arXiv subjects

Wei Lin

Publications and source records attributed to Wei Lin.

At least 91 records · Page 5Linked to original sources

arXiv2026

The Multi-Query Paradox in Zeroth-Order Optimization

Zeroth-order (ZO) optimization provides a powerful framework for problems where explicit gradients are unavailable and have to be approximated using only queries to function value. The prevalent single-query approach is simple, but suffers from high estimation variance, motivating a multi-query paradigm to improve estimation accuracy. This, however, creates a critical trade-off: under a fixed budget of queries (i.e. cost), queries per iteration and the total number of optimization iterations are inversely proportional to one another. How to best allocate this budget is a fundamental, under-explored question. This work systematically resolves this query allocation problem. We analyze two aggregation methods: the de facto simple averaging (ZO-Avg), and a new Projection Alignment method (ZO-Align) we derive from local surrogate minimization. By deriving convergence rates for both methods that make the dependence on the number of queries explicit across strongly convex, convex, non-convex, and stochastic settings, we uncover a stark dichotomy: For ZO-Avg, we prove that using more than one query per iteration is always query-inefficient, rendering the single-query approach optimal. On the contrary, ZO-Align generally performs better with more queries per iteration, resulting in a full-subspace estimation as the optimal approach. Thus, our work clarifies that the multi-query problem boils down to a choice not about an intermediate query size, but between two classic algorithms, a choice dictated entirely by the aggregation method used. These theoretical findings are also consistently validated by extensive experiments.

cs.LG↗

arXiv2026

Your Group-Relative Advantage Is Biased

Reinforcement Learning from Verifier Rewards (RLVR) has emerged as a widely used approach for post-training large language models on reasoning tasks, with group-based methods such as GRPO and its variants gaining broad adoption. These methods rely on group-relative advantage estimation to avoid learned critics, yet its theoretical properties remain poorly understood. In this work, we uncover a fundamental issue of group-based RL: the group-relative advantage estimator is inherently biased relative to the true (expected) advantage. We provide the first theoretical analysis showing that it systematically underestimates advantages for hard prompts and overestimates them for easy prompts, leading to imbalanced exploration and exploitation. To address this issue, we propose History-Aware Adaptive Difficulty Weighting (HA-DW), an adaptive reweighting scheme that adjusts advantage estimates based on an evolving difficulty anchor and training dynamics. Both theoretical analysis and experiments on five mathematical reasoning benchmarks demonstrate that HA-DW consistently improves performance when integrated into GRPO and its variants. Our results suggest that correcting biased advantage estimation is critical for robust and efficient RLVR training.

cs.LG↗

arXiv2026

MedConsultBench: A Full-Cycle, Fine-Grained, Process-Aware Benchmark for Medical Consultation Agents

Current evaluations of medical consultation agents often prioritize outcome-oriented tasks, frequently overlooking the end-to-end process integrity and clinical safety essential for real-world practice. While recent interactive benchmarks have introduced dynamic scenarios, they often remain fragmented and coarse-grained, failing to capture the structured inquiry logic and diagnostic rigor required in professional consultations. To bridge this gap, we propose MedConsultBench, a comprehensive framework designed to evaluate the complete online consultation cycle by covering the entire clinical workflow from history taking and diagnosis to treatment planning and follow-up Q\&A. Our methodology introduces Atomic Information Units (AIUs) to track clinical information acquisition at a sub-turn level, enabling precise monitoring of how key facts are elicited through 22 fine-grained metrics. By addressing the underspecification and ambiguity inherent in online consultations, the benchmark evaluates uncertainty-aware yet concise inquiry while emphasizing medication regimen compatibility and the ability to handle realistic post-prescription follow-up Q\&A via constraint-respecting plan revisions. Systematic evaluation of 19 large language models reveals that high diagnostic accuracy often masks significant deficiencies in information-gathering efficiency and medication safety. These results underscore a critical gap between theoretical medical knowledge and clinical practice ability, establishing MedConsultBench as a rigorous foundation for aligning medical AI with the nuanced requirements of real-world clinical care.

