# Research Papers — 2026-03 · bundle page 159 of 243

> Full-text markdown of papers #7901–7950 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/159
[← Previous bundle](https://aipod.dev/papers/archive/2026-03/page/158.md)
[Next bundle →](https://aipod.dev/papers/archive/2026-03/page/160.md)

### Reference-Guided Machine Unlearning

Authors: Jonas Mirlach, Sonia Laguna, Julia E. Vogt
Published: 2026-03-11
Categories: cs.LG
arXiv: https://arxiv.org/abs/2603.11210v1

Machine unlearning aims to remove the influence of specific data from trained models while preserving general utility. Existing approximate unlearning methods often rely on performance-degradation heuristics, such as loss maximization or random labeling. However, these signals can be poorly conditioned, leading to unstable optimization and harming the model's generalization. We argue that unlearning should instead prioritize distributional indistinguishability, aligning the model's behavior on forget data with its behavior on truly unseen data. Motivated by this, we propose Reference-Guided Unlearning (ReGUn), a framework that leverages a disjoint held-out dataset to provide a principled, class-conditioned reference for distillation. We demonstrate across various model architectures, natural image datasets, and varying forget fractions that ReGUn consistently outperforms standard approximate baselines, achieving a superior forgetting-utility trade-off.

---

### Evidential learning driven Breast Tumor Segmentation with Stage-divided Vision-Language Interaction

Authors: Jingxing Zhong, Qingtao Pan, Xuchang Zhou, Jiazhen Lin, Xinguo Zhuang
Published: 2026-03-11
Categories: cs.CV
arXiv: https://arxiv.org/abs/2603.11206v1

Breast cancer is one of the most common causes of death among women worldwide, with millions of fatalities annually. Magnetic Resonance Imaging (MRI) can provide various sequences for characterizing tumor morphology and internal patterns, and becomes an effective tool for detection and diagnosis of breast tumors. However, previous deep-learning based tumor segmentation methods have limitations in accurately locating tumor contours due to the challenge of low contrast between cancer and normal areas and blurred boundaries. Leveraging text prompt information holds promise in ameliorating tumor segmentation effect by delineating segmentation regions. Inspired by this, we propose text-guided Breast Tumor Segmentation model (TextBCS) with stage-divided vision-language interaction and evidential learning. Specifically, the proposed stage-divided vision-language interaction facilitates information mutual between visual and text features at each stage of down-sampling, further exerting the advantages of text prompts to assist in locating lesion areas in low contrast scenarios. Moreover, the evidential learning is adopted to quantify the segmentation uncertainty of the model for blurred boundary. It utilizes the variational Dirichlet to characterize the distribution of the segmentation probabilities, addressing the segmentation uncertainties of the boundaries. Extensive experiments validate the superiority of our TextBCS over other segmentation networks, showcasing the best breast tumor segmentation performance on publicly available datasets.

---

### Representation Finetuning for Continual Learning

Authors: Haihua Luo, Xuming Ran, Tommi Kärkkäinen, Huiyan Xue, Zhonghua Chen, Qi Xu, Fengyu Cong
Published: 2026-03-11
Categories: cs.LG, cs.AI
arXiv: https://arxiv.org/abs/2603.11201v3

The world is inherently dynamic, and continual learning aims to enable models to adapt to ever-evolving data streams. While pre-trained models have shown powerful performance in continual learning, they still require finetuning to adapt effectively to downstream tasks. However, prevailing Parameter-Efficient Fine-Tuning (PEFT) methods operate through empirical, black-box optimization at the weight level. These approaches lack explicit control over representation drift, leading to sensitivity to domain shifts and catastrophic forgetting in continual learning scenarios. In this work, we introduce Continual Representation Learning (CoRe), a novel framework that for the first time shifts the finetuning paradigm from weight space to representation space. Unlike conventional methods, CoRe performs task-specific interventions within a low-rank linear subspace of hidden representations, adopting a learning process with explicit objectives, which ensures stability for past tasks while maintaining plasticity for new ones. By constraining updates to a low-rank subspace, CoRe achieves exceptional parameter efficiency. Extensive experiments across multiple continual learning benchmarks demonstrate that CoRe not only preserves parameter efficiency but also significantly outperforms existing state-of-the-art methods. Our work introduces representation finetuning as a new, more effective and interpretable paradigm for continual learning.

---

### Representation Finetuning for Continual Learning

Authors: Haihua Luo, Xuming Ran, Tommi Kärkkäinen, Huiyan Xue, Zhonghua Chen, Qi Xu, Fengyu Cong
Published: 2026-03-11
Categories: cs.LG, cs.AI
arXiv: https://arxiv.org/abs/2603.11201v2

The world is inherently dynamic, and continual learning aims to enable models to adapt to ever-evolving data streams. While pre-trained models have shown powerful performance in continual learning, they still require finetuning to adapt effectively to downstream tasks. However, prevailing Parameter-Efficient Fine-Tuning (PEFT) methods operate through empirical, black-box optimization at the weight level. These approaches lack explicit control over representation drift, leading to sensitivity to domain shifts and catastrophic forgetting in continual learning scenarios. In this work, we introduce Continual Representation Learning (CoRe), a novel framework that for the first time shifts the finetuning paradigm from weight space to representation space. Unlike conventional methods, CoRe performs task-specific interventions within a low-rank linear subspace of hidden representations, adopting a learning process with explicit objectives, which ensures stability for past tasks while maintaining plasticity for new ones. By constraining updates to a low-rank subspace, CoRe achieves exceptional parameter efficiency. Extensive experiments across multiple continual learning benchmarks demonstrate that CoRe not only preserves parameter efficiency but also significantly outperforms existing state-of-the-art methods. Our work introduces representation finetuning as a new, more effective and interpretable paradigm for continual learning.

---

### DNS-GT: A Graph-based Transformer Approach to Learn Embeddings of Domain Names from DNS Queries

Authors: Massimiliano Altieri, Ronan Hamon, Roberto Corizzo, Michelangelo Ceci, Ignacio Sanchez
Published: 2026-03-11
Categories: cs.CR, cs.LG
arXiv: https://arxiv.org/abs/2603.11200v1

Network intrusion detection systems play a crucial role in the security strategy employed by organisations to detect and prevent cyberattacks. Such systems usually combine pattern detection signatures with anomaly detection techniques powered by machine learning methods. However, the commonly proposed machine learning methods present drawbacks such as over-reliance on labeled data and limited generalization capabilities. To address these issues, embedding-based methods have been introduced to learn representations from network data, such as DNS traffic, mainly due to its large availability, that generalise effectively to many downstream tasks. However, current approaches do not properly consider contextual information among DNS queries. In this paper, we tackle this issue by proposing DNS-GT, a novel Transformer-based model that learns embeddings for domain names from sequences of DNS queries. The model is first pre-trained in a self-supervised fashion in order to learn the general behavior of DNS activity. Then, it can be finetuned on specific downstream tasks, exploiting interactions with other relevant queries in a given sequence. Our experiments with real-world DNS data showcase the ability of our method to learn effective domain name representations. A quantitative evaluation on domain name classification and botnet detection tasks shows that our approach achieves better results compared to relevant baselines, creating opportunities for further exploration of large-scale language models for intrusion detection systems. Our code is available at: https://github.com/m-altieri/DNS-GT.

---

### Bayesian Optimization of Partially Known Systems using Hybrid Models

Authors: Eike Cramer, Luis Kutschat, Oliver Stollenwerk, Joel A. Paulson, Alexander Mitsos
Published: 2026-03-11
Categories: cs.LG
arXiv: https://arxiv.org/abs/2603.11199v1

Bayesian optimization (BO) has gained attention as an efficient algorithm for black-box optimization of expensive-to-evaluate systems, where the BO algorithm iteratively queries the system and suggests new trials based on a probabilistic model fitted to previous samples. Still, the standard BO loop may require a prohibitively large number of experiments to converge to the optimum, especially for high-dimensional and nonlinear systems. We present a hybrid model-based BO formulation that combines the iterative Bayesian learning of BO with partially known mechanistic physical models. Instead of learning a direct mapping from inputs to the objective, we write all known equations for a physics-based model and infer expressions for variables missing equations using a probabilistic model, in our case, a Gaussian process (GP). The final formulation then includes the GP as a constraint in the hybrid model, thereby allowing other physics-based nonlinear and implicit model constraints. This hybrid model formulation yields a constrained, nonlinear stochastic program, which we discretize using the sample-average approximation. In an in-silico optimization of a single-stage distillation, the hybrid BO model based on mass conservation laws yields significantly better designs than a standard BO loop. Furthermore, the hybrid model converges in as few as one iteration, depending on the initial samples, whereas, the standard BO does not converge within 25 for any of the seeds. Overall, the proposed hybrid BO scheme presents a promising optimization method for partially known systems, leveraging the strengths of both mechanistic modeling and data-driven optimization.

