arXiv subjects

Zhengxin Zhang

Publications and source records attributed to Zhengxin Zhang.

At least 19 recordsLinked to original sources

arXiv2026

Probabilistic Representation and Convergence of Gromov-Wasserstein Gradient Flows

Wasserstein gradient flows are intimately connected with evolution partial differential equations and diffusion processes. We take the first step in developing such connections for inner product Gromov--Wasserstein (IGW) gradient flows by studying the IGW gradient flow of the relative entropy $\mathsf{H}(\cdot\|γ)$ with respect to the standard Gaussian measure $γ$. We first show that $\mathsf{H}(\cdot\|γ)$ fails to be $λ$-convex along generalized or modified generalized IGW geodesics for any $λ\in \mathbb{R}$, and therefore falls outside the scope of the existing IGW gradient flow theory from Zhang et al. (2026). We bridge this gap by establishing a suitable \emph{local} convexity estimate that enables the construction of the gradient flow and its extension to the infinite time horizon. We then obtain increasingly explicit representations of the resulting dynamics. Starting from a partial integro-differential equation, we derive a nonlinear Fokker--Planck equation and show that its second-moment dynamics decouple from the law as they satisfy an autonomous matrix ODE. This reduces the IGW dynamics to a linear, time-inhomogeneous Fokker--Planck equation, yielding a probabilistic representation as the time-marginal flow of a linear stochastic differential equation resembling the Ornstein--Uhlenbeck process. Finally, we study its asymptotic behavior by establishing exponential convergence of the flow to $γ$ in relative entropy.

math.PR↗

arXiv2026

MANTA: Multi-Agent Network Topology Adaptation for Self-Evolving Multi-Agent Systems

Large language model-based multi-agent systems improve complex problem solving through task decomposition, agent specialization, information exchange, and intermediate validation. However, existing systems typically treat communication topology as a fixed design choice or an offline optimization target. We introduce MANTA, a framework for Multi-Agent Network Topology Adaptation that enables communication structures to self-evolve at inference time. Before execution, MANTA initializes a task-conditioned topology from prior structural experience. During deployment, it monitors collaboration traces and applies bounded structural updates when the current organization becomes insufficient. These updates can modify agent roles, communication links, execution order, information visibility, and validation pathways while preserving the task interface and agent budget. We evaluate MANTA against representative single-agent and multi-agent baselines on five benchmarks spanning information seeking, tool use, planning, workflow execution, and mathematical reasoning. MANTA achieves the highest average score of 74.0, outperforming the strongest baseline by 5.8 percentage points and obtaining the best result on PlanCraft. These results show that inference-time self-improvement can extend to the architecture of collaboration itself.

cs.AI↗

arXiv2026

HARP: Efficient Data Selection for Finetuning Large Language Models

Finetuning data selection requires balancing two competing goals: selecting examples that improve the downstream objective, and doing so without repeatedly finetuning models. Train-free selectors are scalable but rely on proxies such as embedding similarity or clustering, which may not match the target objective. Train-based selectors better reflect downstream utility through gradient signals, subset evaluation, or Shapley attribution, but require many costly train--evaluate iterations. We propose Hierarchical Active Region Pruning (HARP), an efficient train-based selector that preserves downstream alignment while reducing selection cost. HARP organizes the training pool into a node--leaf hierarchy, evaluates only representative leaves, and infers unmeasured utilities with empirical Bayes posteriors. It then selects data using two complementary envelopes: HARP-C, which conservatively controls redundancy, and HARP-E, which additively rewards complementary regions. We theoretically show that, under local smoothness and bounded estimation error, HARP controls selection error while reducing train--evaluate cost. We further validate that HARP variants achieve the best result and outperform the strongest baseline by up to $+8.9$ points, while using roughly $7\times$ fewer training examples.

cs.LG↗

arXiv2026

CAPC-CG: A Large-Scale, Expert-Directed LLM-Annotated Corpus of Adaptive Policy Communication in China

We introduce CAPC-CG, the Chinese Adaptive Policy Communication (Central Government) Corpus, the first open dataset of Chinese policy directives annotated with a five-color taxonomy of clear and ambiguous language categories, building on Ang's theory of adaptive policy communication. Spanning 1949-2023, this corpus includes national laws, administrative regulations, and ministerial rules issued by China's top authorities. Each document is segmented into paragraphs, producing a total of 3.3 million units. Alongside the corpus, we release comprehensive metadata, a two-round labeling framework, and a gold-standard annotation set developed by expert and trained coders. Inter-annotator agreement achieves a Fleiss's kappa of K = 0.86 on directive labels, indicating high reliability for supervised modeling. We provide baseline classification results with several large language models (LLMs), together with our annotation codebook, and describe patterns from the dataset. This release aims to support downstream tasks and multilingual NLP research in policy communication.

cs.CL↗

arXiv2026

How Far Are We From True Auto-Research?

