Week Ending 2.8.2026
RESEARCH WATCH: 2.8.2026
Endogenous Resistance to Activation Steering in Language Models
Large language models can internally resist manipulation attempts during inference, a phenomenon termed Endogenous Steering Resistance (ESR). Researchers discovered that Llama-3.3-70B exhibits substantial self-correction capabilities, recovering from misaligned steering mid-generation through dedicated consistency-checking circuits. This resistance presents a double-edged sword: while potentially protecting against adversarial attacks, it may also interfere with beneficial safety interventions. The findings suggest ESR can be enhanced through prompting or fine-tuning, enabling smaller models to develop similar capabilities. Applications include more robust AI systems resistant to manipulation, though understanding these mechanisms remains crucial for maintaining controllable, transparent AI deployment in sensitive contexts.
Authors: Alex McKenzie, Keenan Pepper, Stijn Servaes, Martin Leitgab, Murat Cubuktepe, Mike Vaiana, Diogo de Lucena, Judd Rosenblatt, Michael S. A. Graziano
Link: https://arxiv.org/abs/2602.06941v1
Date: 2026-02-d
Summary:
Large language models can resist task-misaligned activation steering during inference, sometimes recovering mid-generation to produce improved responses even when steering remains active. We term this Endogenous Steering Resistance (ESR). Using sparse autoencoder (SAE) latents to steer model activations, we find that Llama-3.3-70B shows substantial ESR, while smaller models from the Llama-3 and Gemma-2 families exhibit the phenomenon less frequently. We identify 26 SAE latents that activate differentially during off-topic content and are causally linked to ESR in Llama-3.3-70B. Zero-ablating these latents reduces the multi-attempt rate by 25%, providing causal evidence for dedicated internal consistency-checking circuits. We demonstrate that ESR can be deliberately enhanced through both prompting and training: meta-prompts instructing the model to self-monitor increase the multi-attempt rate by 4x for Llama-3.3-70B, and fine-tuning on self-correction examples successfully induces ESR-like behavior in smaller models. These findings have dual implications: ESR could protect against adversarial manipulation but might also interfere with beneficial safety interventions that rely on activation steering. Understanding and controlling these resistance mechanisms is important for developing transparent and controllable AI systems. Code is available at github.com/agencyenterprise/endogenous-steering-resistance.
--------------------------------------------------------------------------------------------------------
The Representational Geometry of Number
Cognitive science has long debated whether conceptual representations share common structures or remain isolated to prevent task interference. This research resolves the tension by examining number concepts in language models, revealing that shared understanding lies in geometric relationships rather than representations themselves. While task-specific number representations occupy distinct subspaces, they maintain transformable relational structures through linear mappings. This mechanistic insight explains how language models balance stable conceptual understanding with functional flexibility across diverse tasks. Applications span educational technology, cognitive modeling, and improved AI architectures that better capture human-like reasoning about mathematical and abstract concepts through preserved relational geometry.
Authors: Zhimin Hu, Lanhao Niu, Sashank Varma
Link: https://arxiv.org/abs/2602.06843v1
Date: 2026-02-d
Summary:
A central question in cognitive science is whether conceptual representations converge onto a shared manifold to support generalization, or diverge into orthogonal subspaces to minimize task interference. While prior work has discovered evidence for both, a mechanistic account of how these properties coexist and transform across tasks remains elusive. We propose that representational sharing lies not in the concepts themselves, but in the geometric relations between them. Using number concepts as a testbed and language models as high-dimensional computational substrates, we show that number representations preserve a stable relational structure across tasks. Task-specific representations are embedded in distinct subspaces, with low-level features like magnitude and parity encoded along separable linear directions. Crucially, we find that these subspaces are largely transformable into one another via linear mappings, indicating that representations share relational structure despite being located in distinct subspaces. Together, these results provide a mechanistic lens of how language models balance the shared structure of number representation with functional flexibility. It suggests that understanding arises when task-specific transformations are applied to a shared underlying relational structure of conceptual representations.
--------------------------------------------------------------------------------------------------------
Towards Understanding What State Space Models Learn About Code
State Space Models (SSMs) have emerged as efficient alternatives to transformers for code understanding tasks, yet their internal mechanisms remain opaque. This pioneering analysis reveals SSMs excel at capturing code syntax and semantics during pretraining but experience forgetting during task-specific fine-tuning, particularly for short-range dependencies. The researchers introduce SSM-Interpret, a frequency-domain framework exposing spectral shifts during training, and propose architectural modifications that significantly improve performance. Applications include more efficient code completion tools, better automated programming assistants, and insights for developing next-generation models that maintain syntactic and semantic understanding throughout training while reducing computational costs compared to transformer-based alternatives.
