arXiv subjects

Yifan Sun

Publications and source records attributed to Yifan Sun.

At least 19 recordsLinked to original sources

arXiv2026

PACE: Plug-and-Play Contextual Embedding for Feature Screening with Pretrained Tabular Foundation Models

In high-dimensional tabular learning, feature screening provides a lightweight, model-agnostic way to remove irrelevant features before model fitting. However, scoring raw values directly can miss nonlinear or distributional structure. We introduce PACE (Plug-and-Play Contextual Embedding), which inserts a frozen tabular foundation model (TFM) column encoder before an existing feature-scoring rule, expanding each feature into a higher-dimensional contextual representation. Across controlled studies, PACE improves raw-space screening of complex nonlinear dependence with only modest additional encoding cost. These gains translate to downstream prediction on TALENT datasets: PACE-DC improves binary AUC by 0.077 and multiclass macro-AUC by 0.064, with a median normalized RMSE improvement of 0.063 across ten learners. Matched random-weight and random-feature controls show that PACE gains from pretrained structure beyond generic dimensional expansion. PACE further achieves favorable performance--time trade-offs against task-fitted selectors and attribution-based methods, positioning pretrained column geometry as a reusable upstream primitive for tabular learning.

stat.ML↗

arXiv2026

PixelFlow: Token-Level Workload Management for Efficient Distributed DiT Serving

Online image generation with Diffusion Transformers (DiTs) must meet latency service-level objectives (SLOs) while using GPU resources efficiently. Existing systems improve GPU utilization by batching multiple requests for joint execution. However, request-level batching offers limited control over batch size: batches may be too small to saturate GPU compute, while larger ones may violate latency SLOs. Globally coordinated scheduling introduces further delays by requiring independently progressing GPUs to synchronize before admitting new work. We present PixelFlow, a distributed DiT serving system that addresses these limitations through token-level workload management. Its key idea is to use image tokens (the units a DiT processes to generate an image) to divide and batch request workloads at a finer granularity. This allows each GPU to take on a portion of additional work under latency constraints. By distributing these portions across GPUs, PixelFlow accommodates more concurrent requests, improving GPU utilization while reducing queueing delays. To realize this flexibility, PixelFlow provides a runtime that splits requests into variable-sized partitions and batches them efficiently on each GPU. To reduce the resulting communication overhead, it optimizes token placement to limit cross-GPU data exchange while balancing GPU workloads. It further exploits similarity across denoising steps to overlap the remaining transfers with computation. An SLO-aware scheduler groups GPUs to share compute resources among requests with compatible latency requirements. Each group progresses independently, synchronizing with others only when their combined resources are needed to admit a new request. Evaluation with Stable Diffusion 3 and FLUX.1-dev on H100 GPUs shows that PixelFlow improves SLO attainment by up to 43% and achieves up to 2.8 times the goodput of state-of-the-art DiT serving systems.

cs.DC↗

arXiv2026

MiCRo: Mixture Modeling and Context-aware Routing for Personalized Preference Learning

Reward modeling is a key step in building safe foundation models when applying reinforcement learning from human feedback (RLHF) to align Large Language Models (LLMs). However, reward modeling based on the Bradley-Terry (BT) model assumes a global reward function, failing to capture the inherently diverse and heterogeneous human preferences. Hence, such oversimplification limits LLMs from supporting personalization and pluralistic alignment. Theoretically, we show that when human preferences follow a mixture distribution of diverse subgroups, a single BT model has an irreducible error. While existing solutions, such as multi-objective learning with fine-grained annotations, help address this issue, they are costly and constrained by predefined attributes, failing to fully capture the richness of human values. In this work, we introduce MiCRo, a two-stage framework that enhances personalized preference learning by leveraging large-scale binary preference datasets without requiring explicit fine-grained annotations. In the first stage, MiCRo introduces context-aware mixture modeling approach to capture diverse human preferences. In the second stage, MiCRo integrates an online routing strategy that dynamically adapts mixture weights based on specific context to resolve ambiguity, allowing for efficient and scalable preference adaptation with minimal additional supervision. Experiments on multiple preference datasets demonstrate that MiCRo effectively captures diverse human preferences and significantly improves downstream personalization.

cs.AI↗

arXiv2026

Cascade: Hierarchical Recoverability Control for Large Language Model Unlearning

Large Language Model (LLM) unlearning is essential for removing sensitive or copyrighted knowledge while preserving general utility. Existing methods often leave residual knowledge in intermediate representations, which can still be recovered. To address this, we propose Cascade, a hierarchical recoverability control framework that minimizes the internal identifiability of target knowledge. Cascade combines three complementary controls: path-level routing to suppress privacy-associated activation routes, representation-level compression to reduce geometric separability, and decoding-level intervention to limit residual recovery. Experiments on TOFU, MUSE-News, and WMDP, including robustness tests with query reformulation and extraction-style prompts, show that Cascade effectively reduces recoverability while maintaining stable model utility.

