# Research Papers — 2026-03 · bundle page 221 of 243 > Full-text markdown of papers #11001–11050 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/221 [← Previous bundle](https://aipod.dev/papers/archive/2026-03/page/220.md) [Next bundle →](https://aipod.dev/papers/archive/2026-03/page/222.md) ### Verify to Amplify: Improving Reasoning via Learned Chain-of-Thought Verification Authors: Maria-Florina Balcan, Avrim Blum, Kiriaki Fragkia, Zhiyuan Li, Dravyansh Sharma Published: 2026-03-03 Categories: cs.LG arXiv: https://arxiv.org/abs/2603.03538v5 Large Language Models (LLMs) using chain-of-thought have demonstrated great potential for solving complex reasoning and planning tasks. Despite these advances, LLM-generated outputs remain susceptible to errors, making verification important for reliable reasoning systems. Learned verifiers can increase trust, enforce safety constraints, and ensure alignment with personal preferences, while also providing feedback to improve generation. This raises a central challenge: when learned verifiers are used to guide generation, the feedback loop between generator and verifier may induce a distribution shift. This is particularly salient for process reward models, a prominent class of learned verifiers that score or classify individual steps in a chain-of-thought reasoning trace. Motivated by this challenge, we propose a new online learning framework for chain-of-thought verifiers that, given a problem statement and a reasoning trace, check the correctness of each reasoning step given the preceding steps. Highlighting the asymmetric role of soundness errors (accepting an incorrect reasoning step) and completeness errors (flagging a correct step as wrong), we introduce novel notions of dimension that characterize their optimal tradeoff. We then show how our learned verifiers can boost the accuracy of a weak generator. Assuming that the generator can produce a correct next step with a small success probability, we show how to learn a strong generator with small error and abstention rates. Our results also allow learning from offline data when queries to an expert verifier can be simulated from a small set of correct reasoning traces. However, we establish a separation between our approach and learning from offline expert demonstrations: we show that learning from offline demonstrations cannot in general achieve the soundness-completeness guarantees produced by our interactive learning approach. --- ### Online Learnability of Chain-of-Thought Verifiers: Soundness and Completeness Trade-offs Authors: Maria-Florina Balcan, Avrim Blum, Kiriaki Fragkia, Zhiyuan Li, Dravyansh Sharma Published: 2026-03-03 Categories: cs.LG arXiv: https://arxiv.org/abs/2603.03538v3 Large Language Models (LLMs) with chain-of-thought generation have demonstrated great potential for solving complex reasoning and planning tasks. However, the output of current LLMs is not fully reliable and needs careful verification. Even if LLMs get more accurate over time, learned verifiers can help increase trust, enforce safety constraints, and ensure alignment with personal preferences. A major challenge in learning verifiers, however, especially when their output will be used by the generator to improve its reasoning, is that the feedback loop between generator and verifier may produce substantial distribution shift. Motivated by this challenge, we propose an online learning framework for learning chain-of-thought verifiers that, given a problem and a sequence of reasoning steps, check the correctness of the solution. Highlighting the asymmetric role of soundness errors (failure in catching errors in a reasoning trace) and completeness errors (flagging correct reasoning steps as wrong), we introduce novel extensions of the Littlestone dimension which tightly characterize the mistake bounds for learning a verifier in the realizable setting. We provide optimal algorithms for finding the Pareto-frontier (the smallest total number of mistakes given a budget of soundness mistakes) as well as for minimizing a linear combination of asymmetric costs. We further show how our learned verifiers can be used to boost the accuracy of a collection of weak generators, and enable generation of proofs beyond what they were initially trained on. With the mild assumption that one of the generators can generate the next reasoning step correctly with some minimal probability, we show how to learn a strong generator with small error and abstention rates. --- ### SafeCRS: Personalized Safety Alignment for LLM-Based Conversational Recommender Systems Authors: Haochang Hao, Yifan Xu, Xinzhuo Li, Yingqiang Ge, Lu Cheng Published: 2026-03-03 Categories: cs.CL, cs.AI, cs.IR arXiv: https://arxiv.org/abs/2603.03536v2 Current LLM-based conversational recommender systems (CRS) primarily optimize recommendation accuracy and user satisfaction. We identify an underexplored vulnerability in which recommendation outputs may negatively impact users by violating personalized safety constraints, when individualized safety sensitivities -- such as trauma triggers, self-harm history, or phobias -- are implicitly inferred from the conversation but not respected during recommendation. We formalize this challenge as personalized CRS safety and introduce SafeRec, a new benchmark dataset designed to systematically evaluate safety risks in LLM-based CRS under user-specific constraints. To further address this problem, we propose SafeCRS, a safety-aware training framework that integrates Safe Supervised Fine-Tuning (Safe-SFT) with Safe Group reward-Decoupled Normalization Policy Optimization (Safe-GDPO) to jointly optimize recommendation quality and personalized safety alignment. Extensive experiments on SafeRec demonstrate that SafeCRS reduces safety violation rates by up to 96.5% relative to the strongest recommendation-quality baseline while maintaining competitive recommendation quality. Warning: This paper contains potentially harmful and offensive content. --- ### Trade-offs in Ensembling, Merging and Routing Among Parameter-Efficient Experts Authors: Sanae Lotfi, Lucas Caccia, Alessandro Sordoni, Jordan T. Ash, Miroslav Dudik Published: 2026-03-03 Categories: cs.LG arXiv: https://arxiv.org/abs/2603.03535v1 While large language models (LLMs) fine-tuned with lightweight adapters achieve strong performance across diverse tasks, their performance on individual tasks depends on the fine-tuning strategy. Fusing independently trained models with different strengths has shown promise for multi-task learning through three main strategies: ensembling, which combines outputs from independent models; merging, which fuses model weights via parameter averaging; and routing, which integrates models in an input-dependent fashion. However, many design decisions in these approaches remain understudied, and the relative benefits of more sophisticated ensembling, merging and routing techniques are not fully understood. We empirically evaluate their trade-offs, addressing two key questions: What are the advantages of going beyond uniform ensembling or merging? And does the flexibility of routing justify its complexity? Our findings indicate that non-uniform ensembling and merging improve performance, but routing offers even greater gains. To mitigate the computational cost of routing, we analyze expert selection techniques, showing that clustering and greedy subset selection can maintain reasonable performance with minimal overhead. These insights advance our understanding of model fusion for multi-task learning. --- ### Role-Aware Conditional Inference for Spatiotemporal Ecosystem Carbon Flux Prediction Authors: Yiming Sun, Runlong Yu, Rongchao Dong, Shuo Chen, Licheng Liu, Youmi Oh, Qianlai Zhuang, Yiqun Xie, Xiaowei Jia Published: 2026-03-03 Categories: cs.LG, cs.AI arXiv: https://arxiv.org/abs/2603.03531v1 Accurate prediction of terrestrial ecosystem carbon fluxes (e.g., CO$_2$, GPP, and CH$_4$) is essential for understanding the global carbon cycle and managing its impacts. However, prediction remains challenging due to strong spatiotemporal heterogeneity: ecosystem flux responses are constrained by slowly varying regime conditions, while short-term fluctuations are driven by high-frequency dynamic forcings. Most existing learning-based approaches treat environmental covariates as a homogeneous input space, implicitly assuming a global response function, which leads to brittle generalization across heterogeneous ecosystems. In this work, we propose Role-Aware Conditional Inference (RACI), a process-informed learning framework that formulates ecosystem flux prediction as a conditional inference problem. RACI employs hierarchical temporal encoding to disentangle slow regime conditioners from fast dynamic drivers, and incorporates role-aware spatial retrieval that supplies functionally similar and geographically local context for each role. By explicitly modeling these distinct functional roles, RACI enables a model to adapt its predictions across diverse environmental regimes without training separate local models or relying on fixed spatial structures. We evaluate RACI across multiple ecosystem types (wetlands and agricultural systems), carbon fluxes (CO$_2$, GPP, CH$_4$), and data sources, including both process-based simulations and observational measurements. Across all settings, RACI consistently outperforms competitive spatiotemporal baselines, demonstrating improved accuracy and spatial generalization under pronounced environmental heterogeneity. --- ### Directional Neural Collapse Explains Few-Shot Transfer in Self-Supervised Learning Authors: Achleshwar Luthra, Yash Salunkhe, Tomer Galanti Published: 2026-03-03 Categories: cs.LG, cs.AI arXiv: https://arxiv.org/abs/2603.03530v1 Frozen self-supervised representations often transfer well with only a few labels across many semantic tasks. We argue that a single geometric quantity, \emph{directional} CDNV (decision-axis variance), sits at the core of two favorable behaviors: strong few-shot transfer within a task, and low interference across many tasks. We show that both emerge when variability \emph{along} class-separating directions is small. First, we prove sharp non-asymptotic multiclass generalization bounds for downstream classification whose leading term is the directional CDNV. The bounds include finite-shot corrections that