Searcharxiv⌕ Search

arXiv subjects

Sen Lin

Publications and source records attributed to Sen Lin.

At least 19 recordsLinked to original sources
arXiv2026

Theory on Attention Dynamics for Out-of-Distribution In-Context Learning

Transformers have demonstrated remarkable in-context learning (ICL) capabilities, enabling them to perform new tasks without additional fine-tuning. However, their performance often deteriorates when encountering out-of-distribution (OOD) inputs that deviate from the training distribution, and the underlying theory remains poorly understood. To fill this gap, we characterize the OOD error under the input distribution shift through the interplay between the dynamics of the so-called $α$-type and $β$-type attention weights, which represent the transformer's confidence in identifying the correct and incorrect features, respectively. Our results indicate that the OOD error for each feature depends on all pairwise interactions between the training features and OOD features, and under certain cases the transformer performs no better than random guessing. To improve the OOD generalization performance, we next investigate the impact of model fine-tuning with the OOD data, and particularly, characterize the model forgetting performance on the source domain. Interestingly, the performance on the source domain may not always degrade after fine-tuning, which highly depends on the nature of the feature shift: fine-tuning on OOD domain keeps enhancing the confidence of identifying correct features from the original distribution, while the interference from other incorrect features may either increase or decrease. Extensive experiments on both synthetic and real data are conducted to corroborate the theoretical insights.

cs.LG↗
arXiv2026

From Generic to Dedicated: A Novel Optimizer for Online Continual Learning

Online Continual Learning (OCL) requires models to learn from non-stationary data streams under a strict single-pass constraint, making them highly susceptible to catastrophic forgetting. While existing studies have explored various strategies from data, architecture, and optimization perspectives, the optimizer often directly uses generic approaches such as SGD and AdamW. In this work, we reveal a new link between optimization dynamics and OCL needs by recasting Polyak-Ruppert averaging as the engine of "plasticity" and Primal averaging as the anchor of "stability". This inspires our novel optimizer, SPIN (Stability-Plasticity INterpolation), which explicitly decouples plasticity and stability by interpolating between a fast-moving "plasticity" sequence and an adaptive "stability" sequence. We theoretically demonstrate that SPIN implicitly employs an inverse Hessian approximation, providing crucial Tikhonov regularization to damp noisy, single-sample Hessian estimates and mitigate forgetting. More importantly, SPIN can be seamlessly integrated into existing OCL methods by taking resulted gradients as input and replacing standard optimizers. Extensive experiments show that using SPIN instead of standard optimizers yields consistent and significant performance gains across multiple OCL approaches, including the challenging rehearsal-free setting and under strict memory constraints.

cs.LG↗
arXiv2026

Training LLMs for Multi-Agent Collaboration with Verifiable Information Rewards

Effective collaboration among large language model (LLM) agents in multi-agent systems requires sharing relevant evidence, requesting missing information, and incorporating teammates' contributions into decisions. Training these behaviors is challenging because final task outcomes provide limited feedback on the value of individual messages. We introduce CoSIR (Collaboration via Surprisal-based Information Rewards), a post-training approach that uses a simple, verifiable environment to teach collaboration behaviors that generalize to more complex, unseen tasks. We propose a Bayesian evaluator that scores shared information by its reduction in team surprisal, defined as the sum of the negative log probabilities assigned to the correct option given each agent's available evidence. Requests are scored by the potential reduction in the same quantity if the requested information were truthfully provided. Separate rewards encourage agents to use the received information in their private choices and calculations. We post-train Qwen3-4B, Qwen3-8B, and Gemma-3-4B-it through self-play and evaluate transfer to seven unseen natural-language collaboration tasks. CoSIR achieves the highest aggregate score, while its performance in general tasks remains competitive. Behavioral analyses show increased disclosure of private evidence and fewer judged failures to use teammates' input.

cs.AI↗
arXiv2026

MeshHeal: Two-Timescale Self-Healing for Gray Failures in Decentralized LLM Agent Networks

