arXiv subjects

Tong Zhang

Publications and source records attributed to Tong Zhang.

At least 109 records · Page 6Linked to original sources

arXiv2026

Beyond Correctness: Harmonizing Process and Outcome Rewards through RL Training

Reinforcement Learning with Verifiable Rewards (RLVR) improves final-answer accuracy on reasoning tasks, but it does not reliably improve reasoning quality. Because outcome rewards only assess final answers, they also reward spurious successes: flawed reasoning can still receive maximal reward when it accidentally reaches the correct outcome. This outcome reward hacking creates biased gradients, making current RLVR insufficient for learning faithful reasoning. Process Reward Models (PRMs) provide step-wise supervision, but directly optimizing PRMs or naively combining them with outcome rewards is unstable under distribution shift during RL training process. We introduce PRocess cOnsistency Filter (PROF), a data curation method that uses PRM--ORM consistency for sample selection rather than direct reward optimization. PROF keeps correct responses with strong process support and incorrect responses with weak process support while maintaining a balanced training ratio. Experiments show that PROF consistently improves both final-answer accuracy and intermediate reasoning quality over strong baselines, with less dependence on strong PRMs.

cs.LG↗

arXiv2026

Adaptive Layerwise Perturbation: Unifying Off-Policy Corrections for LLM RL

Off-policy problems such as policy staleness and training--inference mismatch have become a major bottleneck for training stability and further exploration in LLM RL. The distribution gap between the inference and updated policies grows because of the techniques to enhance inference efficiency, leading to heavy-tailed importance ratios. Heavy-tailed ratios arise when the policy is locally sharp, which further inflates gradients and can push updates outside the trust region. To address this, we propose Adaptive Layerwise Perturbation (ALP), which injects small learnable perturbations into the input hidden states of each layer during updates and uses the resulting perturbed policy as the numerator of the importance ratio against the unchanged inference policy in the objective. Intuitively, by adding controlled noise to intermediate representations, ALP prevents the updated policy from deviating too sharply from the inference policy and enlarges the policy family to cover inference-time mismatch noise. Hence, the flattened distribution can naturally tighten the gap between the updated and inference policies and reduce the tail of importance ratios, thus maintaining training stability. This is further validated empirically. Experiments on single-turn math and multi-turn tool-integrated reasoning tasks show that ALP not only improves final performance, but also avoids blow-up in the importance-ratio tail and KL spikes during iterative training, along with boosted exploration. Ablations show that representation-level perturbations across all layers are most effective, substantially outperforming partial-layer and logits-only variants.

cs.LG↗

arXiv2026

Mixture Prototype Flow Matching for Open-Set Supervised Anomaly Detection

Open-set supervised anomaly detection (OSAD) aims to identify unseen anomalies using limited anomalous supervision. However, existing prototype-based methods typically model normal data via a unimodal Gaussian prior, failing to capture inherent multi-modality and resulting in blurred decision boundaries. To address this, we propose Mixture Prototype Flow Matching (MPFM), a framework that learns a continuous transformation from normal feature distributions to a structured Gaussian mixture prototype space. Departing from traditional flow-based approaches that rely on a single velocity vector, MPFM explicitly models the velocity field as a Gaussian mixture prior where each component corresponds to a distinct normal class. This design facilitates mode-aware and semantically coherent distribution transport. Furthermore, we introduce a Mutual Information Maximization Regularizer (MIMR) to prevent prototype collapse and maximize normal-anomaly separability. Extensive experiments demonstrate that MPFM achieves state-of-the-art performance across diverse benchmarks under both single- and multi-anomaly settings.

cs.CV↗

arXiv2026

Skill-Aligned Annotation for Reliable Evaluation in Text-to-Image Generation

Text-to-image (T2I) generation has advanced rapidly, making reliable evaluation critical as performance differences between models narrow. Existing evaluation practices typically apply uniform annotation mechanisms, such as Likert-scale or binary question answering (BQA), across heterogeneous evaluation skills, despite fundamental differences in their nature. In this work, we revisit T2I evaluation through the lens of skill-aligned annotation, where annotation strategies reflect the underlying characteristics of each evaluation skill. We systematically compare skill-aligned annotation against uniform baselines and show that it produces more consistent evaluation signals, with higher inter-annotator agreement and improved stability across models. Finally, we present an automated pipeline that instantiates the proposed evaluation protocol, enabling scalable and fine-grained evaluation with spatially grounded feedback. Our work highlights that improving the foundations of image evaluation can increase reliability and efficiency without simply scaling annotation effort. We hope this motivates further research on refining evaluation protocols as a central component of reliable model assessment.