---

### DeReason: A Difficulty-Aware Curriculum Improves Decoupled SFT-then-RL Training for General Reasoning

Authors: Hanxu Hu, Yuxuan Wang, Maggie Huan, Jannis Vamvas, Yinya Huang, Zhijiang Guo, Rico Sennrich
Published: 2026-03-11
Categories: cs.CL
arXiv: https://arxiv.org/abs/2603.11193v1

Reinforcement learning with Verifiable Rewards (RLVR) has emerged as a powerful paradigm for eliciting reasoning capabilities in large language models, particularly in mathematics and coding. While recent efforts have extended this paradigm to broader general scientific (STEM) domains, the complex interplay between supervised fine-tuning (SFT) and RL in these contexts remains underexplored. In this paper, we conduct controlled experiments revealing a critical challenge: for general STEM domains, RL applied directly to base models is highly sample-inefficient and is consistently surpassed by supervised fine-tuning (SFT) on moderate-quality responses. Yet sequential SFT followed by RL can further improve performance, suggesting that the two stages play complementary roles, and that how training data is allocated between them matters. Therefore, we propose DeReason, a difficulty-based data decoupling strategy for general reasoning. DeReason partitions training data by reasoning intensity estimated via LLM-based scoring into reasoning-intensive and non-reasoning-intensive subsets. It allocates broad-coverage, non-reasoning-intensive problems to SFT to establish foundational domain knowledge, and reserves a focused subset of difficult problems for RL to cultivate complex reasoning. We demonstrate that this principled decoupling yields better performance than randomly splitting the data for sequential SFT and RL. Extensive experiments on general STEM and mathematical benchmarks demonstrate that our decoupled curriculum training significantly outperforms SFT-only, RL-only, and random-split baselines. Our work provides a systematic study of the interplay between SFT and RL for general reasoning, offering a highly effective and generalized post-training recipe.

---

### PACED: Distillation and On-Policy Self-Distillation at the Frontier of Student Competence

Authors: Yuanda Xu, Hejian Sang, Zhengze Zhou, Ran He, Zhipeng Wang
Published: 2026-03-11
Categories: cs.AI, cs.LG
arXiv: https://arxiv.org/abs/2603.11178v3

Standard LLM distillation treats all training problems equally -- wasting compute on problems the student has already mastered or cannot yet solve. We empirically show that this inefficiency has a precise gradient-level signature: the cross-problem gradient signal-to-noise ratio (SNR) follows a bell curve over student pass rate, collapsing at both extremes. We propose PACED, which weights each problem by $w(p) = p(1{-}p)$ where $p$ is the student's empirical pass rate -- concentrating training on the zone of proximal development. This requires only student rollouts, no architectural changes, and no hyperparameters. We prove the Beta kernel $w(p) = p^α(1{-}p)^β$ is the leading-order optimal weight family arising from the SNR boundary-collapse structure, and is minimax-robust under misspecification (worst-case efficiency loss $O(δ^2)$). Across Qwen3, Qwen2.5, and Llama-3 families, PACED sets a new state of the art in our experimental setting on MATH-500, AIME~2024, and AIME~2025, improving over unweighted distillation by up to $\mathbf{+8.2}$ and over the strong AKL baseline by up to $\mathbf{+3.6}$, while reducing forgetting to $\mathbf{1.4\%}$ and $\mathbf{0.6\%}$ in distillation and self-distillation. A two-stage forward-then-reverse KL schedule pushes gains further to $\mathbf{+5.8}$ over standard forward KL on the hardest benchmark.

---

### GGPT: Geometry Grounded Point Transformer

Authors: Yutong Chen, Yiming Wang, Xucong Zhang, Sergey Prokudin, Siyu Tang
Published: 2026-03-11
Categories: cs.CV
arXiv: https://arxiv.org/abs/2603.11174v1

Recent feed-forward networks have achieved remarkable progress in sparse-view 3D reconstruction by predicting dense point maps directly from RGB images. However, they often suffer from geometric inconsistencies and limited fine-grained accuracy due to the absence of explicit multi-view constraints. We introduce the Geometry-Grounded Point Transformer (GGPT), a framework that augments feed-forward reconstruction with reliable sparse geometric guidance. We first propose an improved Structure-from-Motion pipeline based on dense feature matching and lightweight geometric optimisation to efficiently estimate accurate camera poses and partial 3D point clouds from sparse input views. Building on this foundation, we propose a geometry-guided 3D point transformer that refines dense point maps under explicit partial-geometry supervision using an optimised guidance encoding. Extensive experiments demonstrate that our method provides a principled mechanism for integrating geometric priors with dense feed-forward predictions, producing reconstructions that are both geometrically consistent and spatially complete, recovering fine structures and filling gaps in textureless areas. Trained solely on ScanNet++ with VGGT predictions, GGPT generalises across architectures and datasets, substantially outperforming state-of-the-art feed-forward 3D reconstruction models in both in-domain and out-of-domain settings.

---

### Huntington Disease Automatic Speech Recognition with Biomarker Supervision

Authors: Charles L. Wang, Cady Chen, Ziwei Gong, Julia Hirschberg
Published: 2026-03-11
Categories: cs.LG, cs.CL, cs.SD
arXiv: https://arxiv.org/abs/2603.11168v1

Automatic speech recognition (ASR) for pathological speech remains underexplored, especially for Huntington's disease (HD), where irregular timing, unstable phonation, and articulatory distortion challenge current models. We present a systematic HD-ASR study using a high-fidelity clinical speech corpus not previously used for end-to-end ASR training. We compare multiple ASR families under a unified evaluation, analyzing WER as well as substitution, deletion, and insertion patterns. HD speech induces architecture-specific error regimes, with Parakeet-TDT outperforming encoder-decoder and CTC baselines. HD-specific adaptation reduces WER from 6.99% to 4.95% and we also propose a method for using biomarker-based auxiliary supervision and analyze how error behavior is reshaped in severity-dependent ways rather than uniformly improving WER. We open-source all code and models.

---

### Learning to Unscramble: Simplifying Symbolic Expressions via Self-Supervised Oracle Trajectories

Authors: David Shih
Published: 2026-03-11
Categories: hep-th, cs.LG, cs.SC, hep-ph
arXiv: https://arxiv.org/abs/2603.11164v2

We present a new self-supervised machine learning approach for symbolic simplification of complex mathematical expressions. Training data is generated by scrambling simple expressions and recording the inverse operations, creating oracle trajectories that provide both goal states and explicit paths to reach them. A permutation-equivariant, transformer-based policy network is then trained on this data step-wise to predict the oracle action given the input expression. We demonstrate this approach on two problems in high-energy physics: dilogarithm reduction and spinor-helicity scattering amplitude simplification. In both cases, our trained policy network achieves near perfect solve rates across a wide range of difficulty levels, substantially outperforming prior approaches based on reinforcement learning and end-to-end regression. When combined with contrastive grouping and beam search, our model achieves a 100\% full simplification rate on a representative selection of 5-point gluon tree-level amplitudes in Yang-Mills theory, including expressions with over 200 initial terms.

---

### Algorithmic Task Capture, Computational Complexity, and Inductive Bias of Infinite Transformers

Authors: Orit Davidovich, Zohar Ringel
Published: 2026-03-11
Categories: cs.LG, cond-mat.dis-nn, stat.ML
arXiv: https://arxiv.org/abs/2603.11161v2

We formally define algorithmic capture of combinatorial tasks as the ability of a transformer to extrapolate to arbitrary task sizes with controllable error and logarithmic sample adaptation, providing a sharp scaling criterion for distinguishing logic internalization from statistical interpolation. Empirically, across scaling ranges spanning up to 2.5 orders of magnitude, we observe evidence of capture and non-capture. By analyzing infinite-width transformers in both the lazy and rich regimes, we derive upper bounds on the inference-time computational complexity of the combinatorial tasks these networks can capture. We show that, despite their universal expressivity, transformers possess an inductive bias that disfavors higher-complexity algorithmic procedures within the efficient polynomial-time heuristic scheme class, consistent with successful capture on simpler combinatorial tasks such as induction heads, sort, and string matching.