cs.AI↗

arXiv2026

AWPO: Enhancing Tool-Use of Large Language Models through Adaptive Integration of Reasoning Rewards

While Reinforcement Learning (RL) shows promise in training tool-use Large Language Models (LLMs) using verifiable outcome rewards, existing methods largely overlook the potential of reasoning rewards based on chain-of-thought quality for better tool utilization. Furthermore, naïvely combining reasoning and outcome rewards may yield suboptimal performance or conflict with the primary optimization objective. To address this, we propose Advantage-Weighted Policy Optimization (AWPO), a principled RL framework that adaptively integrates reasoning rewards into advantage estimation to improve tool-use performance. AWPO incorporates variance-aware gating and difficulty-aware weighting to adaptively modulate advantages from reasoning signals based on group-relative statistics, alongside a tailored clipping mechanism for stable optimization. Extensive experiments demonstrate that AWPO achieves state-of-the-art performance across standard tool-use benchmarks, significantly outperforming strong baselines and leading closed-source models in challenging multi-turn scenarios. Notably, with exceptional parameter efficiency, our 4B model surpasses Grok-4 by $16.0\%$ in multi-turn accuracy while preserving generalization capability on the out-of-distribution MMLU-Pro benchmark.

cs.CL↗

arXiv2026

The RoboSense Challenge: Sense Anything, Navigate Anywhere, Adapt Across Platforms

Autonomous systems are increasingly deployed in open and dynamic environments -- from city streets to aerial and indoor spaces -- where perception models must remain reliable under sensor noise, environmental variation, and platform shifts. However, even state-of-the-art methods often degrade under unseen conditions, highlighting the need for robust and generalizable robot sensing. The RoboSense 2025 Challenge is designed to advance robustness and adaptability in robot perception across diverse sensing scenarios. It unifies five complementary research tracks spanning language-grounded decision making, socially compliant navigation, sensor configuration generalization, cross-view and cross-modal correspondence, and cross-platform 3D perception. Together, these tasks form a comprehensive benchmark for evaluating real-world sensing reliability under domain shifts, sensor failures, and platform discrepancies. RoboSense 2025 provides standardized datasets, baseline models, and unified evaluation protocols, enabling large-scale and reproducible comparison of robust perception methods. The challenge attracted 143 teams from 85 institutions across 16 countries, reflecting broad community engagement. By consolidating insights from 23 winning solutions, this report highlights emerging methodological trends, shared design principles, and open challenges across all tracks, marking a step toward building robots that can sense reliably, act robustly, and adapt across platforms in real-world environments.

cs.RO↗

arXiv2025

DaGRPO: Rectifying Gradient Conflict in Reasoning via Distinctiveness-Aware Group Relative Policy Optimization

The evolution of Large Language Models (LLMs) has catalyzed a paradigm shift from superficial instruction following to rigorous long-horizon reasoning. While Group Relative Policy Optimization (GRPO) has emerged as a pivotal mechanism for eliciting such post-training reasoning capabilities due to its exceptional performance, it remains plagued by significant training instability and poor sample efficiency. We theoretically identify the root cause of these issues as the lack of distinctiveness within on-policy rollouts: for routine queries, highly homogeneous samples induce destructive gradient conflicts; whereas for hard queries, the scarcity of valid positive samples results in ineffective optimization. To bridge this gap, we propose Distinctiveness-aware Group Relative Policy Optimization (DaGRPO). DaGRPO incorporates two core mechanisms: (1) Sequence-level Gradient Rectification, which utilizes fine-grained scoring to dynamically mask sample pairs with low distinctiveness, thereby eradicating gradient conflicts at the source; and (2) Off-policy Data Augmentation, which introduces high-quality anchors to recover training signals for challenging tasks. Extensive experiments across 9 mathematical reasoning and out-of-distribution (OOD) generalization benchmarks demonstrate that DaGRPO significantly surpasses existing SFT, GRPO, and hybrid baselines, achieving new state-of-the-art performance (e.g., a +4.7% average accuracy gain on math benchmarks). Furthermore, in-depth analysis confirms that DaGRPO effectively mitigates gradient explosion and accelerates the emergence of long-chain reasoning capabilities.

