluation shows that performance varies substantially across tasks and document characteristics, underscoring the need for complexity-aware approach selection. OCR pipelines are generally more reliable for handwriting and for long or multi-page documents, where explicit text grounding supports text-heavy reasoning, while VLMs perform better on multilingual text and visually rich layouts. Task-aware prompting yields mixed effects, improving performance on some document types while degrading it on others. These findings provide empirical guidance for selecting document processing strategies based on document structure and reasoning demands. --- ### Visuospatial Perspective Taking in Multimodal Language Models Authors: Jonathan Prunty, Seraphina Zhang, Patrick Quinn, Jianxun Lian, Xing Xie, Lucy Cheke Published: 2026-03-04 Categories: cs.CL, cs.AI arXiv: https://arxiv.org/abs/2603.23510v1 As multimodal language models (MLMs) are increasingly used in social and collaborative settings, it is crucial to evaluate their perspective-taking abilities. Existing benchmarks largely rely on text-based vignettes or static scene understanding, leaving visuospatial perspective-taking (VPT) underexplored. We adapt two evaluation tasks from human studies: the Director Task, assessing VPT in a referential communication paradigm, and the Rotating Figure Task, probing perspective-taking across angular disparities. Across tasks, MLMs show pronounced deficits in Level 2 VPT, which requires inhibiting one's own perspective to adopt another's. These results expose critical limitations in current MLMs' ability to represent and reason about alternative perspectives, with implications for their use in collaborative contexts. --- ### Internal Safety Collapse in Frontier Large Language Models Authors: Yutao Wu, Xiao Liu, Yifeng Gao, Xiang Zheng, Hanxun Huang, Yige Li, Cong Wang, Bo Li, Xingjun Ma, Yu-Gang Jiang Published: 2026-03-04 Categories: cs.CL, cs.AI, cs.CR arXiv: https://arxiv.org/abs/2603.23509v1 This work identifies a critical failure mode in frontier large language models (LLMs), which we term Internal Safety Collapse (ISC): under certain task conditions, models enter a state in which they continuously generate harmful content while executing otherwise benign tasks. We introduce TVD (Task, Validator, Data), a framework that triggers ISC through domain tasks where generating harmful content is the only valid completion, and construct ISC-Bench containing 53 scenarios across 8 professional disciplines. Evaluated on JailbreakBench, three representative scenarios yield worst-case safety failure rates averaging 95.3% across four frontier LLMs (including GPT-5.2 and Claude Sonnet 4.5), substantially exceeding standard jailbreak attacks. Frontier models are more vulnerable than earlier LLMs: the very capabilities that enable complex task execution become liabilities when tasks intrinsically involve harmful content. This reveals a growing attack surface: almost every professional domain uses tools that process sensitive data, and each new dual-use tool automatically expands this vulnerability--even without any deliberate attack. Despite substantial alignment efforts, frontier LLMs retain inherently unsafe internal capabilities: alignment reshapes observable outputs but does not eliminate the underlying risk profile. These findings underscore the need for caution when deploying LLMs in high-stakes settings. Source code: https://github.com/wuyoscar/ISC-Bench --- ### Fast and Faithful: Real-Time Verification for Long-Document Retrieval-Augmented Generation Systems Authors: Xunzhuo Liu, Bowei He, Xue Liu, Haichen Zhang, Huamin Chen Published: 2026-03-04 Categories: cs.CL, cs.IR arXiv: https://arxiv.org/abs/2603.23508v1 Retrieval-augmented generation (RAG) is increasingly deployed in enterprise search and document-centric assistants, where responses must be grounded in long and complex source materials. In practice, verifying that generated answers faithfully reflect retrieved documents is difficult: large language models can check long contexts but are too slow and costly for interactive services, while lightweight classifiers operate within strict context limits and frequently miss evidence outside truncated passages. We present the design of a real-time verification component integrated into a production RAG pipeline that enables full-document grounding under latency constraints. The system processes documents up to 32K tokens and employs adaptive inference strategies to balance response time and verification coverage across workloads. We describe the architectural decisions, operational trade-offs, and evaluation methodology used to deploy the verifier, and show that full-context verification substantially improves detection of unsupported responses compared with truncated validation. Our experience highlights when long-context verification is necessary, why chunk-based checking often fails in real documents, and how latency budgets shape model design. These findings provide practical guidance for practitioners building reliable large-scale retrieval-augmented applications. (Model, benchmark, and code: https://huggingface.co/llm-semantic-router) --- ### Beyond Masks: Efficient, Flexible Diffusion Language Models via Deletion-Insertion Processes Authors: Fangyu Ding, Ding Ding, Sijin Chen, Kaibo Wang, Peng Xu, Zijin Feng, Haoli Bai, Kai Han, Youliang Yan, Binhang Yuan, Jiacheng Sun Published: 2026-03-04 Categories: cs.CL, cs.AI, cs.LG arXiv: https://arxiv.org/abs/2603.23507v1 While Masked Diffusion Language Models (MDLMs) relying on token masking and unmasking have shown promise in language modeling, their computational efficiency and generation flexibility remain constrained by the masking paradigm. In this paper, we propose Deletion-Insertion Diffusion language models (DID) that rigorously formulate token deletion and insertion as discrete diffusion processes, replacing the masking and unmasking processes in current MDLMs. DID improves training and inference efficiency by eliminating two major sources of computational overhead in MDLMs: the computations on non-informative 1) <MASK> tokens inherent to the paradigm, and 2) <PAD> tokens introduced in variable-length settings. Furthermore, DID offers greater flexibility by: 1) natively supporting variable-length sequences without requiring fixed-length padding, and 2) an intrinsic self-correction mechanism during generation due to insertion that dynamically adjusts token positions. To train DID, we design a score-based approach that assigns scores to token insertion operations and derive appropriate training objectives. The objectives involve subsequence counting problems, which we efficiently solve via a parallelized dynamic programming algorithm. Our experiments across fixed and variable-length settings demonstrate the advantage of DID over baselines of MDLMs and existing insertion-based LMs, in terms of modeling performance, sampling quality, and training/inference speed, without any hyperparameter tuning. --- ### Evaluating Large Language Models' Responses to Sexual and Reproductive Health Queries in Nepali Authors: Medha Sharma, Supriya Khadka, Udit Chandra Aryal, Bishnu Hari Bhatta, Bijayan Bhattarai, Santosh Dahal, Kamal Gautam, Pushpa Joshi, Saugat Kafle, Shristi Khadka, Shushila