---

### Systematic Scaling Analysis of Jailbreak Attacks in Large Language Models

Authors: Xiangwen Wang, Ananth Balashankar, Varun Chandrasekaran
Published: 2026-03-11
Categories: cs.LG, cs.CR
arXiv: https://arxiv.org/abs/2603.11149v2

Large language models remain vulnerable to jailbreak attacks, yet we still lack a systematic understanding of how jailbreak success scales with attacker effort across methods, model families, and harm types. We initiate a scaling-law framework for jailbreaks by treating each attack as a compute-bounded optimization procedure and measuring progress on a shared FLOPs axis. Our systematic evaluation spans four representative jailbreak paradigms, covering optimization-based attacks, self-refinement prompting, sampling-based selection, and genetic optimization, across multiple model families and scales on a diverse set of harmful goals. We investigate scaling laws that relate attacker budget to attack success score by fitting a simple saturating exponential function to FLOPs--success trajectories, and we derive comparable efficiency summaries from the fitted curves. Empirically, prompting-based paradigms tend to be the most compute-efficient compared to optimization-based methods. To explain this gap, we cast prompt-based updates into an optimization view and show via a same-state comparison that prompt-based attacks more effectively optimize in prompt space. We also show that attacks occupy distinct success--stealthiness operating points with prompting-based methods occupying the high-success, high-stealth region. Finally, we find that vulnerability is strongly goal-dependent: harms involving misinformation are typically easier to elicit than other non-misinformation harms.

---

### Catalogue Grounded Multimodal Attribution for Museum Video under Resource and Regulatory Constraints

Authors: Minsak Nanang, Adrian Hilton, Armin Mustafa
Published: 2026-03-11
Categories: cs.MM, cs.CV, cs.LG
arXiv: https://arxiv.org/abs/2603.11147v3

Audiovisual (AV) archives in museums and galleries are growing rapidly, but much of this material remains effectively locked away because it lacks consistent, searchable metadata. Existing method for archiving requires extensive manual effort. We address this by automating the most labour intensive part of the workflow: catalogue style metadata curation for in gallery video, grounded in an existing collection database. Concretely, we propose catalogue-grounded multimodal attribution for museum AV content using an open, locally deployable video language model. We design a multi pass pipeline that (i) summarises artworks in a video, (ii) generates catalogue style descriptions and genre labels, and (iii) attempts to attribute title and artist via conservative similarity matching to the structured catalogue. Early deployments on a painting catalogue suggest that this framework can improve AV archive discoverability while respecting resource constraints, data sovereignty, and emerging regulation, offering a transferable template for application-driven machine learning in other high-stakes domains.

---

### Attention Gathers, MLPs Compose: A Causal Analysis of an Action-Outcome Circuit in VideoViT

Authors: Sai V R Chereddy
Published: 2026-03-11
Categories: cs.LG, cs.AI, cs.CV
arXiv: https://arxiv.org/abs/2603.11142v1

The paper explores how video models trained for classification tasks represent nuanced, hidden semantic information that may not affect the final outcome, a key challenge for Trustworthy AI models. Through Explainable and Interpretable AI methods, specifically mechanistic interpretability techniques, the internal circuit responsible for representing the action's outcome is reverse-engineered in a pre-trained video vision transformer, revealing that the "Success vs Failure" signal is computed through a distinct amplification cascade. While there are low-level differences observed from layer 0, the abstract and semantic representation of the outcome is progressively amplified from layers 5 through 11. Causal analysis, primarily using activation patching supported by ablation results, reveals a clear division of labor: Attention Heads act as "evidence gatherers", providing necessary low-level information for partial signal recovery, while MLP Blocks function as robust "concept composers", each of which is the primary driver to generate the "success" signal. This distributed and redundant circuit in the model's internals explains its resilience to simple ablations, demonstrating a core computational pattern for processing human-action outcomes. Crucially, the existence of this sophisticated circuit for representing complex outcomes, even within a model trained only for simple classification, highlights the potential for models to develop forms of 'hidden knowledge' beyond their explicit task, underscoring the need for mechanistic oversight for building genuinely Explainable and Trustworthy AI systems intended for deployment.

---

### Procedural Fairness via Group Counterfactual Explanation

Authors: Gideon Popoola, John Sheppard
Published: 2026-03-11
Categories: cs.LG, cs.AI, cs.CY
arXiv: https://arxiv.org/abs/2603.11140v1

Fairness in machine learning research has largely focused on outcome-oriented fairness criteria such as Equalized Odds, while comparatively less attention has been given to procedural-oriented fairness, which addresses how a model arrives at its predictions. Neglecting procedural fairness means it is possible for a model to generate different explanations for different protected groups, thereby eroding trust. In this work, we introduce Group Counterfactual Integrated Gradients (GCIG), an in-processing regularization framework that enforces explanation invariance across groups, conditioned on the true label. For each input, GCIG computes explanations relative to multiple Group Conditional baselines and penalizes cross-group variation in these attributions during training. GCIG formalizes procedural fairness as Group Counterfactual explanation stability and complements existing fairness objectives that constrain predictions alone. We compared GCIG empirically against six state-of-the-art methods, and the results show that GCIG substantially reduces cross-group explanation disparity while maintaining competitive predictive performance and accuracy-fairness trade-offs. Our results also show that aligning model reasoning across groups offers a principled and practical avenue for advancing fairness beyond outcome parity.

---

### H2LooP Spark Preview: Continual Pretraining of Large Language Models for Low-Level Embedded Systems Code

Authors: Amit Singh, Vedant Nipane, Pulkit Agrawal, Jatin Kishnani, Sairanjan Mishra
Published: 2026-03-11
Categories: cs.LG
arXiv: https://arxiv.org/abs/2603.11139v2

Large language models (LLMs) demonstrate strong code generation abilities in general-purpose programming languages but remain limited in specialized domains such as low-level embedded systems programming. This domain involves hardware register manipulation, vendor-specific SDKs, real-time operating system APIs, and hardware abstraction layers that are underrepresented in standard pretraining corpora. We introduce H2LooP Spark Preview, a continual pretraining (CPT) pipeline that adapts the OLMo-3-7B-a fully open language model to the embedded systems domain using BF16 LoRA with rank-stabilized scaling on 8 NVIDIA H100 GPUs. Our training corpus is constructed from repository-datasheet pairs covering 100B tokens of raw embedded systems data across 117 manufacturers, processed using the hierarchical datasheet-to-code mapping approach proposed in SpecMap (Nipane et al., 2026). The resulting curated dataset split contains 23.5B tokens across 13 embedded domains. Continual pretraining with high-rank LoRA (r=512) yields substantial gains, reducing in-domain perplexity by 70.4% and held-out repository perplexity by 66.1%. On generative code completion benchmarks spanning 13 embedded domains, our 7B model outperforms Claude Opus 4.6 and Qwen3-Coder-30B on 8 categories in token accuracy, showing that targeted continual pretraining enables smaller open-weight models to rival frontier systems on specialized technical tasks. We release the production training checkpoint on Huggingface as an open-source artifact.

---

### Deep regression learning from dependent observations with minimum error entropy principle

Authors: William Kengne, Modou Wade
Published: 2026-03-11
Categories: stat.ML, cs.LG, math.ST
arXiv: https://arxiv.org/abs/2603.11138v1

This paper considers nonparametric regression from strongly mixing observations. The proposed approach is based on deep neural networks with minimum error entropy (MEE) principle. We study two estimators: the non-penalized deep neural network (NPDNN) and the sparse-penalized deep neural network (SPDNN) predictors. Upper bounds of the expected excess risk are established for both estimators over the classes of Hölder and composition Hölder functions. For the models with Gaussian error, the rates of the upper bound obtained match (up to a logarithmic factor) with the lower bounds established in \cite{schmidt2020nonparametric}, showing that both the MEE-based NPDNN and SPDNN estimators from strongly mixing data can achieve the minimax optimal convergence rate.