cs.AI↗

arXiv2025

Learning Provably Improves the Convergence of Gradient Descent

Learn to Optimize (L2O) trains deep neural network-based solvers for optimization, achieving success in accelerating convex problems and improving non-convex solutions. However, L2O lacks rigorous theoretical backing for its own training convergence, as existing analyses often use unrealistic assumptions -- a gap this work highlights empirically. We bridge this gap by proving the training convergence of L2O models that learn Gradient Descent (GD) hyperparameters for quadratic programming, leveraging the Neural Tangent Kernel (NTK) theory. We propose a deterministic initialization strategy to support our theoretical results and promote stable training over extended optimization horizons by mitigating gradient explosion. Our L2O framework demonstrates over 50% better optimality than GD and superior robustness over state-of-the-art L2O methods on synthetic datasets. The code of our method can be found from https://github.com/NetX-lab/MathL2OProof-Official.

cs.LG↗

arXiv2025

ToolForge: A Data Synthesis Pipeline for Multi-Hop Search without Real-World APIs

Training LLMs to invoke tools and leverage retrieved information necessitates high-quality, diverse data. However, existing pipelines for synthetic data generation often rely on tens of thousands of real API calls to enhance generalization, incurring prohibitive costs while lacking multi-hop reasoning and self-reflection. To address these limitations, we introduce ToolForge, an automated synthesis framework that achieves strong real-world tool-calling performance by constructing only a small number of virtual tools, eliminating the need for real API calls. ToolForge leverages a (question, golden context, answer) triple to synthesize large-scale tool-learning data specifically designed for multi-hop search scenarios, further enriching the generated data through multi-hop reasoning and self-reflection mechanisms. To ensure data fidelity, we employ a Multi-Layer Validation Framework that integrates both rule-based and model-based assessments. Empirical results show that a model with only 8B parameters, when trained on our synthesized data, outperforms GPT-4o on multiple benchmarks. Our code and dataset are publicly available at https://github.com/Buycar-arb/ToolForge .

cs.AI↗

arXiv2025

From Generated Human Videos to Physically Plausible Robot Trajectories

Video generation models are rapidly improving in their ability to synthesize human actions in novel contexts, holding the potential to serve as high-level planners for contextual robot control. To realize this potential, a key research question remains open: how can a humanoid execute the human actions from generated videos in a zero-shot manner? This challenge arises because generated videos are often noisy and exhibit morphological distortions that make direct imitation difficult compared to real video. To address this, we introduce a two-stage pipeline. First, we lift video pixels into a 4D human representation and then retarget to the humanoid morphology. Second, we propose GenMimic-a physics-aware reinforcement learning policy conditioned on 3D keypoints, and trained with symmetry regularization and keypoint-weighted tracking rewards. As a result, GenMimic can mimic human actions from noisy, generated videos. We curate GenMimicBench, a synthetic human-motion dataset generated using two video generation models across a spectrum of actions and contexts, establishing a benchmark for assessing zero-shot generalization and policy robustness. Extensive experiments demonstrate improvements over strong baselines in simulation and confirm coherent, physically stable motion tracking on a Unitree G1 humanoid robot without fine-tuning. This work offers a promising path to realizing the potential of video generation models as high-level policies for robot control.

cs.RO↗

arXiv2025

Probing the effectiveness of World Models for Spatial Reasoning through Test-time Scaling