cs.CV↗

arXiv2026

VASR: Variance-Aware Systematic Resampling for Reward-Guided Diffusion

Sequential Monte Carlo (SMC) samplers for reward-guided diffusion models often suffer from rapid lineage collapse: a few high-reward particles dominate the population within a handful of resampling steps, destroying diversity and degrading sample quality. We propose a variance-decomposition framework for reward-guided diffusion SMC that separates continuation variance $V_t^{\mathrm{cont}}$ from residual variance $V_t^{\mathrm{res}}$, revealing that high offspring-count variance under the commonly used multinomial resampling drives this collapse. This motivates \textsc{VASR} (Variance-Aware Systematic Resampling), which addresses both variance terms via variance-optimal mass allocation $m_t \propto w_t e^{r_t}$ (minimizing $V_t^{\mathrm{cont}}$) and systematic resampling (controlling $V_t^{\mathrm{res}}$). For latent diffusion models where intermediate rewards are noisy due to stochastic continuations, we propose \textsc{VASR-Max}, a deliberately biased high-selection variant for variance-sensitive reward optimization. Both methods are training-free, fully parallelizable, and add only linear overhead. On MNIST and CIFAR-10, VASR achieves as high as $26\%$ better FID than prior SMC methods while remaining 66 times faster than MCTS-based value methods at matched compute. On text-to-image generation, \textsc{VASR-Max} consistently outperforms the strongest SMC baseline across compute budgets and matches MCTS-based methods within 2.5--3% reward at high budgets while being approximately times faster.

cs.AI↗

arXiv2026

EchoFake: A Replay-Aware Dataset for Practical Speech Deepfake Detection

The growing prevalence of speech deepfakes has raised serious concerns, particularly in real-world scenarios such as telephone fraud and identity theft. While many anti-spoofing systems have demonstrated promising performance on lab-generated synthetic speech, they often fail when confronted with physical replay attacks-a common and low-cost form of attack used in practical settings. Our experiments show that models trained on existing datasets exhibit severe performance degradation, with average accuracy dropping to 59.6% when evaluated on replayed audio. To bridge this gap, we present EchoFake, a comprehensive dataset comprising more than 120 hours of audio from over 13,000 speakers, featuring both cutting-edge zero-shot text-to-speech (TTS) speech and physical replay recordings collected under varied devices and real-world environmental settings. Additionally, we evaluate three baseline detection models and show that models trained on EchoFake achieve lower average EERs across datasets, indicating better generalization. By introducing more practical challenges relevant to real-world deployment, EchoFake offers a more realistic foundation for advancing spoofing detection methods.

eess.AS↗

arXiv2026

Five-Minute Rule 40 Years Later: A First-Principles Revisit for Modern Memory Hierarchy

In 1987, Jim Gray and Gianfranco Putzolu introduced the five-minute rule, a simple, storage-memory-economics-based heuristic for deciding when data should live in DRAM rather than on storage. Subsequent revisits to the rule largely retained that economics-only view, leaving host costs, feasibility limits, and workload behavior out of scope. This paper revisits the rule from first principles, integrating host costs, DRAM bandwidth/capacity, and physics-grounded models of SSD performance and cost, and then embedding these elements in a constraint- and workload-aware framework that yields actionable provisioning guidance. We show that, for modern AI platforms, especially GPU-centric hosts paired with ultra-high-IOPS SSDs engineered for fine-grained random access, the DRAM$\leftrightarrow$flash caching threshold collapses from minutes to a few seconds. This shift reframes NAND flash memory as an \emph{active data tier} and exposes a broad research space across the hardware-software stack. We further introduce MQSim-Next, a calibrated SSD simulator that supports validation and sensitivity analysis and facilitates future architectural and system research. Finally, we present two concrete case studies that showcase the software system design space opened by such memory hierarchy paradigm shift. Overall, we turn a classical heuristic into an actionable, feasibility-aware analysis and provisioning framework and set the stage for further research on AI-era memory hierarchy.

cs.AR↗

arXiv2026

Communication-Semantic-Aware RDMA Loss Recovery for QP-scalable Hyperscale AI Training