Decentralized LLM-based multi-agent systems coordinate through local interactions, but an agent can remain responsive while its task-solving quality persistently degrades. Such gray failures require protecting current tasks before sufficient evidence exists to alter future routing, while still allowing recovered agents to rejoin. We introduce MeshHeal, a fully decentralized self-healing framework that couples ability-matched peer review across two timescales. At the fast timescale, an adaptive hierarchy escalates uncertain or low-scoring outputs from repeated single-reviewer evaluation to committee deliberation and, when needed, correction before use. At the slow timescale, a task- and ability-conditioned peer-relative detector aggregates scores to distinguish persistent degradation from ordinary output variation, trigger mandatory committee review, and eventually exclude degraded agents from ordinary routing; recovery probes provide fresh evidence for reintegration. To faithfully evaluate routing, we introduce Model-Backed MAS Evaluation, which ties ability assignments to execution models, since prompt-based ability assignments alone can leave routing errors hidden. Across BBH, MATH, and MMLU-Pro, MeshHeal achieves 0.839 degraded-phase accuracy using 51k total model tokens per task, versus the strongest baseline Symphony's 0.807 accuracy using 115k per task. Under staggered degradation and recovery, MeshHeal isolates degraded agents, keeps them excluded from ordinary task execution until recovery, and returns them to normal routing.

cs.AI↗
arXiv2026

PRIME-ANC: Path-Ratio-Informed Modeling for Efficient Neural Filter Synthesis in Active Noise Control

Changes in listener acoustics require active noise control (ANC) filters to be redesigned for new acoustic paths. We introduce PRIME-ANC, a shared neural synthesizer that learns a bounded, path-dependent logmagnitude correction to a regularized path-ratio base. Minimum-phase reconstruction and truncation produce finite-impulse-response (FIR) filters; training optimizes their noise-control performance. Across ten random training/test splits per dataset, PRIME-ANC achieves average held-out one-third-octave reductions of 18.81 and 17.76 dB over 50 Hz-5 kHz on a ten-path dataset and a public earphone database, respectively. On original earphone measurements, it improves upon the path-ratio base by 7.89 dB. Ablation studies support the contributions of both the analytic base and the path-dependent correction. Given calibrated paths for a held-out listener condition, PRIME-ANC generates a path-specific FIR without iterative optimization. With three Gauss-Newton updates, it reaches 21.43 dB reduction, approaching direct weighted least-squares design while producing lower amplification and root-mean-square control output.

eess.AS↗
arXiv2026

Rethinking Continual Anomaly Detection on the Edge: Benchmarking Under Realistic Industrial Conditions

Continual anomaly detection (CAD) addresses the need for industrial inspection systems to adapt to evolving production conditions, yet existing methods share three critical gaps: unrealistic evaluation, no systematic comparison, and no consideration of edge deployment constraints. We introduce a unified benchmark combining discrete-task evaluation on structural and logical anomalies, a novel continuous drift protocol, the first head-to-head comparison of all published CAD methods, and computational efficiency profiling on edge hardware. Our results reveal that existing CAD methods do not consistently outperform traditional approaches with simple experience replay. Thus motivated, we propose DINOSaur, a training-free method combining a frozen DINOv3 backbone with spatially-indexed coreset memory and neighborhood-restricted anomaly scoring. DINOSaur achieves zero forgetting by construction, outperforms all evaluated methods across all five protocols, and runs at sub-100\,ms inference on an NVIDIA Jetson Orin Nano, with on-device adaptation to new tasks in under 30 seconds.

cs.LG↗
arXiv2026

Reward Modeling for Reinforcement Learning-Based LLM Reasoning: Design, Challenges, and Evaluation

