arXiv subjects

Zhanwei Zhang

Publications and source records attributed to Zhanwei Zhang.

16 recordsLinked to original sources

arXiv2026

From Inertia to Objectivity: Improving Deep Research Agents with Noise Isolation

Web search agents powered by Large Language Models (LLMs) show strong promise, but deep research tasks expose a recurring failure mode: once an agent has produced a query, plan, or intermediate conclusion, it becomes less objective when later judging the consequences of that same action. We term this phenomenon inertia bias. To make it measurable, we introduce the IBIS benchmark, which controls the search observations while varying whether the model is evaluating the outcome of its own prior action. We find that models are substantially worse when they "own" the preceding search step, showing that self-authored action history can systematically distort subsequent judgment. We further show that this bias propagates into two forms of system-level degradation: search noise at the worker level and contextual noise at the manager level. To address this problem, we propose NIS-Agent, which applies context isolation at the two decision points most vulnerable to inertia bias: webpage triage and final-answer validation. Across GAIA, WebWalkerQA, BrowseComp, and BrowseComp-zh, NIS-Agent achieves competitive performance while reducing token cost by 33% compared to our baseline. We further train an 8B model to be intrinsically more resistant to inertia bias; under the same NIS-Agent framework, it attains average performance comparable to GPT-4o on deep research benchmarks. Our code is publicly available at https://github.com/PangSMPang/NIS-Agent.

cs.AI↗

arXiv2026

DR-Venus: Towards Frontier Edge-Scale Deep Research Agents with Only 10K Open Data

Edge-scale deep research agents based on small language models are attractive for real-world deployment due to their advantages in cost, latency, and privacy. In this work, we study how to train a strong small deep research agent under limited open-data by improving both data quality and data utilization. We present DR-Venus, a frontier 4B deep research agent for edge-scale deployment, built entirely on open data. Our training recipe consists of two stages. In the first stage, we use agentic supervised fine-tuning (SFT) to establish basic agentic capability, combining strict data cleaning with resampling of long-horizon trajectories to improve data quality and utilization. In the second stage, we apply agentic reinforcement learning (RL) to further improve execution reliability on long-horizon deep research tasks. To make RL effective for small agents in this setting, we build on IGPO and design turn-level rewards based on information gain and format-aware regularization, thereby enhancing supervision density and turn-level credit assignment. Built entirely on roughly 10K open-data, DR-Venus-4B significantly outperforms prior agentic models under 9B parameters on multiple deep research benchmarks, while also narrowing the gap to much larger 30B-class systems. Our further analysis shows that 4B agents already possess surprisingly strong performance potential, highlighting both the deployment promise of small models and the value of test-time scaling in this setting. We release our models, code, and key recipes to support reproducible research on edge-scale deep research agents.

cs.LG↗

arXiv2026

ArtiCAD: Articulated CAD Assembly Design via Multi-Agent Code Generation

Parametric Computer-Aided Design (CAD) of articulated assemblies is essential for product development, yet generating these multi-part, movable models from high-level descriptions remains unexplored. To address this, we propose ArtiCAD, the first training-free multi-agent system capable of generating editable, articulated CAD assemblies directly from text or images. Our system divides this complex task among four specialized agents: Design, Generation, Assembly, and Review. One of our key insights is to predict assembly relationships during the initial design stage rather than the assembly stage. By utilizing a Connector that explicitly defines attachment points and joint parameters, ArtiCAD determines these relationships before geometry generation, effectively bypassing the limited spatial reasoning capabilities of current LLMs and VLMs. To further ensure high-quality outputs, we introduce validation steps in the generation and assembly stages, accompanied by a cross-stage rollback mechanism that accurately isolates and corrects design- and code-level errors. Additionally, a self-evolving experience store accumulates design knowledge to continuously improve performance on future tasks. Extensive evaluations on three datasets (ArtiCAD-Bench, CADPrompt, and ACD) validate the effectiveness of our approach. We further demonstrate the applicability of ArtiCAD in requirement-driven conceptual design, physical prototyping, and the generation of embodied AI training assets through URDF export.