---

### Scaling Reasoning Efficiently via Relaxed On-Policy Distillation

Authors: Jongwoo Ko, Sara Abdali, Young Jin Kim, Tianyi Chen, Pashmina Cameron
Published: 2026-03-11
Categories: cs.LG, cs.CL
arXiv: https://arxiv.org/abs/2603.11137v1

On-policy distillation is pivotal for transferring reasoning capabilities to capacity-constrained models, yet remains prone to instability and negative transfer. We show that on-policy distillation can be interpreted, both theoretically and empirically, as a form of policy optimization, where the teacher-student log-likelihood ratio acts as a token reward. From this insight, we introduce REOPOLD (Relaxed On-Policy Distillation) a framework that stabilizes optimization by relaxing the strict imitation constraints of standard on-policy distillation. Specifically, REOPOLD temperately and selectively leverages rewards from the teacher through mixture-based reward clipping, entropy-based token-level dynamic sampling, and a unified exploration-to-refinement training strategy. Empirically, REOPOLD surpasses its baselines with superior sample efficiency during training and enhanced test-time scaling at inference, across mathematical, visual, and agentic tool-use reasoning tasks. Specifically, REOPOLD outperforms recent RL approaches achieving 6.7~12x greater sample efficiency and enables a 7B student to match a 32B teacher in visual reasoning with a ~3.32x inference speedup.

---

### Conformal e-prediction in the presence of confounding

Authors: Vladimir Vovk, Ruodu Wang
Published: 2026-03-11
Categories: math.ST, cs.LG
arXiv: https://arxiv.org/abs/2603.11134v1

This note extends conformal e-prediction to cover the case where there is observed confounding between the random object $X$ and its label $Y$. We consider both the case where the observed data is IID and a case where some dependence between observations is permitted.

---

### Beyond Pairwise Attention: Higher-Order Modular Attention for Efficient Sequence Learning

Authors: Shirin Amiraslani, Xin Gao
Published: 2026-03-11
Categories: cs.LG
arXiv: https://arxiv.org/abs/2603.11133v2

Sequence modeling tasks can involve intrinsic higher-order dependencies, while standard self-attention assigns scores to token pairs and does not explicitly parameterize such interactions. We introduce Higher-Order Modular Attention (HOMA), which fuses pairwise attention with an explicit triadic attention pathway made tractable through overlapping blocks, local windows, and a low-rank projection. We compare HOMA with matched pairwise and purely triadic baselines on controlled PARITY and MATCH3 tasks, as well as TAPE benchmarks. HOMA is competitive with or outperforms the baselines, with its clearest advantages when the underlying dependencies extend beyond the explicitly modeled triadic order. These advantages are accompanied in several settings by faster convergence and improved parameter efficiency, with the learned nonlinear fusion providing an effective mechanism for combining the pairwise and triadic representations. Overall, our results provide empirical evidence that HOMA is an effective attention design when task structure extends beyond pairwise interactions.

---

### Higher-Order Modular Attention: Fusing Pairwise and Triadic Interactions for Protein Sequences

Authors: Shirin Amiraslani, Xin Gao
Published: 2026-03-11
Categories: cs.LG
arXiv: https://arxiv.org/abs/2603.11133v1

Transformer self-attention computes pairwise token interactions, yet protein sequence to phenotype relationships often involve cooperative dependencies among three or more residues that dot product attention does not capture explicitly. We introduce Higher-Order Modular Attention, HOMA, a unified attention operator that fuses pairwise attention with an explicit triadic interaction pathway. To make triadic attention practical on long sequences, HOMA employs block-structured, windowed triadic attention. We evaluate on three TAPE benchmarks for Secondary Structure, Fluorescence, and Stability. Our attention mechanism yields consistent improvements across all tasks compared with standard self-attention and efficient variants including block-wise attention and Linformer. These results suggest that explicit triadic terms provide complementary representational capacity for protein sequence prediction at controllable additional computational cost.

---

### WebWeaver: Breaking Topology Confidentiality in LLM Multi-Agent Systems with Stealthy Context-Based Inference

Authors: Zixun Xiong, Gaoyi Wu, Lingfeng Yao, Miao Pan, Xiaojiang Du, Hao Wang
Published: 2026-03-11
Categories: cs.CR, cs.AI
arXiv: https://arxiv.org/abs/2603.11132v2

Communication topology is a critical factor in the utility and safety of LLM-based multi-agent systems (LLM-MAS), making it a high-value intellectual property (IP) whose confidentiality remains insufficiently studied. Existing topology inference attempts rely on impractical assumptions, including control over the administrative agent and direct identity queries via jailbreaks, which are easily defeated by basic keyword-based defenses. As a result, prior analyses fail to capture the real-world threat of such attacks. To bridge this realism gap, we propose \textit{WebWeaver}, an attack framework that infers the complete LLM-MAS topology by compromising only a single arbitrary agent instead of the administrative agent. Unlike prior approaches, WebWeaver relies solely on agent contexts rather than agent IDs, enabling significantly stealthier inference. WebWeaver further introduces a new covert jailbreak-based mechanism and a novel fully jailbreak-free diffusion design to handle cases where jailbreaks fail. Additionally, we address a key challenge in diffusion-based inference by proposing a masking strategy that preserves known topology during diffusion, with theoretical guarantees of correctness. Extensive experiments show that WebWeaver substantially outperforms state-of-the-art (SOTA) baselines, achieving about 60\% higher inference accuracy under active defenses with negligible overhead.

---

### Beyond Barren Plateaus: A Scalable Quantum Convolutional Architecture for High-Fidelity Image Classification

Authors: Radhakrishnan Delhibabu
Published: 2026-03-11
Categories: cs.LG
arXiv: https://arxiv.org/abs/2603.11131v1

While Quantum Convolutional Neural Networks (QCNNs) offer a theoretical paradigm for quantum machine learning, their practical implementation is severely bottlenecked by barren plateaus -- the exponential vanishing of gradients -- and poor empirical accuracy compared to classical counterparts. In this work, we propose a novel QCNN architecture utilizing localized cost functions and a hardware-efficient tensor-network initialization strategy to provably mitigate barren plateaus. We evaluate our scalable QCNN on the MNIST dataset, demonstrating a significant performance leap. By resolving the gradient vanishing issue, our optimized QCNN achieves a classification accuracy of 98.7\%, a substantial improvement over the baseline QCNN accuracy of 52.32\% found in unmitigated models. Furthermore, we provide empirical evidence of a parameter-efficiency advantage, requiring $\mathcal{O}(\log N)$ fewer trainable parameters than equivalent classical CNNs to achieve $>95\%$ convergence. This work bridges the gap between theoretical quantum utility and practical application, providing a scalable framework for quantum computer vision tasks without succumbing to loss landscape concentration.

---

### Efficient Approximation to Analytic and $L^p$ functions by Height-Augmented ReLU Networks

Authors: ZeYu Li, FengLei Fan, TieYong Zeng
Published: 2026-03-11
Categories: stat.ML, cs.LG, cs.NE
arXiv: https://arxiv.org/abs/2603.11128v1

This work addresses two fundamental limitations in neural network approximation theory. We demonstrate that a three-dimensional network architecture enables a significantly more efficient representation of sawtooth functions, which serves as the cornerstone in the approximation of analytic and $L^p$ functions. First, we establish substantially improved exponential approximation rates for several important classes of analytic functions and offer a parameter-efficient network design. Second, for the first time, we derive a quantitative and non-asymptotic approximation of high orders for general $L^p$ functions. Our techniques advance the theoretical understanding of the neural network approximation in fundamental function spaces and offer a theoretically grounded pathway for designing more parameter-efficient networks.

---

### Enhancing Value Alignment of LLMs with Multi-agent system and Combinatorial Fusion

Authors: Yuanhong Wu, Djallel Bouneffouf, D. Frank Hsu
Published: 2026-03-11
Categories: cs.MA, cs.CL
arXiv: https://arxiv.org/abs/2603.11126v1

Aligning large language models (LLMs) with human values is a central challenge for ensuring trustworthy and safe deployment. While existing methods such as Reinforcement Learning from Human Feedback (RLHF) and its variants have improved alignment, they often rely on a single evaluator or narrowly defined reward signals, limiting their ability to capture ethical pluralism. In this work, we propose the Value Alignment System using Combinatorial Fusion Analysis (VAS-CFA), a framework that operationalizes multi-agent fusion alignment. It instantiates multiple moral agents, each fine-tuned to represent a distinct normative perspective, and fuses their outputs using CFA with both rank- and score-based aggregation. This design leverages cognitive diversity, between agents, to mitigate conflicts and redundancies across multiple agents, producing responses that better reflect human values. Empirical evaluation demonstrates that VAS-CFA outperforms both single agent baselines and prior aggregation approaches on standard metrics, showing that multi-agent fusion provides a robust and effective mechanism for advancing value alignment in LLMs.