cs.CL↗

arXiv2026

GEAR: Generative Expansion and Real Anchoring for Two-Stage Distillation of Tabular Foundation Models

Tabular foundation models (TFMs) achieve strong performance through in-context learning, but context-dependent inference imposes substantial latency and memory costs, hindering large-scale deployment. We propose GEAR (\emph{Generative Expansion and Real Anchoring}), a modular two-stage framework that distills TFMs into lightweight MLP or tree-based predictors that can be deployed on commodity CPUs. Stage 1 uses synthetic covariates solely as teacher-query locations and trains the student on soft TFM targets, expanding coverage beyond observed rows. Stage 2 re-anchors the student to the target distribution using real labels and out-of-fold teacher predictions, whitch avoids self-labeling leakage. We further derive a risk certificate characterizing the trade-off between generated-query volume and generator fidelity. Experiments on TALENT and TabArena demonstrate the broad applicability of GEAR. Two-stage MLPs outperform supervised MLPs by 1.81--2.00 AUC points on binary tasks and 1.19--1.35 points on multiclass tasks, with additional gains over real-data-only distillation of 1.76--2.19 and 2.09--2.40 points, respectively. On binary tasks, the gains also transfer to LightGBM and XGBoost, and all three student families outperform CatBoost, the strongest non-TFM baseline, in mean AUC. Ablations show gains beyond longer training or alternative warm starts, greater stability from staged than mixed optimization, and generator-dependent diminishing returns as query volume increases. Finally, GEAR reduces median inference time by 57--2866 times and peak prediction memory by 1.9--3.3 times, while retaining higher AUC than matched supervised baselines.

cs.LG↗

arXiv2026

Mask-Free Privacy Extraction and Rewriting: A Domain-Aware Approach via Prototype Learning

Client-side privacy rewriting is crucial for deploying LLMs in privacy-sensitive domains. However, existing approaches struggle to balance privacy and utility. Full-text methods often distort context, while span-level approaches rely on impractical manual masks or brittle static dictionaries. Attempts to automate localization via prompt-based LLMs prove unreliable, as they suffer from unstable instruction following that leads to privacy leakage and excessive context scrubbing. To address these limitations, we propose DAMPER (Domain-Aware Mask-free Privacy Extraction and Rewriting). DAMPER operationalizes latent privacy semantics into compact Domain Privacy Prototypes via contrastive learning, enabling precise, autonomous span localization. Furthermore, we introduce a Prototype-Guided Preference Alignment, which leverages learned prototypes as semantic anchors to construct preference pairs, optimizing a domain-compliant rewriting policy without human annotations. At inference time, DAMPER integrates a sampling-based Exponential Mechanism to provide rigorous span-level Differential Privacy (DP) guarantees. Extensive experiments demonstrate that DAMPER significantly outperforms existing baselines, achieving a superior privacy-utility trade-off.

cs.CR↗

arXiv2026

MOON: Multi-Objective OrthoNormalized Updates for Multitask Learning

Multi-objective optimization (MOO) has demonstrated significant success in multi-task learning by mitigating task conflicts through gradient manipulation. However, most existing methods flatten model parameters into vectors and perform gradient manipulation under Euclidean geometry, thereby overlooking the matrix structure prevalent in modern architectures such as Transformers. In this paper, we show that gradient manipulation in Euclidean space does not generally yield the steepest descent direction under matrix geometry, potentially limiting optimization efficiency. Drawing from the theory of steepest descent for matrix-valued parameters, we propose MOON (Multi-Objective OrthoNormalized Updates), which performs gradient manipulation under spectral--nuclear norm geometry and uses the orthonormalized manipulated gradient for parameter updates. Theoretically, for smooth non-convex objectives, we establish convergence of the averaged Pareto-stationarity measure at rates of $\mathcal{O}(T^{-1/2})$ in the deterministic setting and $\mathcal{O}(T^{-1/4})$ under stochastic gradients. Empirical results across various benchmarks show that MOON consistently improves both optimization efficiency and final multi-task performance. Our code is available at https://github.com/KunlinLyu/MOON.

cs.LG↗

arXiv2026

MonitorBench: A Comprehensive Benchmark for Chain-of-Thought Monitorability in Large Language Models