Large Language Models (LLMs) demonstrate transformative potential, yet their reasoning remains inconsistent and unreliable. Reinforcement learning (RL)-based fine-tuning is a key mechanism for improvement, but its effectiveness is fundamentally governed by reward design. Despite its importance, the relationship between reward modeling and core LLM challenges--such as evaluation bias, hallucination, distribution shift, and efficient learning--remains poorly understood. This work argues that reward modeling is not merely an implementation detail but a central architect of reasoning alignment, shaping what models learn, how they generalize, and whether their outputs can be trusted. We introduce Reasoning-Aligned Reinforcement Learning (RARL), a reasoning-centric taxonomic perspective that organizes diverse reward paradigms for multi-step reasoning. Within this perspective, we present a taxonomy of reward mechanisms, analyze reward hacking as a pervasive failure mode, and examine how reward signals unify challenges ranging from inference-time scaling to hallucination mitigation. We further critically evaluate existing benchmarks, highlighting vulnerabilities such as data contamination and reward misalignment, and outline directions for more robust evaluation. By integrating fragmented research threads and clarifying the interplay between reward design and fundamental reasoning capabilities, this work provides a foundational roadmap for building reasoning models that are robust, verifiable, and trustworthy.

cs.LG↗
arXiv2026

Dominant-Layer ZO: A Single Layer Dominates Zeroth-Order Fine-Tuning of LLMs

Zeroth-order (ZO) optimization enables memory-efficient fine-tuning of large language models (LLMs) using only forward passes, but it remains unclear how useful adaptation is distributed across layers. In this work, we reveal a surprising phenomenon: ZO fine-tuning is sharply dominated by a single decoding layer. Across multiple LLM families and downstream tasks, fine-tuning this dominant layer alone consistently matches or even exceeds full-model ZO fine-tuning. We further show that the dominant layer is task-agnostic but model-specific, and can be identified before training through a simple inference-only analysis of activation outliers. Specifically, the dominant layer consistently aligns with the first activation-outlier layer in the pre-trained model. To explain this phenomenon, we analyze how perturbation effects propagate under ZO optimization. We find that the dominant layer combines two key properties: high perturbation sensitivity and early placement in the residual stream, allowing perturbation-induced effects to propagate and accumulate through remaining subsequent decoding layers. As a result, this layer produces disproportionately strong and stable optimization signals under forward-only updates. Extensive experiments on LLaMA2-7B and Qwen3-8B across nine benchmarks show that dominant-layer ZO fine-tuning improves average performance over full-model MeZO and LoRA-based ZO fine-tuning while achieving up to 4.52$\times$ training speedup.

cs.LG↗
arXiv2026

HIPO: Instruction Hierarchy via Constrained Reinforcement Learning

Hierarchical Instruction Following (HIF) refers to the problem of prompting large language models with a priority-ordered stack of instructions. Standard methods like RLHF and DPO typically fail in this problem since they mainly optimize for a single objective, failing to explicitly enforce system prompt compliance. Meanwhile, supervised fine-tuning relies on mimicking filtered, compliant data, which fails to establish the priority asymmetry at the algorithmic level. In this paper, we introduce \textsc{HIPO}, a novel alignment framework that formulates HIF as a Constrained Markov Decision Process. \textsc{HIPO} elevates system prompts from mere input context to strict algorithmic boundaries. Using a primal-dual safe reinforcement learning approach, the algorithm dynamically enforces system prompt compliance as an explicit constraint, maximizing user utility strictly within this feasible region. Extensive evaluations across diverse model architectures (e.g., Qwen, Phi, Llama) demonstrate that \textsc{HIPO} significantly improves both system compliance and user utility. Furthermore, mechanistic analysis reveals that this constrained optimization autonomously drives the model to shift its attention toward long-range system tokens, providing a principled foundation for reliable LLM deployment in complex workflows.

cs.LG↗
arXiv2026

More Than Memory Savings: Zeroth-Order Optimization Mitigates Forgetting in Continual Learning