---

### Co-Diffusion: An Affinity-Aware Two-Stage Latent Diffusion Framework for Generalizable Drug-Target Affinity Prediction

Authors: Yining Qian, Pengjie Wang, Yixiao Li, An-Yang Lu, Cheng Tan, Shuang Li, Lijun Liu
Published: 2026-03-11
Categories: stat.ML, cs.LG
arXiv: https://arxiv.org/abs/2603.11125v1

Predicting drug-target affinity is fundamental to virtual screening and lead optimization. However, existing deep models often suffer from representation collapse in stringent cold-start regimes, where the scarcity of labels and domain shifts prevent the learning of transferable pharmacophores and binding motifs. In this paper, we propose Co-Diffusion, a novel affinity-aware framework that redefines DTA prediction as a constrained latent denoising process to enhance generalization. Co-Diffusion employs a two-stage paradigm: Stage I establishes an affinity-steered latent manifold by aligning drug and target embeddings under an explicit supervised objective, ensuring that the latent space reflects the intrinsic binding landscape. Stage II introduces modality-specific latent diffusion as a stochastic perturb-and-denoise regularizer, forcing the model to recover consistent affinity semantics from noisy structural representations. This approach effectively mitigates the reconstruction-regression conflict common in generative DTA models. Theoretically, we show that Co-Diffusion maximizes a variational lower bound on the joint likelihood of drug structures, protein sequences, and binding strength. Extensive experiments across multiple benchmarks demonstrate that Co-Diffusion significantly outperforms state-of-the-art baselines, particularly yielding superior zero-shot generalization on unseen molecular scaffolds and novel protein families-paving a robust path for in silico drug prioritization in unexplored chemical spaces.

---

### Uni-ASR: Unified LLM-Based Architecture for Non-Streaming and Streaming Automatic Speech Recognition

Authors: Yinfeng Xia, Jian Tang, Junfeng Hou, Gaopeng Xu, Haitao Yao
Published: 2026-03-11
Categories: cs.SD, cs.CL
arXiv: https://arxiv.org/abs/2603.11123v1

Although the deep integration of the Automatic Speech Recognition (ASR) system with Large Language Models (LLMs) has significantly improved accuracy, the deployment of such systems in low-latency streaming scenarios remains challenging. In this paper, we propose Uni-ASR, a unified framework based on LLMs that integrates both non-streaming and streaming speech recognition capabilities. We propose a joint training paradigm that enables the system to seamlessly transition between two recognition modes without any architectural modifications. Furthermore, we introduce a context-aware training paradigm and a co-designed fallback decoding strategy, which can enhance streaming recognition accuracy without introducing additional latency. The experimental results demonstrate that Uni-ASR not only achieves competitive performance within non-streaming mode, but also demonstrates strong effectiveness in streaming scenarios under diverse latency constraints.

---

### High-resolution weather-guided surrogate modeling for data-efficient cross-location building energy prediction

Authors: Piragash Manmatharasan, Girma Bitsuamlak, Katarina Grolinger
Published: 2026-03-11
Categories: cs.LG
arXiv: https://arxiv.org/abs/2603.11121v1

Building design optimization often depends on physics-based simulation tools such as EnergyPlus, which, although accurate, are computationally expensive and slow. Surrogate models provide a faster alternative, yet most are location-specific, and even weather-informed variants require simulations from many sites to generalize to unseen locations. This limitation arises because existing methods do not fully exploit the short-term weather-driven energy patterns shared across regions, restricting their scalability and reusability. This study introduces a high-resolution (weekly) weather-informed surrogate modeling approach that enhances model reusability across locations. By capturing recurring short-term weather-energy demand patterns common to multiple regions, the proposed method produces a generalized surrogate that performs well beyond the training location. Unlike previous weather-informed approaches, it does not require extensive simulations from multiple sites to achieve strong generalization. Experimental results show that when trained on a single location, the model maintains high predictive accuracy for other sites within the same climate zone, with no noticeable performance loss, and exhibits only minimal degradation when applied across different climate zones. These findings demonstrate the potential of climate-informed generalization for developing scalable and reusable surrogate models, supporting more sustainable and optimized building design practices.

---

### Group Resonance Network: Learnable Prototypes and Multi-Subject Resonance for EEG Emotion Recognition

Authors: Renwei Meng
Published: 2026-03-11
Categories: cs.LG
arXiv: https://arxiv.org/abs/2603.11119v2

Electroencephalography (EEG)-based emotion recognition remains challenging in cross-subject settings due to severe inter-subject variability. Existing methods mainly learn subject-invariant features, but often under-exploit stimulus-locked group regularities shared across subjects. To address this issue, we propose the Group Resonance Network (GRN), which integrates individual EEG dynamics with offline group resonance modeling. GRN contains three components: an individual encoder for band-wise EEG features, a set of learnable group prototypes for prototype-induced resonance, and a multi-subject resonance branch that encodes PLV/coherence-based synchrony with a small reference set. A resonance-aware fusion module combines individual and group-level representations for final classification. Experiments on SEED and DEAP under both subject-dependent and leave-one-subject-out protocols show that GRN consistently outperforms competitive baselines, while additional analyses confirm the effects of prototype learning, PLV/coherence resonance, and leakage-safe reference construction.

---

### Group Resonance Network: Learnable Prototypes and Multi-Subject Resonance for EEG Emotion Recognition

Authors: Renwei Meng
Published: 2026-03-11
Categories: cs.LG
arXiv: https://arxiv.org/abs/2603.11119v1

Electroencephalography(EEG)-basedemotionrecognitionre- mains challenging in cross-subject settings due to severe inter-subject variability. Existing methods mainly learn subject-invariant features, but often under-exploit stimulus-locked group regularities shared across sub- jects. To address this issue, we propose the Group Resonance Network (GRN), which integrates individual EEG dynamics with offline group resonance modeling. GRN contains three components: an individual en- coder for band-wise EEG features, a set of learnable group prototypes for prototype-induced resonance, and a multi-subject resonance branch that encodes PLV/coherence-based synchrony with a small reference set. A resonance-aware fusion module combines individual and group-level rep- resentations for final classification. Experiments on SEED and DEAP under both subject-dependent and leave-one-subject-out protocols show that GRN consistently outperforms competitive baselines, while abla- tion studies confirm the complementary benefits of prototype learning and multi-subject resonance modeling.

---

### A Learning-Based Superposition Operator for Non-Renewal Arrival Processes in Queueing Networks

Authors: Eliran Sherzer
Published: 2026-03-11
Categories: cs.LG, math.PR
arXiv: https://arxiv.org/abs/2603.11118v1

The superposition of arrival processes is a fundamental yet analytically intractable operation in queueing networks when inputs are general non-renewal streams. Classical methods either reduce merged flows to renewal surrogates, rely on computationally prohibitive Markovian representations, or focus solely on mean-value performance measures. We propose a scalable data-driven superposition operator that maps low-order moments and autocorrelation descriptors of multiple arrival streams to those of their merged process. The operator is a deep learning model trained on synthetically generated Markovian Arrival Processes (MAPs), for which exact superposition is available, and learns a compact representation that accurately reconstructs the first five moments and short-range dependence structure of the aggregate stream. Extensive computational experiments demonstrate uniformly low prediction errors across heterogeneous variability and correlation regimes, substantially outperforming classical renewal-based approximations. When integrated with learning-based modules for departure-process and steady-state analysis, the proposed operator enables decomposition-based evaluation of feed-forward queueing networks with merging flows. The framework provides a scalable alternative to traditional analytical approaches while preserving higher-order variability and dependence information required for accurate distributional performance analysis.

---

### Learning Tree-Based Models with Gradient Descent

Authors: Sascha Marton
Published: 2026-03-11
Categories: cs.LG
arXiv: https://arxiv.org/abs/2603.11117v1

Tree-based models are widely recognized for their interpretability and have proven effective in various application domains, particularly in high-stakes domains. However, learning decision trees (DTs) poses a significant challenge due to their combinatorial complexity and discrete, non-differentiable nature. As a result, traditional methods such as CART, which rely on greedy search procedures, remain the most widely used approaches. These methods make locally optimal decisions at each node, constraining the search space and often leading to suboptimal tree structures. Additionally, their demand for custom training methods precludes a seamless integration into modern machine learning (ML) approaches. In this thesis, we propose a novel method for learning hard, axis-aligned DTs through gradient descent. Our approach utilizes backpropagation with a straight-through operator on a dense DT representation, enabling the joint optimization of all tree parameters, thereby addressing the two primary limitations of traditional DT algorithms. First, gradient-based training is not constrained by the sequential selection of locally optimal splits but, instead, jointly optimizes all tree parameters. Second, by leveraging gradient descent for optimization, our approach seamlessly integrates into existing ML approaches e.g., for multimodal and reinforcement learning tasks, which inherently rely on gradient descent. These advancements allow us to achieve state-of-the-art results across multiple domains, including interpretable DTs rees for small tabular datasets, advanced models for complex tabular data, multimodal learning, and interpretable reinforcement learning without information loss. By bridging the gap between DTs and gradient-based optimization, our method significantly enhances the performance and applicability of tree-based models across various ML domains.