Large language models (LLMs) can generate chains of thought (CoTs) that are not always causally responsible for their final outputs. When such a mismatch occurs, the CoT no longer faithfully reflects the actual reasons (i.e., decision-critical factors) driving the model's behavior, leading to the reduced CoT monitorability problem. This limits the use of CoTs for reliable oversight. However, a comprehensive and fully open-source benchmark for thoroughly evaluating CoT monitorability remains lacking. To address this gap, we propose MonitorBench, a systematic benchmark for evaluating CoT monitorability in LLMs. MonitorBench provides: (1) a diverse set of 1,514 test instances with carefully designed decision-critical factors across 19 tasks spanning 7 categories to characterize when CoTs can be used to monitor the factors driving LLM behavior; and (2) two prompting stress-test settings to quantify the extent to which CoT monitorability can be degraded. Extensive experiments show that CoT monitorability is a conditional property affected by the evaluated LLM, monitor LLM, and task characteristics. Across these factors, monitorability is higher when decision-critical factors shape the intermediate reasoning process, rather than merely influencing the final answer. Under stress-test prompting, most evaluated LLMs can intentionally reduce monitorability, mainly on tasks where decision-critical factors are not structurally required by the reasoning process. Overall, MonitorBench provides a basis for further research on AI control, reasoning faithfulness, stress-test monitorability, and monitoring scaffords. The code is available at https://github.com/ASTRAL-Group/MonitorBench.

cs.AI↗

arXiv2026

Understanding Reasoning from Pretraining to Post-Training

Reinforcement learning (RL) has become central to improving large language models (LLMs) on complex reasoning tasks, yet RL post-training is largely studied in isolation from the pretraining that precedes it. As a result, two basic questions remain open: (1) how do pretraining choices (model size, data) shape the returns to RL compute, and (2) what does RL actually do to the model? These questions are difficult to study in the standard LLM setting: pretraining corpora are vast and uncontrolled, making it hard to attribute behaviors to pretraining versus RL, and systematic compute sweeps across both stages are prohibitively expensive. To address these challenges, we use chess as a controlled testbed for studying reasoning across the full pretraining-to-post-training pipeline. We follow the standard LLM training pipeline by pretraining language models from 5M to 1B parameters on human chess games, supervised fine-tuning on synthetic reasoning traces, and running RL on chess puzzles with verifiable rewards. Using this framework, we find that the post-RL performance at given RL compute level is well-predicted from the pretraining loss, and slope of the RL reward curves improves approximately linearly with the pretraining tokens. Beyond scaling, we find that RL does not simply sharpen the SFT policy: on easy puzzles it amplifies correct moves the SFT policy already preferred, while on hard puzzles it surfaces correct moves that were nearly absent under SFT. We further test whether our findings transfer beyond chess by training a 1B language model on math-domain text, where the same predictive pattern emerges: longer-pretrained checkpoints reach higher post-RL performance and improve faster under RL. In sum, we provide a quantitative account of the pretraining-to-RL interface and a controlled testbed for studying the science of reasoning across the full pretraining-to-post-training pipeline.

cs.LG↗

arXiv2026

Mitigating Spectral Bias in Neural Operators for Underwater Transmission Loss Prediction

Predicting underwater acoustic transmission loss rapidly and accurately is crucial for real-time ocean acoustic applications. While Fourier Neural Operators (FNO) have emerged as powerful surrogate models due to their global receptive fields, they suffer from spectral bias. The frequency truncation mechanism in FNO filters out high-frequency components, resulting in over-smoothed predictions that fail to capture fine-grained interference patterns. To overcome this limitation, this paper proposes a Spectral-Spatial Residual Learning (S2RL) framework. S2RL decomposes the prediction task into a coarse-to-fine process: a spectral Global Propagator first generates a globally consistent prediction, and a spatial Local Refiner subsequently recovers the high-frequency residuals. Experimental results on a South China Sea dataset show that the proposed method significantly outperforms FNO baselines while maintaining millisecond-level inference speeds.

cs.SD↗

arXiv2026

TheBotCompany: Self-Organizing Multi-agent Systems for Continuous Software Development

Large language model (LLM)-based multi-agent systems have shown promise in automating software development tasks. However, most vibe-coding systems focus on completing small tasks and incremental code changes, leaving persistent, continuous software development largely unexplored. We present TheBotCompany, an open-source orchestration framework for continuous multi-agent software development. TheBotCompany introduces three key innovations: (1) a three-phase state machine (Strategy to Execution to Verification) for milestone-driven development, (2) self-organizing agent teams where manager agents dynamically hire, assign, and retire worker agents based on project needs, and (3) asynchronous human oversight. We evaluate TheBotCompany on real-world software projects over multiple days of continuous development, measuring team adaptation patterns, milestone completion rates, cost efficiency, and code quality. Our results demonstrate that the self-organizing approach enables effective long-term software development with measurable progress, while the verification phase catches defects that would otherwise persist.