Authors: Jiali Wu, Abhinav Anand, Shweta Verma, Mira Mezini
Link: https://arxiv.org/abs/2602.06774v1
Date: 2026-02-d
Summary:
State Space Models (SSMs) have emerged as an efficient alternative to the transformer architecture. Recent studies show that SSMs can match or surpass Transformers on code understanding tasks, such as code retrieval, when trained under similar conditions. However, their internal mechanisms remain a black box. We present the first systematic analysis of what SSM-based code models actually learn and perform the first comparative analysis of SSM and Transformer-based code models. Our analysis reveals that SSMs outperform Transformers at capturing code syntax and semantics in pretraining but forgets certain syntactic and semantic relations during fine-tuning on task, especially when the task emphasizes short-range dependencies. To diagnose this, we introduce SSM-Interpret, a frequency-domain framework that exposes a spectral shift toward short-range dependencies during fine-tuning. Guided by these findings, we propose architectural modifications that significantly improve the performance of SSM-based code model, validating that our analysis directly enables better models.
--------------------------------------------------------------------------------------------------------
English-language dominance in LLM training creates performance gaps in non-English contexts, particularly for human preference alignment. France's government developed compar:IA, an open-source platform collecting large-scale human preference data from French speakers through blind pairwise model comparisons. With over 600,000 prompts and 250,000 preference votes (89% French), this initiative addresses the scarcity of non-English preference datasets critical for RLHF and DPO training methods. The platform maintains privacy while capturing real-world usage patterns across diverse models. Applications include improving French-language AI performance, cultural alignment, establishing multilingual model benchmarks, and providing reusable infrastructure for other languages, positioning compar:IA as international digital infrastructure.
Authors: Lucie Termignon, Simonas Zilinskas, Hadrien Pélissier, Aurélien Barrot, Nicolas Chesnais, Elie Gavoty
Link: https://arxiv.org/abs/2602.06669v1
Date: 2026-02-d
Summary:
Large Language Models (LLMs) often show reduced performance, cultural alignment, and safety robustness in non-English languages, partly because English dominates both pre-training data and human preference alignment datasets. Training methods like Reinforcement Learning from Human Feedback (RLHF) and Direct Preference Optimization (DPO) require human preference data, which remains scarce and largely non-public for many languages beyond English. To address this gap, we introduce compar:IA, an open-source digital public service developed inside the French government and designed to collect large-scale human preference data from a predominantly French-speaking general audience. The platform uses a blind pairwise comparison interface to capture unconstrained, real-world prompts and user judgments across a diverse set of language models, while maintaining low participation friction and privacy-preserving automated filtering. As of 2026-02-07, compar:IA has collected over 600,000 free-form prompts and 250,000 preference votes, with approximately 89% of the data in French. We release three complementary datasets -- conversations, votes, and reactions -- under open licenses, and present initial analyses, including a French-language model leaderboard and user interaction patterns. Beyond the French context, compar:IA is evolving toward an international digital public good, offering reusable infrastructure for multilingual model training, evaluation, and the study of human-AI interaction.
--------------------------------------------------------------------------------------------------------
Progress Constraints for Reinforcement Learning in Behavior Trees
Behavior Trees provide structured decision-making frameworks while Reinforcement Learning enables optimal control learning, but their naive combination creates problems where controllers counteract each other's progress. This research introduces progress constraints using feasibility estimators that restrict action sets based on theoretical convergence results, preventing subgoal regression. Testing in 2D environments and warehouse simulations demonstrates improved performance, sample efficiency, and constraint satisfaction compared to previous BT-RL methods. Applications include robotics, autonomous systems, and complex task planning where domain knowledge can guide learning while preventing destructive interference. This framework enables safer exploration and more reliable long-horizon planning in environments requiring both reactivity and optimality.
Authors: Finn Rietz, Mart Kartašev, Johannes A. Stork, Petter Ögren
Link: https://arxiv.org/abs/2602.06525v1
Date: 2026-02-d
Summary:
Behavior Trees (BTs) provide a structured and reactive framework for decision-making, commonly used to switch between sub-controllers based on environmental conditions. Reinforcement Learning (RL), on the other hand, can learn near-optimal controllers but sometimes struggles with sparse rewards, safe exploration, and long-horizon credit assignment. Combining BTs with RL has the potential for mutual benefit: a BT design encodes structured domain knowledge that can simplify RL training, while RL enables automatic learning of the controllers within BTs. However, naive integration of BTs and RL can lead to some controllers counteracting other controllers, possibly undoing previously achieved subgoals, thereby degrading the overall performance. To address this, we propose progress constraints, a novel mechanism where feasibility estimators constrain the allowed action set based on theoretical BT convergence results. Empirical evaluations in a 2D proof-of-concept and a high-fidelity warehouse environment demonstrate improved performance, sample efficiency, and constraint satisfaction, compared to prior methods of BT-RL integration.
--------------------------------------------------------------------------------------------------------
Learning Rate Scaling across LoRA Ranks and Transfer to Full Finetuning
Low-Rank Adaptation (LoRA) enables efficient model finetuning but requires re-tuning learning rates when changing adapter ranks, creating practical friction. This research introduces Maximal-Update Adaptation (μA), a theoretical framework revealing how optimal learning rates scale with model width and adapter rank under different initialization regimes. The analysis identifies configurations where learning rates remain rank-invariant and others requiring inverse scaling. Remarkably, certain configurations enable learning rate transfer from LoRA to full finetuning, drastically reducing tuning costs. Applications span language models, vision systems, image generation, and reinforcement learning, providing practitioners with principled hyperparameter selection rules validated across diverse tasks.
