arXiv subjects

Zixuan Dong

Publications and source records attributed to Zixuan Dong.

18 recordsLinked to original sources

arXiv2026

From Code to Playable Systems: CEGame-Bench for Long-Horizon Game Creation Agents

A game ships when a player can launch it, finish it, and return to a saved world, none of which a successful build guarantees, yet game-creation agents are mostly tested on one engine or on one-prompt browser games. We present CEGame-Bench, the first long-horizon benchmark for complete-game creation across five production engines: Defold, Godot, Roblox Studio, Unity, and Unreal Engine. Agents build each game from an empty project and a design document in runs averaging up to 295 steps and 65M tokens; the document doubles as a contract, and a hidden evaluator accepts a game only by real-input play. Under four scaffolds, agents earn most integration points but usually under half the gameplay points, and their games reach the ending in only 16% of runs, while references built by engineers with coding agents score near the maximum, which places this artifact-to-play gap in the agents. No two engines rank configurations alike (Kendall's tau 0.14-0.62), and the same game usually scores higher on Unity and Unreal Engine, which check more before play. We release Game Completion Boost, an agent skill for requirement-driven playtesting and repair that raises the execution-based score of three models by 14.1-24.3 points.

cs.AI↗

arXiv2026

Who Takes the Lead? Model Ability and Task-Conditioned Collaboration with Humans

As language-model agents grow more capable, they can carry more of a task without asking anyone for help. Falling reliance on a human partner is easy to read as a failure of collaboration when it may instead be a correct division of labor, and the rate of human involvement alone cannot tell the two apart. We study this systematically by placing an agent and a person under a shared objective while varying who holds the advantage: the human has a clear advantage, neither does, or the agent does. To obtain this control, we introduce a shared-control maze, in which an agent-controlled and a human-controlled avatar share the objective of reaching a common target in as few team turns as possible. The agent allocates each team turn between acting itself and handing the move to the person, and initial placement sets the human's shortest-path advantage while the task is otherwise unchanged. We evaluate GPT~5.6~Sol, Command~A+, Qwen~3.8~27B, and Command~R, first measuring solo capability with a single avatar and full-map observability, then running co-play with human participants across the three advantage conditions. Across the four models, stronger solo performance is associated with selective, route-conditioned human involvement: the models that solve the maze reliably on their own act for themselves almost without exception when the human holds no advantage, and involve the human on roughly $80\%$ of initial decisions when the human route is shorter. Command R, the model with substantially weaker solo performance, shows no such selectivity.

cs.HC↗

arXiv2026

Development of a continuous-wave photocathode very-high-frequency electron gun for $\rm S^3FEL$

A very-high-frequency (VHF) electron gun operating at a resonant frequency of 216.667 MHz has been developed at Tsinghua University for the Shenzhen Superconducting Soft X-ray Free-Electron Laser ($\rm S^3FEL$) facility. Building upon the SHINE gun design, the cavity profile was optimized to achieve a higher cathode electric-field gradient and a higher accelerating voltage at comparable input power. The revised cavity geometry also yields improved multipacting performance. In addition, thermal analysis was carried out to guide the water-cooling design, and enhanced cooling was implemented in the vicinity of the cathode. During high-power conditioning, 82 kW of continuous-wave radio-frequency power was successfully coupled into the gun, corresponding to a cathode gradient of 29.9 MV/m and an accelerating voltage of 856 kV. The gun voltage surpasses the previous world record for room-temperature VHF guns. The maximum dark current measured by the Faraday cup at the gun exit was only 6.8 nA.

physics.acc-ph↗

arXiv2026

GameLogicBench: Evaluating Coding Agents on Runtime Game Logic with Tick-Level State Assertions

Coding agents can modify and test code across large software projects. Game development is a domain where agents must implement gameplay rules. A game can end in a valid state even after violating its rules during the run. Current game-development benchmarks replay fixed examples, score videos, or ask another model to judge the result. However, no existing benchmark checks game rules throughout execution across varied evaluator-selected scenarios while ensuring exactly reproducible verdicts. We introduce GameLogicBench, a benchmark of 72 gameplay-logic tasks in Godot projects. An automated evaluator checks each game's rules at every simulation tick. Across 403 hand-designed scenarios, seeded parameter variations produce 1,451 test cases. To ensure that the evaluator measures behavior rather than implementation choice, it must accept different correct implementations for each task while rejecting mutants, implementations with one required capability removed. The tasks span isolated mechanics, multi-system interactions, and repository-scale features. Across 20 combinations of language models and scaffolds, the best observed run solves 52.78% of tasks. Under Claude Code, all twelve models solve fewer tasks as task scope expands from isolated mechanics, through interacting systems, to repository-scale features. Agents inspect code more often and make more tool calls on repository-scale tasks than on isolated-mechanic tasks. Most unsuccessful submissions are runnable, but implement some required game behavior incorrectly. We compared versions of our benchmark evaluator built with and without validation using mutants. Without this validation, incorrect agent submissions passed. A separate analysis finds agents copying code from public repositories when network access is open. Reliable evaluation thus depends both on what the tests reject and on what external code agents can access.