Current artificial intelligence (AI) infrastructures widely adopt Remote Direct Memory Access (RDMA) to support high-performance communication. Training trillion-parameter models involves frequent collective communication operations, such as All-Reduce and All-to-All, which generate intensive RDMA traffic. Existing RDMA deployments predominantly use the reliable connection (RC) model, where each process pair requires a dedicated queue pair (QP). This leads to poor scalability: since the RDMA-capable network interface card (RNIC) can cache only a few thousand QPs, excess entries trigger PCIe round-trip penalties. Meanwhile, global synchronization makes training sensitive to tail latency, where a few packet losses can delay iteration completion. To address these challenges, we propose Communication-Semantic-Aware Unreliable Datagram (CSA-UD), a novel RDMA loss recovery mechanism that combines scalability and reliability. CSA-UD decouples data transmission from loss recovery and dynamically adjusts the loss detection interval, accelerating tail recovery and exploiting the synchronization semantics of distributed training. It further supports multipath transmission and bitmap-guided reassembly, enabling high throughput without requiring lossless fabrics. Testbed experiments and ns-3 simulations show that CSA-UD significantly reduces tail latency under large-scale collective communication. Under high network load, it achieves better scalability than RC and over 30% lower 99th percentile flow completion times compared with counterparts.

cs.NI↗

arXiv2026

Profiling for Pennies: Unveiling the Privacy Iceberg of LLM Agents

Large Language Models (LLMs) have revolutionized how information are collected, aggregated, and reasoned. However, this enables a novel and accessible vector of privacy intrusion: the automated and in-depth personal profiling; this engenders a chilling effect of "peepers everywhere". Existing research primarily unfolds from the training pipeline of LLM, emphasizing the exposure of Personally Identifiable Information (PII) through memorization, while privacy studies from a human-centric perspective remain underexplored. To fill this void, we empirically investigate privacy perception in the real world through the lens of human awareness and the practices of LLM-integrated platforms, revealing a significant dissonance: platforms fail to technically or policy-wise address public privacy concerns. To facilitate a systematic and quantifiable study of privacy risk, we propose the PrivacyIceberg, which categorizes real-world human privacy risks into three tiers: explicitly searched, contextually inferred, and deeply aggregated, based on the sophistication of LLM exploitation. We developed IcebergExplorer to audit privacy exposure, utilizing minimal PII as a search seed to reconstruct high-fidelity profiles, achieving over 90% factual accuracy within 10 minutes at a cost under $3, for real-world scenarios. Additionally, we identify six root causes contributing to such privacy disclosures and propose multi-stakeholder countermeasures for LLM vendors, individuals, and data publishers.

cs.CR↗

arXiv2026

Optimizer-Model Consistency: Full Finetuning with the Same Optimizer as Pretraining Forgets Less

Optimizers play an important role in both pretraining and finetuning stages when training large language models (LLMs). In this paper, we present an observation that full finetuning with the same optimizer as in pretraining achieves a better learning-forgetting tradeoff, i.e., forgetting less while achieving the same or better performance on the new task, than other optimizers and, possibly surprisingly, LoRA, during the supervised finetuning (SFT) stage. We term this phenomenon optimizer-model consistency. To better understand it, through controlled experiments and theoretical analysis, we show that: 1) optimizers can shape the models by having regularization effects on the activations, leading to different landscapes around the pretrained checkpoints; 2) in response to this regularization effect, the weight update in SFT should follow some specific structures to lower forgetting of the knowledge learned in pretraining, which can be obtained by using the same optimizer. Moreover, we specifically compare Muon and AdamW when they are employed throughout the pretraining and SFT stages and find that Muon performs worse when finetuned for reasoning tasks. With a synthetic language modeling experiment, we demonstrate that this can come from Muon's strong tendency towards rote memorization, which may hurt pattern acquisition with a small amount of data, as for SFT.

cs.LG↗

arXiv2026

WebGym: Scaling Training Environments for Visual Web Agents with Realistic Tasks

We present WebGym, the largest-to-date open-source environment for training realistic visual web agents. Real websites are non-stationary and diverse, making artificial or small-scale task sets insufficient for robust policy learning. WebGym contains nearly 300,000 tasks with rubric-based evaluations across diverse, real-world websites and difficulty levels. We train agents with a simple reinforcement learning (RL) recipe, which trains on the agent's own interaction traces (rollouts), using task rewards as feedback to guide learning. To enable scaling RL, we speed up sampling of trajectories in WebGym by developing a high-throughput asynchronous rollout system, designed specifically for web agents. Our system achieves a 4-5x rollout speedup compared to naive implementations. Second, we scale the task set breadth, depth, and size, which results in continued performance improvement. Fine-tuning a strong base vision-language model, Qwen-3-VL-8B-Instruct, on WebGym results in an improvement in success rate on an out-of-distribution test set from 26.2% to 42.9%, significantly outperforming agents based on proprietary models such as GPT-4o and GPT-5-Thinking that achieve 27.1% and 29.8%, respectively. This improvement is substantial because our test set consists only of tasks on websites never seen during training, unlike many other prior works on training visual web agents.