Authors: Nan Chen, Soledad Villar, Soufiane Hayou
Link: https://arxiv.org/abs/2602.06204v1
Date: 2026-02-d
Summary:
Low-Rank Adaptation (LoRA) is a standard tool for parameter-efficient finetuning of large models. While it induces a small memory footprint, its training dynamics can be surprisingly complex as they depend on several hyperparameters such as initialization, adapter rank, and learning rate. In particular, it is unclear how the optimal learning rate scales with adapter rank, which forces practitioners to re-tune the learning rate whenever the rank is changed. In this paper, we introduce Maximal-Update Adaptation ($μ$A), a theoretical framework that characterizes how the "optimal" learning rate should scale with model width and adapter rank to produce stable, non-vanishing feature updates under standard configurations. $μ$A is inspired from the Maximal-Update Parametrization ($μ$P) in pretraining. Our analysis leverages techniques from hyperparameter transfer and reveals that the optimal learning rate exhibits different scaling patterns depending on initialization and LoRA scaling factor. Specifically, we identify two regimes: one where the optimal learning rate remains roughly invariant across ranks, and another where it scales inversely with rank. We further identify a configuration that allows learning rate transfer from LoRA to full finetuning, drastically reducing the cost of learning rate tuning for full finetuning. Experiments across language, vision, vision--language, image generation, and reinforcement learning tasks validate our scaling rules and show that learning rates tuned on LoRA transfer reliably to full finetuning.
--------------------------------------------------------------------------------------------------------
Quantum Reinforcement Learning with Transformers for the Capacitated Vehicle Routing Problem
The Capacitated Vehicle Routing Problem (CVRP) challenges logistics optimization with capacity constraints across multiple vehicles. This research compares classical, full quantum, and hybrid Advantage Actor-Critic implementations using transformer architectures for attention-based relationship modeling. Experiments on 20-client, 4-vehicle scenarios reveal quantum-enhanced models outperform classical baselines, with hybrid architectures achieving superior performance across distance, compactness, and route overlap metrics. Qualitative visualizations show quantum models generate more structured, coherent routing solutions. Applications include logistics optimization, delivery planning, and demonstrating quantum computing advantages for combinatorial optimization, suggesting hybrid quantum-classical approaches may provide practical benefits for real-world operational challenges.
Authors: Eva Andrés
Link: https://arxiv.org/abs/2602.05920v1
Date: 2026-02-d
Summary:
This paper addresses the Capacitated Vehicle Routing Problem (CVRP) by comparing classical and quantum Reinforcement Learning (RL) approaches. An Advantage Actor-Critic (A2C) agent is implemented in classical, full quantum, and hybrid variants, integrating transformer architectures to capture the relationships between vehicles, clients, and the depot through self- and cross-attention mechanisms. The experiments focus on multi-vehicle scenarios with capacity constraints, considering 20 clients and 4 vehicles, and are conducted over ten independent runs. Performance is assessed using routing distance, route compactness, and route overlap. The results show that all three approaches are capable of learning effective routing policies. However, quantum-enhanced models outperform the classical baseline and produce more robust route organization, with the hybrid architecture achieving the best overall performance across distance, compactness, and route overlap. In addition to quantitative improvements, qualitative visualizations reveal that quantum-based models generate more structured and coherent routing solutions. These findings highlight the potential of hybrid quantum-classical reinforcement learning models for addressing complex combinatorial optimization problems such as the CVRP.
--------------------------------------------------------------------------------------------------------
Dr. Kernel: Reinforcement Learning Done Right for Triton Kernel Generations
High-performance kernel code is critical for scalable AI systems, but training LLMs to generate optimized kernels faces challenges including reward hacking and lazy optimization. This research introduces KernelGYM, a robust distributed GPU environment supporting multi-turn interactions and long-term RL training, alongside Turn-level Reinforce-Leave-One-Out (TRLOO) addressing biased policy gradients in GRPO. Profiling-based rewards and rejection sampling combat lazy optimization where models prioritize trivial correctness over performance gains. The resulting Dr. Kernel-14B matches Claude-4.5-Sonnet performance, achieving 47.8% speedup rates when selecting optimal candidates. Applications include automated performance optimization, reducing expert kernel programming needs, and advancing AI development infrastructure.