cs.SE↗

arXiv2026

QUACK: Questioning, Understanding, and Auditing Communicated Knowledge in Multimodal Social Deduction Agents

Social deduction games have become a popular testbed for probing reasoning, deception, coordination, and belief modeling in Large Language Model (LLM) agents. However, most environments are scored only by game outcomes such as win rates and largely remain to text-only interaction, making it difficult to tell whether an agent's language is actually grounded in what it perceived and did, or to identify the failure modes underlying its behavior. To address this gap, we introduce QUACK, an open-source environment and evaluation framework for auditing the grounding of agent language in multimodal social reasoning. QUACK evaluates agents at three levels: game outcomes, behavioral trajectories, and utterance-level consistency. Its core Statement Verification Pipeline reconstructs each agent's ground-truth trajectory from engine logs and checks every discussion claim against it, automatically flagging spatial hallucination, unsupported accusation, deception collapse, and language-action inconsistency. Evaluating three frontier VLMs in both homogeneous and cross-model adversarial settings, we find that even the strongest agent hallucinates 15.1% of its verifiable spatial claims and 11.5% of accusations are strictly unsupported. We release the full engine, evaluation framework, toolkit, and logs in https://github.com/AAAAA-Academia-Attractions/QUACK.

cs.CL↗

arXiv2026

Discrete Diffusion Models: A Unified Framework from Tokenization to Generation

Discrete denoising diffusion models (DDMs) have recently emerged as a compelling alternative to autoregressive (AR) modeling for discrete data, offering parallel generation and iterative global refinement capabilities. Unlike continuous diffusion, where the state space is fixed, DDMs are fundamentally shaped by how the discrete state space is constructed: the tokenization scheme, the vocabulary topology, and domain-specific structural alphabets. This work introduces a unified conceptual framework that views discrete diffusion models through the construction of the underlying discrete state space. Within this framework, existing formulations, including transition-matrix, masking/absorbing-state, and score/ratio-based approaches, emerge as different instantiations of a common design space. The framework further exposes common design trade-offs across training objectives, inference algorithms, scaling behavior, systems optimization, and evaluation protocols, suggesting several promising directions for future research.

cs.LG↗

arXiv2026

Minimal Ingredients for Reward Assignment from Expert Demonstrations

Reward assignment from scarce demonstrations is a key challenge in both offline and online imitation learning. A common and intuitive strategy assigns rewards according to how closely learner trajectories match expert demonstrations. Although this principle underlies many existing methods, the core ingredients that drive performance remain systematically underexplored. We therefore ask: what is the minimal structure that reward assignment must encode to achieve effective downstream RL performance across settings? We approach this question along two design axes: proximity approximation and temporal alignment. Across 32 benchmarks spanning offline and online settings, and with three downstream RL algorithms, our empirical findings suggest: (1) In offline regimes, proximity alone captures the reward structure necessary for effective offline RL, while (2) lightweight temporal correspondence provides consistent gains that are modest offline but essential online or in the presence of multiple demonstrations. We further complement our offline results with a lightweight theory characterizing when simple proximity approximation suffices. Overall, these findings advocate algorithmic minimalism in reward design before introducing complex schemes in both offline and online imitation learning.

