Searcharxiv⌕ Search

arXiv subjects

Zhentao Zhang

Publications and source records attributed to Zhentao Zhang.

At least 19 recordsLinked to original sources
arXiv2026

State-OPD: State-Grounded On-Policy Distillation for Multi-Turn Agents

On-Policy Distillation (OPD) has emerged as a powerful paradigm for transferring reasoning capabilities from frontier models to smaller models, while extending its success to multi-turn agentic environments remains challenging. We study agentic OPD through interaction states, distinguishing the objective task and executed action-observation record from the internal semantic trajectory of student reasoning. Through this lens, we identify two roles of student reasoning in multi-turn OPD: prior reasoning distorts teacher supervision, while current reasoning shapes the action alternatives sampled at the same state. Specifically, at matched interaction states, removing prior student reasoning improves teacher continuation success and changes the teacher's token-level targets, with larger continuation gains at later turns. Meanwhile, resampling complete reasoning-action responses yields greater action diversity at the same state than resampling actions under a fixed reasoning trace. We introduce State-Grounded On-Policy Distillation, a framework that grounds teacher supervision in interaction states and reallocates the sampling budget to generate multiple complete reasoning-action responses at states with high response entropy, while retaining full-history student generation. Comprehensive evaluations across ALFWorld, ScienceWorld and WebShop demonstrate that State-OPD achieves the highest macro-average success rate among the evaluated methods for both Qwen3-1.7B and Qwen3-4B students. Notably, for Qwen3-4B, State-OPD raises macro-average success from 26.07\% with the strongest baseline to 31.74\% and surpasses its Qwen3-32B teacher on ALFWorld Seen. Further analyses reveal that State-OPD achieves higher task success with fewer supervised turn responses and substantially reduces persistent failed-action repetition; qualitative traces illustrate decision revision based on interaction feedback.

cs.LG↗
arXiv2026

MOPD-Router: Rethinking Teacher Routing in Multi-Teacher On-Policy Distillation

Multi-teacher on-policy distillation (MOPD) integrates specialized capabilities into a single student, but existing practice typically hard-routes each prompt to a domain-matched teacher for the entire rollout. This dependence on prompt-level domain labels restricts using unlabeled training mixtures and leaves complementary signals from other teachers unused. We introduce MOPD-Router, a framework that routes supervision over the full teacher pool at each token, without domain labels or training a separate routing model. Its plug-in interface supports different metrics for selecting and weighting teacher-specific OPD signals. Within this interface, we propose ExpertAlign, which scores each teacher by whether its correction to the student at the current token expresses the specialization that teacher acquired during post-training, and compare it against two reference metrics built on teacher confidence (Entropy) and teacher-student discrepancy (Novelty). Experiments on unlabeled and domain-labeled training mixtures under strong-to-weak and same-size distillation scenarios show that ExpertAlign achieves the strongest overall performance in all four settings. On unlabeled data, it improves the overall score by 5.88 (+12.3%) points over Mean aggregation; on domain-labeled data, it outperforms standard MOPD by 3.95 (+7.8%) points without using available domain labels. These results demonstrate token-level routing can exploit cross-domain complementary supervision, and reduce exclusive reliance on prompt-level domain assignment. Code is available at: https://github.com/TURLEing/MOPD-Router.

cs.LG↗
arXiv2026

Minimal proximal definable flows over the $p$-adics

Let $G$ be a definable group in an NIP theory. We prove that every minimal proximal definable $G$-flow is strongly proximal. Consequently, the universal minimal proximal definable $G$-flow $Π^{\mathrm{def}}(G)$ coincides with the minimal strongly proximal definable $G$-flow $Π^{\mathrm{def}}_{\mathrm{s}}(G)$. Furthermore, for a $p$-adic definable group $G$, we can compute $Π^{\mathrm{def}}(G)$ explicitly. We show that $Π^{\mathrm{def}}(G)$ is exactly $Π^{\mathrm{def}}(S)$ where $S$ is the semisimple part of the definably amenable-semisimple decomposition of $G$. In addition, $Π^{\mathrm{def}}(S)\cong S^*_\mathcal{F}(\mathbb{Q}_p)$, the space of types of full dimension on $\mathbf{F}$, where $\mathcal{F}=\mathbf{F}(\mathbb{Q}_p)$ for a flag variety $\mathbf{F}$ constructed from $S$.