cs.LG↗

arXiv2026

Self-DACE++: Robust Low-Light Enhancement via Efficient Adaptive Curve Estimation

In this paper, we present Self-DACE++, an improved unsupervised and lightweight framework for Low-Light Image Enhancement (LLIE), building upon our previous Self-Reference Deep Adaptive Curve Estimation (Self-DACE). To better address the trade-off between computational efficiency and restoration quality, Self-DACE++ introduces enhanced Adaptive Adjustment Curves (AACs). These curves, governed by minimal trainable parameters, flexibly adjust the dynamic range while preserving the color fidelity, structural integrity, and naturalness of the enhanced images. To achieve an extremely lightweight architecture without sacrificing performance, we propose a randomized order training strategy coupled with a network fusion mechanism, which compresses the model into an efficient iterative inference structure. Furthermore, we formulate a physics-grounded objective function based on Retinex theory and incorporate a dedicated denoising module to effectively estimate and suppress latent noise in dark regions. Extensive qualitative and quantitative evaluations on multiple real-world benchmark datasets demonstrate that Self-DACE++ outperforms existing state-of-the-art methods, delivering superior enhancement quality with real-time inference capability. The code is available at https://github.com/John-Wendell/Self-DACE.

cs.CV↗

arXiv2026

PhysNote: Self-Knowledge Notes for Evolvable Physical Reasoning in Vision-Language Model

Vision-Language Models (VLMs) have demonstrated strong performance on textbook-style physics problems, yet they frequently fail when confronted with dynamic real-world scenarios that require temporal consistency and causal reasoning across frames. We identify two fundamental challenges underlying these failures: (1) spatio-temporal identity drift, where objects lose their physical identity across successive frames and break causal chains, and (2) volatility of inference-time insights, where a model may occasionally produce correct physical reasoning but never consolidates it for future reuse. To address these challenges, we propose PhysNote, an agentic framework that enables VLMs to externalize and refine physical knowledge through self-generated "Knowledge Notes." PhysNote stabilizes dynamic perception through spatio-temporal canonicalization, organizes self-generated insights into a hierarchical knowledge repository, and drives an iterative reasoning loop that grounds hypotheses in visual evidence before consolidating verified knowledge. Experiments on PhysBench demonstrate that PhysNote achieves 56.68% overall accuracy, a 4.96% improvement over the best multi-agent baseline, with consistent gains across all four physical reasoning domains.

cs.AI↗

arXiv2026

Reinforcement Learning with Foundation Priors: Let the Embodied Agent Efficiently Learn on Its Own

Reinforcement learning (RL) is a promising approach for solving robotic manipulation tasks. However, it is challenging to apply the RL algorithms directly in the real world. For one thing, RL is data-intensive and typically requires millions of interactions with environments, which are impractical in real scenarios. For another, it is necessary to make heavy engineering efforts to design reward functions manually. To address these issues, we leverage foundation models in this paper. We propose Reinforcement Learning with Foundation Priors (RLFP) to utilize guidance and feedback from policy, value, and success-reward foundation models. Within this framework, we introduce the Foundation-guided Actor-Critic (FAC) algorithm, which enables embodied agents to explore more efficiently with automatic reward functions. The benefits of our framework are threefold: (1) \textit{sample efficient}; (2) \textit{minimal and effective reward engineering}; (3) \textit{agnostic to foundation model forms and robust to noisy priors}. Our method achieves remarkable performances in various manipulation tasks on both real robots and in simulation. Across 5 dexterous tasks with real robots, FAC achieves an average success rate of 86\% after one hour of real-time learning. Across 8 tasks in the simulated Meta-world, FAC achieves 100\% success rates in 7/8 tasks under less than 100k frames (about 1-hour training), outperforming baseline methods with manual-designed rewards in 1M frames. We believe the RLFP framework can enable future robots to explore and learn autonomously in the physical world for more tasks. Visualizations and code are available at https://yewr.github.io/rlfp.