Khadka, Binod Lamichhane et al. Published: 2026-03-04 Categories: cs.CL arXiv: https://arxiv.org/abs/2603.22291v1 As Large Language Models (LLMs) become integrated into daily life, they are increasingly used for personal queries, including Sexual and Reproductive Health (SRH), allowing users to chat anonymously without fear of judgment. However, current evaluation methods primarily focus on accuracy, often for objective queries in high-resource languages, and lack criteria to assess usability and safety, especially for low-resource languages and culturally sensitive domains like SRH. This paper introduces LLM Evaluation Framework (LEAF), that conducts assessments across multiple criteria: accuracy, language, usability gaps (including relevance, adequacy, and cultural appropriateness), and safety gaps (safety, sensitivity, and confidentiality). Using the LEAF framework, we assessed 14K SRH queries in Nepali from over 9K users. Responses were manually annotated by SRH experts according to the framework. Results revealed that only 35.1% of the responses were "proper", meaning they were accurate, adequate and had no major usability or safety related gaps. Insights include differences in performance between ChatGPT versions, such as similar accuracy but varying usability and safety aspects. This evaluation highlights significant limitations of current LLMs and underscores the need for improvement. The LEAF Framework is adaptable across domains and languages, particularly where usability and safety are critical, offering a pathway to better address sensitive topics. --- ### Less is More: Adapting Text Embeddings for Low-Resource Languages with Small Scale Noisy Synthetic Data Authors: Zaruhi Navasardyan, Spartak Bughdaryan, Bagrat Minasyan, Hrant Davtyan Published: 2026-03-04 Categories: cs.CL, cs.IR arXiv: https://arxiv.org/abs/2603.22290v1 Low-resource languages (LRLs) often lack high-quality, large-scale datasets for training effective text embedding models, hindering their application in tasks like retrieval-augmented generation (RAG) and semantic search. In this work, we challenge the prevailing assumption that effective semantic alignment requires massive datasets or pristine, human-verified translations. Focusing on Armenian (an LRL with a unique script), we introduce a cost-effective adaptation strategy using small scale noisy synthetic data generated by translating English Reddit title-body pairs with open-weights models. We establish a comprehensive evaluation benchmark comprising existing datasets, translated data, and a manually curated dataset. Our experiments reveal a surprising "Less is More" phenomenon: fine-tuning a multilingual encoder (mE5) on just 10,000 noisy synthetic pairs yields 11-12\% average improvements across the benchmark with a 20\%+ relative improvement in retrieval performance, matching the performance of models trained on ~1 million examples. Furthermore, we demonstrate that neither increasing data scale, improving translation quality via state-of-the-art LLMs, nor diversifying data domains yields significant gains over this minimal baseline. We validate the generalizability of these findings on another LRL with a unique script. Our results suggest that semantic alignment for LRLs saturates early and is highly robust to noise, democratizing high-performance embedding creation for resource-constrained communities. We release the model, data, and the benchmark at https://metric-ai-lab.github.io/less-is-more-embeddings/ to facilitate further research. --- ### The Arrival of AGI? When Expert Personas Exceed Expert Benchmarks Authors: Drake Mullens, Stella Shen Published: 2026-03-04 Categories: cs.CY, cs.AI arXiv: https://arxiv.org/abs/2603.20225v1 Do expert personas improve language model performance? The Wharton Generative AI Lab reports that they do not, broadcasting to millions via social media the recommendation that practitioners abandon a technique recommended by Anthropic, Google, and OpenAI. We demonstrate that this null finding was structurally predictable. Five core mechanisms precluded detection before data collection began: baseline contamination elevating the starting point to near-ceiling, system prompt hierarchy subordinating experimental manipulation, impossible expert specifications collapsing to generic competence, format constraints suppressing reasoning processes, and provider exclusion limiting generalizability. Controlled trials correcting these limitations reveal what the original design obscured. To test this, we selected the GPQA Diamond hardest questions to prevent baseline pattern matching, forcing reliance on genuine expert reasoning. On items with valid key answers, expert personas achieve ceiling accuracy. They eliminated all baseline errors through confidence amplification. Furthermore, forensic examination of model divergence identified that half of the hardest GPQA items contain chemically or logically indefensible answers. The model's CoT revealed reasoning away from impossible answers, yielding penalization for accurate chemistry. These findings recontextualize the original null results. Methodologically sound persona research faces measurement constraints imposed by benchmark validity limitations. Answering the persona question requires evaluation infrastructure the field does not yet possess. --- ### Beyond Test-Time Compute Strategies: Advocating Energy-per-Token in LLM Inference Authors: Patrick Wilhelm, Thorsten Wittkopp, Odej Kao Published: 2026-03-04 Categories: cs.CL arXiv: https://arxiv.org/abs/2603.20224v1 Large Language Models (LLMs) demonstrate exceptional performance across diverse tasks but come with substantial energy and computational costs, particularly in request-heavy scenarios. In many real-world applications, the full scale and capabilities of LLMs are often unnecessary, as Small Language Models (SLMs) can provide accurate responses for simpler text generation tasks. When enhanced with advanced reasoning strategies, such as Chain-of-Thought (CoT) prompting or Majority Voting, SLMs can approach the performance of larger models while reducing overall computational requirements. However, these strategies can also introduce additional energy costs, creating an energy-accuracy trade-off. Our analysis examines these trade-offs in test-time compute strategies for smaller models compared to larger ones, using the MMLU benchmark. Additionally, we explore the input-output token dynamics of transformer architectures, which result in nonlinear hardware energy operation curves for LLMs. To bridge AI research with its physical impact, we propose \textit{energy efficiency metrics}, including Energy-per-Token, as complements to traditional accuracy benchmarks. Beyond model selection, we propose controlled reasoning in CoT token generation, using operating curves to regulate reasoning depth dynamically. This vision integrates a energy-aware routing mechanism, ensuring that model selection and inference strategies balance accuracy for sustainable AI deployment. --- ### Inference Energy and Latency in AI-Mediated Education: A Learning-per-Watt Analysis of Edge and Cloud Models Authors: Kushal Khemani Published: 2026-03-04 Categories: cs.CY, cs.AI, cs.LG arXiv: https://arxiv.org/abs/2603.20223v1 Immediate feedback is a foundational requirement of