Vision-Language Models (VLMs) remain limited in spatial reasoning tasks that require multi-view understanding and embodied perspective shifts. Recent approaches such as MindJourney attempt to mitigate this gap through test-time scaling where a world model imagines action-conditioned trajectories and a heuristic verifier selects helpful views from such trajectories. In this work, we systematically examine how such test-time verifiers behave across benchmarks, uncovering both their promise and their pitfalls. Our uncertainty-based analyses show that MindJourney's verifier provides little meaningful calibration, and that random scoring often reduces answer entropy equally well, thus exposing systematic action biases and unreliable reward signals. To mitigate these, we introduce a Verification through Spatial Assertions (ViSA) framework that grounds the test-time reward in verifiable, frame-anchored micro-claims. This principled verifier consistently improves spatial reasoning on the SAT-Real benchmark and corrects trajectory-selection biases through more balanced exploratory behavior. However, on the challenging MMSI-Bench, none of the verifiers, including ours, achieve consistent scaling, suggesting that the current world models form an information bottleneck where imagined views fail to enrich fine-grained reasoning. Together, these findings chart the bad, good, and ugly aspects of test-time verification for world-model-based reasoning. Our code is available at https://github.com/chandar-lab/visa-for-mindjourney.

cs.CV↗

arXiv2025

TTRV: Test-Time Reinforcement Learning for Vision Language Models

Existing methods for extracting reward signals in Reinforcement Learning typically rely on labeled data and dedicated training splits, a setup that contrasts with how humans learn directly from their environment. In this work, we propose TTRV to enhance vision language understanding by adapting the model on the fly at inference time, without the need for any labeled data. Concretely, we enhance the Group Relative Policy Optimization (GRPO) framework by designing rewards based on the frequency of the base model's output, while inferring on each test sample multiple times. Further, we also propose to control the diversity of the model's output by simultaneously rewarding the model for obtaining low entropy of the output empirical distribution. Our approach delivers consistent gains across both object recognition and visual question answering (VQA), with improvements of up to 52.4% and 29.8%, respectively, and average boosts of 24.6% and 10.0% across 16 datasets. Remarkably, on image recognition, TTRV applied to InternVL 8B surpasses GPT-4o by an average of 2.3% over 8 benchmarks, while remaining highly competitive on VQA, demonstrating that test-time reinforcement learning can match or exceed the strongest proprietary models. Finally, we find many interesting properties of test-time RL for VLMs: for example, even in extremely data-constrained scenarios, where adaptation is performed on a single randomly chosen unlabeled test example, TTRV still yields non-trivial improvements of up to 5.5% in recognition tasks.

cs.CV↗

arXiv2025

AR-Med: Automated Relevance Enhancement in Medical Search via LLM-Driven Information Augmentation

Accurate and reliable search on online healthcare platforms is critical for user safety and service efficacy. Traditional methods, however, often fail to comprehend complex and nuanced user queries, limiting their effectiveness. Large language models (LLMs) present a promising solution, offering powerful semantic understanding to bridge this gap. Despite their potential, deploying LLMs in this high-stakes domain is fraught with challenges, including factual hallucinations, specialized knowledge gaps, and high operational costs. To overcome these barriers, we introduce \textbf{AR-Med}, a novel framework for \textbf{A}utomated \textbf{R}elevance assessment for \textbf{Med}ical search that has been successfully deployed at scale on the Online Medical Delivery Platforms. AR-Med grounds LLM reasoning in verified medical knowledge through a retrieval-augmented approach, ensuring high accuracy and reliability. To enable efficient online service, we design a practical knowledge distillation scheme that compresses large teacher models into compact yet powerful student models. We also introduce LocalQSMed, a multi-expert annotated benchmark developed to guide model iteration and ensure strong alignment between offline and online performance. Extensive experiments show AR-Med achieves an offline accuracy of over 93\%, a 24\% absolute improvement over the original online system, and delivers significant gains in online relevance and user satisfaction. Our work presents a practical and scalable blueprint for developing trustworthy, LLM-powered systems in real-world healthcare applications.