math.LO↗
arXiv2026

Newelski's Conjecture for $o$-Minimal and $p$-Adic Groups

Let $M_0$ denote either the field of $p$-adic numbers $\mathbb{Q}_p$ or an $o$-minimal expansion of the real field $\mathbb{R}$. We study the minimal flows and Ellis groups of groups definable over $M_0$ from the viewpoint of definable topological dynamics. The computation of Ellis groups, together with the closely related Newelski conjecture, has so far been available only under restrictive hypotheses: in the $p$-adic setting, the results of \cite{BY-APAL} apply to reductive algebraic groups, while in the $o$-minimal setting the known computations rely on a compact-torsion-free (Iwasawa) decomposition. In this paper we remove both restrictions. We show that every definable group $G$ has a definably amenable radical $D$, with quotient map $π:G\to A$ onto a centreless semisimple group $A$. In $A$, a definably amenable component $V$ is selected: a maximal definably amenable subgroup containing a maximal definably solvable subgroup. Its preimage $B=π^{-1}(V)$ is a definably amenable component of $G$, contains $D$, and is the carrier for the computation of the Ellis group. Our main result is that, for every $M\succ M_0$, the Ellis group of the universal definable flow of $G$ over $M$ is isomorphic to that of $B$ over $M$, and hence to $B/B^{00}$. In particular, the Ellis groups of $G$ are model-independent. As a consequence, when $G$ is definable over a $p$-adically closed field, Newelski's conjecture holds for $G$ if and only if $G$ is definably amenable. The amenable direction is the known result of \cite{CS-Definably-Amenable-NIP-Groups}, while the converse is new.

math.LO↗
arXiv2026

Counterfactual Constraint-Conditioned On-Policy Distillation for Multi-Constraint Instruction Following

Multi-constraint instruction following requires a model to respond to a query under many simultaneously active constraints. Even strong instruction-tuned models still routinely violate some of them. Existing approaches either augment supervision with sequence- or token-level RL rewards from external verifiers or learned graders, or use on-policy distillation (OPD) against a single full-context teacher whose probability mass becomes diluted as more constraints become simultaneously active. We propose CC-OPD (Counterfactual Constraint-Conditioned On-Policy Distillation), which inverts the standard supervision-generation direction in distillation. Rather than enriching the teacher with information beyond what the student sees, CC-OPD ablates each constraint from the teacher's conditioning in turn, and constructs the per-constraint signal from the resulting per-token probability differentials. The resulting per-token leave-one-out log-likelihood shifts are summed, clipped, and added to the vanilla OPD reward as a token-level shaping term. All shaping terms are obtained from the frozen teacher, without an external verifier during distillation, and the reward equals vanilla OPD wherever the aggregate shift is zero. Across two Qwen model pairs and seven benchmarks, CC-OPD achieves the highest average among all evaluated student-training methods. A 1.5B student trained with CC-OPD surpasses its own 7B RL-trained teacher on the MulDimIF benchmark.

cs.LG↗
arXiv2026

PAC: Progress-Augmented Advantage Curriculum for Multi-Task Reinforcement Learning of LLMs

Reinforcement learning (RL) is used to improve the reasoning abilities of LLMs, while training data span heterogeneous tasks. However, most RL post-training pipelines rely on fixed or manually designed task mixtures, even though task usefulness changes as training progresses. Online curriculum methods often define learnability by update magnitude, ignoring whether the update translates into reward gains, which can misallocate rollout budget toward tasks with large but ineffective updates. We propose PAC, a Progress-Augmented Advantage Curriculum for multi-task RL of LLMs that combines two task-level signals: advantage-derived learnability, which measures the magnitude of the policy update a task can induce, and recent reward gains, which show whether those updates have improved task performance. A Bayesian Thompson Sampling controller uses these signals to allocate rollouts across tasks during GRPO training. We evaluate PAC under two settings: a multi-level reasoning setting and a multi-domain reasoning setting. PAC improves sample efficiency and final performance: it reaches comparable validation scores with fewer rollout steps and achieves higher final averages than random sampling and advantage-based curriculum baselines in both settings. These results show that jointly tracking advantage signals and actual reward gains yields an effective online curriculum for LLM post-training.