cs.SE↗

arXiv2026

Towards Privacy-Preserving Federated Prompt Tuning under Data Heterogeneity: A Subspace-Decomposed Expert Approach

Federated prompt tuning (FPT) enables collaborative adaptation of vision--language models (VLMs) using lightweight prompts. Existing methods often address heterogeneity and privacy through a split-prompt design under local differential privacy (DP), combining a shared prompt for global transfer with private prompts for local adaptation. However, a single shared prompt may over-smooth diverse transferable knowledge, weakening the balance between personalization and generalization. Multi-expert prompts (MEPs) can better capture this diversity, but enlarge the communicated space, increasing DP noise and communication cost while making robust expert composition more difficult. We propose FedSEPT, a privacy-preserving Fed}erated Subspace-decomposed Expert Prompt Tuning. Specifically, we employ Subspace-decomposed Expert Modeling (SEM) to parameterize multiple prompt experts with shared low-rank factors, a fixed public basis, and private residuals, thereby confining communication and DP perturbation to a compact factor space while enabling direct server aggregation in a common coordinate system. We further design Instance-aware Expert Fusion (IEF), which adaptively combines semantically complementary experts via on-device routing and performs efficient logit-level fusion using cached expert-specific text features. Extensive experiments on 11 heterogeneous benchmarks show that, under the same privacy constraints, FedSEPT achieves a better trade-off between local adaptation and global generalization than strong baselines.

cs.CV↗

arXiv2026

Parametrically driven Kerr temporal soliton crystals

We theoretically investigate the dynamics of parametrically driven soliton crystals (PDSC) and their associated frequency combs in doubly resonant cavities with quadratic and cubic nonlinearities. We show that, in a regime with strong pump--signal walk-off where the homogeneous state is unstable, noise-seeded dynamics relaxes toward stable, equally spaced multi-soliton states. At fixed detuning, the driving strength acts as the primary control parameter for the soliton number. Pump signal walkoff extends pump depletion from a local perturbation to a global constraint, enabling long-range soliton interactions; in combination with the pump phase, this mechanism stabilizes the crystal's equal spacing and preserves comb coherence. Additionally, we identify a novel nonlinear state in parametric soliton crystals in which circulating solitons periodically alternate their intensities and group velocities, a phenomenon we term the {\it soliton-pursuing} state. Furthermore, due to the phase-selective nature of the optical parametric process, we show how different configurations of soliton phases determine the optical frequency combs, enabling odd-harmonic and subharmonic-like combs.

physics.optics↗

arXiv2026

ReflectWorld-MM: An Entity-Oriented Multimodal Memory System for Open-Ended Video Streams

Building assistants that can continually watch the world, remember what they see, and reason over their accumulated experience is a long-standing goal, and recently multimodal agents equipped with long-term memory over video streams have attracted increasing interest. Unfortunately, existing systems either keep their memory inside the model context or in a flat feature store, and organize it around frames rather than around the persistent entities a stream is really about, which confines them to bounded videos and weakens their ability to track who and what reappears over time. In this paper, we propose ReflectWorld-MM, an entity-oriented multimodal memory system for open-ended video streams. It consists of three parts. The first is a perception front-end that turns an audiovisual stream into entity-resolved observations under a bounded short-term memory. The second is a hierarchical long-term memory, grounded in human memory theory, that couples a multi-scale episodic memory, an evolving entity-centric semantic memory, and a procedural memory. The third is a complete realization, built for real-world operation, that ingests arbitrary streams and plugs into off-the-shelf assistants. Across six long-video and lifelong-memory benchmarks, ReflectWorld-MM achieves the best accuracy on all six, outperforming strong memory agents and a frontier model.

cs.CV↗

arXiv2026

ArchSim: Computer Architecture Simulation as a Service

Conducting a complete computer architecture simulation study is challenging because configuration, execution, and analysis are often encoded implicitly in scripts or directory conventions rather than represented explicitly. As a result, studies are difficult to scale, hard to reproduce, and dependent on custom tooling at every stage. We present ArchSim, which makes the structure of a simulation study explicit. In ArchSim, hardware topologies are described as declarative graphs that automatically generate executable simulation code, eliminating hand-written simulator programs. Stateless runners autonomously claim and execute jobs from a shared experiment store, enabling configuration-benchmark matrices to scale without manual orchestration. Simulation outputs are stored as structured artifacts tied to configurations, benchmarks, and hardware components, enabling systematic result exploration without custom parsers. We evaluate ArchSim on a 12 x 8 = 96-configuration simulation matrix spanning memory-bound, compute-bound, and mixed-intensity GPU workloads. Declarative simulation specifications drive full simulations with a median kernel time error of 0.18% relative to hand-written MGPUSim configurations across 95.8% of configurations. The platform introduces only 1.6 seconds of overhead per simulation, negligible relative to realistic simulation workloads.