Zeroth-order (ZO) optimization has gained attention as a memory-efficient alternative to first-order (FO) methods, particularly in settings where gradient computation is expensive or even impractical. Beyond its memory efficiency, in this work, we investigate ZO optimization for continual learning (CL) as a novel approach to address the plasticity-stability-efficiency trilemma. Through theoretical analysis and empirical evidence, we show that ZO optimization naturally leads to flatter loss landscapes, which in turn reduce forgetting in CL. However, this stability comes at a cost of plasticity: due to its imprecise gradient estimates and slower convergence, ZO optimization tends to be less effective than FO in acquiring new task-specific knowledge, particularly under constrained training budgets. To better understand this trade-off, we conduct a holistic evaluation of ZO optimization applied to various existing CL methods. Our findings reveal that ZO optimization enhances stability but often undermines plasticity, particularly when used with learnable classifiers. Motivated by this insight, we propose ZO-FC, a simple but effective approach that applies ZO optimization to a single adapter-based PEFT module with FO optimized classifier. This design leverages the stability benefits of ZO while preserving the adaptability of FO updates with negligible memory overhead. Experiments demonstrate that ZO-FC achieves an effective balance between stability and plasticity, offering a practical and memory-efficient solution for on-device CL.

cs.LG↗
arXiv2026

Rethinking Continual Learning with Progressive Neural Collapse

Continual Learning (CL) seeks to build an agent that can continuously learn a sequence of tasks, where a key challenge, namely Catastrophic Forgetting, persists due to the potential knowledge interference among different tasks. On the other hand, deep neural networks (DNNs) are shown to converge to a terminal state termed Neural Collapse during training, where all class prototypes geometrically form a static simplex equiangular tight frame (ETF). These maximally and equally separated class prototypes make the ETF an ideal target for model learning in CL to mitigate knowledge interference. Thus inspired, several studies have emerged very recently to leverage a fixed global ETF in CL, which however suffers from key drawbacks, such as impracticability and limited performance.To address these challenges and fully unlock the potential of ETF in CL, we propose Progressive Neural Collapse (ProNC), a novel framework that completely removes the need of a fixed global ETF in CL. Specifically, ProNC progressively expands the ETF target in a principled way by adding new class prototypes as vertices for new tasks, ensuring maximal separability across all encountered classes with minimal shifts from the previous ETF. We next develop a new CL framework by plugging ProNC into commonly used CL algorithm designs, where distillation is further leveraged to balance between target shifting for old classes and target aligning for new classes. Extensive experiments show that our approach significantly outperforms related baselines while maintaining superior flexibility, simplicity, and efficiency. Our code is available at https://github.com/Continue-Edge-AI-Lab/ProNC

cs.LG↗
arXiv2026

Constraint-Rectified Training for Efficient Chain-of-Thought

Chain-of-Thought (CoT) has significantly enhanced the reasoning capabilities of Large Language Models (LLMs), especially when combined with reinforcement learning (RL) based post-training methods. While longer reasoning traces can improve answer quality and unlock abilities such as self-correction, they also incur high inference costs and often introduce redundant steps, known as overthinking. Recent research seeks to develop efficient reasoning strategies that balance reasoning length and accuracy, either through length-aware reward design or prompt-based calibration. However, these heuristic-based approaches may suffer from severe accuracy drop and be very sensitive to hyperparameters. To address these problems, we introduce CRT (Constraint-Rectified Training), a principled post-training framework based on reference-guarded constrained optimization, yielding a more stable and interpretable formulation for efficient reasoning. CRT alternates between minimizing reasoning length and rectifying accuracy only when performance falls below the reference, enabling stable and effective pruning of redundant reasoning. We further extend CRT with a two-stage training scheme that first discovers the shortest reliable reasoning patterns and then refines accuracy under a learnt length budget, preventing the re-emergence of verbose CoT. Our comprehensive evaluation shows that this framework consistently reduces token usage while maintaining answer quality at a robust and reliable level. Further analysis reveals that CRT improves reasoning efficiency not only by shortening responses but also by reducing internal language redundancy, leading to a new evaluation metric. Moreover, CRT-based training naturally yields a sequence of intermediate checkpoints that span a spectrum of explanation lengths while preserving correctness, enabling fine-grained control over reasoning verbosity without retraining.

cs.LG↗
arXiv2026

A Comprehensive Survey of Mixture-of-Experts: Algorithms, Theory, and Applications