cs.LG↗
arXiv2026

On groups definable in $p$-adically closed fields

This paper is about the $dfg$/$fsg$ decomposition for groups $G$ definable in $p$-adically closed fields. It is proved that for $G$ definably amenable, $G$ has a definable normal $dfg$ subgroup $H$ such that the quotient $G/H$ is a definable $fsg$ group. The result was known for groups definable in $o$-minimal expansions of real closed fields (see \cite{C-P-o-mini}). We also give a version for arbitrary (not necessarily definably amenable) groups $G$ definable in $p$-adically closed fields: there is a definable $dfg$ subgroup $H$ of $G$ such that the homogeneous space $G/H$ is definable and definably compact. (In the $o$-minimal case this is Fact 3.25 of \cite{Peterzil-Starchenko-mutypes}). Finally, we also isolate the definably amenable part of $G$, which we call the definably amenable component. Note that $dfg$ stands for ``has a definable $f$-generic type", and $fsg$ for ``has finitely satisfiable generics", which will be discussed together with various equivalences. We will need to understand something about groups of the form $G(k)$ where $k$ is a $p$-adically closed field and $G$ a semisimple algebraic group over $k$, and as part of the analysis we will prove the Kneser-Tits conjecture over $p$-adically closed fields.

math.LO↗
arXiv2026

Weak regularity and one-parameter subgroups of definable $p$-adic Lie groups

We prove that every group $G$ definable in the pure $p$-adic field $\mathbb{Q}_p$ is weakly regular. We show that every one-parameter subgroup is definable. We also show that its one-parameter core $G_u$, together with a suitable regular open subgroup $G_Ω$, is definable. Finally, we show that if $H$ is a dfg component of $G$, then $G_u=\langle(H^g)_u:g\in G\rangle$. In fact, $G_u$ is a finite product of the one-parameter cores of finitely many conjugates of $H$. In particular, when $G$ is definably amenable, $G_u=H_u$.

math.LO↗
arXiv2026

LQCDMaster: Agentic Scientific Computing for Lattice Quantum Chromodynamics Research

Lattice quantum chromodynamics (LQCD) provides a first-principles framework for computing hadronic observables, but its practical use remains limited by the substantial expertise required to turn research motivation into reliable computing workflows. Here we present \textsc{LQCDMaster}, a tool-augmented, skill-guided and domain-specialized scientific computing agent that converts natural-language LQCD research tasks into executable PyQUDA computing workflows, including measurement scripts, job-submission artifacts, execution logs and numerical outputs. The system combines agentic planning, expert-annotated LQCD skills and a deterministic Wick-contraction tool to constrain the algebraically fragile components of code generation. We evaluate \textsc{LQCDMaster} on a benchmark at the forefront of scientific research, comprising 70 LQCD computing tasks, with observables covering local and nonlocal two-point functions, Wilson loops, meson and baryon three-point functions. The generated workflows exactly reproduce expert-written implementations in 63 of 70 tasks at machine precision, with three additional discrepancies attributable to convention mismatches. Across representative observables, the agent reduces implementation time from hours to minutes while preserving end-to-end numerical validation. Further, we present a typical case of \textsc{LQCDMaster}-driven exploration: a lattice computation of light-cone distribution amplitudes with diagonal Wilson-line, a quantity accessible with standard methods but never before computed, and computation of the spectrum of proton, deuteron, triton, hyperon, hyperdeuteron and hypertriton. This work pioneers the paradigm of agentic scientific computing by automating the end-to-end scientific computing workflows in lattice QCD research, lowering its barrier and facilitating the exploration and verification of non-standard scientific ideas.