cs.CL↗

arXiv2025

Privacy-Preserving Reasoning with Knowledge-Distilled Parametric Retrieval Augmented Generation

The current RAG system requires uploading plaintext documents to the cloud, risking private data leakage. Parametric RAG (PRAG) encodes documents as LoRA parameters within LLMs, offering a possible way to reduce exposure of raw content. However, it still faces two issues: (1) PRAG demands synthesizing QA pairs and fine-tuning LLM for each individual document to create its corresponding LoRA, leading to unacceptable inference latency. (2) The performance of PRAG relies solely on synthetic QA data while lacking internal alignment with standard RAG, resulting in poor generalization on out-of-distribution(OOD) inputs. Therefore, achieving high-efficiency parameterization while maintaining RAG-level performance remains a critical challenge for privacy-preserving reasoning. In this paper, we propose DistilledPRAG, a generalizable knowledge-distilled parametric RAG model aligned with standard RAG in document structure and parameter activation. We first synthesize QA pairs from single and multi-documents to enhance cross-document reasoning. Then, we mask the plaintext documents with a special token and translate them to LoRA via a parameter generator, maintaining the standard RAG document structure. Finally, guided by synthetic QA data, we train the parameter generator to match standard RAG's hidden states and output logits, enabling RAG-style reasoning without original documents. Experiments on four QA datasets show that DistilledPRAG outperforms baselines in accuracy and generalizes well on OOD data.

cs.CL↗

arXiv2025

Arithmetic-Mean $μ$P for Modern Architectures: A Unified Learning-Rate Scale for CNNs and ResNets

Choosing an appropriate learning rate remains a key challenge in scaling depth of modern deep networks. The classical maximal update parameterization ($μ$P) enforces a fixed per-layer update magnitude, which is well suited to homogeneous multilayer perceptrons (MLPs) but becomes ill-posed in heterogeneous architectures where residual accumulation and convolutions introduce imbalance across layers. We introduce Arithmetic-Mean $μ$P (AM-$μ$P), which constrains not each individual layer but the network-wide average one-step pre-activation second moment to a constant scale. Combined with a residual-aware He fan-in initialization - scaling residual-branch weights by the number of blocks ($\mathrm{Var}[W]=c/(K\cdot \mathrm{fan\text{-}in})$) - AM-$μ$P yields width-robust depth laws that transfer consistently across depths. We prove that, for one- and two-dimensional convolutional networks, the maximal-update learning rate satisfies $η^\star(L)\propto L^{-3/2}$; with zero padding, boundary effects are constant-level as $N\gg k$. For standard residual networks with general conv+MLP blocks, we establish $η^\star(L)=Θ(L^{-3/2})$, with $L$ the minimal depth. Empirical results across a range of depths confirm the $-3/2$ scaling law and enable zero-shot learning-rate transfer, providing a unified and practical LR principle for convolutional and deep residual networks without additional tuning overhead.

cs.LG↗

arXiv2025

From Experience to Strategy: Empowering LLM Agents with Trainable Graph Memory

Large Language Models (LLMs) based agents have demonstrated remarkable potential in autonomous task-solving across complex, open-ended environments. A promising approach for improving the reasoning capabilities of LLM agents is to better utilize prior experiences in guiding current decisions. However, LLMs acquire experience either through implicit memory via training, which suffers from catastrophic forgetting and limited interpretability, or explicit memory via prompting, which lacks adaptability. In this paper, we introduce a novel agent-centric, trainable, multi-layered graph memory framework and evaluate how context memory enhances the ability of LLMs to utilize parametric information. The graph abstracts raw agent trajectories into structured decision paths in a state machine and further distills them into high-level, human-interpretable strategic meta-cognition. In order to make memory adaptable, we propose a reinforcement-based weight optimization procedure that estimates the empirical utility of each meta-cognition based on reward feedback from downstream tasks. These optimized strategies are then dynamically integrated into the LLM agent's training loop through meta-cognitive prompting. Empirically, the learnable graph memory delivers robust generalization, improves LLM agents' strategic reasoning performance, and provides consistent benefits during Reinforcement Learning (RL) training.

