arXiv subjects

Zili Wang

Publications and source records attributed to Zili Wang.

At least 19 recordsLinked to original sources

arXiv2026

Online Draft Co-Training for Speculative Decoding in Large-Scale, Long-Context RL Post-Training

Speculative decoding accelerates rollout generation, which dominates the cost of reinforcement learning (RL) post-training. Online co-training can further increase the draft's accuracy, yielding greater speedups. However, scaling this approach to co-training on large models with long contexts poses two obstacles: (1) branch attention is unsupported by standard causal context-parallel (CP) implementations, and (2) target features span across pipeline-parallel (PP) stages. We address both with an end-to-end system for large-scale online draft co-training. For CP, we extend packed, load-balanced zigzag ring attention by merging rank-local branch attention with causal main-sequence attention. For PP, TapChannel transports intermediate target features across stages via a separate path, leaving the pipeline schedule unaffected. Experiments demonstrate that co-trained drafts closely track the policy baseline while delivering substantial rollout and end-to-end speedups across model scales up to 122B. Our CP design achieves strong scaling at 256K tokens with significant memory savings over prior work, and our PP transport incurs modest overhead. Code can be found at https://github.com/NVIDIA-NeMo/RL/issues/3698.

cs.LG↗

arXiv2026

Flip-graph non-convexity for once-punctured polygons

The set of the triangulations with vertex set $X$ of a simple polygon $\mathrm{P}$ can be structured into a flip-graph $\mathcal{F}(\mathrm{P},X)$ whose edges connect two triangulations that differ by a single arc. The geometry of flip-graphs has been thoroughly studied and it is known that the subgraph $\mathcal{F}_\varepsilon(\mathrm{P},X)$ induced by the triangulations that contain a given arc $\varepsilon$ is strongly convex in $\mathcal{F}(\mathrm{P},X)$ when $\mathrm{P}$ is convex and $X$ contains no puncture (points in the interior of $\mathrm{P}$) and at most one flat vertex (points in the interior of an edge). When $X$ contains at least two punctures or flat vertices, it is also known that this strong convexity property fails. Here, we close the last open case by showing that, for any convex polygon with sufficiently many vertices, one can always place a single puncture in $X$ in such a way that $\mathcal{F}_\varepsilon(\mathrm{P},X)$ is not strongly convex in $\mathcal{F}(\mathrm{P},X)$. We prove a similar result for simple polygons with a single reflex vertex. The main ingredients in our proofs are a decomposition lemma for a class of $3$-dimensional triangulations and a hyperbolic volume argument regarding their embedding into $\mathbb{H}^3$.

math.MG↗

arXiv2026

Audio-Visual Segmentation via Depth-Guided Collaborative Modeling

Audio-Visual Segmentation (AVS) is a fundamental task in multimodal perception that performs pixel-level segmentation of sounding objects in videos by leveraging both visual and audio cues. It has broad applications in video understanding, human-computer interaction, and autonomous driving. However, most existing AVS methods do not explicitly model geometric cues such as relative distance and occlusion, thereby limiting the robustness of cross-modal alignment. In human perception, spatial structure is naturally integrated with audio-visual evidence to accurately localize sounding objects. Motivated by this, we incorporate estimated depth as a spatial structural cue for AVS and propose DGCM-AVS, a tri-modal framework that jointly models audio, visual, and depth information. Specifically, we design a Depth-Aware Dynamic Modulator to improve the separation of adjacent objects while preserving intra-object feature consistency. Furthermore, we propose Depth-Guided Progressive Fusion, which uses depth as an intermediate bridge to progressively align audio cues with visual features. Compared to state-of-the-art methods, DGCM-AVS achieves relative improvements of 10.2 percent in M_J and 8.7 percent in M_F on the AVSS dataset. We believe our study highlights depth as a promising yet underexplored modality for AVS and may encourage further research in this direction.