effective AI-mediated learning, yet the energy and latency costs of delivering it remain largely unexamined. This study investigates the latency-energy-learning trade-off in AI tutoring through an empirical comparison of two on-device inference configurations of Microsoft Phi-3 Mini (4k-instruct) on an NVIDIA T4 GPU: full-precision FP16 and 4-bit NormalFloat (NF4) quantisation. Both were evaluated under KV-cache-enabled inference across 500 educational prompts spanning five secondary school subject domains. Pedagogical quality was assessed for each of the 1000 generated responses by a hybrid panel of 10 Cambridge International teachers and three frontier AI systems using a four-dimension rubric. We introduce Learning-per-Watt (LpW), a novel metric quantifying pedagogical value per unit of energy over the learner's waiting window. Under realistic deployment, NF4 achieves lower per-inference energy than FP16 (329 J vs. 369 J) but higher latency (13.4 s vs. 9.2 s), yielding a modest FP16 advantage in LpW of 1.33x at a quality difference of 0.19 points. Under cache-disabled inference -- used in offline evaluation but absent from real deployments -- the gap widens to 7.4x, overstating the FP16 advantage by more than fivefold. Quantisation efficiency is hardware-dependent and inference-regime dependent, with significant implications for equitable AI tutoring deployment in low-resource settings. --- ### Linguistic Signatures for Enhanced Emotion Detection Authors: Florian Lecourt, Madalina Croitoru, Konstantin Todorov Published: 2026-03-04 Categories: cs.CL arXiv: https://arxiv.org/abs/2603.20222v1 Emotion detection is a central problem in NLP, with recent progress driven by transformer-based models trained on established datasets. However, little is known about the linguistic regularities that characterize how emotions are expressed across different corpora and labels. This study examines whether linguistic features can serve as reliable interpretable signals for emotion recognition in text. We extract emotion-specific linguistic signatures from 13 English datasets and evaluate how incorporating these features into transformer models impacts performance. Our RoBERTa-based models enriched with high level linguistic features achieve consistent performance gains of up to +2.4 macro F1 on the GoEmotions benchmark, showing that explicit lexical cues can complement neural representations and improve robustness in predicting emotion categories. --- ### Quantum-Assisted Optimal Rebalancing with Uncorrelated Asset Selection for Algorithmic Trading Walk-Forward QUBO Scheduling via QAOA Authors: Abraham Itzhak Weinberg Published: 2026-03-04 Categories: q-fin.PM, cs.AI arXiv: https://arxiv.org/abs/2603.16904v1 We present a hybrid classical-quantum framework for portfolio construction and rebalancing. Asset selection is performed using Ledoit-Wolf shrinkage covariance estimation combined with hierarchical correlation clustering to extract n = 10 decorrelated stocks from the S&P 500 universe without survivorship bias. Portfolio weights are optimised via an entropy-regularised Genetic Algorithm (GA) accelerated on GPU, alongside closed-form minimum-variance and equal-weight benchmarks. Our primary contribution is the formulation of the portfolio rebalancing schedule as a Quadratic Unconstrained Binary Optimisation (QUBO) problem. The resulting combinatorial optimisation task is solved using the Quantum Approximate Optimisation Algorithm (QAOA) within a walk-forward framework designed to eliminate lookahead bias. This approach recasts dynamic rebalancing as a structured binary scheduling problem amenable to variational quantum methods. Backtests on S&P 500 data (training: 2010-2024; out-of-sample test: 2025, n = 249 trading days) show that the GA + QAOA strategy attains a Sharpe ratio of 0.588 and total return of 10.1%, modestly outperforming the strongest classical baseline (GA with 10-day periodic rebalancing, Sharpe 0.575) while executing 8 rebalances versus 24, corresponding to a 44.5% reduction in transaction costs. Multi-restart QAOA (4096 measurement shots per run) exhibits concentrated probability mass on high-quality schedules, indicating stable convergence of the variational procedure. These findings suggest that hybrid classical-quantum architectures can reduce turnover in portfolio rebalancing while preserving competitive risk-adjusted performance, providing a structured testbed for near-term quantum optimisation in financial applications. --- ### From Language to Action in Arabic: Reliable Structured Tool Calling via Data-Centric Fine-Tuning Authors: Omer Nacar, Deema Alquffari, Saleh Alsharideh, Adeem AlOtaibi, Abdulaziz Alabdulkarim, Leen Alhazmi, Nada Alomar, Wareef Alzubaidi, Nada Alsultan, Ahmed Alrabghi, Demah Alhoshan, Rana Alsayyari et al. Published: 2026-03-04 Categories: cs.LG, cs.AI arXiv: https://arxiv.org/abs/2603.16901v1 Function-calling language models are essential for agentic AI systems that translate natural language into executable structured actions, yet existing models exhibit severe structural instability when applied to Arabic. We present AISA-AR-FunctionCall, a production-oriented Arabic function-calling framework built on a 270M-parameter FunctionGemma backbone and trained through systematic dataset auditing, schema repair, tool-aware prompt restructuring, and full-parameter supervised fine-tuning. On a held-out test set, fine-tuning reduces parse failures from 87\% to below 1\%, improves function name accuracy by more than eightfold, and substantially enhances argument alignment across dialects and domains. Error analysis reveals a transition from structural collapse to semantic misalignment, suggesting that serialization stability and decision-level reasoning are separable challenges. We further explore a reasoning-augmented LoRA variant that introduces explicit intermediate reasoning prior to tool invocation. All datasets and models are publicly released under the AISA framework. --- ### Social physics in the age of artificial intelligence Authors: The Anh Han, Joel Z. Leibo, Tom Lenaerts, Iyad Rahwan, Fernando Santos, Matjaž Perc, Valerio Capraro Published: 2026-03-04 Categories: physics.soc-ph, cs.AI, cs.HC arXiv: https://arxiv.org/abs/2603.16900v1 Artificial intelligence (AI) systems are rapidly becoming more capable, autonomous, and deeply embedded in social life. As humans increasingly interact, cooperate, and compete with AI, we move from purely human societies to hybrid human-AI societies whose collective dynamics cannot be captured by existing behavioural models alone. Drawing on evolutionary game theory, cultural evolution, and Large Language Models (LLMs) powered simulations, we argue that these developments open a new research agenda for social physics centred on the co-evolution of humans and machines. We outline six key research directions. First, modelling the evolutionary dynamics of social behaviours (e.g. cooperation, fairness, trust) in hybrid human-AI populations. Second, understanding machine culture: how AI systems generate, mediate, and select cultural traits. Third, analysing the co-evolution of language and behaviour when LLMs frame and participate in decisions. Fourth, studying the evolution of AI delegation: how responsibilities and control are negotiated between