cs.AR↗

arXiv2026

Akita: A High Usability Simulation Framework for Computer Architecture

Computer architecture simulation is essential for evaluating new designs without the need for costly tapeout. The community has developed dozens of valuable simulators that have enabled significant architectural advances. However, using and developing simulators remains a major barrier due to ad-hoc component interfaces, strict deployment requirements, the burden of managing performance optimizations like parallelization at the component level, and limited monitoring and visualization capabilities. The root cause of these limitations is the systematic neglect of user and developer experience in favor of technical functionality. We believe that only by separating technical concerns from user and developer experience concerns -- through a dedicated simulation engine decoupled from hardware models -- can the community overcome these fundamental obstacles and enable more productive architectural research. Akita embodies this philosophy as a dedicated simulation engine that cleanly separates infrastructure from architectural models. Smart Ticking and Availability Backpropagation let developers write simple cycle-based code while achieving event-driven performance. Parallel simulation happens transparently -- developers write single-threaded code while Akita handles multi-core execution. Akita's simple, uniform, yet powerful simulation tracing support enables real-time monitoring and post-simulation visualization. We demonstrate the flexibility of Akita through case studies, including the development of a trace-based DNN simulation and a RISC-V CPU simulation, showing how prioritizing developer experience accelerates architectural research.

cs.DC↗

arXiv2026

Replication in Visual Diffusion Models: A Survey and Outlook

Visual diffusion models have revolutionized the field of creative AI, producing high-quality and diverse content. However, they inevitably memorize training images or videos, subsequently replicating their concepts, content, or styles during inference. This phenomenon raises significant concerns about privacy, security, and copyright within generated outputs. In this survey, we provide the first comprehensive review of replication in visual diffusion models, marking a novel contribution to the field by systematically categorizing the existing studies into unveiling, understanding, and mitigating this phenomenon. Specifically, unveiling mainly refers to the methods used to detect replication instances. Understanding involves analyzing the underlying mechanisms and factors that contribute to this phenomenon. Mitigation focuses on developing strategies to reduce or eliminate replication. Beyond these aspects, we also review papers focusing on its real-world influence. For instance, in the context of healthcare, replication is critically worrying due to privacy concerns related to patient data. Finally, the paper concludes with a discussion of the ongoing challenges, such as the difficulty in detecting and benchmarking replication, and outlines future directions including the development of more robust mitigation techniques. By synthesizing insights from diverse studies, this paper aims to equip researchers and practitioners with a deeper understanding at the intersection between AI technology and social good. We release this project at https://github.com/WangWenhao0716/Awesome-Diffusion-Replication.

cs.CV↗

arXiv2026

Thermophysical and mechanical properties of UFe$_2$ fabricated by spark plasma sintering

Following the accident at the Fukushima Daiichi Nuclear Power Plant in 2011, core meltdown produced fuel debris whose safe retrieval and management require reliable thermophysical and mechanical property data. Among the metallic phases identified in the debris, the U-Fe system is particularly important because of the abundant iron originating from in-vessel stainless steel structures. However, within this system, the high-temperature thermophysical properties of UFe$_2$ have received relatively little attention, with most prior studies focusing on its magnetic and electronic properties. To fill this data gap in the literature, we fabricated dense, nearly single-phase polycrystalline UFe$_2$ by arc melting followed by spark plasma sintering, and characterized its thermal and mechanical properties from room temperature to 1073 K. Results show that the thermal conductivity of UFe$_2$ increased monotonically from 10 Wm$^{-1}$K$^{-1}$ at 306 K to 25 Wm$^{-1}$K$^{-1}$ at 1073 K, surpassing those of the iron intermetallics Fe$_2$Zr and Fe$_2$B at high temperatures. In addition, UFe$_2$ is mechanically more compliant, displaying a Young's modulus $E$ of 69 GPa, a shear modulus $G$ of 24 GPa, and a Vickers hardness $H_{\mathrm{V}}$ of 5.6 GPa, all well below those of both Fe intermetallics. Consequently, during decommissioning, thermal-management and structural evaluations should take into account the comparatively high-conductivity and mechanically compliant nature of UFe$_2$ within the heterogeneous fuel debris.

cond-mat.mtrl-sci↗