cs.CV↗

arXiv2026

Invariant Pretraining for Robust Code Representations

Encoder-based code representation models remain widely deployed for discriminative tasks such as clone detection and code classification, where their small size and low inference cost are decisive. Their robustness, however, is fragile: under invariant programs, semantically equivalent code written in different syntactic forms, learned representations degrade substantially even though program behavior is unchanged. We present an empirical study of this robustness gap across four encoder baselines, two downstream tasks, and four datasets, together with a minimal code-only continued pretraining recipe that closes much of it. Our method, invariant pretraining (InvPT), applies semantic-preserving transformations to the corpus and combines masked language modeling with multi-positive supervised contrastive learning that treats all augmentations of the same source function as positives, mixing self-contrast pairs (same code, different masks) with invariant-contrast pairs (transformed code) for positives of varying difficulty. Unlike prior contrastive code encoders, InvPT does not require paired natural-language data. Across our evaluation, InvPT improves robustness on transformed test sets by up to 11 percentage points on clone detection and 19 on code classification while matching or improving standard accuracy, and our ablations isolate multi-positive invariant contrast as the main source of the gains. Our aim is not a new objective but a careful measurement of where encoder robustness breaks and how far a simple, code-only recipe can recover it.

cs.LG↗

arXiv2026

Dynamic Surveys: Using LLMs to Blend Qualitative Depth,Quantitative Structure, and Collaborative Interaction

Surveys are a powerful tool for collecting data and eliciting insights on social phenomena, and are critical in product design, marketing, scientific research. However, traditional open-ended and closed-ended question formats limit researchers' ability to capture data that combines both the richness of qualitative insights and the analytical rigor of quantitative data. To address these problems, we propose Dynamic Surveys, a survey platform that uses Large Language Models (LLMs) to dynamically cluster qualitative responses in real time and to elicit quantitative ratings and rankings on those clusters and qualitative reflections on how their views compare to broader respondent trends, especially helpful in early-stage or exploratory research settings. This process generates a report showing survey creators and respondents the clustered responses as well as each cluster's rank, rating distribution, and follow-up reflections. To evaluate Dynamic Surveys, we conducted two field studies with 93 participants over a 2-month period. In the first study, 52 students provided input for a career workshop, while in the second, 41 students gave feedback on gaps in their academic curriculum. Of these, 44 respondents filled out a survey on their experience using Dynamic Surveys. We also shared the generated report with 4 individuals who were interested in the insights for their work, and interviewed them to understand their perspectives on the results and any contextual risks they saw in the platform design. Our findings suggest that Dynamic Surveys not only provide richer and deeper insights into responses compared with traditional survey tools, but also increase engagement and foster a sense of community. We discuss broader implications for the design of survey platforms that blend qualitative depth with quantitative structure, facilitating richer insights and offering more collaborative interactions.

cs.HC↗

arXiv2026

CLI-Universe: Towards Verifiable Task Synthesis Engine for Terminal Agents

While recent LLM-based terminal agents have demonstrated promising capabilities, the scarcity of high-quality, executable training data remains a critical bottleneck. Existing synthesis pipelines typically scale by retrofitting surface-level artifacts into tasks, frequently yielding ambiguous instructions, shallow execution paths, and brittle tests that provide weak learning signals. To overcome this, we introduce CLI-Universe, a principled synthesis engine that constructs terminal-agent tasks. CLI-Universe generates candidate tasks by sampling combinations across a multi-dimensional capability taxonomy (domain, skill type, capability, and engineering pillar), then grounds each candidate through evidence-guided deep research over real-world technical materials. To ensure rigorous supervision, validated blueprints are instantiated into Dockerized environments and subjected to a multi-stage executable verification pipeline featuring rubric-gated test construction, hint-conditional filtering, and strict fail-to-pass checking. Across the full pipeline, from candidate generation to verification, approximately two-thirds of candidates are discarded, retaining only those that are genuine, verifiable, and non-trivially challenging. To validate our framework, we instantiate a highly distilled dataset of 6,000 trajectories called CLI-Universe-6K. Remarkably, fine-tuning Qwen3-32B on CLI-Universe-6K achieves 33.4% on Terminal-Bench 2.0. This sets a new state-of-the-art for models trained on open-source data at or below 32B parameters, and outperforms several models an order of magnitude larger, demonstrating the profound data efficiency of structured, high-fidelity synthesis.