---

### Task-Conditioned Routing Signatures in Sparse Mixture-of-Experts Transformers

Authors: Mynampati Sri Ranganadha Avinash
Published: 2026-03-11
Categories: cs.LG, cs.AI
arXiv: https://arxiv.org/abs/2603.11114v1

Sparse Mixture-of-Experts (MoE) architectures enable efficient scaling of large language models through conditional computation, yet the routing mechanisms responsible for expert selection remain poorly understood. In this work, we introduce routing signatures, a vector representation summarizing expert activation patterns across layers for a given prompt, and use them to study whether MoE routing exhibits task-conditioned structure. Using OLMoE-1B-7B-0125-Instruct as an empirical testbed, we show that prompts from the same task category induce highly similar routing signatures, while prompts from different categories exhibit substantially lower similarity. Within-category routing similarity (0.8435 +/- 0.0879) significantly exceeds across-category similarity (0.6225 +/- 0.1687), corresponding to Cohen's d = 1.44. A logistic regression classifier trained solely on routing signatures achieves 92.5% +/- 6.1% cross-validated accuracy on four-way task classification. To ensure statistical validity, we introduce permutation and load-balancing baselines and show that the observed separation is not explained by sparsity or balancing constraints alone. We further analyze layer-wise signal strength and low-dimensional projections of routing signatures, finding that task structure becomes increasingly apparent in deeper layers. These results suggest that routing in sparse transformers is not merely a balancing mechanism, but a measurable task-sensitive component of conditional computation. We release MOE-XRAY, a lightweight toolkit for routing telemetry and analysis.

---

### Partition-Based Functional Ridge Regression for High-Dimensional Data

Authors: Shaista Ashraf, Ismail Shah, Farrukh Javed
Published: 2026-03-11
Categories: stat.ME, math.ST, stat.ML
arXiv: https://arxiv.org/abs/2603.11113v1

This paper proposes a partition-based functional ridge regression framework to address multicollinearity, overfitting, and interpretability in high-dimensional functional linear models. The coefficient function vector \( \boldsymbolβ(s) \) is decomposed into two components, \( \boldsymbolβ_1(s) \) and \( \boldsymbolβ_2(s) \), representing dominant and weaker functional effects. This partition enables differential ridge penalization across functional blocks, so that important signals are preserved while less informative components are more strongly shrunk. The resulting approach improves numerical stability and enhances interpretability without relying on explicit variable selection. We develop three estimators: the Functional Ridge Estimator (FRE), the Functional Ridge Full Model (FRFM), and the Functional Ridge Sub-Model (FRSM). Under standard regularity conditions, we establish consistency and asymptotic normality for all estimators. Simulation results reveal a clear bias--variance trade-off where FRSM performs best in small samples through strong variance reduction, whereas FRFM achieves superior accuracy in moderate to large samples by retaining informative functional structure through adaptive penalization. An empirical application to Canadian weather data further demonstrates improved predictive performance, reduced variance inflation, and clearer identification of influential functional effects. Overall, partition-based ridge regularization provides a practical and theoretically grounded method for high-dimensional functional regression.

---

### ResWM: Residual-Action World Model for Visual RL

Authors: Jseen Zhang, Gabriel Adineera, Jinzhou Tan, Jinoh Kim
Published: 2026-03-11
Categories: cs.RO, cs.AI
arXiv: https://arxiv.org/abs/2603.11110v1

Learning predictive world models from raw visual observations is a central challenge in reinforcement learning (RL), especially for robotics and continuous control. Conventional model-based RL frameworks directly condition future predictions on absolute actions, which makes optimization unstable: the optimal action distributions are task-dependent, unknown a priori, and often lead to oscillatory or inefficient control. To address this, we introduce the Residual-Action World Model (ResWM), a new framework that reformulates the control variable from absolute actions to residual actions -- incremental adjustments relative to the previous step. This design aligns with the inherent smoothness of real-world control, reduces the effective search space, and stabilizes long-horizon planning. To further strengthen the representation, we propose an Observation Difference Encoder that explicitly models the changes between adjacent frames, yielding compact latent dynamics that are naturally coupled with residual actions. ResWM is integrated into a Dreamer-style latent dynamics model with minimal modifications and no extra hyperparameters. Both imagination rollouts and policy optimization are conducted in the residual-action space, enabling smoother exploration, lower control variance, and more reliable planning. Empirical results on the DeepMind Control Suite demonstrate that ResWM achieves consistent improvements in sample efficiency, asymptotic returns, and control smoothness, significantly surpassing strong baselines such as Dreamer and TD-MPC. Beyond performance, ResWM produces more stable and energy-efficient action trajectories, a property critical for robotic systems deployed in real-world environments. These findings suggest that residual action modeling provides a simple yet powerful principle for bridging algorithmic advances in RL with the practical requirements of robotics.

---

### RC-NF: Robot-Conditioned Normalizing Flow for Real-Time Anomaly Detection in Robotic Manipulation

Authors: Shijie Zhou, Bin Zhu, Jiarui Yang, Xiangyu Zhao, Jingjing Chen, Yu-Gang Jiang
Published: 2026-03-11
Categories: cs.CV, cs.RO
arXiv: https://arxiv.org/abs/2603.11106v1

Recent advances in Vision-Language-Action (VLA) models have enabled robots to execute increasingly complex tasks. However, VLA models trained through imitation learning struggle to operate reliably in dynamic environments and often fail under Out-of-Distribution (OOD) conditions. To address this issue, we propose Robot-Conditioned Normalizing Flow (RC-NF), a real-time monitoring model for robotic anomaly detection and intervention that ensures the robot's state and the object's motion trajectory align with the task. RC-NF decouples the processing of task-aware robot and object states within the normalizing flow. It requires only positive samples for unsupervised training and calculates accurate robotic anomaly scores during inference through the probability density function. We further present LIBERO-Anomaly-10, a benchmark comprising three categories of robotic anomalies for simulation evaluation. RC-NF achieves state-of-the-art performance across all anomaly types compared to previous methods in monitoring robotic tasks. Real-world experiments demonstrate that RC-NF operates as a plug-and-play module for VLA models (e.g., pi0), providing a real-time OOD signal that enables state-level rollback or task-level replanning when necessary, with a response latency under 100 ms. These results demonstrate that RC-NF noticeably enhances the robustness and adaptability of VLA-based robotic systems in dynamic environments.

---

### Thousand-GPU Large-Scale Training and Optimization Recipe for AI-Native Cloud Embodied Intelligence Infrastructure

Authors: Yongjian Guo, Yunxuan Ma, Haoran Sun, Zhong Guan, Shuai Di, Jing Long, Wanting Xu, Xiaodong Bai, Wen Huang, Yucheng Guo, Chen Zhou, Qiming Yang et al.
Published: 2026-03-11
Categories: cs.RO, cs.AI, cs.DC
arXiv: https://arxiv.org/abs/2603.11101v2

Embodied intelligence is a key step towards Artificial General Intelligence (AGI), yet its development faces multiple challenges including data, frameworks, infrastructure, and evaluation systems. To address these issues, we have, for the first time in the industry, launched a cloud-based, thousand-GPU distributed training platform for embodied intelligence, built upon the widely adopted LeRobot framework, and have systematically overcome bottlenecks across the entire pipeline. At the data layer, we have restructured the data pipeline to optimize the flow of embodied training data. In terms of training, for the GR00T-N1.5 model, utilizing thousand-GPU clusters and data at the scale of hundreds of millions, the single-round training time has been reduced from 15 hours to just 22 minutes, achieving a 40-fold speedup. At the model layer, by combining variable-length FlashAttention and Data Packing, we have moved from sample redundancy to sequence integration, resulting in a 188% speed increase; π-0.5 attention optimization has accelerated training by 165%; and FP8 quantization has delivered a 140% speedup. On the infrastructure side, relying on high-performance storage, a 3.2T RDMA network, and a Ray-driven elastic AI data lake, we have achieved deep synergy among data, storage, communication, and computation. We have also built an end-to-end evaluation system, creating a closed loop from training to simulation to assessment. This framework has already been fully validated on thousand-GPU clusters, laying a crucial technical foundation for the development and application of next-generation autonomous intelligent robots, and is expected to accelerate the arrival of the era of human-machine integration.