hep-lat↗
arXiv2026

Independence relations induced by ideals

Inspired by the non-meager independence $\downarrow^{nm}$ introduced by Krupiński [3], we study further possible independence relations induced by ideals and provide a general framework for this topic. We show that the independence relation $\downarrow^\mathrm{Haar}$ induced by Haar null ideals is a good example for locally compact groups and provide an example showing $\downarrow^\mathrm{Haar}\neq \downarrow^{nm}$. Moreover, we introduce an order on the collection of all well-behaved independence relations and conjecture that $\downarrow^{nm}$ is the maximum one for Polish groups. We prove the conjecture under the extra hypothesis of $σ$-compactness. For Lie groups, we discuss $\mathrm{SO}(3)$ and $\mathrm{SE}(2)$ as examples, and prove that the independence relation is unique for nilpotent Lie groups.

math.LO↗
arXiv2026

UME: A Unified Meta-Generalization Framework for Cross-Domain ETA

Accurate Estimated Time of Arrival (ETA) prediction on checkout page is crucial in instant logistics for enhancing user satisfaction, optimizing dispatching, and controlling operational costs. In international on-demand delivery platforms, where ETA data originates from diverse countries or regions with different patterns, multi-domain modeling is of great importance and has been widely adopted. However, existing methods still face three critical challenges in real-world deployment. First, current multi-domain models struggle to generalize to completely unseen domains, failing to achieve zero-shot prediction during the initial cold-start phase. Second, cross-domain feature spaces are often assumed to be consistent, whereas new domains commonly suffer from structural missingness of offline (statistical) features due to the lack of historical data. Third, such feature missingness often compels industrial systems to model mature and cold-start domains separately, hindering knowledge transfer and increasing maintenance overhead. To address these challenges, we propose \textbf{UME}, a \textbf{U}nified \textbf{M}eta-generalization framework for \textbf{E}TA. Specifically, UME integrates a unified dual-branch architecture with a novel meta-learning mechanism that employs a hypernetwork-based meta learner. By leveraging domain-level knowledge and instance-level context, the meta learner empowers three meta modules to dynamically modulate feature gating, expert attention, and final prediction, capturing cross-domain correlations and facilitating intra-domain adaptation. A knowledge distillation strategy is further introduce to enhance performance. UME has now been deployed in Meituan-keeta delivery platform (the largest international food delivery platform in China). Extensive offline experiments and online A/B tests demonstrate that UME significantly outperforms existing baselines.

cs.LG↗
arXiv2026

USIM and U0: A Vision-Language-Action Dataset and Model for General Underwater Robots

Underwater environments pose unique challenges for robotic navigation and manipulation. While existing research has primarily focused on task-specific methods, studies on general-purpose intelligence for multi-task execution remain scarce. To address this gap, we propose a unified framework for general-purpose underwater robots that integrates perception and action driven by language instructions. First, we develop a data synthesis pipeline to construct USIM, a simulation-based dataset which comprises over 905K frames from 2275 trajectories, totaling approximately 25 hours of BlueROV2 interactions. Furthermore, we propose U0, a vision-language-action (VLA) model capable of executing various tasks from obstacle-avoidance navigation to three-dimensional mobile manipulation. The model features a convolution-attention-based perception (CAP) module, which incorporates target pose estimation as an auxiliary task to explicitly bolster the model's spatial awareness. For evaluation, we establish a systematic assessment framework and an automated pipeline encompassing both offline metrics and online task execution. Experimental results demonstrate that the USIM dataset significantly empowers existing VLA models to adapt to underwater scenarios. Notably, our U0 model achieves state-of-the-art performance: it reduces the offline mean action prediction error to 0.0359 and achieves an overall online success rate of 43.1%, marking a 5.5% improvement over existing competitive baselines (below 37.6%), with navigation tasks reaching as high as 87.5%. These results validate the feasibility of general-purpose intelligence in underwater robotics, providing a foundation for scalable dataset synthesis and aquatic embodied agents.