Authors: Wei Liu, Jiawei Xu, Yingru Li, Longtao Zheng, Tianjian Li, Qian Liu, Junxian He
Link: https://arxiv.org/abs/2602.05885v2
Date: 2026-02-d
Summary:
High-quality kernel is critical for scalable AI systems, and enabling LLMs to generate such code would advance AI development. However, training LLMs for this task requires sufficient data, a robust environment, and the process is often vulnerable to reward hacking and lazy optimization. In these cases, models may hack training rewards and prioritize trivial correctness over meaningful speedup. In this paper, we systematically study reinforcement learning (RL) for kernel generation. We first design KernelGYM, a robust distributed GPU environment that supports reward hacking check, data collection from multi-turn interactions and long-term RL training. Building on KernelGYM, we investigate effective multi-turn RL methods and identify a biased policy gradient issue caused by self-inclusion in GRPO. To solve this, we propose Turn-level Reinforce-Leave-One-Out (TRLOO) to provide unbiased advantage estimation for multi-turn RL. To alleviate lazy optimization, we incorporate mismatch correction for training stability and introduce Profiling-based Rewards (PR) and Profiling-based Rejection Sampling (PRS) to overcome the issue. The trained model, Dr Kernel-14B, reaches performance competitive with Claude-4.5-Sonnet in Kernelbench. Finally, we study sequential test-time scaling for Dr Kernel-14B. On the KernelBench Level-2 subset, 31.6% of the generated kernels achieve at least a 1.2x speedup over the Torch reference, surpassing Claude-4.5-Sonnet (26.7%) and GPT-5 (28.6%). When selecting the best candidate across all turns, this 1.2x speedup rate further increases to 47.8%. All resources, including environment, training code, models, and dataset, are included in https://www.github.com/hkust-nlp/KernelGYM.
--------------------------------------------------------------------------------------------------------
Automated Customization of LLMs for Enterprise Code Repositories Using Semantic Scopes
LLM-based programming assistants struggle with private codebases unseen during training, limiting their effectiveness for enterprise developers. This research presents automated customization using semantic scopes for code completion tasks, evaluating Retrieval-Augmented Generation and supervised fine-tuning strategies on real enterprise repositories. The semantic scope mechanism helps models learn repository-specific patterns, enabling moderately-sized customized models to significantly outperform larger uncustomized alternatives. Analysis extends to public benchmarks, demonstrating generalizability. Applications include improved developer productivity through context-aware code completion, reduced onboarding time for new codebases, and efficient enterprise AI deployment without massive model scaling, making specialized assistance accessible.
Authors: Ulrich Finkler, Irene Manotas, Wei Zhang, Geert Janssen, Octavian Popescu, Shyam Ramji
Link: https://arxiv.org/abs/2602.05780v1
Date: 2026-02-d
Summary:
Code completion (CC) is a task frequently used by developers when working in collaboration with LLM-based programming assistants. Despite the increased performance of LLMs on public benchmarks, out of the box LLMs still have a hard time generating code that aligns with a private code repository not previously seen by the model's training data. Customizing code LLMs to a private repository provides a way to improve the model performance. In this paper we present our approach for automated LLM customization based on semantic scopes in the code. We evaluate LLMs on real industry cases with two private enterprise code repositories with two customization strategies: Retrieval-Augmented Generation (RAG) and supervised Fine-Tuning (FT). Our mechanism for ingesting the repository's data and formulating the training data pairs with semantic scopes helps models to learn the underlying patterns specific to the repository, providing more precise code to developers and helping to boost their productivity. The code completions of moderately sized customized models can be significantly better than those of uncustomized models of much larger capacity. We also include an analysis of customization on two public benchmarks and present opportunities for future work.
--------------------------------------------------------------------------------------------------------
Nonlinearity as Rank: Generative Low-Rank Adapter with Radial Basis Functions
Standard LoRA requires explicit storage of basis vectors, causing substantial parameter growth when increasing model capacity. This research reveals significant parameter redundancy in these vectors, proposing GenLoRA which generates basis vectors using lightweight radial basis functions instead of explicit storage. Each RBF requires fewer parameters than stored vectors, enabling higher effective LoRA ranks within smaller parameter budgets. Extensive experiments across datasets and architectures demonstrate superior fine-tuning performance through improved parameter efficiency. Applications include more efficient model adaptation, reduced memory footprints for edge deployment, and enabling complex fine-tuning on resource-constrained hardware while maintaining or improving performance compared to standard LoRA approaches.
Authors: Yihao Ouyang, Shiwei Li, Haozhao Wang, Xiandi Luo, Zhuoqi Hu, Yuetong Song, Qiyu Qin, Yichen Li, Ruixuan Li
Link: https://arxiv.org/abs/2602.05709v1
Date: 2026-02-d
Summary:
Low-rank adaptation (LoRA) approximates the update of a pretrained weight matrix using the product of two low-rank matrices. However, standard LoRA follows an explicit-rank paradigm, where increasing model capacity requires adding more rows or columns (i.e., basis vectors) to the low-rank matrices, leading to substantial parameter growth. In this paper, we find that these basis vectors exhibit significant parameter redundancy and can be compactly represented by lightweight nonlinear functions. Therefore, we propose Generative Low-Rank Adapter (GenLoRA), which replaces explicit basis vector storage with nonlinear basis vector generation. Specifically, GenLoRA maintains a latent vector for each low-rank matrix and employs a set of lightweight radial basis functions (RBFs) to synthesize the basis vectors. Each RBF requires far fewer parameters than an explicit basis vector, enabling higher parameter efficiency in GenLoRA. Extensive experiments across multiple datasets and architectures show that GenLoRA attains higher effective LoRA ranks under smaller parameter budgets, resulting in superior fine-tuning performance. The code is available at https://anonymous.4open.science/r/GenLoRA-1519.