cs.CL↗

arXiv2025

CAT-ID$^2$: Category-Tree Integrated Document Identifier Learning for Generative Retrieval In E-commerce

Generative retrieval (GR) has gained significant attention as an effective paradigm that integrates the capabilities of large language models (LLMs). It generally consists of two stages: constructing discrete semantic identifiers (IDs) for documents and retrieving documents by autoregressively generating ID tokens. The core challenge in GR is how to construct document IDs (DocIDS) with strong representational power. Good IDs should exhibit two key properties: similar documents should have more similar IDs, and each document should maintain a distinct and unique ID. However, most existing methods ignore native category information, which is common and critical in E-commerce. Therefore, we propose a novel ID learning method, CAtegory-Tree Integrated Document IDentifier (CAT-ID$^2$), incorporating prior category information into the semantic IDs. CAT-ID$^2$ includes three key modules: a Hierarchical Class Constraint Loss to integrate category information layer by layer during quantization, a Cluster Scale Constraint Loss for uniform ID token distribution, and a Dispersion Loss to improve the distinction of reconstructed documents. These components enable CAT-ID$^2$ to generate IDs that make similar documents more alike while preserving the uniqueness of different documents' representations. Extensive offline and online experiments confirm the effectiveness of our method, with online A/B tests showing a 0.33% increase in average orders per thousand users for ambiguous intent queries and 0.24% for long-tail queries.

cs.IR↗

arXiv2025

A Soft-partitioned Semi-supervised Collaborative Transfer Learning Approach for Multi-Domain Recommendation

In industrial practice, Multi-domain Recommendation (MDR) plays a crucial role. Shared-specific architectures are widely used in industrial solutions to capture shared and unique attributes via shared and specific parameters. However, with imbalanced data across different domains, these models face two key issues: (1) Overwhelming: Dominant domain data skews model performance, neglecting non-dominant domains. (2) Overfitting: Sparse data in non-dominant domains leads to overfitting in specific parameters. To tackle these challenges, we propose Soft-partitioned Semi-supervised Collaborative Transfer Learning (SSCTL) for multi-domain recommendation. SSCTL generates dynamic parameters to address the overwhelming issue, thus shifting focus towards samples from non-dominant domains. To combat overfitting, it leverages pseudo-labels with weights from dominant domain instances to enhance non-dominant domain data. We conduct comprehensive experiments, both online and offline, to validate the efficacy of our proposed method. Online tests yielded significant improvements across various domains, with increases in GMV ranging from 0.54% to 2.90% and enhancements in CTR ranging from 0.22% to 1.69%.

cs.IR↗

arXiv2025

Data-Juicer 2.0: Cloud-Scale Adaptive Data Processing for and with Foundation Models

Foundation models demand advanced data processing for their vast, multimodal datasets. However, traditional frameworks struggle with the unique complexities of multimodal data. In response, we present Data-Juicer 2.0, a data processing system backed by 100+ data processing operators spanning text, image, video, and audio modalities, supporting more critical tasks including data analysis, synthesis, annotation, and foundation model post-training. With seamless compatibility and dedicated optimization for popular dataset hubs like Hugging Face and computing engines like Ray, it improves upon its predecessor in terms of usability, efficiency, and programmability. It features an easily accessible user interface layer that supports decoupled Python interactions, RESTful APIs, and conversational commands. Its new runtime layer offers adaptive execution across diverse scales and environments, abstracting away system complexities. Extensive empirical evaluations demonstrate Data-Juicer 2.0's remarkable performance and scalability, highlighting its capability to efficiently process TB-level data with 10k+ CPU cores. The system is publicly available and has been widely adopted in diverse research fields and real-world products such as Alibaba Cloud PAI. We actively maintain the system and share practical insights to foster research and applications of next-generation foundation models.

cs.DC↗