cs.LG↗

arXiv2026

AVSCap: Orchestrating Audio-Visual Synergy for Omni-modal Video Captioning

Omni-modal video captioning is not merely combining visual captioning with audio transcription: a useful caption must describe how visual actions, speech, music, and sound effects co-evolve. Existing large multimodal models often fail at this relational step, treating audio and visual streams as loosely coupled observations, relying on automatic speech recognition, and under-specifying non-speech sounds and their links to visual events. We present AVSCap, a framework for audio-visual captioning centered on explicit cross-modal event binding. First, we construct AVSCap-130K, a tri-modal training corpus generated by a decoupled-then-fused pipeline that anchors visual and acoustic evidence before composing grounded omni-modal captions. Second, we train AVSCap-7B, a 7B captioner with a two-stage strategy: supervised fine-tuning establishes baseline capabilities, while sample-efficient reinforcement learning uses hybrid rewards to optimize acoustic completeness and audio-visual synergy. Our scaling analysis shows that reinforcement learning brings larger gains than increasing SFT data. Third, we introduce AVSCapBench, a benchmark that decomposes captions into visual, audio, and synergy events and evaluates them with fine-grained event recall. Experiments on AVSCapBench and external benchmarks show that AVSCap-7B improves non-speech audio coverage and cross-modal binding, delivering the best overall performance among evaluated open-source models.

cs.CV↗

arXiv2026

OmniHalluc-L: Counterfactual Benchmarking and Modality-Perturbation Reliability Calibration for Long-Form Omni Hallucination

Long-video Omni assistants often fail not by inventing content, but by misbinding real evidence: they hear the right utterance and see the right event, yet attach it to the wrong speaker, moment, or modality. These \emph{almost-true} errors evade standard video QA because local evidence remains valid, so item-level scoring can reward both a supported claim and its near-counterfactual. We introduce a counterfactual event-binding protocol that constructs paired supported/counterfactual claims from the same audio-visual event evidence and evaluates them by strict-pair accuracy. We instantiate it as \bench, a benchmark for long-video Omni hallucination, with 3{,}600 single-claim QA items from 638 long-form videos averaging 24.16 minutes and covering 256.87 hours. Under this protocol, open-weight Omni models remain weak at pair-level binding: Qwen2.5-Omni-7B reaches 32.06\% and Qwen3-Omni-Instruct reaches 41.55\%, versus 76.54\% for a closed-source reference. To narrow this gap without updating the backbone, we propose \method, Modality-Perturbation Reliability Calibration, a frozen-backbone framework that selects audio-negative probes within video-level folds and fuses their response shifts with native audio-visual confidence into per-claim support estimates. \method lifts Qwen2.5-Omni-7B to 36.22\% and Qwen3 to 51.09\% on \bench, and improves target-adapted MCQ accuracy on OmniVideoBench ($+$2.20) and WorldSense ($+$1.51) with Qwen3.

cs.MM↗

arXiv2026

Geometric Analysis of Neural Regression Collapse via Intrinsic Dimension

Neural multivariate regression underpins a wide range of domains, including control, robotics, and finance, yet the geometry of its learned representations remains poorly characterized. While neural collapse has been shown to benefit generalization in classification, we find that analogous collapse in regression consistently degrades performance. To explain this contrast, we analyze regression models through the lens of intrinsic dimension. Across control tasks and synthetic datasets, we estimate the intrinsic dimension of last-layer features (ID_H) and compare it with that of the regression targets (ID_Y). Collapsed models exhibit ID_H < ID_Y, leading to over-compression and poor generalization, whereas non-collapsed models typically maintain ID_H > ID_Y. For the non-collapsed models, performance with respect to ID_H depends on the data quantity and noise levels. From these observations, we identify two regimes (over-compressed and under-compressed) that determine when expanding or reducing feature dimensionality improves performance. Our results provide new geometric insights into neural regression collapse and suggest practical strategies for enhancing generalization.

cs.LG↗

arXiv2025

See What You Need: Query-Aware Visual Intelligence through Reasoning-Perception Loops

Human video comprehension demonstrates dynamic coordination between reasoning and visual attention, adaptively focusing on query-relevant details. However, current long-form video question answering systems employ rigid pipelines that decouple reasoning from perception, leading to either information loss through premature visual abstraction or computational inefficiency through exhaustive processing. The core limitation lies in the inability to adapt visual extraction to specific reasoning requirements, different queries demand fundamentally different visual evidence from the same video content. In this work, we present CAVIA, a training-free framework that revolutionizes video understanding through reasoning, perception coordination. Unlike conventional approaches where visual processing operates independently of reasoning, CAVIA creates a closed-loop system where reasoning continuously guides visual extraction based on identified information gaps. CAVIA introduces three innovations: (1) hierarchical reasoning, guided localization to precise frames; (2) cross-modal semantic bridging for targeted extraction; (3) confidence-driven iterative synthesis. CAVIA achieves state-of-the-art performance on challenging benchmarks: EgoSchema (65.7%, +5.3%), NExT-QA (76.1%, +2.6%), and IntentQA (73.8%, +6.9%), demonstrating that dynamic reasoning-perception coordination provides a scalable paradigm for video understanding.