Recent auto-research systems can produce complete papers, but feasibility is not the same as quality, and the field still lacks a systematic study of how good agent-generated papers actually are. We introduce ResearchArena, a minimal scaffold that lets off-the-shelf agents (Claude Code using Opus 4.6, Codex using GPT-5.4, and Kimi Code using K2.5) carry out the full research loop themselves (ideation, experimentation, paper writing, self-refinement) under only lightweight guidance. Across 13 computer science seeds and 3 trials per agent-domain pair, ResearchArena yields 117 agent-generated papers, each evaluated under three complementary lenses: a manuscript-only reviewer (SAR), an artifact-aware peer review (PR) in which agents inspect the workspace alongside the manuscript, and an human conducted meta-review. Under SAR alone the picture is optimistic: Claude Code obtains the highest score, outperforms Analemma's FARS, and matches the weighted-average human ICLR 2025 submission, suggesting that minimally scaffolded agents can produce papers that look competitive on manuscript-only review. Manual inspection, however, reveals this picture is overstated: SAR scores are poorly aligned with its actual acceptance decisions and reward plausible framing without verifying experimental substance. Under artifact-aware PR scores drop sharply, and manual auditing identifies experimental rigor as the major bottleneck, decomposing into three failure modes (fabricated results, underpowered experiments, and plan/execution mismatch) that are highly agent-dependent: Codex 5%/8% paper-vs-artifact mismatch / fabricated references versus Kimi Code 77%/72%, a $\sim$15$\times$ spread that tracks distinct research personas the agents develop. None of the 117 agent-generated papers reaches the acceptance bar of a top-tier venue. This suggests that we are still gapped from the true auto-research.

cs.AI↗

arXiv2026

Bootstrapping Post-training Signals for Open-ended Tasks via Rubric-based Self-play on Pre-training Text

Self-play has recently emerged as a promising paradigm for post-training Large Language Models (LLMs). In self-play, the target LLM creates the task input (e.g., a question), which it then addresses itself by producing a task output (e.g., an answer). A reward model evaluates the output, and the rewards are used to train the LLM, typically via Reinforcement Learning (RL). A key benefit of self-play for post-training LLMs is its minimal supervision costs: self-play avoids the need for high-quality input-output pairs traditionally constructed by humans or expensive proprietary models. Existing work, however, explores self-play only for verifiable tasks, such as math and coding, for which objective ground truth is available and easily checkable. In this paper, we seek to extend self-play to more realistic open-ended tasks. We propose POP, a self-play framework that uses the same LLM to synthesize evaluation rubrics along with each input-output pair. The rubric is used to evaluate outputs and train the model. Crucially, we ground the framework on a content-rich pretraining corpus to (1) enable an exploitable generation-verification gap and reduce reward hacking, and (2) prevent mode collapse. On Qwen-2.5-7B, POP increases performance of both the pretrained base model and instruction-tuned model on multiple tasks ranging from long-form healthcare QA to creative writing and instruction following.

cs.CL↗

arXiv2026

AppellateGen: A Benchmark for Appellate Legal Judgment Generation

Legal judgment generation is a critical task in legal intelligence. However, existing research in legal judgment generation has predominantly focused on first-instance trials, relying on static fact-to-verdict mappings while neglecting the dialectical nature of appellate (second-instance) review. To address this, we introduce AppellateGen, a benchmark for second-instance legal judgment generation comprising 7,351 case pairs. The task requires models to draft legally binding judgments by reasoning over the initial verdict and evidentiary updates, thereby modeling the causal dependency between trial stages. We further propose a judicial Standard Operating Procedure (SOP)-based Legal Multi-Agent System (SLMAS) to simulate judicial workflows, which decomposes the generation process into discrete stages of issue identification, retrieval, and drafting. Experimental results indicate that while SLMAS improves logical consistency, the complexity of appellate reasoning remains a substantial challenge for current LLMs. The dataset and code are publicly available at: https://anonymous.4open.science/r/AppellateGen-5763.

cs.CY↗

arXiv2026

Discussing Your Needs in VR: A Novel Approach through Persona-based Stakeholder Role-Playing

In this study, we propose a novel approach that supports requirements discussions in virtual environments by automatically generating personas from real-time speech-to-text data. In our pilot experiment, 18 participants (14 from universities and 4 from IT companies) used the generated personas to discuss accessibility requirements within the virtual environment. Participants reported a relatively high level of satisfaction with the social presence and usability of the VR system. We also found that requirements discussions based on personas have a lower workload. Finally, we outline the main directions for future work.