cs.RO↗
arXiv2026

SkillGenBench: Benchmarking Skill Generation Pipelines for LLM Agents

As LLM agents are increasingly built around reusable skills, a central challenge is no longer only whether agents can use provided skills, but whether they can generate correct, reusable, and executable skills from repositories and documents. Existing benchmarks primarily evaluate the efficacy of given skills or the ability of agents to solve downstream tasks from raw context, but they do not isolate skill generation itself as the object of study. We introduce SkillGenBench, a benchmark for evaluating skill generation pipelines under a unified and controlled protocol. In SkillGenBench, a generator receives raw corpora and produces standardized skill artifacts, which are then executed under fixed harnesses and assessed with unified evaluation procedures. The benchmark covers two generation regimes: task-conditioned generation, where a task-specific skill is synthesized after the task is revealed, and task-agnostic generation, where a reusable skill library must be distilled before downstream tasks are known. It also spans two complementary procedural sources: repository-grounded instances, where procedures are distributed across code, configuration, and scripts, and document-grounded instances, where procedures and constraints must be distilled from long-form text. We provide standardized task specifications, pinned environments, and evaluation protocols centered on deterministic execution-based checks, supplemented by auxiliary signals for diagnosis. Experiments across a range of skill-generation methods and backbones show substantial performance variation, highlight the difficulty of reusable skill distillation, and reveal distinct failure modes in skill generation from software repositories versus long-form documents. SkillGenBench establishes a reproducible testbed for studying skill generation as an independent research problem in agent systems.

cs.AI↗
arXiv2026

Rubrics to Tokens: Bridging Response-level Rubrics and Token-level Rewards in Instruction Following Tasks

Rubric-based Reinforcement Learning (RL) has emerged as a promising approach for aligning Large Language Models (LLMs) with complex, open-domain instruction following tasks. However, existing methods predominantly rely on response-level rewards, introducing severe reward sparsity and reward ambiguity problems. To address these issues, we propose Rubrics to Tokens (RTT), a novel rubric-based RL framework that bridges coarse response-level scores and fine-grained token-level credit assignment. RTT introduces a Token-Level Relevance Discriminator to predict which tokens in the response are responsible for a specific constraint, and optimizes the policy model via RTT-GRPO, which integrates response-level and token-level advantages within a unified framework. Furthermore, when transitioning from one-dimensional, outcome-level reward to three-dimensional reward space in the token-level rubric-based RL, we propose a novel group normalization method, called Intra-sample Token Group Normalization, to accommodate this shift. Extensive experiments and benchmarks demonstrate that RTT consistently outperforms other baselines in both instruction- and rubric-level accuracy across different models.

cs.CL↗
arXiv2026

Story2Proposal: A Scaffold for Structured Scientific Paper Writing

Generating scientific manuscripts requires maintaining alignment between narrative reasoning, experimental evidence, and visual artifacts across the document lifecycle. Existing language-model generation pipelines rely on unconstrained text synthesis with validation applied only after generation, often producing structural drift, missing figures or tables, and cross-section inconsistencies. We introduce Story2Proposal, a contract-governed multi-agent framework that converts a research story into a structured manuscript through coordinated agents operating under a persistent shared visual contract. The system organizes architect, writer, refiner, and renderer agents around a contract state that tracks section structure and registered visual elements, while evaluation agents supply feedback in a generate evaluate adapt loop that updates the contract during generation. Experiments on tasks derived from the Jericho research corpus show that Story2Proposal achieved an expert evaluation score of 6.145 versus 3.963 for DirectChat (+2.182) across GPT, Claude, Gemini, and Qwen backbones. Compared with the structured generation baseline Fars, Story2Proposal obtained an average score of 5.705 versus 5.197, indicating improved structural consistency and visual alignment.

cs.CL↗
arXiv2026

Idea2Story: An Automated Pipeline for Transforming Research Concepts into Complete Scientific Narratives