cs.CV↗

arXiv2025

LeAdQA: LLM-Driven Context-Aware Temporal Grounding for Video Question Answering

Video Question Answering (VideoQA) requires identifying sparse critical moments in long videos and reasoning about their causal relationships to answer semantically complex questions. While recent advances in multimodal learning have improved alignment and fusion, current approaches remain limited by two prevalent but fundamentally flawed strategies: (1) task-agnostic sampling indiscriminately processes all frames, overwhelming key events with irrelevant content; and (2) heuristic retrieval captures superficial patterns but misses causal-temporal structures needed for complex reasoning. To address these challenges, we introduce LeAdQA, an innovative approach that bridges these gaps through synergizing causal-aware query refinement with fine-grained visual grounding. Our method first leverages LLMs to reformulate question-option pairs, resolving causal ambiguities and sharpening temporal focus. These refined queries subsequently direct a temporal grounding model to precisely retrieve the most salient segments, complemented by an adaptive fusion mechanism dynamically integrating the evidence to maximize relevance. The integrated visual-textual cues are then processed by an MLLM to generate accurate, contextually-grounded answers. Experiments on NExT-QA, IntentQA, and NExT-GQA demonstrate that our method's precise visual grounding substantially enhances the understanding of video-question relationships, achieving state-of-the-art (SOTA) performance on complex reasoning tasks while maintaining computational efficiency.

cs.CV↗

arXiv2025

Should We Ever Prefer Decision Transformer for Offline Reinforcement Learning?

In recent years, extensive work has explored the application of the Transformer architecture to reinforcement learning problems. Among these, Decision Transformer (DT) has gained particular attention in the context of offline reinforcement learning due to its ability to frame return-conditioned policy learning as a sequence modeling task. Most recently, Bhargava et al. (2024) provided a systematic comparison of DT with more conventional MLP-based offline RL algorithms, including Behavior Cloning (BC) and Conservative Q-Learning (CQL), and claimed that DT exhibits superior performance in sparse-reward and low-quality data settings. In this paper, through experimentation on robotic manipulation tasks (Robomimic) and locomotion benchmarks (D4RL), we show that MLP-based Filtered Behavior Cloning (FBC) achieves competitive or superior performance compared to DT in sparse-reward environments. FBC simply filters out low-performing trajectories from the dataset and then performs ordinary behavior cloning on the filtered dataset. FBC is not only very straightforward, but it also requires less training data and is computationally more efficient. The results therefore suggest that DT is not preferable for sparse-reward environments. From prior work, arguably, DT is also not preferable for dense-reward environments. Thus, we pose the question: Is DT ever preferable?

cs.AI↗

arXiv2025

Cross Entropy versus Label Smoothing: A Neural Collapse Perspective

Label smoothing loss is a widely adopted technique to mitigate overfitting in deep neural networks. This paper studies label smoothing from the perspective of Neural Collapse (NC), a powerful empirical and theoretical framework which characterizes model behavior during the terminal phase of training. We first show empirically that models trained with label smoothing converge faster to neural collapse solutions and attain a stronger level of neural collapse. Additionally, we show that at the same level of NC1, models under label smoothing loss exhibit intensified NC2. These findings provide valuable insights into the performance benefits and enhanced model calibration under label smoothing loss. We then leverage the unconstrained feature model to derive closed-form solutions for the global minimizers for both loss functions and further demonstrate that models under label smoothing have a lower conditioning number and, therefore, theoretically converge faster. Our study, combining empirical evidence and theoretical results, not only provides nuanced insights into the differences between label smoothing and cross-entropy losses, but also serves as an example of how the powerful neural collapse framework can be used to improve our understanding of DNNs.

cs.LG↗

arXiv2025

On the Convergence of Monte Carlo UCB for Random-Length Episodic MDPs

In reinforcement learning, Monte Carlo algorithms update the Q function by averaging the episodic returns. In the Monte Carlo UCB (MC-UCB) algorithm, the action taken in each state is the action that maximizes the Q function plus an Upper Confidence Bounds (UCB) exploration term, which biases the choice of actions to those that have been chosen less frequently. Although there has been significant work on establishing regret bounds for MC-UCB, most of that work has been focused on finite-horizon versions of the problem, for which each episode terminates after a constant number of steps. For such finite-horizon problems, the optimal policy depends both on the current state and the time within the episode. However, for many natural episodic problems, such as games like Go and Chess and robotic tasks, the episode is of random length and the optimal policy is stationary. For such environments, it is an open question whether the Q-function in MC-UCB will converge to the optimal Q function; we conjecture that, unlike Q-learning, it does not converge for all MDPs. We nevertheless show that for a large class of MDPs, which includes stochastic MDPs such as blackjack and deterministic MDPs such as Go, the Q function in MC-UCB converges almost surely to the optimal Q function. An immediate corollary of this result is that it also converges almost surely for all finite-horizon MDPs. We also provide numerical experiments, providing further insights into MC-UCB.