cs.CV↗

arXiv2025

A Mathematics-Guided Approach to Floating-Point Error Detection

Floating-point program errors can lead to severe consequences, particularly in critical domains such as military applications. Only a small subset of inputs may induce substantial floating-point errors, prompting researchers to develop methods for identifying these error-inducing inputs. Although existing approaches have achieved some success, they still suffer from two major limitations: (1) High computational cost: The evaluation of error magnitude for candidate inputs relies on high-precision programs, which are prohibitively time-consuming. (2) Limited long-range convergence capability: Current methods exhibit inefficiency in search, making the process akin to finding a needle in a haystack. To address these two limitations, we propose a novel method, named MGDE, to detect error-inducing inputs based on mathematical guidance. By employing the Newton-Raphson method, which exhibits quadratic convergence properties, we achieve highly effective and efficient results. Since the goal of identifying error-inducing inputs is to uncover the underlying bugs, we use the number of bugs detected in floating-point programs as the primary evaluation metric in our experiments. As FPCC represents the most effective state-of-the-art approach to date, we use it as the baseline for comparison. The dataset of FPCC consists of 88 single-input floating-point programs. FPCC is able to detect 48 bugs across 29 programs, whereas our method successfully identifies 89 bugs across 44 programs. Moreover, FPCC takes 6.4096 times as long as our proposed method. We also deploy our method to multi-input programs, identifying a total of nine bugs with an average detection time of 0.6443 seconds per program. In contrast, FPCC fails to detect any bugs while requiring an average computation time of 100 seconds per program.

cs.SE↗

arXiv2025

EviNote-RAG: Enhancing RAG Models via Answer-Supportive Evidence Notes

Retrieval-Augmented Generation (RAG) has advanced open-domain question answering by incorporating external information into model reasoning. However, effectively leveraging external information to enhance reasoning presents the following challenges: (1) low signal-to-noise ratio, where answer-supportive external information is diluted by irrelevant material, and (2) error accumulation, which arises in multi-hop reasoning when incomplete or misleading information is incorporated. To address these challenges, we introduce EviNote-RAG, a framework that follows a retrieve-note-answer workflow. Instead of reasoning directly over raw external information, the model first produces Supportive-Evidence Notes (SENs), which concisely preserve answer-critical information and explicitly mark key and uncertainty information to improve accuracy. We further design an entailment-based Evidence Quality Reward (EQR) to ensure that SENs are logically sufficient to derive the final answer, thereby enhancing SENs' quality. Experiments on both in-domain and out-of-domain QA benchmarks show that EviNote-RAG achieves state-of-the-art performance, improving answer accuracy, training stability, robustness, and efficiency. In particular, it yields relative F1 gains of 20% on HotpotQA (+0.093), 40% on Bamboogle (+0.151), and 91% on 2Wiki (+0.256), benefiting from improvements in the reasoning process.

cs.CL↗

arXiv2025

Atom-Searcher: Enhancing Agentic Deep Research via Fine-Grained Atomic Thought Reward

Large language models (LLMs) exhibit remarkable problem-solving abilities, but struggle with complex tasks due to static internal knowledge. Retrieval-Augmented Generation (RAG) enhances access to external information, yet remains limited in multi-hop reasoning and strategic search due to rigid workflows. Recent advancements in agentic deep research empower LLMs to autonomously reason, search, and synthesize information. However, current approaches relying on outcome-based reinforcement learning (RL) face critical issues such as conflicting gradients and reward sparsity, limiting performance gains and training efficiency. To address these, we first propose Atomic Thought, a novel LLM thinking paradigm that decomposes reasoning into fine-grained functional units. These units are supervised by Reasoning Reward Models (RRMs), which provide Atomic Thought Rewards (ATR) for fine-grained guidance. Building on this, we propose Atom-Searcher, a novel RL framework for agentic deep research that integrates Atomic Thought and ATR. Atom-Searcher uses a curriculum-inspired reward schedule, prioritizing process-level ATR early and transitioning to outcome rewards, accelerating convergence on effective reasoning paths. Experiments on seven benchmarks show consistent improvements over the state-of-the-art. Key advantages include: (1) Atom-Searcher scales computation at test-time. (2) Atomic Thought provides supervision anchors for RRMs, bridging deep research tasks and RRMs. (3) Atom-Searcher exhibits more interpretable, human-like reasoning patterns.