---

### Graph Tokenization for Bridging Graphs and Transformers

Authors: Zeyuan Guo, Enmao Diao, Cheng Yang, Chuan Shi
Published: 2026-03-11
Categories: cs.LG, cs.AI
arXiv: https://arxiv.org/abs/2603.11099v1

The success of large pretrained Transformers is closely tied to tokenizers, which convert raw input into discrete symbols. Extending these models to graph-structured data remains a significant challenge. In this work, we introduce a graph tokenization framework that generates sequential representations of graphs by combining reversible graph serialization, which preserves graph information, with Byte Pair Encoding (BPE), a widely adopted tokenizer in large language models (LLMs). To better capture structural information, the graph serialization process is guided by global statistics of graph substructures, ensuring that frequently occurring substructures appear more often in the sequence and can be merged by BPE into meaningful tokens. Empirical results demonstrate that the proposed tokenizer enables Transformers such as BERT to be directly applied to graph benchmarks without architectural modifications. The proposed approach achieves state-of-the-art results on 14 benchmark datasets and frequently outperforms both graph neural networks and specialized graph transformers. This work bridges the gap between graph-structured data and the ecosystem of sequence models. Our code is available at \href{https://github.com/BUPT-GAMMA/Graph-Tokenization-for-Bridging-Graphs-and-Transformers}{\color{blue}here}.

---

### Fingerprinting Concepts in Data Streams with Supervised and Unsupervised Meta-Information

Authors: Ben Halstead, Yun Sing Koh, Patricia Riddle, Mykola Pechenizkiy, Albert Bifet, Russel Pears
Published: 2026-03-11
Categories: cs.LG
arXiv: https://arxiv.org/abs/2603.11094v1

Streaming sources of data are becoming more common as the ability to collect data in real-time grows. A major concern in dealing with data streams is concept drift, a change in the distribution of data over time, for example, due to changes in environmental conditions. Representing concepts (stationary periods featuring similar behaviour) is a key idea in adapting to concept drift. By testing the similarity of a concept representation to a window of observations, we can detect concept drift to a new or previously seen recurring concept. Concept representations are constructed using meta-information features, values describing aspects of concept behaviour. We find that previously proposed concept representations rely on small numbers of meta-information features. These representations often cannot distinguish concepts, leaving systems vulnerable to concept drift. We propose FiCSUM, a general framework to represent both supervised and unsupervised behaviours of a concept in a fingerprint, a vector of many distinct meta-information features able to uniquely identify more concepts. Our dynamic weighting strategy learns which meta-information features describe concept drift in a given dataset, allowing a diverse set of meta-information features to be used at once. FiCSUM outperforms state-of-the-art methods over a range of 11 real world and synthetic datasets in both accuracy and modeling underlying concept drift.

---

### A Survey of Reasoning in Autonomous Driving Systems: Open Challenges and Emerging Paradigms

Authors: Kejin Yu, Yuhan Sun, Taiqiang Wu, Ruixu Zhang, Zhiqiang Lin, Yuxin Meng, Junjie Wang, Yujiu Yang
Published: 2026-03-11
Categories: cs.AI, cs.RO
arXiv: https://arxiv.org/abs/2603.11093v1

The development of high-level autonomous driving (AD) is shifting from perception-centric limitations to a more fundamental bottleneck, namely, a deficit in robust and generalizable reasoning. Although current AD systems manage structured environments, they consistently falter in long-tail scenarios and complex social interactions that require human-like judgment. Meanwhile, the advent of large language and multimodal models (LLMs and MLLMs) presents a transformative opportunity to integrate a powerful cognitive engine into AD systems, moving beyond pattern matching toward genuine comprehension. However, a systematic framework to guide this integration is critically lacking. To bridge this gap, we provide a comprehensive review of this emerging field and argue that reasoning should be elevated from a modular component to the system's cognitive core. Specifically, we first propose a novel Cognitive Hierarchy to decompose the monolithic driving task according to its cognitive and interactive complexity. Building on this, we further derive and systematize seven core reasoning challenges, such as the responsiveness-reasoning trade-off and social-game reasoning. Furthermore, we conduct a dual-perspective review of the state-of-the-art, analyzing both system-centric approaches to architecting intelligent agents and evaluation-centric practices for their validation. Our analysis reveals a clear trend toward holistic and interpretable "glass-box" agents. In conclusion, we identify a fundamental and unresolved tension between the high-latency, deliberative nature of LLM-based reasoning and the millisecond-scale, safety-critical demands of vehicle control. For future work, a primary objective is to bridge the symbolic-to-physical gap by developing verifiable neuro-symbolic architectures, robust reasoning under uncertainty, and scalable models for implicit social negotiation.

---

### Interventional Time Series Priors for Causal Foundation Models

Authors: Dennis Thumm, Ying Chen
Published: 2026-03-11
Categories: cs.LG, stat.ME
arXiv: https://arxiv.org/abs/2603.11090v2

Prior-data fitted networks (PFNs) have emerged as powerful foundation models for tabular causal inference, yet their extension to time series remains limited by the absence of synthetic data generators that provide interventional targets. Existing time series benchmarks generate observational data with ground-truth causal graphs but lack the interventional data required for training causal foundation models. To address this, we propose \textbf{CausalTimePrior}, a principled framework for generating synthetic temporal structural causal models (TSCMs) with paired observational and interventional time series. Our prior supports configurable causal graph structures, nonlinear autoregressive mechanisms, regime-switching dynamics, and multiple intervention types (hard, soft, time-varying). We demonstrate that PFNs trained on CausalTimePrior can perform in-context causal effect estimation on held-out TSCMs, establishing a pathway toward foundation models for time series causal inference.

---

### The Attack and Defense Landscape of Agentic AI: A Comprehensive Survey

Authors: Juhee Kim, Xiaoyuan Liu, Zhun Wang, Shi Qiu, Bo Li, Wenbo Guo, Dawn Song
Published: 2026-03-11
Categories: cs.CR, cs.AI
arXiv: https://arxiv.org/abs/2603.11088v1

AI agents that combine large language models with non-AI system components are rapidly emerging in real-world applications, offering unprecedented automation and flexibility. However, this unprecedented flexibility introduces complex security challenges fundamentally different from those in traditional software systems. This paper presents the first systematic and comprehensive survey of AI agent security, including an analysis of the design space, attack landscape, and defense mechanisms for secure AI agent systems. We further conduct case studies to point out existing gaps in securing agentic AI systems and identify open challenges in this emerging domain. Our work also introduces the first systematic framework for understanding the security risks and defense strategies of AI agents, serving as a foundation for building both secure agentic systems and advancing research in this critical area.

---

### Edge-Assisted Multi-Robot Visual-Inertial SLAM with Efficient Communication

Authors: Xin Liu, Shuhuan Wen, Jing Zhao, Tony Z. Qiu, Hong Zhang
Published: 2026-03-11
Categories: cs.RO, cs.CV, cs.MA
arXiv: https://arxiv.org/abs/2603.11085v1

The integration of cloud computing and edge computing is an effective way to achieve global consistent and real-time multi-robot Simultaneous Localization and Mapping (SLAM). Cloud computing effectively solves the problem of limited computing, communication and storage capacity of terminal equipment. However, limited bandwidth and extremely long communication links between terminal devices and the cloud result in serious performance degradation of multi-robot SLAM systems. To reduce the computational cost of feature tracking and improve the real-time performance of the robot, a lightweight SLAM method of optical flow tracking based on pyramid IMU prediction is proposed. On this basis, a centralized multi-robot SLAM system based on a robot-edge-cloud layered architecture is proposed to realize real-time collaborative SLAM. It avoids the problems of limited on-board computing resources and low execution efficiency of single robot. In this framework, only the feature points and keyframe descriptors are transmitted and lossless encoding and compression are carried out to realize real-time remote information transmission with limited bandwidth resources. This design reduces the actual bandwidth occupied in the process of data transmission, and does not cause the loss of SLAM accuracy caused by data compression. Through experimental verification on the EuRoC dataset, compared with the current most advanced local feature compression method, our method can achieve lower data volume feature transmission, and compared with the current advanced centralized multi-robot SLAM scheme, it can achieve the same or better positioning accuracy under low computational load.