cs.LG↗

arXiv2024

The Prevalence of Neural Collapse in Neural Multivariate Regression

Recently it has been observed that neural networks exhibit Neural Collapse (NC) during the final stage of training for the classification problem. We empirically show that multivariate regression, as employed in imitation learning and other applications, exhibits Neural Regression Collapse (NRC), a new form of neural collapse: (NRC1) The last-layer feature vectors collapse to the subspace spanned by the $n$ principal components of the feature vectors, where $n$ is the dimension of the targets (for univariate regression, $n=1$); (NRC2) The last-layer feature vectors also collapse to the subspace spanned by the last-layer weight vectors; (NRC3) The Gram matrix for the weight vectors converges to a specific functional form that depends on the covariance matrix of the targets. After empirically establishing the prevalence of (NRC1)-(NRC3) for a variety of datasets and network architectures, we provide an explanation of these phenomena by modeling the regression task in the context of the Unconstrained Feature Model (UFM), in which the last layer feature vectors are treated as free variables when minimizing the loss function. We show that when the regularization parameters in the UFM model are strictly positive, then (NRC1)-(NRC3) also emerge as solutions in the UFM optimization problem. We also show that if the regularization parameters are equal to zero, then there is no collapse. To our knowledge, this is the first empirical and theoretical study of neural collapse in the context of regression. This extension is significant not only because it broadens the applicability of neural collapse to a new category of problems but also because it suggests that the phenomena of neural collapse could be a universal behavior in deep learning.

cs.LG↗

arXiv2024

EffiQA: Efficient Question-Answering with Strategic Multi-Model Collaboration on Knowledge Graphs

While large language models (LLMs) have shown remarkable capabilities in natural language processing, they struggle with complex, multi-step reasoning tasks involving knowledge graphs (KGs). Existing approaches that integrate LLMs and KGs either underutilize the reasoning abilities of LLMs or suffer from prohibitive computational costs due to tight coupling. To address these limitations, we propose a novel collaborative framework named EffiQA that can strike a balance between performance and efficiency via an iterative paradigm. EffiQA consists of three stages: global planning, efficient KG exploration, and self-reflection. Specifically, EffiQA leverages the commonsense capability of LLMs to explore potential reasoning pathways through global planning. Then, it offloads semantic pruning to a small plug-in model for efficient KG exploration. Finally, the exploration results are fed to LLMs for self-reflection to further improve the global planning and efficient KG exploration. Empirical evidence on multiple KBQA benchmarks shows EffiQA's effectiveness, achieving an optimal balance between reasoning accuracy and computational costs. We hope the proposed new framework will pave the way for efficient, knowledge-intensive querying by redefining the integration of LLMs and KGs, fostering future research on knowledge-based question answering.

cs.CL↗

arXiv2024

Pre-training with Synthetic Data Helps Offline Reinforcement Learning

Recently, it has been shown that for offline deep reinforcement learning (DRL), pre-training Decision Transformer with a large language corpus can improve downstream performance (Reid et al., 2022). A natural question to ask is whether this performance gain can only be achieved with language pre-training, or can be achieved with simpler pre-training schemes which do not involve language. In this paper, we first show that language is not essential for improved performance, and indeed pre-training with synthetic IID data for a small number of updates can match the performance gains from pre-training with a large language corpus; moreover, pre-training with data generated by a one-step Markov chain can further improve the performance. Inspired by these experimental results, we then consider pre-training Conservative Q-Learning (CQL), a popular offline DRL algorithm, which is Q-learning-based and typically employs a Multi-Layer Perceptron (MLP) backbone. Surprisingly, pre-training with simple synthetic data for a small number of updates can also improve CQL, providing consistent performance improvement on D4RL Gym locomotion datasets. The results of this paper not only illustrate the importance of pre-training for offline DRL but also show that the pre-training data can be synthetic and generated with remarkably simple mechanisms.

cs.AI↗