cs.CL↗

arXiv2025

Careful Queries, Credible Results: Teaching RAG Models Advanced Web Search Tools with Reinforcement Learning

Retrieval-Augmented Generation (RAG) enhances large language models (LLMs) by integrating up-to-date external knowledge, yet real-world web environments present unique challenges. These limitations manifest as two key challenges: pervasive misinformation in the web environment, which introduces unreliable or misleading content that can degrade retrieval accuracy, and the underutilization of web tools, which, if effectively employed, could enhance query precision and help mitigate this noise, ultimately improving the retrieval results in RAG systems. To address these issues, we propose WebFilter, a novel RAG framework that generates source-restricted queries and filters out unreliable content. This approach combines a retrieval filtering mechanism with a behavior- and outcome-driven reward strategy, optimizing both query formulation and retrieval outcomes. Extensive experiments demonstrate that WebFilter improves answer quality and retrieval precision, outperforming existing RAG methods on both in-domain and out-of-domain benchmarks.

cs.IR↗

arXiv2025

Computing Floating-Point Errors by Injecting Perturbations

Floating-point programs form the foundation of modern science and engineering, providing the essential computational framework for a wide range of applications, such as safety-critical systems, aerospace engineering, and financial analysis. Floating-point errors can lead to severe consequences. Although floating-point errors widely exist, only a subset of inputs may trigger significant errors in floating-point programs. Therefore, it is crucial to determine whether a given input could produce such errors. Researchers tend to take the results of high-precision floating-point programs as oracles for detecting floating-point errors, which introduces two main limitations: (1) difficulty of implementation and (2) prolonged execution time. The two recent tools, ATOMU and FPCC, can partially address these issues. However, ATOMU suffers from false positives; while FPCC, though eliminating false positives, operates at a considerably slower speed. To address these two challenges, we propose a novel approach named PI-detector to computing floating-point errors effectively and efficiently. Our approach is based on the observation that floating-point errors stem from large condition numbers in atomic operations (such as addition and subtraction), which then propagate and accumulate. PI-detector injects small perturbations into the operands of individual atomic operations within the program and compares the outcomes of the original program with the perturbed version to compute floating-point errors. We evaluate PI-detector with datasets from ATOMU and HSED, as well as a complex linear system-solving program. Experimental results demonstrate that PI-detector can perform efficient and accurate floating-point error computation.

cs.SE↗

arXiv2025

Efficient Reasoning Through Suppression of Self-Affirmation Reflections in Large Reasoning Models

While recent advances in large reasoning models have demonstrated remarkable performance, efficient reasoning remains critical due to the rapid growth of output length. Existing optimization approaches highlights a tendency toward "overthinking", yet lack fine-grained analysis. In this work, we focus on Self-Affirmation Reflections: redundant reflective steps that affirm prior content and often occurs after the already correct reasoning steps. Observations of both original and optimized reasoning models reveal pervasive self-affirmation reflections. Notably, these reflections sometimes lead to longer outputs in optimized models than their original counterparts. Through detailed analysis, we uncover an intriguing pattern: compared to other reflections, the leading words (i.e., the first word of sentences) in self-affirmation reflections exhibit a distinct probability bias. Motivated by this insight, we can locate self-affirmation reflections and conduct a train-free experiment demonstrating that suppressing self-affirmation reflections reduces output length without degrading accuracy across multiple models (R1-Distill-Models, QwQ-32B, and Qwen3-32B). Furthermore, we also improve current train-based method by explicitly suppressing such reflections. In our experiments, we achieve length compression of 18.7\% in train-free settings and 50.2\% in train-based settings for R1-Distill-Qwen-1.5B. Moreover, our improvements are simple yet practical and can be directly applied to existing inference frameworks, such as vLLM. We believe that our findings will provide community insights for achieving more precise length compression and step-level efficient reasoning.