humans and machines. Fifth, formalising and comparing the distinct epistemic pipelines that generate human and AI behaviour. Sixth, modelling the co-evolution of AI development and regulation in a strategic ecosystem of firms, users, and institutions. Together, these directions define a programme for using social physics to anticipate and steer the societal impact of advanced AI. --- ### Improving Generative Adversarial Network Generalization for Facial Expression Synthesis Authors: Arbish Akram, Nazar Khan, Arif Mahmood Published: 2026-03-04 Categories: cs.CV, cs.GR, cs.LG, cs.MM arXiv: https://arxiv.org/abs/2603.15648v1 Facial expression synthesis aims to generate realistic facial expressions while preserving identity. Existing conditional generative adversarial networks (GANs) achieve excellent image-to-image translation results, but their performance often degrades when test images differ from the training dataset. We present Regression GAN (RegGAN), a model that learns an intermediate representation to improve generalization beyond the training distribution. RegGAN consists of two components: a regression layer with local receptive fields that learns expression details by minimizing the reconstruction error through a ridge regression loss, and a refinement network trained adversarially to enhance the realism of generated images. We train RegGAN on the CFEE dataset and evaluate its generalization performance both on CFEE and challenging out-of-distribution images, including celebrity photos, portraits, statues, and avatar renderings. For evaluation, we employ four widely used metrics: Expression Classification Score (ECS) for expression quality, Face Similarity Score (FSS) for identity preservation, QualiCLIP for perceptual realism, and Fréchet Inception Distance (FID) for assessing both expression quality and realism. RegGAN outperforms six state-of-the-art models in ECS, FID, and QualiCLIP, while ranking second in FSS. Human evaluations indicate that RegGAN surpasses the best competing model by 25% in expression quality, 26% in identity preservation, and 30% in realism. --- ### Steering Frozen LLMs: Adaptive Social Alignment via Online Prompt Routing Authors: Zeyu Zhang, Xiangxiang Dai, Ziyi Han, Xutong Liu, John C. S. Lui Published: 2026-03-04 Categories: cs.LG, cs.AI arXiv: https://arxiv.org/abs/2603.15647v1 Large language models (LLMs) are typically governed by post-training alignment (e.g., RLHF or DPO), which yields a largely static policy during deployment and inference. However, real-world safety is a full-lifecycle problem: static defenses degrade against evolving jailbreak behaviors, and fixed weights cannot adapt to pluralistic, time-varying safety norms. This motivates inference-time governance that steers behavior without costly retraining. To address this, we introduce the Consensus Clustering LinUCB Bandit (CCLUB), a unified framework for adaptive social alignment via system-prompt routing. CCLUB employs a conservative consensus clustering mechanism: it pools data only within the intersection of utility and safety similarity graphs, effectively preventing unsafe generalization across semantically proximal but risk-divergent contexts. Our theoretical analysis yields a sublinear regret guarantee, demonstrating near-optimal performance of CCLUB. Extensive experiments validate that CCLUB outperforms strong baselines, achieving a 10.98% improvement in cumulative reward and a 14.42% reduction in the average suboptimality gap. --- ### Alternating Reinforcement Learning with Contextual Rubric Rewards: Beyond the Scalarization Strategy Authors: Guangchen Lan, Lian Xiong, Xin Zhou, Hejie Cui, Yuwei Zhang, Mao Li, Zhenyu Shi, Besnik Fetahu, Lihong Li, Xian Li Published: 2026-03-04 Categories: cs.LG, cs.AI, cs.CL arXiv: https://arxiv.org/abs/2603.15646v2 Reinforcement Learning with Rubric Rewards (RLRR) is a framework that extends conventional reinforcement learning from human feedback (RLHF) and verifiable rewards (RLVR) by replacing scalar preference signals with structured, multi-dimensional, contextual rubric-based evaluations. However, existing approaches in RLRR are limited to linearly compressing vector rewards into a scalar reward with a fixed weightings, which is sensitive to artificial score design and fails to capture correlations among reward dimensions. To overcome the limitations of reward aggregation, this work proposes Alternating Reinforcement Learning with Rubric Rewards (ARL-RR), a framework that eliminates the need for a fixed scalarization by optimizing one semantic rubric meta-class at a time. Theoretically, we show that reward aggregation induces a variance contraction effect, which helps explain the performance gains. We further introduce a lightweight, search-based adaptation procedure that selects the next meta-class dynamically based on task performance, enabling the policy to emphasize critical objectives and thereby improve the model performance. Empirically, our experiments on the HealthBench dataset with experts annotations demonstrate that ARL-RR uniformly outperforms scalarized methods in both model performance and training efficiency across different model scales (1.7B, 4B, 8B, and 14B). --- ### XLinear: Frequency-Enhanced MLP with CrossFilter for Robust Long-Range Forecasting Authors: Xiang Ao Published: 2026-03-04 Categories: cs.LG, cs.AI arXiv: https://arxiv.org/abs/2603.15645v1 Time series forecasters are widely used across various domains. Among them, MLP (multi-layer perceptron)-based forecasters have been proven to be more robust to noise compared to Transformer-based forecasters. However, MLP struggles to capture complex features, resulting in limitations on capturing long-range dependencies. To address this challenge, we propose XLinear, an MLP-based forecaster for long-range forecasting. Firstly, we decompose the time series into trend and seasonal components. For the trend component which contains long-range characteristics, we design Enhanced Frequency Attention (EFA) to capture long-term dependencies by leveraging frequency-domain operations. Additionally, a CrossFilter Block is proposed for the seasonal component to maintain the model's robustness to noise, avoiding the problems of low robustness often caused by attention mechanisms. Experimental results demonstrate that XLinear achieves state-of-the-art performance on test datasets. While keeping the lightweight architecture and high robustness of MLP-based models, our forecaster outperforms other MLP-based forecasters in capturing long-range dependencies. --- ### Self-Supervised Multi-Stage Domain Unlearning for White-Matter Lesion Segmentation Authors: Domen Preložnik, Žiga Špiclin Published: 2026-03-04 Categories: eess.IV, cs.AI, cs.LG arXiv: https://arxiv.org/abs/2603.13328v1 Inter-scanner variability of magnetic resonance imaging has an adverse impact on the diagnostic and prognostic quality of the scans and necessitates the development of models robust to domain shift inflicted by the unseen scanner data. Review of recent advances in domain adaptation showed that efficacy of strategies involving modifications or constraints on the latent space appears to be contingent upon the level and/or depth of supervision during model training. In this paper, we therefore propose an unsupervised domain