Artificial intelligence (AI) has achieved astonishing successes in many domains, especially with the recent breakthroughs in the development of foundational large models. These large models, leveraging their extensive training data, provide versatile solutions for a wide range of downstream tasks. However, as modern datasets become increasingly diverse and complex, the development of large AI models faces two major challenges: (1) the enormous consumption of computational resources and deployment difficulties, and (2) the difficulty in fitting heterogeneous and complex data, which limits the usability of the models. Mixture of Experts (MoE) models has recently attracted much attention in addressing these challenges, by dynamically selecting and activating the most relevant sub-models to process input data. It has been shown that MoEs can significantly improve model performance and efficiency with fewer resources, particularly excelling in handling large-scale, multimodal data. Given the tremendous potential MoE has demonstrated across various domains, it is urgent to provide a comprehensive summary of recent advancements of MoEs in many important fields. Existing surveys on MoE have their limitations, e.g., being outdated or lacking discussion on certain key areas, and we aim to address these gaps. In this paper, we first introduce the basic design of MoE, including gating functions, expert networks, routing mechanisms, training strategies, and system design. We then explore the algorithm design of MoE in important machine learning paradigms such as continual learning, meta-learning, multi-task learning, and reinforcement learning. Additionally, we summarize theoretical studies aimed at understanding MoE and review its applications in computer vision and natural language processing. Finally, we discuss promising future research directions.

cs.LG↗
arXiv2026

Towards Fast Safe Online Reinforcement Learning via Policy Finetuning

The high costs and risks involved in extensive environment interactions hinder the practical application of current online safe reinforcement learning (RL) methods. While offline safe RL addresses this by learning policies from static datasets, the performance therein is usually limited due to reliance on data quality and challenges with out-of-distribution (OOD) actions. Inspired by recent successes in offline-to-online (O2O) RL, it is crucial to explore whether offline safe RL can be leveraged to facilitate faster and safer online policy learning, a direction that has yet to be fully investigated. To fill this gap, we first demonstrate that naively applying existing O2O algorithms from standard RL would not work well in the safe RL setting due to two unique challenges: \emph{erroneous Q-estimations}, resulted from offline-online objective mismatch and offline cost sparsity, and \emph{Lagrangian mismatch}, resulted from difficulties in aligning Lagrange multipliers between offline and online policies. To address these challenges, we introduce \textbf{Marvel}, a novel framework for O2O safe RL, comprising two key components that work in concert: \emph{Value Pre-Alignment} to align the Q-functions with the underlying truth before online learning, and \emph{Adaptive PID Control} to effectively adjust the Lagrange multipliers during online finetuning. Extensive experiments demonstrate that Marvel significantly outperforms existing baselines in both reward maximization and safety constraint satisfaction. By introducing the first policy-finetuning based framework for O2O safe RL, which is compatible with many offline and online safe RL methods, our work has the great potential to advance the field towards more efficient and practical safe RL solutions.

cs.LG↗
arXiv2025

Instantaneous velocity during quantum tunnelling

Quantum tunnelling, a hallmark phenomenon of quantum mechanics, allows particles to pass through the classically forbidden region. It underpins fundamental processes ranging from nuclear fusion and photosynthesis to the operation of superconducting qubits. Yet the underlying dynamics of particle motion during tunnelling remain subtle and are still the subject of active debate. Here, by analyzing the temporal evolution of the tunnelling process, we show that the particle velocity inside the barrier continuously relaxes from a large initial value toward a smaller one, and may even approach zero in the evanescent regime. Meanwhile, the probability density within the barrier gradually builds up before reaching its stationary profile, in contrast to existing inherently. In addition, starting from the steady-state equations, we derive an explicit relation between the particle velocity and the barrier width, and show that the velocity in evanescent states approaches zero when the barrier is sufficiently wide. These findings resolve the apparent paradox of a vanishing steady-state velocity coexisting with a finite particle density. We point out that defining an effective speed from the probability density, rather than from the probability current, can lead to spuriously nonzero "stationary speed," as appears to be the case in Ref. [Nature 643, 67 (2025)]. Our work establishes a clear dynamical picture for the formation of tunnelling flow and provides a theoretical foundation for testing time-resolved tunnelling phenomena.