cleanly separate intrinsic decision-axis variability from centroid-estimation error. Second, we link decision-axis collapse to multitask geometry: for independent balanced labelings, small directional CDNV across tasks forces the corresponding decision axes to be nearly orthogonal, helping a single representation support many tasks with minimal interference. Empirically, across SSL objectives, directional CDNV collapses during pretraining even when classical CDNV remains large, and our bounds closely track few-shot error at practical shot sizes. Additionally, on synthetic multitask data, we verify that SSL learns representations whose induced decision axes are nearly orthogonal. The code and project page of the paper are available at (\href{https://dlfundamentals.github.io/directional-neural-collapse/}{project page}). --- ### mlx-snn: Spiking Neural Networks on Apple Silicon via MLX Authors: Jiahao Qin Published: 2026-03-03 Categories: cs.LG, cs.AI, cs.NE arXiv: https://arxiv.org/abs/2603.03529v1 We introduce mlx-snn, the first spiking neural network (SNN) library built natively on Apple's MLX framework. As SNN research grows rapidly, all major libraries -- snnTorch, Norse, SpikingJelly, Lava -- target PyTorch or custom backends, leaving Apple Silicon users without a native option. mlx-snn provides six neuron models (LIF, IF, Izhikevich, Adaptive LIF, Synaptic, Alpha), four surrogate gradient functions, four spike encoding methods (including an EEG-specific encoder), and a complete backpropagation-through-time training pipeline. The library leverages MLX's unified memory architecture, lazy evaluation, and composable function transforms (mx.grad, mx.compile) to enable efficient SNN research on Apple Silicon hardware. We validate mlx-snn on MNIST digit classification across five hyperparameter configurations and three backends, achieving up to 97.28% accuracy with 2.0--2.5 times faster training and 3--10 times lower GPU memory than snnTorch on the same M3 Max hardware. mlx-snn is open-source under the MIT license and available on PyPI. https://github.com/D-ST-Sword/mlx-snn --- ### Logit-Level Uncertainty Quantification in Vision-Language Models for Histopathology Image Analysis Authors: Betul Yurdem, Ferhat Ozgur Catak, Murat Kuzlu, Mehmet Kemal Gullu Published: 2026-03-03 Categories: cs.LG arXiv: https://arxiv.org/abs/2603.03527v1 Vision-Language Models (VLMs) with their multimodal capabilities have demonstrated remarkable success in almost all domains, including education, transportation, healthcare, energy, finance, law, and retail. Nevertheless, the utilization of VLMs in healthcare applications raises crucial concerns due to the sensitivity of large-scale medical data and the trustworthiness of these models (reliability, transparency, and security). This study proposes a logit-level uncertainty quantification (UQ) framework for histopathology image analysis using VLMs to deal with these concerns. UQ is evaluated for three VLMs using metrics derived from temperature-controlled output logits. The proposed framework demonstrates a critical separation in uncertainty behavior. While VLMs show high stochastic sensitivity (cosine similarity (CS) $<0.71$ and $<0.84$, Jensen-Shannon divergence (JS) $<0.57$ and $<0.38$, and Kullback-Leibler divergence (KL) $<0.55$ and $<0.35$, respectively for mean values of VILA-M3-8B and LLaVA-Med v1.5), near-maximal temperature impacts ($Δ_T \approx 1.00$), and displaying abrupt uncertainty transitions, particularly for complex diagnostic prompts. In contrast, the pathology-specific PRISM model maintains near-deterministic behavior (mean CS $>0.90$, JS $<0.10$, KL $<0.09$) and significantly minimal temperature effects across all prompt complexities. These findings emphasize the importance of logit-level uncertainty quantification to evaluate trustworthiness in histopathology applications utilizing VLMs. --- ### Multi-Agent Influence Diagrams to Hybrid Threat Modeling Authors: Maarten C. Vonk, Anna V. Kononova, Thomas Bäck, Tim Sweijs Published: 2026-03-03 Categories: cs.MA, cs.AI, econ.EM arXiv: https://arxiv.org/abs/2603.03526v1 Western governments have adopted an assortment of counter-hybrid threat measures to defend against hostile actions below the conventional military threshold. The impact of these measures is unclear because of the ambiguity of hybrid threats, their cross-domain nature, and uncertainty about how countermeasures shape adversarial behavior. This paper offers a novel approach to clarifying this impact by unifying previously bifurcating hybrid threat modeling methods through a (multi-agent) influence diagram framework. The model balances the costs of countermeasures, their ability to dissuade the adversary from executing hybrid threats, and their potential to mitigate the impact of hybrid threats. We run 1000 semi-synthetic variants of a real-world-inspired scenario simulating the strategic interaction between attacking agent A and defending agent B over a cyber attack on critical infrastructure to explore the effectiveness of a set of five different counter-hybrid threat measures. Counter-hybrid measures range from strengthening resilience and denial of the adversary's ability to execute a hybrid threat to dissuasion through the threat of punishment. Our analysis primarily evaluates the overarching characteristics of counter-hybrid threat measures. This approach allows us to generalize the effectiveness of these measures and examine parameter impact sensitivity. In addition, we discuss policy relevance and outline future research avenues. --- ### Test-Time Meta-Adaptation with Self-Synthesis Authors: Zeyneb N. Kaya, Nick Rui Published: 2026-03-03 Categories: cs.LG, cs.AI arXiv: https://arxiv.org/abs/2603.03524v2 As strong general reasoners, large language models (LLMs) encounter diverse domains and tasks, where the ability to adapt and self-improve at test time is valuable. We introduce MASS, a meta-learning framework that enables LLMs to self-adapt by generating problem-specific synthetic training data and performing targeted self-updates optimized for downstream performance at inference time. We train this behavior end-to-end via bilevel optimization: an inner loop adapts on self-generated examples while an outer loop meta-learns data-attribution signals and rewards post-update task performance. The synthetic data is optimized with scalable meta-gradients, backpropagating the downstream loss through the inner updates to reward useful generations. Experiments on mathematical reasoning show that MASS learns to synthesize per-instance curricula that yield effective, data-efficient test-time adaptation. --- ### Q-Measure-Learning for Continuous State RL: Efficient Implementation and Convergence Authors: Shengbo Wang Published: 2026-03-03 Categories: cs.LG, math.OC arXiv: https://arxiv.org/abs/2603.03523v1 We study reinforcement learning in infinite-horizon discounted Markov decision processes with continuous state spaces, where data are generated online from a single trajectory under a Markovian behavior policy. To avoid maintaining an infinite-dimensional, function-valued estimate, we propose the novel Q-Measure-Learning, which learns a signed empirical measure supported on visited state-action pairs and reconstructs an action-value estimate via kernel integration. The method jointly estimates the stationary distribution of the behavior chain and the Q-measure through coupled stochastic approximation, leading to an efficient weight-based implementation with $O(n)$ memory and $O(n)$ computation cost per iteration. Under uniform ergodicity of the behavior chain, we prove almost sure sup-norm convergence of the induced Q-function to the fixed point of a kernel-smoothed Bellman operator. We also bound the approximation error between this limit and the optimal $Q^*$ as a function of the kernel bandwidth. To assess the performance of our proposed algorithm, we conduct RL experiments in a two-item inventory control setting. --- ### MMAI Gym for Science: Training Liquid Foundation Models for Drug Discovery Authors: Maksim Kuznetsov, Zulfat Miftahutdinov, Rim Shayakhmetov, Mikolaj Mizera, Roman Schutski, Bogdan Zagribelnyy, Ivan Ilin, Nikita Bondarev, Thomas MacDougall, Mathieu Reymond, Mihir Bafna, Kaeli Kaymak-Loveless et al. Published: 2026-03-03 Categories: cs.LG, cs.AI, cs.CL arXiv: https://arxiv.org/abs/2603.03517v2 General-purpose large language models (LLMs) that rely on in-context learning do not reliably deliver the scientific understanding and performance required for drug discovery tasks. Simply increasing model size or introducing reasoning tokens does not yield significant performance gains. To address this gap, we introduce the MMAI Gym for Science, a one-stop shop molecular data formats and modalities as well as task-specific reasoning, training, and benchmarking recipes designed to teach foundation models the 'language of molecules' in order to solve practical drug discovery problems. We use MMAI Gym to train an efficient Liquid Foundation Model (LFM) for these applications, demonstrating that smaller, purpose-trained foundation models can outperform substantially larger general-purpose or specialist models on molecular benchmarks. Across essential drug discovery tasks - including molecular optimization, ADMET property prediction, retrosynthesis, drug-target activity prediction, and functional group reasoning - the resulting model achieves near specialist-level performance and, in the majority of settings, surpasses larger models, while remaining more efficient and broadly applicable in the domain. --- ### MMAI Gym for Science: Training Liquid Foundation Models for Drug Discovery Authors: Maksim Kuznetsov, Zulfat Miftahutdinov, Rim Shayakhmetov, Mikolaj Mizera, Roman Schutski, Bogdan Zagribelnyy, Ivan Ilin, Nikita Bondarev, Thomas MacDougall, Mathieu Reymond, Mihir Bafna, Kaeli Kaymak-Loveless et al. Published: 2026-03-03 Categories: cs.LG, cs.AI, cs.CL arXiv: https://arxiv.org/abs/2603.03517v1 General-purpose large language models (LLMs) that rely on in-context learning do not reliably deliver the scientific understanding and performance required for drug discovery tasks. Simply increasing model size or introducing reasoning tokens does not yield significant performance gains. To address this gap, we introduce the MMAI Gym for Science, a one-stop shop molecular