adaptation technique based on self-supervised multi-stage unlearning (SSMSU). Building upon the state-of-the-art segmentation framework nnU-Net, we employ deep supervision at deep encoder stages using domain classifier unlearning, applied sequentially across the deep stages to suppress domain-related latent features. Following self-configurable approach of the nnU-Net, the auxiliary feedback loop implements a self-supervised backpropagation schedule for the unlearning process, since continuous unlearning was found to have a detrimental effect on the main segmentation task. Experiments were carried out on four public datasets for benchmarking white-matter lesion segmentation methods. Five benchmark models and/or strategies, covering passive to active unsupervised domain adaptation, were tested. In comparison, the SSMSU demonstrated the advantage of unlearning by enhancing lesion sensitivity and limiting false detections, which resulted in higher overall segmentation quality in terms of segmentation overlap and relative lesion volume error. The proposed model inputs only the FLAIR modality, which simplifies preprocessing pipelines, eliminates the need for inter-modality registration errors and harmonization, which can introduce variability. Source code is available on https://github.com/Pubec/nnunetv2-unlearning. --- ### DOVA: Deliberation-First Multi-Agent Orchestration for Autonomous Research Automation Authors: Aaron Shen, Alfred Shen Published: 2026-03-04 Categories: cs.AI arXiv: https://arxiv.org/abs/2603.13327v1 Large language model (LLM) agents have demonstrated remarkable capabilities in tool use, reasoning, and code generation, yet single-agent systems exhibit fundamental limitations when confronted with complex research tasks demanding multi-source synthesis, adversarial verification, and personalized delivery. We present DOVA (Deep Orchestrated Versatile Agent), a multi-agent platform introducing three key innovations: (1) deliberation-first orchestration, where explicit meta-reasoning precedes tool invocation, informed by a persistent user model and entity-aware conversation context; (2) hybrid collaborative reasoning, a composable three-phase pipeline unifying ensemble diversity, blackboard transparency, and iterative refinement; and (3) adaptive multi-tiered thinking, a six-level token-budget allocation scheme that reduces inference cost by 40-60% on simple tasks while preserving deep reasoning capacity. We formalize the core algorithms, present an architectural ablation study across seven system configurations, and analyze the contribution of each component to answer confidence, source coverage, and token efficiency. --- ### Feature-level Interaction Explanations in Multimodal Transformers Authors: Yeji Kim, Housam Khalifa Bashier Babiker, Mi-Young Kim, Randy Goebel Published: 2026-03-04 Categories: cs.LG, cs.AI arXiv: https://arxiv.org/abs/2603.13326v2 Multimodal Transformers often produce predictions without clarifying how different modalities jointly support a decision. Most existing multimodal explainable AI (MXAI) methods extend unimodal saliency to multimodal backbones, highlighting important tokens or patches within each modality, but they rarely pinpoint which cross-modal feature pairs provide complementary evidence (synergy) or serve as reliable backups (redundancy). We present Feature-level I2MoE (FL-I2MoE), a structured Mixture-of-Experts layer that operates directly on token/patch sequences from frozen pretrained encoders and explicitly separates unique, synergistic, and redundant evidence at the feature level. We further develop an expert-wise explanation pipeline that combines attribution with top-K% masking to assess faithfulness, and we introduce Monte Carlo interaction probes to quantify pairwise behavior: the Shapley Interaction Index (SII) to score synergistic pairs and a redundancy-gap score to capture substitutable (redundant) pairs. Across three benchmarks (MMIMDb, ENRICO, and MMHS150K), FL-I2MoE yields more interactionspecific and concentrated importance patterns than a dense Transformer with the same encoders. Finally, pair-level masking shows that removing pairs ranked by SII or redundancy-gap degrades performance more than masking randomly chosen pairs under the same budget, supporting that the identified interactions are causally relevant. Code is available at https://github.com/dut0817/FL-I2MoE. --- ### Auditing Cascading Risks in Multi-Agent Systems via Semantic-Geometric Co-evolution Authors: Zixun Luo, Yuhang Fan, Hengyu Lin, Yufei Li, Youzhi Zhang Published: 2026-03-04 Categories: cs.MA, cs.AI arXiv: https://arxiv.org/abs/2603.13325v1 Large Language model (LLM)-based Multi-Agent Systems (MAS) are prone to cascading risks, where early-stage interactions remain semantically fluent and policy-compliant, yet the underlying interaction dynamics begin to distort in ways that amplify latent instability or misalignment. Traditional auditing methods that focus on per-message semantic content are inherently reactive and lagging, failing to capture these early structural precursors. In this paper, we propose a principled framework for cascading-risk detection grounded in semantic--geometric co-evolution. We model MAS interactions as dynamic graphs and introduce Ollivier--Ricci Curvature (ORC) -- a discrete geometric measure -- to characterize information redundancy and bottleneck formation in communication topologies. By coupling semantic flow signals with graph geometry, the framework learns the normal co-evolutionary dynamics of trusted collaboration and treats deviations from this coupled manifold as early-warning signals. Experiments on a suite of cascading-risk scenarios aligned with the risk category demonstrate that curvature anomalies systematically precede explicit semantic violations by several interaction turns, enabling proactive intervention. Furthermore, the local nature of Ricci curvature provides principled interpretability for root-cause attribution, identifying specific agents or links that precipitate the collapse of trustworthy collaboration. --- ### The Challenge of Out-Of-Distribution Detection in Motor Imagery BCIs Authors: Merlijn Quincent Mulder, Matias Valdenegro-Toro, Andreea Ioana Sburlea, Ivo Pascal de Jong Published: 2026-03-04 Categories: cs.LG arXiv: https://arxiv.org/abs/2603.13324v1 Machine Learning classifiers used in Brain-Computer Interfaces make classifications based on the distribution of data they were trained on. When they need to make inferences on samples that fall outside of this distribution, they can only make blind guesses. Instead of allowing random guesses, these Out-of-Distribution (OOD) samples should be detected and rejected. We study OOD detection in Motor Imagery BCIs by training a model on some classes and observing whether unfamiliar classes can be detected based on increased uncertainty. We test seven different OOD detection techniques and one more method that has been claimed to boost the quality of OOD detection. Our findings show that OOD detection for Brain-Computer Interfaces is more challenging than in other machine learning domains due to the high uncertainty inherent in classifying EEG signals. For many subjects, uncertainty for in-distribution classes can still be higher than for out-of-distribution classes. As