cs.CL↗

arXiv2025

GeoCAD: Local Geometry-Controllable CAD Generation with Large Language Models

Local geometry-controllable computer-aided design (CAD) generation aims to modify local parts of CAD models automatically, enhancing design efficiency. It also ensures that the shapes of newly generated local parts follow user-specific geometric instructions (e.g., an isosceles right triangle or a rectangle with one corner cut off). However, existing methods encounter challenges in achieving this goal. Specifically, they either lack the ability to follow textual instructions or are unable to focus on the local parts. To address this limitation, we introduce GeoCAD, a user-friendly and local geometry-controllable CAD generation method. Specifically, we first propose a complementary captioning strategy to generate geometric instructions for local parts. This strategy involves vertex-based and VLLM-based captioning for systematically annotating simple and complex parts, respectively. In this way, we caption $\sim$221k different local parts in total. In the training stage, given a CAD model, we randomly mask a local part. Then, using its geometric instruction and the remaining parts as input, we prompt large language models (LLMs) to predict the masked part. During inference, users can specify any local part for modification while adhering to a variety of predefined geometric instructions. Extensive experiments demonstrate the effectiveness of GeoCAD in generation quality, validity and text-to-CAD consistency. Code will be available at https://github.com/Zhanwei-Z/GeoCAD.

cs.CV↗

arXiv2025

FlexCAD: Unified and Versatile Controllable CAD Generation with Fine-tuned Large Language Models

Recently, there is a growing interest in creating computer-aided design (CAD) models based on user intent, known as controllable CAD generation. Existing work offers limited controllability and needs separate models for different types of control, reducing efficiency and practicality. To achieve controllable generation across all CAD construction hierarchies, such as sketch-extrusion, extrusion, sketch, face, loop and curve, we propose FlexCAD, a unified model by fine-tuning large language models (LLMs). First, to enhance comprehension by LLMs, we represent a CAD model as a structured text by abstracting each hierarchy as a sequence of text tokens. Second, to address various controllable generation tasks in a unified model, we introduce a hierarchy-aware masking strategy. Specifically, during training, we mask a hierarchy-aware field in the CAD text with a mask token. This field, composed of a sequence of tokens, can be set flexibly to represent various hierarchies. Subsequently, we ask LLMs to predict this masked field. During inference, the user intent is converted into a CAD text with a mask token replacing the part the user wants to modify, which is then fed into FlexCAD to generate new CAD models. Comprehensive experiments on public dataset demonstrate the effectiveness of FlexCAD in both generation quality and controllability. Code will be available at https://github.com/microsoft/FlexCAD.

cs.CV↗

arXiv2025

DELA: A Novel Approach for Detecting Errors Induced by Large Atomic Condition Numbers

Numerical programs form the foundation of modern science and engineering, providing essential solutions to complex mathematical problems. Therefore, errors in numerical results would lead to harmful consequences, especially in safety-critical applications. Since only a few inputs may lead to substantial errors for numerical programs, it is essential to determine whether a given input could result in a significant error. Existing researchers tend to use the results of high-precision programs to assess whether there is a substantial error, which introduces three main challenges: difficulty of implementation, existence of potential faults in the detection of numerical errors, and long execution time. To address these limitations, we propose a novel approach named DELA. Our approach is based on the observation that most numerical errors stem from large condition numbers in atomic operations (such as subtraction), which then propagate and accumulate. DELA injects small perturbations into the results of individual atomic operations within the program and compares the outcomes of the original program with the perturbed version to detect errors. We evaluate DELA with datasets from ATOMU and HSED, as well as data from a complex linear system-solving program. Experimental results demonstrate that we can detect all the significant errors that were reported by prior research. DELA shows strong alignment with high-precision programs of ATOMU and HSED, with average Pearson and Spearman correlations of 0.86 and 0.61. Additionally, DELA effectively detects significant errors in complex programs, achieving correlation scores of 0.9763 and 0.8993. More importantly, in experiments with ATOMU and HSED, DELA's perturbed programs run within only 0.13% of the time needed by high-precision versions; while for the linear system-solving programs, DELA is 73.46 times faster than the high-precision programs.

cs.SE↗

arXiv2024

Large Language Models Based JSON Parser Fuzzing for Bug Discovery and Behavioral Analysis