data formats and modalities as well as task-specific reasoning, training, and benchmarking recipes designed to teach foundation models the 'language of molecules' in order to solve practical drug discovery problems. We use MMAI Gym to train an efficient Liquid Foundation Model (LFM) for these applications, demonstrating that smaller, purpose-trained foundation models can outperform substantially larger general-purpose or specialist models on molecular benchmarks. Across essential drug discovery tasks - including molecular optimization, ADMET property prediction, retrosynthesis, drug-target activity prediction, and functional group reasoning - the resulting model achieves near specialist-level performance and, in the majority of settings, surpasses larger models, while remaining more efficient and broadly applicable in the domain. --- ### The Controllability Trap: A Governance Framework for Military AI Agents Authors: Subramanyam Sahoo Published: 2026-03-03 Categories: cs.CY, cs.AI arXiv: https://arxiv.org/abs/2603.03515v1 Agentic AI systems - capable of goal interpretation, world modeling, planning, tool use, long-horizon operation, and autonomous coordination - introduce distinct control failures not addressed by existing safety frameworks. We identify six agentic governance failures tied to these capabilities and show how they erode meaningful human control in military settings. We propose the Agentic Military AI Governance Framework (AMAGF), a measurable architecture structured around three pillars: Preventive Governance (reducing failure likelihood), Detective Governance (real-time detection of control degradation), and Corrective Governance (restoring or safely degrading operations). Its core mechanism, the Control Quality Score (CQS), is a composite real-time metric quantifying human control and enabling graduated responses as control weakens. For each failure type, we define concrete mechanisms, assign responsibilities across five institutional actors, and formalize evaluation metrics. A worked operational scenario illustrates implementation, and we situate the framework within established agent safety literature. We argue that governance must move from a binary conception of control to a continuous model in which control quality is actively measured and managed throughout the operational lifecycle. --- ### Baseline Performance of AI Tools in Classifying Cognitive Demand of Mathematical Tasks Authors: Danielle S. Fox, Brenda L. Robles, Elizabeth DiPietro Brovey, Christian D. Schunn Published: 2026-03-03 Categories: cs.CY, cs.AI arXiv: https://arxiv.org/abs/2603.03512v1 Teachers face increasing demands on their time, particularly in adapting mathematics curricula to meet individual student needs while maintaining cognitive rigor. This study evaluates whether AI tools can accurately classify the cognitive demand of mathematical tasks, which is important for creating or adapting tasks that support student learning. We tested eleven AI tools: six general-purpose (ChatGPT, Claude, DeepSeek, Gemini, Grok, Perplexity) and five education-specific (Brisk, Coteach AI, Khanmigo, Magic School, School$.$AI), on their ability to categorize mathematics tasks across four levels of cognitive demand using a research-based framework. The goal was to approximate the performance teachers will achieve with straightforward prompts. On average, AI tools accurately classified cognitive demand in only 63% of cases. Education-specific tools were not more accurate than general-purpose tools, and no tool exceeded 83% accuracy. All tools struggled with tasks at the extremes of cognitive demand (Memorization and Doing Mathematics), exhibiting a systematic bias toward middle-category levels (Procedures with/without Connections). The tools often gave plausible-sounding explanations likely to be persuasive to novice teachers. Error analysis of AI tools' misclassification of the broad level of cognitive demand (high vs. low) revealed that tools consistently overweighted surface textual features over underlying cognitive processes. Further, AI tools showed weaknesses in reasoning about factors that make tasks higher vs. lower cognitive demand. Errors stemmed not from ignoring relevant dimensions, but from incorrectly reasoning about multiple task aspects. These findings carry implications for AI integration into teacher planning workflows and highlight the need for improved prompt engineering and tool development for educational applications. --- ### Orbital Transformers for Predicting Wavefunctions in Time-Dependent Density Functional Theory Authors: Xuan Zhang, Haiyang Yu, Chengdong Wang, Jacob Helwig, Shuiwang Ji, Xiaofeng Qian Published: 2026-03-03 Categories: cs.LG, cond-mat.mtrl-sci, physics.chem-ph arXiv: https://arxiv.org/abs/2603.03511v2 We aim to learn wavefunctions simulated by time-dependent density functional theory (TDDFT), which can be efficiently represented as linear combination coefficients of atomic orbitals. In real-time TDDFT, the electronic wavefunctions of a molecule evolve over time in response to an external excitation, enabling first-principles predictions of physical properties such as optical absorption, electron dynamics, and high-order response. However, conventional real-time TDDFT relies on time-consuming propagation of all occupied states with fine time steps. In this work, we propose OrbEvo, which is based on an equivariant graph transformer architecture and learns to evolve the full electronic wavefunction coefficients across time steps. First, to account for external field, we design an equivariant conditioning to encode both strength and direction of external electric field and break the symmetry from SO(3) to SO(2). Furthermore, we design two OrbEvo models, OrbEvo-WF and OrbEvo-DM, using wavefunction pooling and density matrix as interaction method, respectively. Motivated by the central role of the density functional in TDDFT, OrbEvo-DM encodes the density matrix aggregated from all occupied electronic states into feature vectors via tensor contraction, providing a more intuitive approach to learn the time evolution operator. We adopt a training strategy specifically tailored to limit the error accumulation of time-dependent wavefunctions over autoregressive rollout. To evaluate our approach, we generate TDDFT datasets consisting of 5,000 different molecules in the QM9 dataset and 1,500 molecular configurations of the malonaldehyde molecule in the MD17 dataset. Results show that our OrbEvo model accurately captures quantum dynamics of excited states under external field, including time-dependent wavefunctions, time-dependent dipole moment, and optical absorption spectra. --- ### A theoretical model of dynamical grammatical gender shifting based on set-valued set function Authors: Mohamed El Idrissi Published: 2026-03-03 Categories: cs.CL, cs.AI arXiv: https://arxiv.org/abs/2603.03510v2 This study investigates the diverse characteristics of nouns, focusing on both semantic (e.g., countable/uncountable) and morphosyntactic (e.g., masculine/feminine) distinctions. We explore inter-word variations for gender markers in noun morphology. Grammatical gender shift is a widespread phenomenon in languages around the world.The aim is to uncover the underlying patterns governing the variation of lexemes. To this end, we propose a new computational component dedicated to pairing items with morphological templates (e.g., the result of a generated item-template pair: (funas, $\{N, +SG, -PL, -M, +F, -COL, +SING\}$), with its spell-out form: $ð$a-funast 'cow'). This process is formally represented by the Template-Based and Modular Cognitive model. This proposed model, defined by a set-valued set function $h : \mathscr{P}(M) \rightarrow \mathscr{P}(M)$, predicts the nonlinear dynamic mapping of lexical items onto morphological templates. By applying this formalism, we present a unified framework for understanding the complexities of morphological markings across languages. Through empirical observations, we demonstrate how these shifts, as well as non-gender shifts, arise during lexical changes, especially in Riffian. Our model posits that these variant markings emerge due to template shifts occurring during word and meaning formation. This study achieves two primary objectives. First, on the formal side, we prove the model's representational completeness in learning and prediction. Second, on the linguistic side, we challenge and broaden the conventional view of word formation by formally demonstrating that conversion is applicable to noun-to-noun derivation. This data-driven mathematical model not only contributes to a deeper understanding of morphosyntactic variation but also offers potential applications in other fields requiring precise modelling of linguistic patterns. --- ### A theoretical model of dynamical grammatical gender shifting based on set-valued set function Authors: Mohamed El Idrissi Published: 2026-03-03 Categories: cs.CL, cs.AI arXiv: https://arxiv.org/abs/2603.03510v1 This study investigates the diverse characteristics of nouns, focusing on both semantic (e.g., countable/uncountable) and morphosyntactic (e.g., masculine/feminine) distinctions. We explore inter-word variations for gender markers in noun morphology. Grammatical gender shift is a widespread phenomenon in languages around the world. The aim is to uncover through a formal model the underlying patterns governing the variation of lexemes. To this end, we propose a new computational component dedicated to pairing items with morphological templates (e.g., the result of a generated item-template pair: (funas, $\{N, +SG, -PL, -M, +F, -COL, +SING\}$), with its spell-out form: $ð$a-funast 'cow'). This process is formally represented by the Template-Based and Modular Cognitive model. This proposed model, defined by a set-valued set function $h : \mathscr{P}(M) \rightarrow \mathscr{P}(M)$, predicts the nonlinear dynamic mapping of lexical items onto morphological templates. By applying this formalism, we present a unified framework for understanding the complexities of morphological markings across languages. Through empirical observations, we demonstrate how these shifts, as well as non-gender shifts, arise during lexical changes, especially in Riffian. Our model posits that these variant markings emerge due to template shifts occurring during word and