---

### COMIC: Agentic Sketch Comedy Generation

Authors: Susung Hong, Brian Curless, Ira Kemelmacher-Shlizerman, Steve Seitz
Published: 2026-03-11
Categories: cs.CV, cs.AI, cs.CL, cs.MA, cs.NE
arXiv: https://arxiv.org/abs/2603.11048v1

We propose a fully automated AI system that produces short comedic videos similar to sketch shows such as Saturday Night Live. Starting with character references, the system employs a population of agents loosely based on real production studio roles, structured to optimize the quality and diversity of ideas and outputs through iterative competition, evaluation, and improvement. A key contribution is the introduction of LLM critics aligned with real viewer preferences through the analysis of a corpus of comedy videos on YouTube to automatically evaluate humor. Our experiments show that our framework produces results approaching the quality of professionally produced sketches while demonstrating state-of-the-art performance in video generation.

---

### LiTo: Surface Light Field Tokenization

Authors: Jen-Hao Rick Chang, Xiaoming Zhao, Dorian Chan, Oncel Tuzel
Published: 2026-03-11
Categories: cs.CV, cs.AI, cs.GR
arXiv: https://arxiv.org/abs/2603.11047v1

We propose a 3D latent representation that jointly models object geometry and view-dependent appearance. Most prior works focus on either reconstructing 3D geometry or predicting view-independent diffuse appearance, and thus struggle to capture realistic view-dependent effects. Our approach leverages that RGB-depth images provide samples of a surface light field. By encoding random subsamples of this surface light field into a compact set of latent vectors, our model learns to represent both geometry and appearance within a unified 3D latent space. This representation reproduces view-dependent effects such as specular highlights and Fresnel reflections under complex lighting. We further train a latent flow matching model on this representation to learn its distribution conditioned on a single input image, enabling the generation of 3D objects with appearances consistent with the lighting and materials in the input. Experiments show that our approach achieves higher visual quality and better input fidelity than existing methods.

---

### Neural Field Thermal Tomography: A Differentiable Physics Framework for Non-Destructive Evaluation

Authors: Tao Zhong, Yixun Hu, Dongzhe Zheng, Aditya Sood, Christine Allen-Blanchette
Published: 2026-03-11
Categories: cs.LG, cond-mat.mtrl-sci, cs.AI, cs.CV, physics.ins-det
arXiv: https://arxiv.org/abs/2603.11045v2

Inverse problems for stiff parabolic partial differential equations (PDEs), such as the inverse heat conduction problem (IHCP), are severely ill-posed: the forward map rapidly damps high-frequency interior structure before it reaches the boundary. Soft-constrained physics-informed neural networks (PINNs), which embed the PDE as a residual penalty, suffer from gradient pathology in this regime and tend to fit boundary measurements while leaving the interior field essentially untouched. We propose Neural Field Thermal Tomography (NeFTY), a hard-constrained neural field framework for label-free three-dimensional inverse heat conduction. NeFTY represents the unknown diffusivity as a continuous coordinate-based neural network, and at every optimization step passes the candidate field through a differentiable implicit-Euler heat solver with harmonic-mean interface flux, so that the governing PDE holds exactly on the discretization rather than as a soft penalty. Adjoint gradients propagate the surface reconstruction error back to the network weights at solver-level memory cost, making test-time inversion tractable on a single GPU. Across synthetic 3D benchmarks, NeFTY substantially outperforms soft-constrained PINN variants and a voxel-grid baseline on label-free volumetric recovery, and it transfers to real thermography data, surpassing classical signal-processing baselines in both defect segmentation and depth estimation. Additional details at https://cab-lab-princeton.github.io/nefty/

---

### LingDT-VL-OCR: Structure-Aware Document-Level Parsing with Fine-Grained Visual Reference

Authors: Siyi Qian, Xiongfei Bai, Bingtao Fu, Yichen Lu, Gaoyang Zhang, Xudong Yang, Peng Zhang
Published: 2026-03-11
Categories: cs.CV
arXiv: https://arxiv.org/abs/2603.11044v2

In this paper, we propose LingDT-VL-OCR, a document parsing system tailored to financial-domain documents, transforming ultra-long financial PDFs into semantically consistent, highly accurate, structured outputs with auditing-grade provenance. To address finance-specific challenges such as complex layouts, cross-page structural discontinuities, and cell-level referencing capability, LingDT-VL-OCR combines (1) a Cross-page Contents Consolidation algorithm to restore continuity across pages and a Document-level Heading Hierarchy Reconstruction (DHR) module to build a globally consistent Table of Contents (TOC) tree for structure-aware retrieval, and (2) a difficulty-adaptive curriculum learning training strategy for table parsing, together with a CellBBoxRegressor module that uses structural anchor tokens to localize table cells from decoder hidden states without external detectors. Experiments demonstrate that our model shows high performance on the Overall metric of OmniDocBench. To enable realistic evaluation in the financial vertical, we further introduce FinDocBench, a benchmark that includes six financial document categories with expert-verified annotations and evaluation metrics including Table of Contents edit-distance-based similarity (TocEDS), cross-page concatenated TEDS, and Table Cell Intersection over Union (C-IoU). We evaluate a wide range of state-of-the-art models on FinDocBench to assess their capabilities and remaining limitations on financial documents. Overall, LingDT-VL-OCR and FinDocBench provide a practical foundation for reliable downstream financial document applications.

---

### V2M-Zero: Zero-Pair Time-Aligned Video-to-Music Generation

Authors: Yan-Bo Lin, Jonah Casebeer, Long Mai, Aniruddha Mahapatra, Gedas Bertasius, Nicholas J. Bryan
Published: 2026-03-11
Categories: cs.CV, cs.AI, cs.LG, cs.MM, cs.SD
arXiv: https://arxiv.org/abs/2603.11042v2

Generating music that temporally aligns with video events is challenging for existing text-to-music models, which lack fine-grained temporal control. We introduce V2M-ZERO, a video-to-music generation approach that generates time-aligned music with disentangled time synchronization and semantic control (e.g., genre, mood) from video while requiring zero video-music pairs at training time. Our method is motivated by a key observation: temporal synchronization requires matching when and how much change occurs, not what changes. While musical and visual events differ semantically, they exhibit shared temporal structure that can be captured independently within each modality. We capture this structure through event curves computed from intra-modal similarity using pretrained music and video encoders. By measuring temporal change within each modality independently, these curves provide comparable representations across modalities. This enables a simple training strategy: fine-tune a text-to-music model on music-event curves, then substitute video-event curves at inference without cross-modal training or paired data. Across OES-Pub, MovieGenBench-Music, and AIST++, V2M-ZERO achieves state-of-the-art performance without any paired music-video data, surpassing the strongest prior baselines per metric with 5-9% higher audio quality, 13-15% better semantic alignment, 21-52% improved temporal synchronization, and 28% higher beat alignment on dance videos. We find similar results via a large crowd-source subjective listening test. Our results validate that temporal alignment through within-modality features is not only effective for video-to-music generation but also leads to better performance than paired cross-modal supervision. Furthermore, our approach enables independent controls for timing and music style (e.g., genre, mood) for more controllable generation.

---

### DynVLA: Learning World Dynamics for Action Reasoning in Autonomous Driving

Authors: Shuyao Shang, Bing Zhan, Yunfei Yan, Yuqi Wang, Yingyan Li, Yasong An, Xiaoman Wang, Jierui Liu, Lu Hou, Lue Fan, Zhaoxiang Zhang, Tieniu Tan
Published: 2026-03-11
Categories: cs.CV, cs.RO
arXiv: https://arxiv.org/abs/2603.11041v2

We propose DynVLA, a driving VLA model that introduces a new CoT paradigm termed Dynamics CoT. DynVLA forecasts compact world dynamics before action generation, enabling more informed and physically grounded decision-making. To obtain compact dynamics representations, DynVLA introduces a Dynamics Tokenizer that compresses future evolution into a small set of dynamics tokens. Considering the rich environment dynamics in interaction-intensive driving scenarios, DynVLA decouples ego-centric and environment-centric dynamics, yielding more accurate world dynamics modeling. We then train DynVLA to generate dynamics tokens before actions through SFT and RFT, improving decision quality while maintaining latency-efficient inference. Compared to Textual CoT, which lacks fine-grained spatiotemporal understanding, and Visual CoT, which introduces substantial redundancy due to dense image prediction, Dynamics CoT captures the evolution of the world in a compact, interpretable, and efficient form. Extensive experiments on NAVSIM, Bench2Drive, and a large-scale in-house dataset demonstrate that DynVLA consistently outperforms Textual CoT and Visual CoT methods, validating the effectiveness and practical value of Dynamics CoT. Project Page: https://yaoyao-jpg.github.io/dynvla.

---