cs.HC↗

arXiv2026

Better LLM Reasoning via Dual-Play

Large Language Models (LLMs) have achieved remarkable progress through Reinforcement Learning with Verifiable Rewards (RLVR), yet still rely heavily on external supervision (e.g., curated labels). Adversarial learning, particularly through self-play, offers a promising alternative that enables models to iteratively learn from themselves - thus reducing reliance on external supervision. Dual-play extends adversarial learning by assigning specialized roles to two models and training them against each other, fostering sustained competition and mutual evolution. Despite its promise, adapting dual-play training to LLMs remains limited, largely due to their susceptibility to reward hacking and training instability. In this paper, we introduce PasoDoble, a novel LLM dual-play framework. PasoDoble adversarially trains two models initialized from the same base model: a Proposer, which generates challenging questions with ground-truth answers, and a Solver, which attempts to solve them. We enrich the Proposer with knowledge from a pre-training dataset to ensure the questions' quality and diversity. To avoid reward hacking, the Proposer is rewarded for producing only valid questions that push the Solver's limit, while the Solver is rewarded for solving them correctly, and both are updated jointly. To further enhance training stability, we introduce an optional offline paradigm that decouples Proposer and Solver updates, alternately updating each for several steps while holding the other fixed. Notably, PasoDoble operates without supervision during training. Experimental results show that PasoDoble can improve the reasoning performance of LLMs. Our project page is available at https://hcy123902.github.io/PasoDoble.

cs.LG↗

arXiv2026

How Long Does It Take to Alleviate Discomfort? A Preliminary Study on Reducing Cybersickness in Novice Users

Cybersickness significantly impacts the user experience in VR applications. Locomotion tunneling is a widely adopted technique for mitigating cybersickness in susceptible users. However, there is a lack of research investigating the effects of prolonged use of locomotion tunneling among novice users. To fill this gap, we used VRChat as our experimental platform. We recruited 24 novice VR users, defined as participants with no prior experience using immersive virtual environments. We collected five days of data within a one-week period. The results indicated that participants exhibited significant mitigation to cybersickness by Day 4. However, a change in the VR scene on Day 5 led to a notable increase in cybersickness symptoms. Qualitative feedback revealed participant-perceived causes of cybersickness and suggested that the effectiveness of locomotion tunneling was limited in some scenarios. Finally, we discussed the limitations of the study and proposed directions for future research.

cs.HC↗

arXiv2025

HAPO: Training Language Models to Reason Concisely via History-Aware Policy Optimization

While scaling the length of responses at test-time has been shown to markedly improve the reasoning abilities and performance of large language models (LLMs), it often results in verbose outputs and increases inference cost. Prior approaches for efficient test-time scaling, typically using universal budget constraints or query-level length optimization, do not leverage historical information from previous encounters with the same problem during training. We hypothesize that this limits their ability to progressively make solutions more concise over time. To address this, we present History-Aware Policy Optimization (HAPO), which keeps track of a history state (e.g., the minimum length over previously generated correct responses) for each problem. HAPO employs a novel length reward function based on this history state to incentivize the discovery of correct solutions that are more concise than those previously found. Crucially, this reward structure avoids overly penalizing shorter incorrect responses with the goal of facilitating exploration towards more efficient solutions. By combining this length reward with a correctness reward, HAPO jointly optimizes for correctness and efficiency. We use HAPO to train DeepSeek-R1-Distill-Qwen-1.5B, DeepScaleR-1.5B-Preview, and Qwen-2.5-1.5B-Instruct, and evaluate HAPO on several math benchmarks that span various difficulty levels. Experiment results demonstrate that HAPO effectively induces LLMs' concise reasoning abilities, producing length reductions of 33-59% with accuracy drops of only 2-5%.

cs.CL↗

arXiv2025

Towards More Robust Retrieval-Augmented Generation: Evaluating RAG Under Adversarial Poisoning Attacks