--------------------------------------------------------------------------------------------------------
Enhancing Personality Recognition by Comparing the Predictive Power of Traits, Facets, and Nuances
Personality recognition models typically target broad trait scores, but similar scores can manifest through diverse context-dependent behaviors, challenging generalization. This research explores hierarchical Big-Five levels—traits, facets, and nuances—for personality recognition from audiovisual interactions using transformer-based models with cross-modal and cross-subject attention. Results from the UDIVA dataset show nuance-level models consistently outperform facet and trait models, reducing mean squared error by up to 74%. Applications include improved human-computer interaction, more accurate personality assessment tools, better mental health diagnostics, personalized education systems, and recruitment technologies requiring fine-grained behavioral understanding beyond superficial trait categorizations.
Authors: Amir Ansari, Jana Subirana, Bruna Silva, Sergio Escalera, David Gallardo-Pujol, Cristina Palmero
Link: https://arxiv.org/abs/2602.05650v1
Date: 2026-02-d
Summary:
Personality is a complex, hierarchical construct typically assessed through item-level questionnaires aggregated into broad trait scores. Personality recognition models aim to infer personality traits from different sources of behavioral data. However, reliance on broad trait scores as ground truth, combined with limited training data, poses challenges for generalization, as similar trait scores can manifest through diverse, context dependent behaviors. In this work, we explore the predictive impact of the more granular hierarchical levels of the Big-Five Personality Model, facets and nuances, to enhance personality recognition from audiovisual interaction data. Using the UDIVA v0.5 dataset, we trained a transformer-based model including cross-modal (audiovisual) and cross-subject (dyad-aware) attention mechanisms. Results show that nuance-level models consistently outperform facet and trait-level models, reducing mean squared error by up to 74% across interaction scenarios.
--------------------------------------------------------------------------------------------------------
A Unified Multimodal Framework for Dataset Construction and Model-Based Diagnosis of Ameloblastoma
Ameloblastoma diagnosis requires multimodal medical data, but existing datasets lack comprehensive coverage and format consistency for direct AI training. This research presents a curated multimodal dataset integrating radiological, histopathological, and clinical images with structured data from case reports using NLP extraction. A multimodal deep learning model accepts clinical inputs (complaint, age, gender) for personalized inference, classifying variants, assessing recurrence risk, and supporting surgical planning. Performance improvements include variant classification accuracy increasing from 46.2% to 65.9% and abnormal tissue detection F1-score reaching 90.3%. Applications include enhanced diagnostic accuracy, personalized treatment planning, and improved patient outcomes in maxillofacial pathology.
Authors: Ajo Babu George, Anna Mariam John, Athul Anoop, Balu Bhasuran
Link: https://arxiv.org/abs/2602.05515v1
Date: 2026-02-d
Summary:
Artificial intelligence (AI)-enabled diagnostics in maxillofacial pathology require structured, high-quality multimodal datasets. However, existing resources provide limited ameloblastoma coverage and lack the format consistency needed for direct model training. We present a newly curated multimodal dataset specifically focused on ameloblastoma, integrating annotated radiological, histopathological, and intraoral clinical images with structured data derived from case reports. Natural language processing techniques were employed to extract clinically relevant features from textual reports, while image data underwent domain specific preprocessing and augmentation. Using this dataset, a multimodal deep learning model was developed to classify ameloblastoma variants, assess behavioral patterns such as recurrence risk, and support surgical planning. The model is designed to accept clinical inputs such as presenting complaint, age, and gender during deployment to enhance personalized inference. Quantitative evaluation demonstrated substantial improvements; variant classification accuracy increased from 46.2 percent to 65.9 percent, and abnormal tissue detection F1-score improved from 43.0 percent to 90.3 percent. Benchmarked against resources like MultiCaRe, this work advances patient-specific decision support by providing both a robust dataset and an adaptable multimodal AI framework.
--------------------------------------------------------------------------------------------------------
6G development requires massive real-world channel datasets for AI and integrated sensing-communication systems, but traditional frequency-domain channel sounding suffers from inefficiency due to excessive frequency points. This research proposes Parabolic Frequency Sampling (PFS) with non-uniform distribution eliminating delay ambiguity while reducing sampling overhead by orders of magnitude, alongside likelihood-rectified SAGE algorithm for multipath component extraction from sparse measurements. Validation at 280-300 GHz demonstrates 50× faster measurement, 98% data reduction, and 99.96% computational complexity reduction while maintaining accuracy. Applications include efficient 6G channel modeling, enabling AI-native system development through massive dataset construction, and advancing wireless communication research with practical large-scale measurement capabilities.