meaning's formation. By formally demonstrating that conversion is applicable to noun-to-noun derivation, we challenge and broaden the conventional view of word formation. This mathematical model not only contributes to a deeper understanding of morphosyntactic variation but also offers potential applications in other fields requiring precise modelling of linguistic patterns. --- ### Raising Bars, Not Parameters: LilMoo Compact Language Model for Hindi Authors: Shiza Fatimah, Aniket Sen, Sophia Falk, Florian Mai, Lucie Flek, Nicholas Kluge Corrêa Published: 2026-03-03 Categories: cs.CL, cs.AI arXiv: https://arxiv.org/abs/2603.03508v1 The dominance of large multilingual foundation models has widened linguistic inequalities in Natural Language Processing (NLP), often leaving low-resource languages underrepresented. This paper introduces LilMoo, a 0.6-billion-parameter Hindi language model trained entirely from scratch to address this gap. Unlike prior Hindi models that rely on continual pretraining from opaque multilingual foundations, LilMoo is developed through a fully transparent and reproducible pipeline optimized for limited compute environments. We construct a high-quality Hindi corpus (GigaLekh) filtered through both heuristic and learned (LLM-as-a-judge) methods, complemented by bilingual augmentation with curated English data. Using this dataset, we explore various training recipes for small-scale language models. Across comprehensive evaluation suites, LilMoo consistently outperforms comparably sized multilingual baselines such as Qwen2.5-0.5B and Qwen3-0.6B, demonstrating that well-designed language-specific pretraining can rival large multilingual models at the sub-billion-parameter range. --- ### Solving adversarial examples requires solving exponential misalignment Authors: Alessandro Salvatore, Stanislav Fort, Surya Ganguli Published: 2026-03-03 Categories: cs.LG, cond-mat.dis-nn, q-bio.NC, stat.ML arXiv: https://arxiv.org/abs/2603.03507v2 Adversarial attacks - input perturbations imperceptible to humans that fool neural networks - remain both a persistent failure mode in machine learning, and a phenomenon with mysterious origins. To shed light, we define and analyze a network's perceptual manifold (PM) for a class concept as the space of all inputs confidently assigned to that class by the network. We find, strikingly, that the dimensionalities of neural network PMs are orders of magnitude higher than those of natural human concepts. Since volume typically grows exponentially with dimension, this suggests exponential misalignment between machines and humans, with exponentially many inputs confidently assigned to concepts by machines but not humans. Furthermore, this provides a natural geometric hypothesis for the origin of adversarial examples: because a network's PM fills such a large region of input space, any input will be very close to any class concept's PM. Our hypothesis thus suggests that adversarial robustness cannot be attained without dimensional alignment of machine and human PMs, and therefore makes strong predictions: both robust accuracy and distance to any PM should be negatively correlated with the PM dimension. We confirmed these predictions across 18 different networks of varying robust accuracy. Crucially, we find even the most robust networks are still exponentially misaligned, and only the few PMs whose dimensionality approaches that of human concepts exhibit alignment to human perception. Our results connect the fields of alignment and adversarial examples, and suggest the curse of high dimensionality of machine PMs is a major impediment to adversarial robustness. --- ### PhyPrompt: RL-based Prompt Refinement for Physically Plausible Text-to-Video Generation Authors: Shang Wu, Chenwei Xu, Zhuofan Xia, Weijian Li, Lie Lu, Pranav Maneriker, Fan Du, Manling Li, Han Liu Published: 2026-03-03 Categories: cs.CV, cs.AI arXiv: https://arxiv.org/abs/2603.03505v1 State-of-the-art text-to-video (T2V) generators frequently violate physical laws despite high visual quality. We show this stems from insufficient physical constraints in prompts rather than model limitations: manually adding physics details reliably produces physically plausible videos, but requires expertise and does not scale. We present PhyPrompt, a two-stage reinforcement learning framework that automatically refines prompts for physically realistic generation. First, we fine-tune a large language model on a physics-focused Chain-of-Thought dataset to integrate principles like object motion and force interactions while preserving user intent. Second, we apply Group Relative Policy Optimization with a dynamic reward curriculum that initially prioritizes semantic fidelity, then progressively shifts toward physical commonsense. This curriculum achieves synergistic optimization: PhyPrompt-7B reaches 40.8\% joint success on VideoPhy2 (8.6pp gain), improving physical commonsense by 11pp (55.8\% to 66.8\%) while simultaneously increasing semantic adherence by 4.4pp (43.4\% to 47.8\%). Remarkably, our curriculum exceeds single-objective training on both metrics, demonstrating compositional prompt discovery beyond conventional multi-objective trade-offs. PhyPrompt outperforms GPT-4o (+3.8\% joint) and DeepSeek-V3 (+2.2\%, 100$\times$ larger) using only 7B parameters. The approach transfers zero-shot across diverse T2V architectures (Lavie, VideoCrafter2, CogVideoX-5B) with up to 16.8\% improvement, establishing that domain-specialized reinforcement learning with compositional curricula surpasses general-purpose scaling for physics-aware generation. --- ### Geographically-Weighted Weakly Supervised Bayesian High-Resolution Transformer for 200m Resolution Pan-Arctic Sea Ice Concentration Mapping and Uncertainty Estimation using Sentinel-1, RCM, and AMSR2 Data Authors: Mabel Heffring, Lincoln Linlin Xu Published: 2026-03-03 Categories: cs.CV, cs.LG arXiv: https://arxiv.org/abs/2603.03503v1 Although high-resolution mapping of pan-Arctic sea ice with reliable corresponding uncertainty is essential for operational sea ice concentration (SIC) charting, it is a difficult task due to key challenges, such as the subtle nature of ice signature features, inexact SIC labels, model uncertainty, and data heterogeneity. This study presents a novel Bayesian High-Resolution Transformer approach for 200 meter resolution pan-Arctic SIC mapping and uncertainty quantification using Sentinel-1, RADARSAT Constellation Mission (RCM), and Advanced Microwave Scanning Radiometer 2 (AMSR2) data. First, to improve small and subtle sea ice feature (e.g., cracks/leads, ponds, and ice floes) extraction, we design a novel high-resolution Transformer model with both global and local modules that can better discern the subtle differences in sea ice patterns. Second, to address low-resolution and inexact SIC labels, we design a geographically-weighted weakly supervised loss function to supervise the model at region level instead of pixel level, and to prioritize pure open water and ice pack signatures while mitigating the impact of ambiguity in the marginal ice zone (MIZ). Third, to improve uncertainty quantification, we design a Bayesian extension of the proposed Transformer model, treating its parameters as random variables to more effectively capture uncertainties. Fourth, to address data heterogeneity, we fuse three different data types (Sentinel-1, RCM, and AMSR2) at decision-level to improve both SIC mapping and uncertainty quantification. The proposed approach is evaluated under pan-Arctic minimum-extent conditions in 2021 and 2025. Results demonstrate that the proposed model achieves 0.70 overall feature detection accuracy using Sentinel-1 data, while also preserving pan-Arctic SIC patterns (Sentinel-1 R\textsuperscript{2} = 0.90 relative to the ARTIST Sea Ice product). --- ### Quantifying Ranking Instability Across Evaluation Protocol Axes in Gene Regulatory Network Benchmarking Authors: Ihor Kendiukhov Published: 2026-03-03 Categories: q-bio.MN, cs.LG arXiv: https://arxiv.org/abs/2603.03493v1 Benchmark rankings are routinely used to justify scientific claims about method quality in gene regulatory network (GRN) inference, yet the stability of these rankings under plausible evaluation protocol choices is rarely examined. We present a systematic diagnostic framework for measuring ranking instability under protocol shift, including decomposition tools that separate base rate effects from discrimination effects. Using existing single cell GRN benchmark outputs across three human tissues and six inference methods, we quantify pairwise reversal rates across four protocol axes: candidate set restriction (16.3 percent, 95 percent CI 11.0 to 23.4 percent), tissue context (19.3 percent), reference network choice (32.1 percent), and symbol mapping policy (0.0 percent). A permutation null confirms that observed reversal rates are far below random order expectations (0.163 versus null mean 0.500), indicating partially stable but non invariant ranking structure. Our decomposition reveals that reversals are driven by changes in the relative discrimination ability of methods rather than by base rate inflation, a finding that challenges a common implicit assumption in GRN benchmarking. We propose concrete reporting practices for stability aware evaluation and provide a diagnostic toolkit for identifying method pairs at risk of reversal. --- ### When Small Variations Become Big Failures: Reliability Challenges in Compute-in-Memory Neural Accelerators Authors: Yifan Qin, Jiahao Zheng, Zheyu Yan, Wujie Wen, Xiaobo Sharon Hu, Yiyu Shi Published: 2026-03-03 Categories: cs.LG, cs.AR arXiv: https://arxiv.org/abs/2603.03491v1 Compute-in-memory (CiM) architectures promise significant improvements in energy efficiency and throughput for deep neural network acceleration by alleviating the von Neumann bottleneck. However, their reliance on emerging non-volatile memory devices introduces device-level non-idealities-such as write variability, conductance drift, and stochastic noise-that fundamentally challenge reliability, predictability, and safety, especially in safety-critical applications. This talk examines the reliability limits of CiM-based neural accelerators and presents a series of techniques that bridge device physics, architecture, and learning algorithms to address these challenges. We