Fuzzing has been incredibly successful in uncovering bugs and vulnerabilities across diverse software systems. JSON parsers play a vital role in modern software development, and ensuring their reliability is of great importance. This research project focuses on leveraging Large Language Models (LLMs) to enhance JSON parser testing. The primary objectives are to generate test cases and mutants using LLMs for the discovery of potential bugs in open-source JSON parsers and the identification of behavioral diversities among them. We aim to uncover underlying bugs, plus discovering (and overcoming) behavioral diversities.

cs.SE↗

arXiv2024

G2LTraj: A Global-to-Local Generation Approach for Trajectory Prediction

Predicting future trajectories of traffic agents accurately holds substantial importance in various applications such as autonomous driving. Previous methods commonly infer all future steps of an agent either recursively or simultaneously. However, the recursive strategy suffers from the accumulated error, while the simultaneous strategy overlooks the constraints among future steps, resulting in kinematically infeasible predictions. To address these issues, in this paper, we propose G2LTraj, a plug-and-play global-to-local generation approach for trajectory prediction. Specifically, we generate a series of global key steps that uniformly cover the entire future time range. Subsequently, the local intermediate steps between the adjacent key steps are recursively filled in. In this way, we prevent the accumulated error from propagating beyond the adjacent key steps. Moreover, to boost the kinematical feasibility, we not only introduce the spatial constraints among key steps but also strengthen the temporal constraints among the intermediate steps. Finally, to ensure the optimal granularity of key steps, we design a selectable granularity strategy that caters to each predicted trajectory. Our G2LTraj significantly improves the performance of seven existing trajectory predictors across the ETH, UCY and nuScenes datasets. Experimental results demonstrate its effectiveness. Code will be available at https://github.com/Zhanwei-Z/G2LTraj.

cs.CV↗

arXiv2024

Pseudo Label Refinery for Unsupervised Domain Adaptation on Cross-dataset 3D Object Detection

Recent self-training techniques have shown notable improvements in unsupervised domain adaptation for 3D object detection (3D UDA). These techniques typically select pseudo labels, i.e., 3D boxes, to supervise models for the target domain. However, this selection process inevitably introduces unreliable 3D boxes, in which 3D points cannot be definitively assigned as foreground or background. Previous techniques mitigate this by reweighting these boxes as pseudo labels, but these boxes can still poison the training process. To resolve this problem, in this paper, we propose a novel pseudo label refinery framework. Specifically, in the selection process, to improve the reliability of pseudo boxes, we propose a complementary augmentation strategy. This strategy involves either removing all points within an unreliable box or replacing it with a high-confidence box. Moreover, the point numbers of instances in high-beam datasets are considerably higher than those in low-beam datasets, also degrading the quality of pseudo labels during the training process. We alleviate this issue by generating additional proposals and aligning RoI features across different domains. Experimental results demonstrate that our method effectively enhances the quality of pseudo labels and consistently surpasses the state-of-the-art methods on six autonomous driving benchmarks. Code will be available at https://github.com/Zhanwei-Z/PERE.

cs.CV↗

arXiv2024

UniHDA: A Unified and Versatile Framework for Multi-Modal Hybrid Domain Adaptation

Recently, generative domain adaptation has achieved remarkable progress, enabling us to adapt a pre-trained generator to a new target domain. However, existing methods simply adapt the generator to a single target domain and are limited to a single modality, either text-driven or image-driven. Moreover, they cannot maintain well consistency with the source domain, which impedes the inheritance of the diversity. In this paper, we propose UniHDA, a \textbf{unified} and \textbf{versatile} framework for generative hybrid domain adaptation with multi-modal references from multiple domains. We use CLIP encoder to project multi-modal references into a unified embedding space and then linearly interpolate the direction vectors from multiple target domains to achieve hybrid domain adaptation. To ensure \textbf{consistency} with the source domain, we propose a novel cross-domain spatial structure (CSS) loss that maintains detailed spatial structure information between source and target generator. Experiments show that the adapted generator can synthesise realistic images with various attribute compositions. Additionally, our framework is generator-agnostic and versatile to multiple generators, e.g., StyleGAN, EG3D, and Diffusion Models.

cs.CV↗