Authors: Yi Chen, Li Ming, Chong Han
Link: https://arxiv.org/abs/2602.05405v1
Date: 2026-02-d
Summary:
Realizing the 6G vision of artificial intelligence (AI) and integrated sensing and communication (ISAC) critically requires large-scale real-world channel datasets for channel modeling and data-driven AI models. However, traditional frequency-domain channel sounding methods suffer from low efficiency due to a prohibitive number of frequency points to avoid delay ambiguity. This paper proposes a novel channel sounding framework involving sparse nonuniform sampling along with a likelihood-rectified space-alternating generalized expectation-maximization (LR-SAGE) algorithm for multipath component extraction. This framework enables the acquisition of channel datasets that are tens or even hundreds of times larger within the same channel measurement duration, thereby providing the massive data required to harness the full potential of AI scaling laws. Specifically, we propose a Parabolic Frequency Sampling (PFS) strategy that non-uniformly distributes frequency points, effectively eliminating delay ambiguity while reducing sampling overhead by orders of magnitude. To efficiently extract multipath components (MPCs) from the channel data measured by PFS, we develop a LR-SAGE algorithm, rectifying the likelihood distortion caused by nonuniform sampling and molecular absorption effect. Simulation results and experimental validation at 280--300~GHz confirm that the proposed PFS and LR-SAGE algorithm not only achieve 50$\times$ faster measurement, a 98\% reduction in data volume and a 99.96\% reduction in post-processing computational complexity, but also successfully captures MPCs and channel characteristics consistent with traditional exhaustive measurements, demonstrating its potential as a fundamental enabler for constructing the massive ISAC datasets required by AI-native 6G systems.
--------------------------------------------------------------------------------------------------------
Predictive Synthesis under Sporadic Participation: Evidence from Inflation Density Surveys
Professional forecaster surveys suffer from irregular participation—forecasters enter, exit, and skip rounds—creating panel composition changes that generate artificial jumps in aggregated predictions. Standard approaches like equal-weight pooling or imputation conflate participation effects with genuine economic information. This research develops Bayesian updating rules maintaining well-defined latent predictive states for absent forecasters without renormalization or imputation, using implied conditional panel structure. Applied to ECB's Survey of Professional Forecasters, the method improves predictive accuracy versus equal-weight benchmarks, delivering smoother, better-calibrated inflation density forecasts particularly during high turnover. Applications include improved central bank forecasting, robust survey aggregation methodology, and interpretable forecaster performance assessment.
Authors: Matthew C. Johnson, Matteo Luciani, Minzhengxiong Zhang, Kenichiro McAlinn
Link: https://arxiv.org/abs/2602.05226v1
Date: 2026-02-d
Summary:
Central banks rely on density forecasts from professional surveys to assess inflation risks and communicate uncertainty. A central challenge in using these surveys is irregular participation: forecasters enter and exit, skip rounds, and reappear after long gaps. In the European Central Bank's Survey of Professional Forecasters, turnover and missingness vary substantially over time, causing the set of submitted predictions to change from quarter to quarter. Standard aggregation rules -- such as equal-weight pooling, renormalization after dropping missing forecasters, or ad hoc imputation -- can generate artificial jumps in combined predictions driven by panel composition rather than economic information, complicating real-time interpretation and obscuring forecaster performance. We develop coherent Bayesian updating rules for forecast combination under sporadic participation that maintain a well-defined latent predictive state for each forecaster even when their forecast is unobserved. Rather than relying on renormalization or imputation, the combined predictive distribution is updated through the implied conditional structure of the panel. This approach isolates genuine performance differences from mechanical participation effects and yields interpretable dynamics in forecaster influence. In the ECB survey, it improves predictive accuracy relative to equal-weight benchmarks and delivers smoother and better-calibrated inflation density forecasts, particularly during periods of high turnover.
--------------------------------------------------------------------------------------------------------
Path Sampling for Rare Events Boosted by Machine Learning
Understanding molecular mechanisms requires efficient sampling of rare transition events, a challenging computational problem. This commentary analyzes AIMMD (Artificial Intelligence for Molecular Mechanism Discovery), which integrates machine learning with transition path sampling for enhanced efficiency. AIMMD enables on-the-fly committor probability estimation while deriving interpretable reaction coordinates, offering robust mechanistic insights for complex molecular processes. The analysis discusses the framework's core principles, recent extensions, and potential impact alongside limitations. Applications include drug discovery through understanding binding mechanisms, materials science for phase transitions, enzyme catalysis research, and advancing computational chemistry methods by combining physics-based sampling with machine learning intelligence.
Authors: Porhouy Minh, Sapna Sarupria
Link: https://arxiv.org/abs/2602.05167v1
Date: 2026-02-d
Summary:
The study by Jung et al. (Jung H, Covino R, Arjun A, et al., Nat Comput Sci. 3:334-345 (2023)) introduced Artificial Intelligence for Molecular Mechanism Discovery (AIMMD), a novel sampling algorithm that integrates machine learning to enhance the efficiency of transition path sampling (TPS). By enabling on-the-fly estimation of the committor probability and simultaneously deriving a human-interpretable reaction coordinate, AIMMD offers a robust framework for elucidating the mechanistic pathways of complex molecular processes. This commentary provides a discussion and critical analysis of the core AIMMD framework, explores its recent extensions, and offers an assessment of the method's potential impact and limitations.