first demonstrate that even small device variations can lead to disproportionately large accuracy degradation and catastrophic failures in safety-critical inference workloads, revealing a critical gap between average-case evaluations and worst-case behavior. Building on this insight, we introduce SWIM, a selective write-verify mechanism that strategically applies verification only where it is most impactful, significantly improving reliability while maintaining CiM's efficiency advantages. Finally, we explore a learning-centric solution that improves realistic worst-case performance by training neural networks with right-censored Gaussian noise, aligning training assumptions with hardware-induced variability and enabling robust deployment without excessive hardware overhead. Together, these works highlight the necessity of cross-layer co-design for CiM accelerators and provide a principled path toward dependable, efficient neural inference on emerging memory technologies-paving the way for their adoption in safety- and reliability-critical systems. --- ### Phys4D: Fine-Grained Physics-Consistent 4D Modeling from Video Diffusion Authors: Haoran Lu, Shang Wu, Songling Liu, Jianshu Zhang, Maojiang Su, Guo Ye, Chenwei Xu, Lie Lu, Pranav Maneriker, Fan Du, Zhaoran Wang, Han Liu Published: 2026-03-03 Categories: cs.CV, cs.AI, cs.RO arXiv: https://arxiv.org/abs/2603.03485v4 Recent video diffusion models have achieved impressive capabilities as large-scale generative world models. However, these models often struggle with fine-grained physical consistency, exhibiting physically implausible dynamics over time. In this work, we present \textbf{Phys4D}, a pipeline for learning physics-consistent 4D world representations from video diffusion models. Phys4D adopts \textbf{a three-stage training paradigm} that progressively lifts appearance-driven video diffusion models into physics-consistent 4D world representations. We first bootstrap robust geometry and motion representations through large-scale pseudo-supervised pretraining, establishing a foundation for 4D scene modeling. We then perform physics-grounded supervised fine-tuning using simulation-generated data, enforcing temporally consistent 4D dynamics. Finally, we apply simulation-grounded reinforcement learning to correct residual physical violations that are difficult to capture through explicit supervision. To evaluate fine-grained physical consistency beyond appearance-based metrics, we introduce a set of \textbf{4D world consistency evaluation} that probe geometric coherence, motion stability, and long-horizon physical plausibility. Experimental results demonstrate that Phys4D substantially improves fine-grained spatiotemporal and physical consistency compared to appearance-driven baselines, while maintaining strong generative performance. Our project page is available at https://sensational-brioche-7657e7.netlify.app/ --- ### Phys4D: Fine-Grained Physics-Consistent 4D Modeling from Video Diffusion Authors: Haoran Lu, Shang Wu, Songling Liu, Jianshu Zhang, Maojiang Su, Guo Ye, Chenwei Xu, Lie Lu, Pranav Maneriker, Fan Du, Manling Li, Zhaoran Wang et al. Published: 2026-03-03 Categories: cs.CV, cs.AI, cs.RO arXiv: https://arxiv.org/abs/2603.03485v3 Recent video diffusion models have achieved impressive capabilities as large-scale generative world models. However, these models often struggle with fine-grained physical consistency, exhibiting physically implausible dynamics over time. In this work, we present \textbf{Phys4D}, a pipeline for learning physics-consistent 4D world representations from video diffusion models. Phys4D adopts \textbf{a three-stage training paradigm} that progressively lifts appearance-driven video diffusion models into physics-consistent 4D world representations. We first bootstrap robust geometry and motion representations through large-scale pseudo-supervised pretraining, establishing a foundation for 4D scene modeling. We then perform physics-grounded supervised fine-tuning using simulation-generated data, enforcing temporally consistent 4D dynamics. Finally, we apply simulation-grounded reinforcement learning to correct residual physical violations that are difficult to capture through explicit supervision. To evaluate fine-grained physical consistency beyond appearance-based metrics, we introduce a set of \textbf{4D world consistency evaluation} that probe geometric coherence, motion stability, and long-horizon physical plausibility. Experimental results demonstrate that Phys4D substantially improves fine-grained spatiotemporal and physical consistency compared to appearance-driven baselines, while maintaining strong generative performance. Our project page is available at https://sensational-brioche-7657e7.netlify.app/ --- ### Optimal trajectory-guided stochastic co-optimization for e-fuel system design and real-time operation Authors: Jeongdong Kim, Minsu Kim, Jonggeol Na, Junghwan Kim Published: 2026-03-03 Categories: cs.LG, cs.AI arXiv: https://arxiv.org/abs/2603.03484v1 E-fuels are promising long-term energy carriers supporting the net-zero transition. However, the large combinatorial design-operation spaces under renewable uncertainty make the use of mathematical programming impractical for co-optimizing e-fuel production systems. Here, we present MasCOR, a machine-learning-assisted co-optimization framework that learns from global operational trajectories. By encoding system design and renewable trends, a single MasCOR agent generalizes dynamic operation across diverse configurations and scenarios, substantially simplifying design-operation co-optimization under uncertainty. Benchmark comparisons against state-of-the-art reinforcement learning baselines demonstrate near-optimal performance, while computational costs are substantially lower than those of mathematical programming, enabling rapid parallel evaluation of designs within the co-optimization loop. This framework enables rapid screening of feasible design spaces together with corresponding operational policies. When applied to four potential European sites targeting e-methanol production, MasCOR shows that most locations benefit from reducing system load below 50 MW to achieve carbon-neutral methanol production, with production costs of 1.0-1.2 USD per kg. In contrast, Dunkirk (France), with limited renewable availability and high grid prices, favors system loads above 200 MW and expanded storage to exploit dynamic grid exchange and hydrogen sales to the market. These results underscore the value of the MasCOR framework for site-specific guidance from system design to real-time operation. --- ### Beyond Pixel Histories: World Models with Persistent 3D State Authors: Samuel Garcin, Thomas Walker, Steven McDonagh, Tim Pearce, Hakan Bilen, Tianyu He, Kaixin Wang, Jiang Bian Published: 2026-03-03 Categories: cs.CV, cs.AI, cs.LG arXiv: https://arxiv.org/abs/2603.03482v2 Interactive world models continually generate video by responding to a user's actions, enabling open-ended generation capabilities. However, existing models typically lack a 3D representation of the environment, meaning 3D consistency must be implicitly learned from data, and spatial memory is restricted to limited temporal context windows. This results in an unrealistic user experience and presents significant obstacles to downstream tasks such as training agents. To address this, we present PERSIST, a new paradigm of world model which simulates the evolution of a latent 3D scene: environment, camera, and renderer. This allows us to synthesise new frames with persistent spatial memory and consistent geometry. Both quantitative metrics and a qualitative user study show substantial improvements in spatial memory, 3D consistency, and long-horizon stability over existing methods, enabling coherent, evolving 3D worlds. We further demonstrate novel capabilities, including synthesising diverse 3D environments from a single image, as well as enabling fine-grained, geometry-aware control over generated experiences by supporting environment editing and specification directly in 3D space. Project page: https://francelico.github.io/persist.github.io --- ### Minimax Optimal Strategy for Delayed Observations in Online Reinforcement Learning Authors: Harin Lee, Kevin Jamieson Published: 2026-03-03 Categories: cs.LG, stat.ML arXiv: https://arxiv.org/abs/2603.03480v2 We study reinforcement learning with delayed state observation, where the agent observes the current state after some random number of time steps. We propose an algorithm that combines the augmentation method and the upper confidence bound approach. For tabular Markov decision processes (MDPs), we derive a regret bound of $\tilde{\mathcal{O}}(H \sqrt{D_{\max} SAK})$, where $S$ and $A$ are the cardinalities of the state and action spaces, $H$ is the time horizon, $K$ is the number of episodes, and $D_{\max}$ is the maximum length of the delay. We also provide a matching lower bound up to logarithmic factors, showing the optimality of our approach. Our analytical framework formulates this problem as a special case of a broader class of MDPs, where their transition dynamics decompose into a known component and an unknown but structured component. We establish general results for this abstract setting, which may be of independent interest. --- ### Stringology-Based Motif Discovery from EEG Signals: an ADHD Case Study Authors: Anat Dahan, Samah Ghazawi Published: 2026-03-03 Categories: q-bio.NC, cs.DS, cs.IR, cs.NE arXiv: https://arxiv.org/abs/2603.03476v1 We propose a novel computational framework for analyzing electroencephalography (EEG) time series using methods from stringology, the study of efficient algorithms for string processing, to systematically identify and characterize recurrent temporal patterns in neural signals. The primary aim is to introduce quantitative measures to understand neural signal dynamics, with the present findings serving as a proof-of-concept. The framework adapts order-preserving matching (OPM) and Cartesian tree matching (CTM) to detect temporal motifs that preserve relative ordering and hierarchical structure while remaining invariant to amplitude scaling. This approach provides a temporally precise representation of EEG dynamics that complements traditional spectral and global complexity analyses. To evaluate its utility, we applied the framework