Retrieval-Augmented Generation (RAG) systems have emerged as a promising solution to mitigate LLM hallucinations and enhance their performance in knowledge-intensive domains. However, these systems are vulnerable to adversarial poisoning attacks, where malicious passages injected into the retrieval corpus can mislead models into producing factually incorrect outputs. In this paper, we present a rigorously controlled empirical study of how RAG systems behave under such attacks and how their robustness can be improved. On the generation side, we introduce a structured taxonomy of context types-adversarial, untouched, and guiding-and systematically analyze their individual and combined effects on model outputs. On the retrieval side, we evaluate several retrievers to measure how easily they expose LLMs to adversarial contexts. Our findings also reveal that "skeptical prompting" can activate LLMs' internal reasoning, enabling partial self-defense against adversarial passages, though its effectiveness depends strongly on the model's reasoning capacity. Together, our experiments (code available at https://github.com/JinyanSu1/eval_PoisonRaG) and analysis provide actionable insights for designing safer and more resilient RAG systems, paving the way for more reliable real-world deployments.

cs.IR↗

arXiv2025

EEG-fused Digital Twin Brain for Autonomous Driving in Virtual Scenarios

Current methodologies typically integrate biophysical brain models with functional magnetic resonance imaging(fMRI) data - while offering millimeter-scale spatial resolution (0.5-2 mm^3 voxels), these approaches suffer from limited temporal resolution (>0.5 Hz) for tracking rapid neural dynamics during continuous tasks. Conversely, Electroencephalogram (EEG) provides millisecond-scale temporal precision (<=1 ms sampling rate) for real-time guidance of continuous task execution, albeit constrained by low spatial resolution. To reconcile these complementary modalities, we present a generalizable Bayesian inference framework that integrates high-spatial-resolution structural MRI(sMRI) with high-temporal-resolution EEG to construct a biologically realistic digital twin brain(DTB) model. The framework establishes voxel-wise mappings between millisecond-scale EEG and sMRI-derived spiking networks, while demonstrating its translational potential through a brain-inspired autonomous driving simulation. Our EEG-DTB model achieves capabilities: (1) Biologically-plausible EEG signal generation (0.88 resting-state,0.60 task-state correlation), with simulated signals in task-state yielding steering predictions outperforming both chance and empirical signals (p<0.05); (2) Successful autonomous driving in the CARLA simulator using decoded steering angles. The proposed approach pioneers a new paradigm for studying sensorimotor integration and for mechanistic studies of perception-action cycles and the development of brain-inspired control systems.

q-bio.NC↗

arXiv2025

Gradient Flows and Riemannian Structure in the Gromov-Wasserstein Geometry

The Wasserstein space of probability measures is known for its intricate Riemannian structure, which underpins the Wasserstein geometry and enables gradient flow algorithms. However, the Wasserstein geometry may not be suitable for certain tasks or data modalities. Motivated by scenarios where the global structure of the data needs to be preserved, this work initiates the study of gradient flows and Riemannian structure in the Gromov-Wasserstein (GW) geometry, which is particularly suited for such purposes. We focus on the inner product GW (IGW) distance between distributions on $\mathbb{R}^d$. Given a functional $\mathsf{F}:\mathcal{P}_2(\mathbb{R}^d)\to\mathbb{R}$ to optimize, we present an implicit IGW minimizing movement scheme that generates a sequence of distributions $\{ρ_i\}_{i=0}^n$, which are close in IGW and aligned in the 2-Wasserstein sense. Taking the time step to zero, we prove that the discrete solution converges to an IGW generalized minimizing movement (GMM) $(ρ_t)_t$ that follows the continuity equation with a velocity field $v_t\in L^2(ρ_t;\mathbb{R}^d)$, specified by a global transformation of the Wasserstein gradient of $\mathsf{F}$. The transformation is given by a mobility operator that modifies the Wasserstein gradient to encode not only local information, but also global structure. Our gradient flow analysis leads us to identify the Riemannian structure that gives rise to the intrinsic IGW geometry, using which we establish a Benamou-Brenier-like formula for IGW. We conclude with a formal derivation, akin to the Otto calculus, of the IGW gradient as the inverse mobility acting on the Wasserstein gradient. Numerical experiments validating our theory and demonstrating the global nature of IGW interpolations are provided.

math.AP↗

arXiv2025

LLM Evaluation Based on Aerospace Manufacturing Expertise: Automated Generation and Multi-Model Question Answering

Aerospace manufacturing demands exceptionally high precision in technical parameters. The remarkable performance of Large Language Models (LLMs), such as GPT-4 and QWen, in Natural Language Processing has sparked industry interest in their application to tasks including process design, material selection, and tool information retrieval. However, LLMs are prone to generating "hallucinations" in specialized domains, producing inaccurate or false information that poses significant risks to the quality of aerospace products and flight safety. This paper introduces a set of evaluation metrics tailored for LLMs in aerospace manufacturing, aiming to assess their accuracy by analyzing their performance in answering questions grounded in professional knowledge. Firstly, key information is extracted through in-depth textual analysis of classic aerospace manufacturing textbooks and guidelines. Subsequently, utilizing LLM generation techniques, we meticulously construct multiple-choice questions with multiple correct answers of varying difficulty. Following this, different LLM models are employed to answer these questions, and their accuracy is recorded. Experimental results demonstrate that the capabilities of LLMs in aerospace professional knowledge are in urgent need of improvement. This study provides a theoretical foundation and practical guidance for the application of LLMs in aerospace manufacturing, addressing a critical gap in the field.