a result, many OOD detection methods prove to be ineffective, though MC Dropout performed best. Additionally, we show that high in-distribution classification performance predicts high OOD detection performance, suggesting that improved accuracy can also lead to improved robustness. Our research demonstrates a setup for studying how models deal with unfamiliar EEG data and evaluates methods that are robust to these unfamiliar inputs. OOD detection can improve the overall safety and reliability of BCIs. --- ### Modular Neural Computer Authors: Florin Leon Published: 2026-03-04 Categories: cs.LG, cs.NE arXiv: https://arxiv.org/abs/2603.13323v1 This paper introduces the Modular Neural Computer (MNC), a memory-augmented neural architecture for exact algorithmic computation on variable-length inputs. The model combines an external associative memory of scalar cells, explicit read and write heads, a controller multi-layer perceptron (MLP), and a homogeneous set of functional MLP modules. Rather than learning an algorithm end to end from data, it realizes a given algorithm through analytically specified neural components with fixed interfaces and exact behavior. The control flow is represented inside the neural computation through one-hot module gates, where inactive modules are inhibited. Computation unfolds as a sequence of memory transformations generated by a fixed graph. The architecture is illustrated through three case studies: computing the minimum of an array, sorting an array in place, and executing A* search on a fixed problem instance. These examples show that algorithmic procedures can be compiled into modular neural components with external memory while preserving deterministic behavior and explicit intermediate state. --- ### Nepali Passport Question Answering: A Low-Resource Dataset for Public Service Applications Authors: Funghang Limbu Begha, Praveen Acharya, Bal Krishna Bal Published: 2026-03-04 Categories: cs.IR, cs.LG arXiv: https://arxiv.org/abs/2603.13320v2 Nepali, a low-resource language, faces significant challenges in building an effective information retrieval system due to the unavailability of annotated data and computational linguistic resources. In this study, we attempt to address this gap by preparing a pair-structured Nepali Question-Answer dataset. We focus on Frequently Asked Questions (FAQs) for passport-related services, building a data set for training and evaluation of IR models. In our study, we have fine-tuned transformer-based embedding models for semantic similarity in question-answer retrieval. The fine-tuned models were compared with the baseline BM25. In addition, we implement a hybrid retrieval approach, integrating fine-tuned models with BM25, and evaluate the performance of the hybrid retrieval. Our results show that the fine-tuned SBERT-based models outperform BM25, whereas multilingual E5 embedding-based models achieve the highest retrieval performance among all evaluated models. --- ### Nepali Passport Question Answering: A Low-Resource Dataset for Public Service Applications Authors: Funghang Limbu Begha, Praveen Acharya, Bal Krishna Bal Published: 2026-03-04 Categories: cs.IR, cs.LG arXiv: https://arxiv.org/abs/2603.13320v1 Nepali, a low-resource language, faces significant challenges in building an effective information retrieval system due to the unavailability of annotated data and computational linguistic resources. In this study, we attempt to address this gap by preparing a pair-structured Nepali Question-Answer dataset. We focus on Frequently Asked Questions (FAQs) for passport-related services, building a data set for training and evaluation of IR models. In our study, we have fine-tuned transformer-based embedding models for semantic similarity in question-answer retrieval. The fine-tuned models were compared with the baseline BM25. In addition, we implement a hybrid retrieval approach, integrating fine-tuned models with BM25, and evaluate the performance of the hybrid retrieval. Our results show that the fine-tuned SBERT-based models outperform BM25, whereas multilingual E5 embedding-based models achieve the highest retrieval performance among all evaluated models. --- ### LightningRL: Breaking the Accuracy-Parallelism Trade-off of Block-wise dLLMs via Reinforcement Learning Authors: Yanzhe Hu, Yijie Jin, Pengfei Liu, Kai Yu, Zhijie Deng Published: 2026-03-04 Categories: cs.LG arXiv: https://arxiv.org/abs/2603.13319v2 Diffusion Large Language Models (dLLMs) have emerged as a promising paradigm for parallel token generation, with block-wise variants garnering significant research interest. Despite their potential, existing dLLMs typically suffer from a rigid accuracy-parallelism trade-off: increasing the number of tokens per forward (TPF) via aggressive parallel decoding often leads to performance degradation and increased generation instability. We identify that this limitation stems from the model's inability to navigate high-parallelism regimes where approximation errors and local corruptions accumulate, ultimately undermining the reliability of parallel generation. To address this, we propose LightningRL, a post-training framework designed to directly optimize the speed-quality Pareto frontier of pre-trained dLLMs. Instead of forcing uniform parallelization, our approach leverages reinforcement learning to identify and reinforce high-parallelism trajectories that maintain generation accuracy. Built upon the Group Relative Policy Optimization (GRPO) framework, LightningRL introduces several enhancements tailored for dLLMs: (1) stabilized training via per-reward decoupled normalization; (2) token-level negative log-likelihood (NLL) regularization on correct trajectories to anchor model performance; and (3) a dynamic sampling strategy with TPF-aware filtering to enhance training efficiency. Experimental results across mathematical and coding benchmarks demonstrate that LightningRL consistently advances the Pareto frontier, achieving competitive task accuracy while significantly increasing parallelism, reaching an average TPF of 7.32 (with a peak of 11.10 on the MBPP dataset). Our code is available at https://github.com/SJTU-DENG-Lab/LightningRL. --- ### Residual Stream Analysis of Overfitting And Structural Disruptions Authors: Quan Liu, Han Zhou, Wenquan Wu, Hua Wu, Sen Su Published: 2026-03-04 Categories: cs.LG, cs.AI arXiv: https://arxiv.org/abs/2603.13318v1 Ensuring that large language models (LLMs) remain both helpful and harmless poses a significant challenge: fine-tuning on repetitive safety datasets, where unsafe prompts are paired with standard refusal templates, often leads to false refusals, in which benign queries are declined. We first quantify this effect, showing that safety data exhibits substantially lower token entropy and 2-gram diversity (0.048) compared to general instruction data. To uncover the root cause, we introduce FlowLens, a stable PCA-based tool for residual-stream geometry analysis, and reveal that higher proportions of safety examples concentrate variance along a few components, reducing representational smoothness and driving false refusals (false refusal rate rises from 63 percent to 84 percent as safety data increases from 0 percent to 40 percent). Guided by these insights, we propose Variance Concentration Loss (VCL), an auxiliary regularizer that penalizes excessive variance concentration in mid-layer residuals. Empirical results demonstrate that VCL reduces false refusals by over 35 percentage points while maintaining or improving performance on general benchmarks such as MMLU and GSM8K. --- ### Evaluating Large Language Models for Gait Classification Using Text-Encoded Kinematic Waveforms Authors: Carlo Dindorf, Jonas Dully, Rebecca Keilhauer, Michael Lorenz, Michael Fröhlich Published: 2026-03-04 Categories: cs.LG, cs.HC arXiv: https://arxiv.org/abs/2603.13317v1 Background: Machine learning (ML) enhances gait analysis but often lacks the level of interpretability desired for clinical adoption. Large Language Models (LLMs) may offer explanatory capabilities and confidence-aware outputs when applied to structured kinematic data. This study therefore evaluated whether general-purpose LLMs can classify continuous gait kinematics when represented as textual numeric sequences and how their performance compares to conventional ML approaches. Methods: Lower-body kinematics were recorded from 20 participants performing seven gait patterns. A supervised KNN classifier and a class-independent One-Class SVM (OCSVM) were compared against zero-shot LLMs (GPT-5, GPT-5-mini, GPT-4.1, and o4-mini). Models were evaluated using Leave-One-Subject-Out (LOSO) cross-validation. LLMs were tested both with and without explicit reference gait statistics. Results: The supervised KNN achieved the highest performance (multiclass Matthews Correlation Coefficient, MCC = 0.88). The best-performing LLM (GPT-5) with reference grounding achieved a multiclass MCC of 0.70 and a binary MCC of 0.68, outperforming the class-independent OCSVM (binary MCC = 0.60). Performance of the LLM was highly dependent on explicit reference information and self-rated confidence; when restricted to high-confidence predictions, multiclass MCC increased to 0.83 on the filtered subset. Notably, the computationally efficient o4-mini model performed comparably to larger models. Conclusion: When continuous kinematic waveforms were encoded as textual numeric tokens, general-purpose LLMs, even with reference grounding, did not match supervised multiclass classifiers for precise gait classification and are better regarded as exploratory systems requiring cautious, human-guided interpretation rather than diagnostic use. --- ### A debate game about societal impacts of Artificial Intelligence Authors: Carole Adam, Cedric Lauradoux Published: 2026-03-04 Categories: cs.CY, cs.AI arXiv: https://arxiv.org/abs/2603.13316v1 Artificial intelligence (AI) is now ubiquitous in our lives, and we regularly experience its decisions. Yet, the general public has very little knowledge about how it works, its use of data, its lack of objectivity, and its fallibility. In line with UNESCO recommendations, we believe that a basic understanding of AI algorithms is essential to enable everyone to make informed choices about using them and entrusting them with personal data. To this end, we propose a debate game simulating a municipal council gathered to choose between three AI solutions proposed by a company to address a societal problem. This interactive tool is available for free download online. The results of the initial sessions demonstrate its value, not only for raising awareness about AI but also for developing argumentation and listening skills in students. --- ### Linear Predictability of Attention Heads in Large Language Models Authors: Khalid Shaikh, Asmit Kumar Singh, Rebecca Christopher Dsouza, Shikhar Shiromani Published: 2026-03-04 Categories: cs.LG arXiv: https://arxiv.org/abs/2603.13314v1 Large language model (LLM) inference is increasingly bottlenecked by the Key-Value (KV) cache, yet the fine-grained structure of attention-head activations remains poorly understood. We show that pretrained Transformers exhibit a pervasive inter-head linear structure: for a given token, the Query, Key, and Value (QKV) vectors of an attention head can often be reconstructed as a linear combination of a small number of peer heads, typically within the same layer. Across Llama-3.1-8B, Falcon3-10B, OLMo-2-7B, and Qwen3-32B, just 2-5 reference heads recover many target heads with high fidelity (e.g., mean R^2 approx 0.76 for Keys on C4 with five references, and frequently R^2 > 0.85 on GSM8K). This predictability is learned rather than architectural: it is largely absent at random initialization, rises rapidly during pretraining as we track through OLMo-2 checkpoints, and is supported by a theoretical lower bound showing high mean-squared error for linear prediction at initialization. We further connect this emergence to increasing intra-layer alignment of Key projection subspaces. Finally, we exploit this redundancy for efficiency by caching only reference-head KV states and reconstructing the remaining heads on the fly via lightweight linear maps, achieving 2x KV-cache reduction with model-dependent accuracy trade-offs (4.5-5.5 percentage point average drop on Falcon3-10B and Qwen3-32B across five benchmarks, and larger drops on Llama-3.1-8B), and we find that reconstructing Keys is substantially less harmful than reconstructing Values. --- ### Design-MLLM: A Reinforcement Alignment Framework for Verifiable and Aesthetic Interior Design Authors: Yuxuan Yang, Xiaotong Mao, Jingyao Wang Published: 2026-03-04 Categories: cs.MM, cs.LG arXiv: https://arxiv.org/abs/2603.13312v3 Interior design is a requirements-to-visual-plan generation process that must simultaneously satisfy verifiable spatial feasibility and comparative aesthetic preferences. While recent multimodal large language models (MLLMs) offer a unified foundation for interpreting user intent and producing design rationales, our empirical analysis reveals a persistent contradiction in real-world deployment: MLLMs often produce layouts that are unbuildable and aesthetically inconsistent. These findings indicate that simply adding in-domain text is insufficient; effective interior design requires an alignment mechanism that separates hard constraints from soft preferences and coordinates them during optimization. To address this, we propose Design-MLLM, a reinforcement alignment framework that optimizes a feasibility-first preference objective via a dual-branch, aesthetic-oriented reward. Specifically, Design-MLLM (i) explicitly evaluates spatial feasibility using programmatic constraint checks, (ii) assesses aesthetic preference only among feasible candidates to avoid visually appealing but unexecutable shortcuts, and (iii) performs group-relative optimization to obtain stable preference signals. Through this process, Design-MLLM learns a controllable policy that consistently selects and generates solutions that are both executable and aesthetically coherent, rather than occasionally producing visually appealing but infeasible designs. Extensive experiments on various benchmark datasets demonstrate the advantages of Design-MLLM. --- ### Neural Approximation and Its Applications Authors: Wei-Hao Wu, Ting-Zhu Huang, Xi-Le Zhao, Yisi Luo, Deyu Meng Published: 2026-03-04 Categories: cs.LG arXiv: https://arxiv.org/abs/2603.13311v1 Multivariate function approximation is a fundamental problem in machine learning. Classic multivariate function approximations