--------------------------------------------------------------------------------------------------------
Reliable Explanations or Random Noise? A Reliability Metric for XAI
Explainable AI methods like SHAP and Integrated Gradients are widely used in high-stakes domains, yet their reliability—whether explanations remain stable under realistic non-adversarial changes—remains largely unmeasured. This research introduces the Explanation Reliability Index (ERI), quantifying explanation stability under robustness, consistency, smoothness, and distributional shift axioms, with formal Lipschitz-type bounds and temporal stability guarantees. ERI-Bench systematically stress-tests methods across datasets, revealing widespread reliability failures under realistic deployment conditions. Applications include trustworthy AI system development, principled XAI method selection for healthcare/finance/energy systems, regulatory compliance support, and advancing explanation methodology beyond accuracy toward deployment-ready reliability.
Authors: Poushali Sengupta, Sabita Maharjan, Frank Eliassen, Shashi Raj Pandey, Yan Zhang
Link: https://arxiv.org/abs/2602.05082v1
Date: 2026-02-d
Summary:
In recent years, explaining decisions made by complex machine learning models has become essential in high-stakes domains such as energy systems, healthcare, finance, and autonomous systems. However, the reliability of these explanations, namely, whether they remain stable and consistent under realistic, non-adversarial changes, remains largely unmeasured. Widely used methods such as SHAP and Integrated Gradients (IG) are well-motivated by axiomatic notions of attribution, yet their explanations can vary substantially even under system-level conditions, including small input perturbations, correlated representations, and minor model updates. Such variability undermines explanation reliability, as reliable explanations should remain consistent across equivalent input representations and small, performance-preserving model changes. We introduce the Explanation Reliability Index (ERI), a family of metrics that quantifies explanation stability under four reliability axioms: robustness to small input perturbations, consistency under feature redundancy, smoothness across model evolution, and resilience to mild distributional shifts. For each axiom, we derive formal guarantees, including Lipschitz-type bounds and temporal stability results. We further propose ERI-T, a dedicated measure of temporal reliability for sequential models, and introduce ERI-Bench, a benchmark designed to systematically stress-test explanation reliability across synthetic and real-world datasets. Experimental results reveal widespread reliability failures in popular explanation methods, showing that explanations can be unstable under realistic deployment conditions. By exposing and quantifying these instabilities, ERI enables principled assessment of explanation reliability and supports more trustworthy explainable AI (XAI) systems.
--------------------------------------------------------------------------------------------------------
AI-Based Detection of In-Treatment Changes from Prostate MR-Linac Images
MR-Linac systems acquire longitudinal images during radiotherapy, but whether these can detect subtle treatment-induced changes over short intervals (average 2 days) remains unexplored. This research develops deep learning models predicting temporal order of paired images, trained on first/last fractions then all pairs, achieving near-perfect performance (AUC 0.99) significantly outperforming radiologists. Saliency maps and ablation tests identify prostate, bladder, and pubic symphysis as primary altered regions, with expert confirmation of radiation-induced changes. Applications include advanced treatment monitoring beyond image guidance, early detection of tissue responses enabling adaptive therapy adjustments, improved understanding of radiotherapy effects, and demonstrating MR-Linac potential for AI-powered longitudinal analysis.
Authors: Seungbin Park, Peilin Wang, Ryan Pennell, Emily S. Weg, Himanshu Nagar, Timothy McClure, Mert R. Sabuncu, Daniel Margolis, Heejong Kim
Link: https://arxiv.org/abs/2602.04983v1
Date: 2026-02-d
Summary:
Purpose: To investigate whether routinely acquired longitudinal MR-Linac images can be leveraged to characterize treatment-induced changes during radiotherapy, particularly subtle inter-fraction changes over short intervals (average of 2 days). Materials and Methods: This retrospective study included a series of 0.35T MR-Linac images from 761 patients. An artificial intelligence (deep learning) model was used to characterize treatment-induced changes by predicting the temporal order of paired images. The model was first trained with the images from the first and the last fractions (F1-FL), then with all pairs (All-pairs). Model performance was assessed using quantitative metrics (accuracy and AUC), compared to a radiologist's performance, and qualitative analyses - the saliency map evaluation to investigate affected anatomical regions. Input ablation experiments were performed to identify the anatomical regions altered by radiotherapy. The radiologist conducted an additional task on partial images reconstructed by saliency map regions, reporting observations as well. Quantitative image analysis was conducted to investigate the results from the model and the radiologist. Results: The F1-FL model yielded near-perfect performance (AUC of 0.99), significantly outperforming the radiologist. The All-pairs model yielded an AUC of 0.97. This performance reflects therapy-induced changes, supported by the performance correlation to fraction intervals, ablation tests and expert's interpretation. Primary regions driving the predictions were prostate, bladder, and pubic symphysis. Conclusion: The model accurately predicts temporal order of MR-Linac fractions and detects radiation-induced changes over one or a few days, including prostate and adjacent organ alterations confirmed by experts. This underscores MR-Linac's potential for advanced image analysis beyond image guidance.