quant-ph↗
arXiv2025

Uncertainty Guided Online Ensemble for Non-stationary Data Streams in Fusion Science

Machine Learning (ML) is poised to play a pivotal role in the development and operation of next-generation fusion devices. Fusion data shows non-stationary behavior with distribution drifts, resulted by both experimental evolution and machine wear-and-tear. ML models assume stationary distribution and fail to maintain performance when encountered with such non-stationary data streams. Online learning techniques have been leveraged in other domains, however it has been largely unexplored for fusion applications. In this paper, we present an application of online learning to continuously adapt to drifting data stream for prediction of Toroidal Field (TF) coils deflection at the DIII-D fusion facility. The results demonstrate that online learning is critical to maintain ML model performance and reduces error by 80% compared to a static model. Moreover, traditional online learning can suffer from short-term performance degradation as ground truth is not available before making the predictions. As such, we propose an uncertainty guided online ensemble method to further improve the performance. The Deep Gaussian Process Approximation (DGPA) technique is leveraged for calibrated uncertainty estimation and the uncertainty values are then used to guide a meta-algorithm that produces predictions based on an ensemble of learners trained on different horizon of historical data. The DGPA also provides uncertainty estimation along with the predictions for decision makers. The online ensemble and the proposed uncertainty guided online ensemble reduces predictions error by about 6%, and 10% respectively over standard single model based online learning.

cs.LG↗
arXiv2025

Mixture-of-Transformers Learn Faster: A Theoretical Study on Classification Problems

Mixture-of-Experts (MoE) models improve transformer efficiency but lack a unified theoretical explanation, especially when both feed-forward and attention layers are allowed to specialize. To this end, we study the Mixture-of-Transformers (MoT), a tractable theoretical framework in which each transformer block acts as an expert governed by a continuously trained gating network. This design allows us to isolate and study the core learning dynamics of expert specialization and attention alignment. In particular, we develop a three-stage training algorithm with continuous training of the gating network, and show that each transformer expert specializes in a distinct class of tasks and that the gating network accurately routes data samples to the correct expert. Our analysis shows how expert specialization reduces gradient conflicts and makes each subtask strongly convex. We prove that the training drives the expected prediction loss to near zero in $O(\log(ε^{-1}))$ iteration steps, significantly improving over the $O(ε^{-1})$ rate for a single transformer. We further validate our theoretical findings through extensive real-data experiments, demonstrating the practical effectiveness of MoT. Together, these results offer the first unified theoretical account of transformer-level specialization and learning dynamics, providing practical guidance for designing efficient large-scale models.

cs.LG↗
arXiv2025

Knowledge-Guided Machine Learning for Stabilizing Near-Shortest Path Routing

We propose a simple algorithm that needs only a few data samples from a single graph for learning local routing policies that generalize across a rich class of geometric random graphs in Euclidean metric spaces. We thus solve the all-pairs near-shortest path problem by training deep neural networks (DNNs) that let each graph node efficiently and scalably route (i.e., forward) packets by considering only the node's state and the state of the neighboring nodes. Our algorithm design exploits network domain knowledge in the selection of input features and design of the policy function for learning an approximately optimal policy. Domain knowledge also provides theoretical assurance that the choice of a ``seed graph'' and its node data sampling suffices for generalizable learning. Remarkably, one of these DNNs we train -- using distance-to-destination as the only input feature -- learns a policy that exactly matches the well-known Greedy Forwarding policy, which forwards packets to the neighbor with the shortest distance to the destination. We also learn a new policy, which we call GreedyTensile routing -- using both distance-to-destination and node stretch as the input features -- that almost always outperforms greedy forwarding. We demonstrate the explainability and ultra-low latency run-time operation of Greedy Tensile routing by symbolically interpreting its DNN in low-complexity terms of two linear actions.

cs.LG↗