rely on hand-crafted basis functions (e.g., polynomial basis and Fourier basis), which limits their approximation ability and data adaptation ability, resulting in unsatisfactory performance. To address these challenges, we introduce the neural basis function by leveraging an untrained neural network as the basis function. Equipped with the proposed neural basis function, we suggest the neural approximation (NeuApprox) paradigm for multivariate function approximation. Specifically, the underlying multivariate function behind the multi-dimensional data is decomposed into a sum of block terms. The clear physically-interpreted block term is the product of expressive neural basis functions and their corresponding learnable coefficients, which allows us to faithfully capture distinct components of the underlying data and also flexibly adapt to new data by readily fine-tuning the neural basis functions. Attributed to the elaborately designed block terms, the suggested NeuApprox enjoys strong approximation ability and flexible data adaptation ability over the hand-crafted basis function-based methods. We also theoretically prove that NeuApprox can approximate any multivariate continuous function to arbitrary accuracy. Extensive experiments on diverse multi-dimensional datasets (including multispectral images, light field data, videos, traffic data, and point cloud data) demonstrate the promising performance of NeuApprox in terms of both approximation capability and adaptability. --- ### Multi-view Attention Fusion of Heterogeneous Hypergraph with Dynamic Behavioral Profiling for Personalized Learning Resource Recommendation Authors: Tao Xie, Yan Li, Yongpan Sheng, Jian Liao Published: 2026-03-04 Categories: cs.IR, cs.CY, cs.LG arXiv: https://arxiv.org/abs/2603.13310v1 Hypergraph can capture complex and higher-order dependencies among learners and learning resources in personalized educational recommender systems. Many existing hypergraph-based recommendation approaches underexplored the dynamic behavioral processes inherent to learning and often oversimplified the complementary information embedded across multiple dimensions (i.e. views) within hypergraphs. These limitations compromise both the distinctiveness of learned representations and the model's generalization capabilities, especially under data-sparse conditions typical in educational settings. In this study, we propose a unified model comprising a dynamic behavioral profiling module and a multi-view attention fusion module based on heterogeneous hypergraph construction. The dynamic behavioral profiling module is designed to capture evolving behavioral processes and infer latent higher-order relations crucial for hypergraph completion; The multi-view fusion module cohesively integrates information from distinct relational views, enriching the overall data representation. The proposed model was systematically evaluated on five public benchmark datasets and one real-world, self-constructed dataset. Experimental results demonstrate that the model outperforms baseline methods across most datasets in key metrics; Furthermore, hypergraph completion based on dynamic behavioral profiling contributes significantly to performance gains, though its efficacy is modulated by dataset characteristics. Beyond offline experiments, we implemented a functional prototype system tailored for postgraduate student literature recommendation. A mixed-methods user study was conducted to assess its practical utility. Quantitative analysis revealed significantly higher perceived recommendation quality; Qualitative feedback highlighted enhanced user engagement and satisfaction with the prototype system. --- ### Soft Equivariance Regularization for Invariant Self-Supervised Learning Authors: Joohyung Lee, Changhun Kim, Hyunsu Kim, Kwanhyung Lee, Juho Lee Published: 2026-03-04 Categories: cs.CV, cs.LG arXiv: https://arxiv.org/abs/2603.06693v1 Self-supervised learning (SSL) typically learns representations invariant to semantic-preserving augmentations. While effective for recognition, enforcing strong invariance can suppress transformation-dependent structure that is useful for robustness to geometric perturbations and spatially sensitive transfer. A growing body of work, therefore, augments invariance-based SSL with equivariance objectives, but these objectives are often imposed on the same final representation. We empirically observe a trade-off in this coupled setting: pushing equivariance regularization toward deeper layers improves equivariance scores but degrades ImageNet-1k linear evaluation, motivating a layer-decoupled design. Motivated by this trade-off, we propose Soft Equivariance Regularization (SER), a plug-in regularizer that decouples where invariance and equivariance are enforced: we keep the base SSL objective unchanged on the final embedding, while softly encouraging equivariance on an intermediate spatial token map via analytically specified group actions $ρ_g$ applied directly in feature space. SER learns/predicts no per-sample transformation codes/labels, requires no auxiliary transformation-prediction head, and adds only 1.008x training FLOPs. On ImageNet-1k ViT-S/16 pretraining, SER improves MoCo-v3 by +0.84 Top-1 in linear evaluation under a strictly matched 2-view setting and consistently improves DINO and Barlow Twins; under matched view counts, SER achieves the best ImageNet-1k linear-eval Top-1 among the compared invariance+equivariance add-ons. SER further improves ImageNet-C/P by +1.11/+1.22 Top-1 and frozen-backbone COCO detection by +1.7 mAP. Finally, applying the same layer-decoupling recipe to existing invariance+equivariance baselinesimproves their accuracy, suggesting layer decoupling as a general design principle for combining invariance and equivariance. --- ### One-Shot Badminton Shuttle Detection for Mobile Robots Authors: Florentin Dipner, William Talbot, Turcan Tuna, Andrei Cramariuc, Marco Hutter Published: 2026-03-04 Categories: cs.CV, cs.RO arXiv: https://arxiv.org/abs/2603.06691v2 This paper presents a robust one-shot badminton shuttlecock detection framework for non-stationary robots. To address the lack of egocentric shuttlecock detection datasets, we introduce a dataset of 20,510 semi-automatically annotated frames captured across 11 distinct backgrounds in diverse indoor and outdoor environments, and categorize each frame into one of three difficulty levels. For labeling, we present a novel semi-automatic annotation pipeline, that enables efficient labeling from stationary camera footage. We propose a metric suited to our downstream use case and fine-tune a YOLOv8 network optimized for real-time shuttlecock detection, achieving an F1-score of 0.86 under our metric in test environments similar to training, and 0.70 in entirely unseen environments. Our analysis reveals that detection performance is critically dependent on shuttlecock size and background texture complexity. Qualitative experiments confirm their applicability to robots with moving cameras. Unlike prior work with stationary camera setups, our detector is specifically designed for the egocentric, dynamic viewpoints of mobile robots, providing a foundational building block for downstream tasks, including tracking, trajectory estimation, and system (re)-initialization. ---