cs.AI↗

arXiv2026

Synergizing Global Pattern Learning and Time Order Characterization in Mobile Channel Prediction: An RWKV-Based Approach

Owing to the potential to reduce pilot overhead and mitigate channel aging, channel prediction is emerging as an important research topic in wireless communications. Meanwhile, deep neural networks are becoming a foundational technology for high-precision prediction thanks to their excellent non-linear representation capabilities. In this paper, we conceive a task-driven prediction network, which aims to deeply synergize the following two functions: learning global patterns for shareable features across adjacent time slots and structurally encoding time order to characterize the inherent causality within the channel dynamics. To implement channel prediction accuracy, we employ RWKV (receptance weighted key value) as network backbone and adapt it to the task's specific characteristics, utilizing its deep interleaved learning architecture to extract global patterns across multiple channel samples and leveraging its unique exponential decay to characterize temporal order. These task-driven unique designs significantly improve the learning efficiency of prediction network. Comprehensive experimental evaluations demonstrate the superiority of the proposed method over current data-driven methods, such as long short-term memory and Transformer, in the channel prediction task, including 1.84~4.29 dB gains in normalized mean squared error and 2.6~10.5 percentage point gains in cosine correlation.

eess.SP↗

arXiv2026

Joint Training of Multi-Token Prediction in Reinforcement Learning via Optimal Coefficient Calibration

Reinforcement Learning from Verifiable Rewards (RLVR) has emerged as the standard paradigm for improving reasoning capability of large language models, while Multi-Token Prediction (MTP) has been a widely adopted module in pretraining. Combining them is a natural approach, yet current RL practices detach MTP gradients because joint training degrades the performance. We revisit this failure from an optimization perspective. We show that the per-step effect of MTP on the RL objective can be decomposed into two terms: a first-order correlation and a second-order perturbation penalty. This decomposition unifies three MTP training regimes: Detach, Cross-Entropy loss, and Policy loss, and explains why each succeeds or fails. Further analysis of policy loss reveals that, although it aligns with intuition, performance still degrades: the correlation term decays while the quadratic penalty persists. Guided by the analysis, we propose Optimal Coefficient Calibration (OCC), an adaptive scheme that tracks the optimal coefficient online via a log-probability proxy at negligible cost. Across six competition-level mathematical reasoning benchmarks, OCC consistently matches or exceeds the detach baseline, delivering improved joint MTP-RL training performance.

cs.LG↗

arXiv2026

Design for Manufacturing: A Manufacturability Knowledge-Integrated Reinforcement Learning Framework for Free-Form Pipe Routing in Aeroengines

Design for manufacturing plays a critical role in advanced aeroengine development, where complex components necessitate careful consideration of manufacturability. However, current practices in pipe routing remain largely decoupled from down-stream manufacturing, leading to labor-intensive, trial-and-error iterations to achieve manufacturable designs. To address this problem, this study proposes the Frenet-based pipe routing optimization (FPRO) framework, a manufacturability knowledge-integrated reinforcement learning approach for free-form pipe design in aeroengines. FPRO formulates the routing problem as a boundary value problem in the Frenet frame. In this framework, the pipe path is represented by curvature and torsion profiles, which are generated using cubic Hermite interpolation. To integrate design and manufacturing, domain-specific manufacturing knowledge is embedded as constraints on the permissible ranges of curvature and torsion. The path optimization is performed using the proximal policy optimization algorithm with stochastic exploration and a stage-guided reward mechanism. A unified mapping formulation then translates the optimized path into motion trajectories for the bending die, enabling direct fabrication on a six-axis free-bending machine. Experimental results demonstrate that FPRO consistently generates collision-free, manufacturable paths with smoother geometric profiles compared to Cartesian-based methods. It also achieves faster convergence and superior performance in terminal alignment, path length, obstacle avoidance, and manufacturability compared to state-of-the-art reinforcement learning baselines. Real-world validation confirms the close geometric correspondence between the manufactured pipe and its digital design, validating the practical feasibility of FPRO.