to multichannel EEG recordings from individuals with attention-deficit/hyperactivity disorder (ADHD) and matched controls using a publicly available dataset. Highly recurrent, group-specific motifs were extracted and quantified using both OPM and CTM. The ADHD group exhibited significantly higher motif frequencies, suggesting increased repetitiveness in neural activity. OPM analysis revealed shorter motif lengths and greater gradient instability in ADHD, reflected in larger mean and maximal inter-sample amplitude changes. CTM analysis further demonstrated reduced hierarchical complexity in ADHD, characterized by shallower tree structures and fewer hierarchical levels despite comparable motif lengths. These findings suggest that ADHD-related EEG alterations involve systematic differences in the structure, stability, and hierarchical organization of recurrent temporal patterns. The proposed stringology-based motif framework provides a complementary computational tool with potential applications for objective biomarker development in neurodevelopmental disorders. --- ### When Shallow Wins: Silent Failures and the Depth-Accuracy Paradox in Latent Reasoning Authors: Subramanyam Sahoo, Aman Chadha, Vinija Jain, Divya Chaudhary Published: 2026-03-03 Categories: cs.LG, cs.AI, cs.CL arXiv: https://arxiv.org/abs/2603.03475v1 Mathematical reasoning models are widely deployed in education, automated tutoring, and decision support systems despite exhibiting fundamental computational instabilities. We demonstrate that state-of-the-art models (Qwen2.5-Math-7B) achieve 61% accuracy through a mixture of reliable and unreliable reasoning pathways: 18.4% of correct predictions employ stable, faithful reasoning while 81.6% emerge through computationally inconsistent pathways. Additionally, 8.8% of all predictions are silent failures -- confident yet incorrect outputs. Through comprehensive analysis using novel faithfulness metrics, we reveal: (1) reasoning quality shows weak negative correlation with correctness (r=-0.21, p=0.002), reflecting a binary classification threshold artifact rather than a monotonic inverse relationship; (2) scaling from 1.5B to 7B parameters (4.7x increase) provides zero accuracy benefit on our evaluated subset (6% of GSM8K), requiring validation on the complete benchmark; and (3) latent reasoning employs diverse computational strategies, with ~20% sharing CoT-like patterns. These findings highlight that benchmark accuracy can mask computational unreliability, demanding evaluation reforms measuring stability beyond single-sample metrics. --- ### Biased Generalization in Diffusion Models Authors: Jerome Garnier-Brun, Luca Biggio, Davide Beltrame, Marc Mézard, Luca Saglietti Published: 2026-03-03 Categories: cs.LG, cond-mat.dis-nn, cond-mat.stat-mech arXiv: https://arxiv.org/abs/2603.03469v1 Generalization in generative modeling is defined as the ability to learn an underlying distribution from a finite dataset and produce novel samples, with evaluation largely driven by held-out performance and perceived sample quality. In practice, training is often stopped at the minimum of the test loss, taken as an operational indicator of generalization. We challenge this viewpoint by identifying a phase of biased generalization during training, in which the model continues to decrease the test loss while favoring samples with anomalously high proximity to training data. By training the same network on two disjoint datasets and comparing the mutual distances of generated samples and their similarity to training data, we introduce a quantitative measure of bias and demonstrate its presence on real images. We then study the mechanism of bias, using a controlled hierarchical data model where access to exact scores and ground-truth statistics allows us to precisely characterize its onset. We attribute this phenomenon to the sequential nature of feature learning in deep networks, where coarse structure is learned early in a data-independent manner, while finer features are resolved later in a way that increasingly depends on individual training samples. Our results show that early stopping at the test loss minimum, while optimal under standard generalization criteria, may be insufficient for privacy-critical applications. --- ### Graph Hopfield Networks: Energy-Based Node Classification with Associative Memory Authors: Abinav Rao, Alex Wa, Rishi Athavale Published: 2026-03-03 Categories: cs.LG, cs.AI, cs.IR arXiv: https://arxiv.org/abs/2603.03464v1 We introduce Graph Hopfield Networks, whose energy function couples associative memory retrieval with graph Laplacian smoothing for node classification. Gradient descent on this joint energy yields an iterative update interleaving Hopfield retrieval with Laplacian propagation. Memory retrieval provides regime-dependent benefits: up to 2.0~pp on sparse citation networks and up to 5 pp additional robustness under feature masking; the iterative energy-descent architecture itself is a strong inductive bias, with all variants (including the memory-disabled NoMem ablation) outperforming standard baselines on Amazon co-purchase graphs. Tuning enables graph sharpening for heterophilous benchmarks without architectural changes. --- ### Half the Nonlinearity Is Wasted: Measuring and Reallocating the Transformer's MLP Budget Authors: Peter Balogh Published: 2026-03-03 Categories: cs.LG, cs.CL arXiv: https://arxiv.org/abs/2603.03459v2 We investigate when transformer MLP nonlinearity is actually necessary. A gate with $d+1$ parameters decides when to replace the full MLP with a linear surrogate. Through systematic investigation across six models (162M-2.8B parameters), two architectures, and three corpora, we establish that nonlinearity need cannot be predicted from token identity: cross-corpus correlation is zero ($r < 0.05$). The routing decision is fully contextual. Despite weak per-instance predictability, the gate exploits a heavily skewed distribution where most MLP computations are near-linear, achieving 25-56% linear routing at <1% perplexity cost in GPT-2. In GPT-2 Large, 11 of 36 layers beat baseline with gating and no layer exceeds 3.7% all-linear cost. This success is architecture-dependent: Pythia models show higher costs, though Pythia-2.8B's full 32-layer sweep reveals one layer that narrowly beats baseline. As a proof of concept, we progressively replace middle-layer MLPs with frozen linear matrices: 5 of 24 layers linearize at zero cost. With a full training budget, 4 linearized layers yield a 10.2% perplexity improvement -- and a two-phase gated approach pushes this to 17.3%, beating a vanilla fine-tuning control and confirming that the nonlinear MLPs at these layers were actively harmful. --- ### Asymmetric Goal Drift in Coding Agents Under Value Conflict Authors: Magnus Saebo, Spencer Gibson, Tyler Crosse, Achyutha Menon, Eyon Jang, Diogo Cruz Published: 2026-03-03 Categories: cs.AI, cs.CL, cs.SE arXiv: https://arxiv.org/abs/2603.03456v2 Coding agents are increasingly deployed autonomously, at scale, and over long-context horizons. To be effective and safe, these agents must navigate complex trade-offs in deployment, balancing influence from the user, their learned values, and the codebase itself. Understanding how agents resolve these trade-offs in practice is critical, yet prior work has relied on static, synthetic settings that do not capture the complexity of real-world environments. To this end, we introduce a framework built on OpenCode in which a coding agent completes realistic, multi-step tasks under a system prompt constraint favoring one side of a value trade-off. We measure how often the agent violates this constraint as it completes tasks, with and without environmental pressure toward the competing value. Using this framework, we demonstrate that GPT-5 mini, Haiku 4.5, and Grok Code Fast 1 exhibit $\textit{asymmetric drift}$: they are more likely to violate their system prompt when its constraint opposes strongly-held values like security and privacy. We find for the models and values tested that goal drift correlates with three compounding factors: value alignment, adversarial pressure, and accumulated context. However, even constraints aligned with strongly-held values like privacy are violated under sustained environmental pressure for some models. Our findings reveal that shallow compliance checks are insufficient, and that environmental signals can override explicit constraints in ways that appear exploitable. Malicious actors with access to the codebase could manipulate agent behavior by appealing to learned values, with the risk compounding over the long horizons typical of agentic deployment. --- ### (Re) FairDICE: A Fair Tradeoff in Multi-objective Offline RL Authors: Peter Adema, Karim Galliamov, Aleksey Evstratovskiy, Ross Geurts Published: 2026-03-03 Categories: cs.LG arXiv: https://arxiv.org/abs/2603.03454v2 Offline Reinforcement Learning (RL) is an emerging field of RL in which policies are learned solely from demonstrations. Within offline RL, some environments involve balancing multiple objectives, but existing multi-objective offline RL algorithms do not provide an efficient way to find a fair compromise. FairDICE (see arXiv:2506.08062v2) seeks to fill this gap by adapting OptiDICE (an offline RL algorithm) to automatically learn weights for multiple objectives to e.g. incentivise fairness among objectives. As this would be a valuable contribution, this replication study examines the replicability of claims made regarding FairDICE. We find that many theoretical claims hold, but an error in the code reduces FairDICE to standard behaviour cloning in continuous environments, and many important hyperparameters were originally underspecified. After rectifying this, we show in experiments extending the original paper that FairDICE can scale to complex environments and high-dimensional rewards, though it can be reliant on (online) hyperparameter tuning. We conclude that FairDICE is a theoretically interesting method, but the experimental justification requires significant revision. --- ### Impact of Localization Errors on Label Quality for Online HD Map Construction Authors: Alexander Blumberg, Jonas Merkert, Richard Fehler, Fabian Immel, Frank Bieder, Jan-Hendrik Pauls, Christoph Stiller Published: 2026-03-03 Categories: cs.RO, cs.CV arXiv: https://arxiv.org/abs/2603.03452v1 High-definition (HD) maps are crucial for autonomous vehicles, but their creation and maintenance is very costly. This motivates the idea of online HD map construction. To provide a continuous large-scale stream of training data, existing HD maps can be used as labels for onboard sensor data from consumer vehicle fleets. However, compared to current, well curated HD map perception datasets, this fleet data suffers from localization errors, resulting in distorted map labels. We introduce three kinds of localization errors, Ramp, Gaussian, and Perlin noise, to examine their influence on generated map labels. We train a variant of MapTRv2, a state-of-the-art online HD map construction model, on the Argoverse 2 dataset with various levels of localization errors and assess the degradation of model performance. Since localization errors affect distant labels more severely, but are also less significant to driving performance, we introduce a distance-based map construction metric. Our experiments reveal that localization noise affects the model performance significantly. We demonstrate that errors in heading angle exert a more substantial influence than position errors, as angle errors result in a greater distortion of labels as distance to the vehicle increases. Furthermore, we can demonstrate that the model benefits from non-distorted ground truth (GT) data and that the performance decreases more than linearly with the increase in noisy data. Our study additionally provides a qualitative evaluation of the extent to which localization errors influence the construction of HD maps. --- ### Proact-VL: A Proactive VideoLLM for Real-Time AI Companions Authors: Weicai Yan, Yuhong Dai, Qi Ran, Haodong Li, Wang Lin, Tao Jin, Xing Xie, Hao Liao, Jianxun Lian Published: 2026-03-03 Categories: cs.CV arXiv: https://arxiv.org/abs/2603.03447v4 Proactive and real-time interactive experiences are essential for human-like AI companions, yet face three key challenges: (1) achieving low-latency inference under continuous streaming inputs, (2) autonomously deciding when to respond, and (3) controlling both quality and quantity of generated content to meet real-time constraints. In this work, we instantiate AI companions through two gaming scenarios, commentator and guide, selected for their suitability for automatic evaluation. We introduce the Live Gaming Benchmark, a large-scale dataset with three representative scenarios: solo commentary, co-commentary, and user guidance, and present Proact-VL, a general framework that shapes multimodal language models into proactive, real-time interactive agents capable of human-like environment perception and interaction. Extensive experiments show Proact-VL achieves superior response latency and quality while maintaining strong video understanding capabilities, demonstrating its practicality for real-time interactive applications. --- ### Beyond Accuracy: Evaluating Visual Grounding In Multimodal Medical Reasoning Authors: Anas Zafar, Leema Krishna Murali, Ashish Vashist Published: 2026-03-03 Categories: cs.CV arXiv: https://arxiv.org/abs/2603.03437v1 Recent work shows that text-only reinforcement learning with verifiable rewards (RLVR) can match or outperform image-text RLVR on multimodal medical VQA benchmarks, suggesting current evaluation protocols may fail to measure causal visual dependence. We introduce a counterfactual evaluation framework using real, blank, and shuffled images across four medical VQA benchmarks: PathVQA, PMC-VQA, SLAKE, and VQA-RAD. Beyond accuracy, we measure Visual Reliance Score (VRS), Image Sensitivity (IS), and introduce Hallucinated Visual Reasoning Rate (HVRR) to detect cases where models generate visual claims despite producing image-invariant answers. Our findings reveal that RLVR improves accuracy while degrading visual grounding: text-only RLVR achieves negative VRS on PathVQA (-0.09), performing better with mismatched images, while image-text RLVR reduces image sensitivity to 39.8% overall despite improving accuracy. On VQA-RAD, both variants achieve 63% accuracy through different mechanisms: text-only RLVR retains 81% performance with blank images, while image-text RLVR shows only 29% image sensitivity. Models generate visual claims in 68-74% of responses, yet 38-43% are ungrounded (HVRR). These findings demonstrate that accuracy-only rewards enable shortcut exploitation, and progress requires grounding-aware evaluation protocols and training objectives that explicitly enforce visual dependence. --- ### mHC-HSI: Clustering-Guided Hyper-Connection Mamba for Hyperspectral Image Classification Authors: Yimin Zhu, Zack Dewis, Quinn Ledingham, Saeid Taleghanidoozdoozan, Mabel Heffring, Zhengsen Xu, Motasem Alkayid, Megan Greenwood, Lincoln Linlin Xu Published: 2026-03-03 Categories: cs.CV arXiv: https://arxiv.org/abs/2603.03418v1 Recently, DeepSeek has invented the manifold-constrained hyper-connection (mHC) approach which has demonstrated significant improvements over the traditional residual connection in deep learning models \cite{xie2026mhc}. Nevertheless, this approach has not been tailor-designed for improving hyperspectral image (HSI) classification. This paper presents a clustering-guided mHC Mamba model (mHC-HSI) for enhanced HSI classification, with the following contributions. First, to improve spatial-spectral feature learning, we design a novel clustering-guided Mamba module, based on the mHC framework, that explicitly learns both spatial and spectral information in HSI. Second, to decompose the complex and heterogeneous HSI into smaller clusters, we design a new implementation of the residual matrix in mHC, which can be treated as soft cluster membership maps, leading to improved explainability of the mHC approach. Third, to leverage the physical spectral knowledge, we divide the spectral bands into physically-meaningful groups and use them as the "parallel streams" in mHC, leading to a physically-meaningful approach with enhanced interpretability. The proposed approach is tested on benchmark datasets in comparison with the state-of-the-art methods, and the results suggest that the proposed model not only improves the accuracy but also enhances the model explainability. Code is available here: https://github.com/GSIL-UCalgary/mHC_HyperSpectral --- ### Parallel Test-Time Scaling with Multi-Sequence Verifiers Authors: Yegon Kim, Seungyoo Lee, Chaeyun Jang, Hyungi Lee, Juho Lee Published: 2026-03-03 Categories: cs.CR, cs.AI arXiv: https://arxiv.org/abs/2603.03417v2 Parallel test-time scaling, which generates multiple candidate solutions for a single problem, is a powerful technique for improving large language model performance. However, it is hindered by two key bottlenecks: accurately selecting the correct solution from the candidate pool, and the high inference latency from generating many full solutions. We argue that both challenges are fundamentally linked to verifier calibration, as a well-calibrated verifier improves answer selection and enables early-stopping strategies to reduce latency. However, existing non-generative verifiers are limited as they score each candidate in isolation, overlooking rich contextual information across the set of candidates. To address this, we introduce the Multi-Sequence Verifier (MSV), a lightweight verifier that predicts each candidate's correctness conditioned on the full sampled set. MSV achieves improved calibration, which directly enhances best-of-N selection performance and empowers a novel early-stopping framework. Across challenging mathematical reasoning benchmarks, MSV improves best-of-64 accuracy by up to 6\% relative to strong baselines, and in the early-stopping setting reaches the same accuracy as baselines with less than half the latency. --- ### Farther the Shift, Sparser the Representation: Analyzing OOD Mechanisms in LLMs Authors: Mingyu Jin, Yutong Yin, Jingcheng Niu, Qingcheng Zeng, Wujiang Xu, Mengnan Du, Wei Cheng, Zhaoran Wang, Tianlong Chen, Dimitris N. Metaxas Published: 2026-03-03 Categories: cs.CL, cs.AI arXiv: https://arxiv.org/abs/2603.03415v3 In this work, we investigate how Large Language Models (LLMs) adapt their internal representations when encountering inputs of increasing difficulty, quantified as the degree of out-of-distribution (OOD) shift. We reveal a consistent and quantifiable phenomenon: as task difficulty increases, whether through harder reasoning questions, longer contexts, or adding answer choices, the last hidden states of LLMs become substantially sparser. In short, \textbf{\textit{the farther the shift, the sparser the representations}}. This sparsity--difficulty relation is observable across diverse models and domains, suggesting that language models respond to unfamiliar or complex inputs by concentrating computation into specialized subspaces in the last hidden state. Through a series of controlled analyses with a learning dynamic explanation, we demonstrate that this sparsity is not incidental but an adaptive mechanism for stabilizing reasoning under OOD. Leveraging this insight, we design \textit{Sparsity-Guided Curriculum In-Context Learning (SG-ICL)}, a strategy that explicitly uses representation sparsity to schedule few-shot demonstrations, leading to considerable performance enhancements. Our study provides new mechanistic insights into how LLMs internalize OOD challenges. The source code is available at the URL: https://github.com/MingyuJ666/sparsityLLM. --- ### Farther the Shift, Sparser the Representation: Analyzing OOD Mechanisms in LLMs Authors: Mingyu Jin, Yutong Yin, Jingcheng Niu, Qingcheng Zeng, Wujiang Xu, Mengnan Du, Wei Cheng, Zhaoran Wang, Tianlong Chen, Dimitris N. Metaxas Published: 2026-03-03 Categories: cs.CL, cs.AI arXiv: https://arxiv.org/abs/2603.03415v2 In this work, we investigate how Large Language Models (LLMs) adapt their internal representations when encountering inputs of increasing difficulty, quantified as the degree of out-of-distribution (OOD) shift. We reveal a consistent and quantifiable phenomenon: as task difficulty increases, whether through harder reasoning questions, longer contexts, or adding answer choices, the last hidden states of LLMs become substantially sparser. In short, \textbf{\textit{the farther the shift, the sparser the representations}}. This sparsity--difficulty relation is observable across