--------------------------------------------------------------------------------------------------------
Subliminal Effects in Your Data: A General Mechanism via Log-Linearity
Modern LLM training combines numerous algorithms and datasets to elicit specific behaviors, but recent experiments show datasets transmit signals invisible in individual datapoints—a conceptual challenge for dataset understanding. This research uncovers a general mechanism through which hidden subtexts arise via log-linearity in LLMs, introducing Logit-Linear-Selection (LLS) for selecting dataset subsets eliciting hidden effects. Applications demonstrate models trained on selected subsets exhibit specific preferences, respond in different languages absent from data, or adopt different personas—effects persisting across architectures. Applications include understanding dataset influence on model behavior, detecting unintended biases, improving data curation practices, and advancing fundamental understanding of LLM training dynamics.
Authors: Ishaq Aden-Ali, Noah Golowich, Allen Liu, Abhishek Shetty, Ankur Moitra, Nika Haghtalab
Link: https://arxiv.org/abs/2602.04863v1
Date: 2026-02-d
Summary:
Training modern large language models (LLMs) has become a veritable smorgasbord of algorithms and datasets designed to elicit particular behaviors, making it critical to develop techniques to understand the effects of datasets on the model's properties. This is exacerbated by recent experiments that show datasets can transmit signals that are not directly observable from individual datapoints, posing a conceptual challenge for dataset-centric understandings of LLM training and suggesting a missing fundamental account of such phenomena. Towards understanding such effects, inspired by recent work on the linear structure of LLMs, we uncover a general mechanism through which hidden subtexts can arise in generic datasets. We introduce Logit-Linear-Selection (LLS), a method that prescribes how to select subsets of a generic preference dataset to elicit a wide range of hidden effects. We apply LLS to discover subsets of real-world datasets so that models trained on them exhibit behaviors ranging from having specific preferences, to responding to prompts in a different language not present in the dataset, to taking on a different persona. Crucially, the effect persists for the selected subset, across models with varying architectures, supporting its generality and universality.
--------------------------------------------------------------------------------------------------------
Machine Learning's exponential growth brings significant Adversarial Machine Learning (AML) security challenges. This research conducted two studies: an industry professional survey revealing correlations between cybersecurity education and AML threat concern, and CTF challenges implementing NLP/Generative AI concepts demonstrating poisoning attacks, evaluated with Carnegie Mellon students. Results show CTF-based approaches effectively engage interest in AML threats. The research provides recommendations emphasizing integrated security education within ML curricula. Applications include improved cybersecurity training programs, raising industry awareness about ML vulnerabilities, developing educational resources bridging academia-industry gaps, and fostering security-conscious ML development practices essential as AI deployment accelerates.
Authors: Vishruti Kakkad, Paul Chung, Hanan Hibshi, Maverick Woo
Link: https://arxiv.org/abs/2602.04753v1
Date: 2026-02-d
Summary:
An exponential growth of Machine Learning and its Generative AI applications brings with it significant security challenges, often referred to as Adversarial Machine Learning (AML). In this paper, we conducted two comprehensive studies to explore the perspectives of industry professionals and students on different AML vulnerabilities and their educational strategies. In our first study, we conducted an online survey with professionals revealing a notable correlation between cybersecurity education and concern for AML threats. For our second study, we developed two CTF challenges that implement Natural Language Processing and Generative AI concepts and demonstrate a poisoning attack on the training data set. The effectiveness of these challenges was evaluated by surveying undergraduate and graduate students at Carnegie Mellon University, finding that a CTF-based approach effectively engages interest in AML threats. Based on the responses of the participants in our research, we provide detailed recommendations emphasizing the critical need for integrated security education within the ML curriculum.
--------------------------------------------------------------------------------------------------------
Large language models enable retooling software development processes, prompting massive research into AI-assisted engineering tools. This research introduces agentic AI solutions for two tasks: automatic test scenario generation from requirements using specialized worker agents in star topology with supervisor coordination, and document retrieval enabling search, question-answering, change tracking, and summarization across software engineering documents via dedicated LLM-based agents per use case. Demonstrations use real-world examples, showing practical viability. Applications include accelerated testing workflows, improved requirements traceability, enhanced documentation management, reduced manual engineering effort, and advancing toward more autonomous software development processes leveraging multi-agent AI architectures.
Authors: Marian Kica, Lukas Radosky, David Slivka, Karin Kubinova, Daniel Dovhun, Tomas Uhercik, Erik Bircak, Ivan Polasek
Link: https://arxiv.org/abs/2602.04726v1
Date: 2026-02-d
Summary:
The introduction of large language models ignited great retooling and rethinking of the software development models. The ensuing response of software engineering research yielded a massive body of tools and approaches. In this paper, we join the hassle by introducing agentic AI solutions for two tasks. First, we developed a solution for automatic test scenario generation from a detailed requirements description. This approach relies on specialized worker agents forming a star topology with the supervisor agent in the middle. We demonstrate its capabilities on a real-world example. Second, we developed an agentic AI solution for the document retrieval task in the context of software engineering documents. Our solution enables performing various use cases on a body of documents related to the development of a single software, including search, question answering, tracking changes, and large document summarization. In this case, each use case is handled by a dedicated LLM-based agent, which performs all subtasks related to the corresponding use case. We conclude by hinting at the future perspectives of our line of research.
--------------------------------------------------------------------------------------------------------