cs.LG↗

arXiv2026

ContractBench: Can LLM Agents Preserve Observation Contracts?

Tool-augmented LLM agents call APIs whose intermediate outputs, such as presigned URLs, session tokens, and OAuth state parameters, are observation contracts: artifacts whose later use is constrained by the external system that produced them. We show that observation contract compliance (preserving the temporal validity and byte-level integrity) is an emergent, regression-prone capability: it is neither guaranteed by general tool-use ability nor consistently improved by larger or newer models. To measure this, we introduce ContractBench, a benchmark of 33 dual-axis tasks that probe two orthogonal failure modes no existing benchmark evaluates: validity failures (using an artifact after expiry) and integrity failures (corrupting an artifact's bytes through the observation-to-action pipeline). Our evaluation is deterministic and programmatic, with a virtual clock controlling time and SHA-256 hashes verifying byte integrity. We assign each outcome a failure label drawn from real-world API specifications. We evaluate 38 models and report four findings: (i) no evaluated model clears 80%, with Claude-Opus-4.6 leading at 77.8%, revealing that current frontier models still fail to comply with observation contracts; (ii) a sharp within-family capability cliff in Qwen 3.5 between 4B (0%) and 9B (56.6%), smoothing to 70.7% at 397B-A17B: what emerges across the cliff is mid-trajectory restraint, not tool-call competence; (iii) non-monotonic scaling across the GPT-5 family: agentic post-training can erode compliance through sycophancy-driven regression; (iv) our failure taxonomy works as an actionable in-context reward signal, yielding +7.1 pp on 42 paired GPT-5.1 failures.

cs.SE↗

arXiv2026

OProver: A Unified Framework for Agentic Formal Theorem Proving

Recent progress in formal theorem proving has benefited from large-scale proof generation and verifier-aware training, but agentic proving is rarely integrated into prover training, appearing only at inference time. We present OProver, a unified framework for agentic formal theorem proving in Lean 4, in which failed proof attempts are iteratively revised using retrieved compiler verified proofs and Lean compiler feedback. OProver is trained through continued pretraining followed by iterative post-training: each iteration runs agentic proving, indexes newly verified proofs into OProofs and the retrieval memory, uses repair trajectories as SFT data, and uses unresolved hard cases for RL. OProofs is built from public Lean resources, large-scale proof synthesis, and agentic proving traces, containing 1.77M Lean statements, 6.86M compiler-verified proofs, and serialized trajectories with retrieved context, failed attempts, feedback, and repairs. Across five benchmarks, OProver-32B attains the best Pass@32 on MiniF2F (93.3%), ProverBench (58.2%), and PutnamBench (11.3%), and ranks second on MathOlympiad (22.8%) and ProofNet (33.2%) more top placements than any prior open-weight whole-proof prover.

cs.CL↗

arXiv2026

Visual Generation in the New Era: An Evolution from Atomic Mapping to Agentic World Modeling