Autonomous scientific discovery with large language model (LLM)-based agents has recently made substantial progress, demonstrating the ability to automate end-to-end research workflows. However, existing systems largely rely on runtime-centric execution paradigms, repeatedly reading, summarizing, and reasoning over large volumes of scientific literature online. This on-the-spot computation strategy incurs high computational cost, suffers from context window limitations, and often leads to brittle reasoning and hallucination. We propose Idea2Story, a pre-computation-driven framework for autonomous scientific discovery that shifts literature understanding from online reasoning to offline knowledge construction. Idea2Story continuously collects peer-reviewed papers together with their review feedback, extracts core methodological units, composes reusable research patterns, and organizes them into a structured methodological knowledge graph. At runtime, underspecified user research intents are aligned to established research paradigms, enabling efficient retrieval and reuse of high-quality research patterns instead of open-ended generation and trial-and-error. By grounding research planning and execution in a pre-built knowledge graph, Idea2Story alleviates the context window bottleneck of LLMs and substantially reduces repeated runtime reasoning over literature. We conduct qualitative analyses and preliminary empirical studies demonstrating that Idea2Story can generate coherent, methodologically grounded, and novel research patterns, and can produce several high-quality research demonstrations in an end-to-end setting. These results suggest that offline knowledge construction provides a practical and scalable foundation for reliable autonomous scientific discovery.

cs.CE↗
arXiv2025

MimicDreamer: Aligning Human and Robot Demonstrations for Scalable VLA Training

Vision Language Action (VLA) models derive their generalization capability from diverse training data, yet collecting embodied robot interaction data remains prohibitively expensive. In contrast, human demonstration videos are far more scalable and cost-efficient to collect, and recent studies confirm their effectiveness in training VLA models. However, a significant domain gap persists between human videos and robot-executed videos, including unstable camera viewpoints, visual discrepancies between human hands and robotic arms, and differences in motion dynamics. To bridge this gap, we propose MimicDreamer, a framework that turns fast, low-cost human demonstrations into robot-usable supervision by jointly aligning vision, viewpoint, and actions to directly support policy training. For visual alignment, we propose H2R Aligner, a video diffusion model that generates high-fidelity robot demonstration videos by transferring motion from human manipulation footage. For viewpoint stabilization, EgoStabilizer is proposed, which canonicalizes egocentric videos via homography and inpaints occlusions and distortions caused by warping. For action alignment, we map human hand trajectories to the robot frame and apply a constrained inverse kinematics solver to produce feasible, low-jitter joint commands with accurate pose tracking. Empirically, VLA models trained purely on our synthesized human-to-robot videos achieve few-shot execution on real robots. Moreover, scaling training with human data significantly boosts performance compared to models trained solely on real robot data; our approach improves the average success rate by 14.7\% across six representative manipulation tasks.

cs.RO↗
arXiv2025

MathReal: We Keep It Real! A Real Scene Benchmark for Evaluating Math Reasoning in Multimodal Large Language Models

Multimodal Large Language Models (MLLMs) have demonstrated remarkable capabilities in visual mathematical reasoning across various existing benchmarks. However, these benchmarks are predominantly based on clean or processed multimodal inputs, without incorporating the images provided by real-world Kindergarten through 12th grade (K-12) educational users. To address this gap, we introduce MathReal, a meticulously curated dataset comprising 2,000 mathematical questions with images captured by handheld mobile devices in authentic scenarios. Each question is an image, containing the question text and visual element. We systematically classify the real images into three primary categories: image quality degradation, perspective variation, and irrelevant content interference, which are further delineated into 14 subcategories. Additionally, MathReal spans five core knowledge and ability categories, which encompass three question types and are divided into three difficulty levels. To comprehensively evaluate the multimodal mathematical reasoning abilities of state-of-the-art MLLMs in real-world scenarios, we design six experimental settings that enable a systematic analysis of their performance. Through extensive experimentation, we find that the problem-solving abilities of existing MLLMs are significantly challenged in realistic educational contexts. Based on this, we conduct a thorough analysis of their performance and error patterns, providing insights into their recognition, comprehension, and reasoning capabilities, and outlining directions for future improvements. Data and code: https://github.com/junfeng0288/MathReal.

cs.CV↗