cs.RO↗

arXiv2026

Evolution of Optimization Methods: Algorithms, Scenarios, and Evaluations

Balancing convergence speed, generalization capability, and computational efficiency remains a core challenge in deep learning optimization. First-order gradient descent methods, epitomized by stochastic gradient descent (SGD) and Adam, serve as the cornerstone of modern training pipelines. However, large-scale model training, stringent differential privacy requirements, and distributed learning paradigms expose critical limitations in these conventional approaches regarding privacy protection and memory efficiency. To mitigate these bottlenecks, researchers explore second-order optimization techniques to surpass first-order performance ceilings, while zeroth-order methods reemerge to alleviate memory constraints inherent to large-scale training. Despite this proliferation of methodologies, the field lacks a cohesive framework that unifies underlying principles and delineates application scenarios for these disparate approaches. In this work, we retrospectively analyze the evolutionary trajectory of deep learning optimization algorithms and present a comprehensive empirical evaluation of mainstream optimizers across diverse model architectures and training scenarios. We distill key emerging trends and fundamental design trade-offs, pinpointing promising directions for future research. By synthesizing theoretical insights with extensive empirical evidence, we provide actionable guidance for designing next-generation highly efficient, robust, and trustworthy optimization methods. The code is available at https://github.com/APRIL-AIGC/Awesome-Optimizer.

cs.LG↗

arXiv2026

AgentSPEX: An Agent SPecification and EXecution Language

Language-model agent systems commonly rely on reactive prompting, in which a single instruction guides the model through an open-ended sequence of reasoning and tool-use steps, leaving control flow and intermediate state implicit and making agent behavior potentially difficult to control. Orchestration frameworks such as LangGraph, DSPy, and CrewAI impose greater structure through explicit workflow definitions, but tightly couple workflow logic with Python, making agents difficult to maintain and modify. In this paper, we introduce AgentSPEX, an Agent SPecification and EXecution Language for specifying LLM-agent workflows with explicit control flow and modular structure, along with a customizable agent harness. AgentSPEX supports typed steps, branching and loops, parallel execution, reusable submodules, and explicit state management, and these workflows execute within an agent harness that provides tool access, a sandboxed virtual environment, and support for checkpointing, verification, and logging. Furthermore, we provide a visual editor with synchronized graph and workflow views for authoring and inspection. We include ready-to-use agents for deep research and scientific research, and we evaluate AgentSPEX on 7 benchmarks. Finally, we show through a user study that AgentSPEX provides a more interpretable and accessible workflow-authoring paradigm than a popular existing agent framework.

cs.CL↗

arXiv2026

Group Relative Policy Optimization for Robust Blind Interference Alignment with Fluid Antennas

Fluid antenna system (FAS) leverages dynamic reconfigurability to unlock spatial degrees of freedom and reshape wireless channels. Blind interference alignment (BIA) aligns interference through antenna switching. This paper proposes, for the first time, a robust fluid antenna-driven BIA framework for a K-user MISO downlink under imperfect channel state information (CSI). We formulate a robust sum-rate maximization problem through optimizing fluid antenna positions (switching positions). To solve this challenging non-convex problem, we employ group relative policy optimization (GRPO), a novel deep reinforcement learning algorithm that eliminates the critic network. This robust design reduces model size and floating point operations (FLOPs) by nearly half compared to proximal policy optimization (PPO) while significantly enhancing performance through group-based exploration that escapes bad local optima. Simulation results demonstrate that GRPO outperforms PPO by 4.17%, and a 100K-step pre-trained PPO by 30.29%. Due to error distribution learning, GRPO exceeds heuristic MaximumGain and RandomGain by 200.78% and 465.38%, respectively.

cs.IT↗

arXiv2026

Microscopic evidence of spin-driven multiferroicity and topological spin textures in monolayer NiI2

In type II multiferroics, noncollinear spin textures are expected to induce electric polarization directly, leading to strong magnetoelectric coupling. Realizing such spin driven multiferroicity in two-dimensional systems, and elucidating the interplay between local spins and electric polarization, are of both fundamental and technological importance. Here, using vectorial spin polarized scanning tunneling microscopy, we investigated the spin-driven multiferroicity in monolayer NiI2 at atomic scale. We identify a canted spin-spiral state with fully determined spin rotation plane, accompanied by a 2Q charge modulation. At spin spiral domain walls, we discover topological spin textures that composed of meron/antimeron pairs. These textures are associated with distinct charge pattern and notable band shifts, indicating local bound charges induced by variations of ferroelectricity at domain wall. Our observations are well captured by a realistic spin model incorporating Kitaev interactions and generalized spin-current model of type II multiferroicity. The findings provide microscopic evidence of spin-driven multiferroicity in an extreme 2D system and establish a platform for low-dissipation, electric-field control of topological spin textures.

cond-mat.mes-hall↗