Recent visual generation models have made major progress in photorealism, typography, instruction following, and interactive editing, yet they still struggle with spatial reasoning, persistent state, long-horizon consistency, and causal understanding. We argue that the field should move beyond appearance synthesis toward intelligent visual generation: plausible visuals grounded in structure, dynamics, domain knowledge, and causal relations. To frame this shift, we introduce a five-level taxonomy: Atomic Generation, Conditional Generation, In-Context Generation, Agentic Generation, and World-Modeling Generation, progressing from passive renderers to interactive, agentic, world-aware generators. We analyze key technical drivers, including flow matching, unified understanding-and-generation models, improved visual representations, post-training, reward modeling, data curation, synthetic data distillation, and sampling acceleration. We further show that current evaluations often overestimate progress by emphasizing perceptual quality while missing structural, temporal, and causal failures. By combining benchmark review, in-the-wild stress tests, and expert-constrained case studies, this roadmap offers a capability-centered lens for understanding, evaluating, and advancing the next generation of intelligent visual generation systems.

cs.CV↗

arXiv2026

Prospect Theoretic Approach to Pursuit-evasion Differential Games with Risk Aversion and Probability Sensitivity

This paper considers for the first time pursuit-evasion (PE) differential games with irrational perceptions of both pursuer and evader on probabilistic characteristics of environmental uncertainty. Firstly, the irrational perceptions of risk aversion and probability sensitivity are modeled and incorporated within a Bayesian PE differential game framework by using Cumulative Prospect Theory (CPT) approach; Secondly, several sufficient conditions of capturability are established in terms of system dynamics and irrational parameters; Finally, the existence of CPT-Nash equilibria is rigorously analyzed by invoking Brouwer's fixed-point theorem. The new results reveal that irrational behaviors benefit the pursuer in some cases and the evader in others. Certain captures that are unachievable under rational behaviors can be achieved under irrational ones. By bridging irrational behavioral theory with game-theoretic control, this framework establishes a rigorous theoretical foundation for practical control engineering within complex human-machine systems.

eess.SY↗

arXiv2026

Decomposing Centrally Symmetric Convex Polyhedral Surfaces into Parallelograms

Let $\mathcal{M}_{2N}(\delta_1, \delta_2,\dots, \delta_N)$ be the moduli space of centrally symmetric convex polyhedral surfaces with $2N$ labeled vertices and prescribed cone-deficits $\delta_1$, $\delta_2$, $\dots$, $\delta_N$. We show that $\mathcal{M}_{2N}(\delta_1, \delta_2,\dots, \delta_N)$ has the structure of a real hyperbolic manifold of dimension $2N-3$. When $N=4$ and $5$, we show that every surface in $\mathcal{M}_{2N}(\delta_1, \delta_2,\dots, \delta_N)$ can be decomposed into at most $2\binom{2N-2}{2}$ parallelograms, and the decomposition is invariant under the antipodal map. Using the edge-lengths of these parallelograms as coordinates, we show that the moduli space of centrally symmetric polyhedral surfaces with $8$ unlabeled vertices and cone-deficits $\frac{\pi}{2}$ is isometric to the quotient of a real hyperbolic regular ideal $5$-simplex by the dihedral group $D_6$.

math.GT↗

arXiv2026

Beyond Sequential Distance: Inter-Modal Distance Invariant Position Encoding

Despite the remarkable capabilities of Multimodal Large Language Models (MLLMs), they still suffer from visual fading in long-context scenarios. Specifically, the attention to visual tokens diminishes as the text sequence lengthens, leading to text generation detached from visual constraints. We attribute this degradation to the inherent inductive bias of Multimodal RoPE, which penalizes inter-modal attention as the distance between visual and text tokens increases. To address this, we propose inter-modal Distance Invariant Position Encoding (DIPE), a simple but effective mechanism that disentangles position encoding based on modality interactions. DIPE retains the natural relative positioning for intra-modal interactions to preserve local structure, while enforcing an anchored perceptual proximity for inter-modal interactions. This strategy effectively mitigates the inter-modal distance-based penalty, ensuring that visual signals remain perceptually consistent regardless of the context length. Experimental results demonstrate that by integrating DIPE with Multimodal RoPE, the model maintains stable visual grounding in long-context scenarios, significantly alleviating visual fading while preserving performance on standard short-context benchmarks. Code is available at https://github.com/lchen1019/DIPE.