cs.CL↗

arXiv2024

SpecInfer: Accelerating Generative Large Language Model Serving with Tree-based Speculative Inference and Verification

This paper introduces SpecInfer, a system that accelerates generative large language model (LLM) serving with tree-based speculative inference and verification. The key idea behind SpecInfer is leveraging small speculative models to predict the LLM's outputs; the predictions are organized as a token tree, whose nodes each represent a candidate token sequence. The correctness of all candidate token sequences represented by a token tree is verified against the LLM in parallel using a novel tree-based parallel decoding mechanism. SpecInfer uses an LLM as a token tree verifier instead of an incremental decoder, which significantly reduces the end-to-end latency and computational requirement for serving generative LLMs while provably preserving model quality. Our evaluation shows that SpecInfer outperforms existing LLM serving systems by 1.5-2.8x for distributed LLM inference and by 2.6-3.5x for offloading-based LLM inference, while preserving the same generative performance. SpecInfer is publicly available at https://github.com/flexflow/FlexFlow/

cs.CL↗

arXiv2024

Quantized Side Tuning: Fast and Memory-Efficient Tuning of Quantized Large Language Models

Finetuning large language models (LLMs) has been empirically effective on a variety of downstream tasks. Existing approaches to finetuning an LLM either focus on parameter-efficient finetuning, which only updates a small number of trainable parameters, or attempt to reduce the memory footprint during the training phase of the finetuning. Typically, the memory footprint during finetuning stems from three contributors: model weights, optimizer states, and intermediate activations. However, existing works still require considerable memory and none can simultaneously mitigate memory footprint for all three sources. In this paper, we present Quantized Side Tuing (QST), which enables memory-efficient and fast finetuning of LLMs by operating through a dual-stage process. First, QST quantizes an LLM's model weights into 4-bit to reduce the memory footprint of the LLM's original weights; QST also introduces a side network separated from the LLM, which utilizes the hidden states of the LLM to make task-specific predictions. Using a separate side network avoids performing backpropagation through the LLM, thus reducing the memory requirement of the intermediate activations. Furthermore, QST leverages several low-rank adaptors and gradient-free downsample modules to significantly reduce the trainable parameters, so as to save the memory footprint of the optimizer states. Experiments show that QST can reduce the total memory footprint by up to 2.3 $\times$ and speed up the finetuning process by up to 3 $\times$ while achieving competent performance compared with the state-of-the-art. When it comes to full finetuning, QST can reduce the total memory footprint up to 7 $\times$.

cs.LG↗

arXiv2023

Gromov-Wasserstein Distances: Entropic Regularization, Duality, and Sample Complexity

The Gromov-Wasserstein (GW) distance, rooted in optimal transport (OT) theory, quantifies dissimilarity between metric measure spaces and provides a framework for aligning heterogeneous datasets. While computational aspects of the GW problem have been widely studied, a duality theory and fundamental statistical questions concerning empirical convergence rates remained obscure. This work closes these gaps for the quadratic GW distance over Euclidean spaces of different dimensions $d_x$ and $d_y$. We treat both the standard and the entropically regularized GW distance, and derive dual forms that represent them in terms of the well-understood OT and entropic OT (EOT) problems, respectively. This enables employing proof techniques from statistical OT based on regularity analysis of dual potentials and empirical process theory, using which we establish the first GW empirical convergence rates. The derived two-sample rates are $n^{-2/\max\{\min\{d_x,d_y\},4\}}$ (up to a log factor when $\min\{d_x,d_y\}=4$) for standard GW and $n^{-1/2}$ for EGW, which matches the corresponding rates for standard and entropic OT. The parametric rate for EGW is evidently optimal, while for standard GW we provide matching lower bounds, which establish sharpness of the derived rates. We also study stability of EGW in the entropic regularization parameter and prove approximation and continuity results for the cost and optimal couplings. Lastly, the duality is leveraged to shed new light on the open problem of the one-dimensional GW distance between uniform distributions on $n$ points, illuminating why the identity and anti-identity permutations may not be optimal. Our results serve as a first step towards a comprehensive statistical theory as well as computational advancements for GW distances, based on the discovered dual formulations.

math.ST↗