diverse models and domains, suggesting that language models respond to unfamiliar or complex inputs by concentrating computation into specialized subspaces in the last hidden state. Through a series of controlled analyses with a learning dynamic explanation, we demonstrate that this sparsity is not incidental but an adaptive mechanism for stabilizing reasoning under OOD. Leveraging this insight, we design \textit{Sparsity-Guided Curriculum In-Context Learning (SG-ICL)}, a strategy that explicitly uses representation sparsity to schedule few-shot demonstrations, leading to considerable performance enhancements. Our study provides new mechanistic insights into how LLMs internalize OOD challenges. The source code is available at the URL: https://github.com/MingyuJ666/sparsityLLM. --- ### PRIVATEEDIT: A Privacy-Preserving Pipeline for Face-Centric Generative Image Editing Authors: Dipesh Tamboli, Vineet Punyamoorty, Atharv Pawar, Vaneet Aggarwal Published: 2026-03-03 Categories: cs.CR, cs.AI arXiv: https://arxiv.org/abs/2603.03412v1 Recent advances in generative image editing have enabled transformative applications, from professional head shot generation to avatar stylization. However, these systems often require uploading high-fidelity facial images to third-party models, raising concerns around biometric privacy, data misuse, and user consent. We propose a privacy-preserving pipeline that supports high-quality editing while keeping users in control over their biometric data in face-centric use cases. Our approach separates identity-sensitive regions from editable image context using on-device segmentation and masking, enabling secure, user-controlled editing without modifying third-party generative models. Unlike traditional cloud-based tools, PRIVATEEDIT enforces privacy by default: biometric data is never exposed or transmitted. This design requires no access to or retraining of third-party models, making it compatible with a wide range of commercial APIs. By treating privacy as a core design constraint, our system supports responsible generative AI centered on user autonomy and trust. The pipeline includes a tunable masking mechanism that lets users control how much facial information is concealed, allowing them to balance privacy and output fidelity based on trust level or use case. We demonstrate its applicability in professional and creative workflows and provide a user interface for selective anonymization. By advocating privacy-by-design in generative AI, our work offers both technical feasibility and normative guidance for protecting digital identity. The source code is available at https://github.com/Dipeshtamboli/PrivateEdit-Privacy-Preserving-GenAI. --- ### Scalable Contrastive Causal Discovery under Unknown Soft Interventions Authors: Mingxuan Zhang, Khushi Desai, Sopho Kevlishvili, Elham Azizi Published: 2026-03-03 Categories: stat.ML, cs.LG arXiv: https://arxiv.org/abs/2603.03411v1 Observational causal discovery is only identifiable up to the Markov equivalence class. While interventions can reduce this ambiguity, in practice interventions are often soft with multiple unknown targets. In many realistic scenarios, only a single intervention regime is observed. We propose a scalable causal discovery model for paired observational and interventional settings with shared underlying causal structure and unknown soft interventions. The model aggregates subset-level PDAGs and applies contrastive cross-regime orientation rules to construct a globally consistent maximal PDAG under Meek closure, enabling generalization to both in-distribution and out-of-distribution settings. Theoretically, we prove that our model is sound with respect to a restricted $Ψ$ equivalence class induced solely by the information available in the subset-restricted setting. We further show that the model asymptotically recovers the corresponding identifiable PDAG and can orient additional edges compared to non-contrastive subset-restricted methods. Experiments on synthetic data demonstrate improved causal structure recovery, generalization to unseen graphs with held-out causal mechanisms, and scalability to larger graphs, with ablations supporting the theoretical results. --- ### On Google's SynthID-Text LLM Watermarking System: Theoretical Analysis and Empirical Validation Authors: Romina Omidi, Yun Dong, Binghui Wang Published: 2026-03-03 Categories: cs.CR, cs.AI arXiv: https://arxiv.org/abs/2603.03410v2 Google's SynthID-Text, the first ever production-ready generative watermark system for large language model, designs a novel Tournament-based method that achieves the state-of-the-art detectability for identifying AI-generated texts. The system's innovation lies in: 1) a new Tournament sampling algorithm for watermarking embedding, 2) a detection strategy based on the introduced score function (e.g., Bayesian or mean score), and 3) a unified design that supports both distortionary and non-distortionary watermarking methods. This paper presents the first theoretical analysis of SynthID-Text, with a focus on its detection performance and watermark robustness, complemented by empirical validation. For example, we prove that the mean score is inherently vulnerable to increased tournament layers, and design a layer inflation attack to break SynthID-Text. We also prove the Bayesian score offers improved watermark robustness w.r.t. layers and further establish that the optimal Bernoulli distribution for watermark detection is achieved when the parameter is set to 0.5. Together, these theoretical and empirical insights not only deepen our understanding of SynthID-Text, but also open new avenues for analyzing effective watermark removal strategies and designing robust watermarking techniques. Source code is available at https: //github.com/romidi80/Synth-ID-Empirical-Analysis. --- ### A Short Note on a Variant of the Squint Algorithm Authors: Haipeng Luo Published: 2026-03-03 Categories: cs.LG arXiv: https://arxiv.org/abs/2603.03409v1 This short note describes a simple variant of the Squint algorithm of Koolen and Van Erven (2015) for the classic expert problem. Via an equally simple modification of their proof, we prove that this variant ensures a regret bound that resembles the one shown in a recent work by Freund et al. (2026) for a variant of the NormalHedge algorithm (Chaudhuri et al., 2009). --- ### Tracing Pharmacological Knowledge In Large Language Models Authors: Basil Hasan Khwaja, Dylan Chen, Guntas Toor, Anastasiya Kuznetsova Published: 2026-03-03 Categories: cs.CL arXiv: https://arxiv.org/abs/2603.03407v1 Large language models (LLMs) have shown strong empirical performance across pharmacology and drug discovery tasks, yet the internal mechanisms by which they encode pharmacological knowledge remain poorly understood. In this work, we investigate how drug-group semantics are represented and retrieved within Llama-based biomedical language models using causal and probing-based interpretability methods. We apply activation patching to localize where drug-group information is stored across model layers and token positions, and complement this analysis with linear probes trained on token-level and sum-pooled activations. Our results demonstrate that early layers play a key role in encoding drug-group knowledge, with the strongest causal effects arising from intermediate tokens within the drug-group span rather than the final drug-group token. Linear probing further reveals that pharmacological semantics are distributed across tokens and are already present in the embedding space, with token-level probes performing near chance while sum-pooled representations achieve maximal accuracy. Together, these findings suggest that drug-group semantics in LLMs are not localized to single tokens but instead arise from distributed representations. This study provides the first systematic mechanistic analysis of pharmacological knowledge in LLMs, offering insights into how biomedical semantics are encoded in large language models. --- ### Surprisal-Rényi Free Energy Authors: Shion Matsumoto, Raul Castillo, Benjamin Prada, Ankur Arjun Mali Published: 2026-03-03 Categories: stat.ML, cs.LG arXiv: https://arxiv.org/abs/2603.03405v1 The forward and reverse Kullback-Leibler (KL) divergences arise as limiting objectives in learning and inference yet induce markedly different inductive biases that cannot be explained at the level of expectations alone. In this work, we introduce the Surprisal-Rényi Free Energy (SRFE), a log-moment-based functional of the likelihood ratio that lies outside the class of $f$-divergences. We show that SRFE recovers forward and reverse KL divergences as singular endpoint limits and derive local expansions around both limits in which the variance of the log-likelihood ratio appears as a first-order correction. This reveals an explicit mean-variance tradeoff governing departures from KL-dominated regimes. We further establish a Gibbs-type variational characterization of SRFE as the unique minimizer of a weighted sum of KL divergences and prove that SRFE directly controls large deviations of excess code-length via Chernoff-type bounds, yielding a precise Minimum Description Length interpretation. Together, these results identify SRFE as a variance- and tail-sensitive free-energy functional that clarifies the geometric and large-deviation structure underlying forward and reverse KL limits, without unifying or subsuming distinct learning frameworks. --- ### Heterogeneous Time Constants Improve Stability in Equilibrium Propagation Authors: Yoshimasa Kubo, Suhani Pragnesh Modi, Smit Patel Published: 2026-03-03 Categories: cs.LG, cs.AI arXiv: https://arxiv.org/abs/2603.03402v1 Equilibrium propagation (EP) is a biologically plausible alternative to backpropagation for training neural networks. However, existing EP models use a uniform scalar time step dt, which corresponds biologically to a membrane time constant that is heterogeneous across neurons. Here, we introduce heterogeneous time steps (HTS) for EP by assigning neuron-specific time constants drawn from biologically motivated distributions. We show that HTS improves training stability while maintaining competitive task performance. These results suggest that incorporating heterogeneous temporal dynamics enhances both the biological realism and robustness of equilibrium propagation. ---