cs.CV↗

arXiv2026

Beyond Next-Token Alignment: Distilling Multimodal Large Language Models via Token Interactions

Multimodal Large Language Models (MLLMs) demonstrate impressive cross-modal capabilities, yet their substantial size poses significant deployment challenges. Knowledge distillation (KD) is a promising solution for compressing these models, but existing methods primarily rely on static next-token alignment, neglecting the dynamic token interactions, which embed essential capabilities for multimodal understanding and generation. To this end, we introduce Align-TI, a novel KD framework designed from the perspective of Token Interactions. Our approach is motivated by the insight that MLLMs rely on two primary interactions: vision-instruction token interactions to extract relevant visual information, and intra-response token interactions for coherent generation. Accordingly, Align-TI introduces two components: IVA enables the student model to imitate the teacher's instruction-relevant visual information extract capability by aligning on salient visual regions. TPA captures the teacher's dynamic generative logic by aligning the sequential token-to-token transition probabilities. Extensive experiments demonstrate Align-TI's superiority. Notably, our approach achieves $2.6\%$ relative improvement over Vanilla KD, and our distilled Align-TI-2B even outperforms LLaVA-1.5-7B (a much larger MLLM) by $7.0\%$, establishing a new state-of-the-art distillation framework for training parameter-efficient MLLMs. Code is available at https://github.com/lchen1019/Align-TI.

cs.CV↗

arXiv2025

CodeSimpleQA: Scaling Factuality in Code Large Language Models

Large language models (LLMs) have made significant strides in code generation, achieving impressive capabilities in synthesizing code snippets from natural language instructions. However, a critical challenge remains in ensuring LLMs generate factually accurate responses about programming concepts, technical implementations, etc. Most previous code-related benchmarks focus on code execution correctness, overlooking the factual accuracy of programming knowledge. To address this gap, we present CodeSimpleQA, a comprehensive bilingual benchmark designed to evaluate the factual accuracy of code LLMs in answering code-related questions, which contains carefully curated question-answer pairs in both English and Chinese, covering diverse programming languages and major computer science domains. Further, we create CodeSimpleQA-Instruct, a large-scale instruction corpus with 66M samples, and develop a post-training framework combining supervised fine-tuning and reinforcement learning. Our comprehensive evaluation of diverse LLMs reveals that even frontier LLMs struggle with code factuality. Our proposed framework demonstrates substantial improvements over the base model, underscoring the critical importance of factuality-aware alignment in developing reliable code LLMs.

cs.CL↗

arXiv2025

A Comprehensive Survey on Long Context Language Modeling

Efficient processing of long contexts has been a persistent pursuit in Natural Language Processing. With the growing number of long documents, dialogues, and other textual data, it is important to develop Long Context Language Models (LCLMs) that can process and analyze extensive inputs in an effective and efficient way. In this paper, we present a comprehensive survey on recent advances in long-context modeling for large language models. Our survey is structured around three key aspects: how to obtain effective and efficient LCLMs, how to train and deploy LCLMs efficiently, and how to evaluate and analyze LCLMs comprehensively. For the first aspect, we discuss data strategies, architectural designs, and workflow approaches oriented with long context processing. For the second aspect, we provide a detailed examination of the infrastructure required for LCLM training and inference. For the third aspect, we present evaluation paradigms for long-context comprehension and long-form generation, as well as behavioral analysis and mechanism interpretability of LCLMs. Beyond these three key aspects, we thoroughly explore the diverse application scenarios where existing LCLMs have been deployed and outline promising future development directions. This survey provides an up-to-date review of the literature on long-context LLMs, which we wish to serve as a valuable resource for both researchers and engineers. An associated GitHub repository collecting the latest papers and repos is available at: \href{https://github.com/LCLM-Horizon/A-Comprehensive-Survey-For-Long-Context-Language-Modeling}{\color[RGB]{175,36,67}{LCLM-Horizon}}.

cs.CL↗