Kevin Wang

Kevin Wang, Anna Thöni, Benjamin Kempinski et al.

Large language models (LLMs) are increasingly deployed as interactive agents, yet their capacity for social and strategic reasoning over extended interaction remains poorly understood. Existing evaluations rely on static vignettes or single-game benchmarks that cannot capture the sustained, multi-faceted reasoning that real-world multi-agent settings demand. We introduce Mindgames, a multi-game arena and evaluation platform for LLM agents that operationalizes complementary reasoning demands relevant to ``theory of mind'': belief attribution under hidden information, opponent modeling through repeated strategic interaction, cooperative inference under knowledge asymmetries, and sustained deception in social deduction. Built on TextArena, Mindgames provides a unified interaction interface, TrueSkill-based rating, and full trajectory logging across four game environments. We instantiate Mindgames through a 2025 competition cycle hosted at a major AI conference, which assessed 944 submitted agents from 76 teams across four games: Colonel Blotto, Iterated Prisoner's Dilemma, Codenames, and Secret Mafia. Our analysis surfaces both agent-level and evaluation-level limitations: brittle rule adherence remains a major bottleneck, top-performing systems repeatedly rely on explicit structural scaffolding, and leaderboard validity differs sharply across environments. In particular, failure-heavy environments can reward robustness to opponent errors as much as strategic ability, with Secret Mafia exhibiting a pronounced error-survival confound in this cycle. We release a dataset of 29,571 multi-agent games with turn-level observations, actions, and rewards, together with MG-Ref, a deterministic offline tournament protocol that scores new agents against a frozen reference pool of top-ranked, low-error Stage~II submissions under the same error-attribution lens used in this analysis.

Yang Yang, Kevin Wang, Yuanhai Luo et al.

The rise of LLM-as-a-Service and other confidential cloud workloads demands cryptographic proof that user data is processed in a trusted, untampered environment. Existing solutions, notably Confidential Containers (CoCo), enforce a strict "one Pod per VM" model that attests only the Guest OS stack, leaving container-level identity unverified and incurring prohibitive per-VM resource overhead. We present dstack-capsule, a Kubernetes platform that enables Pod-level remote attestation on Intel TDX by allowing multiple Pods to share a single Confidential VM while each retains independent, hardware-backed proof of identity. Our key insight is a two-layer attestation architecture: static platform measurements are frozen in RTMR[3] via an irreversible privilege fuse, while dynamic Pod identities (pod_uid, pod_spec_hash, workload_id) are embedded in the TDX Quote's report_data field and signed by hardware on every request. dstack-capsule introduces (1) a Pod-level attestation protocol binding Pod spec digests to hardware-signed Quotes; (2) a privilege fuse mechanism that atomically transitions a node from setup mode to secure mode; (3) a multi-layer sandbox spanning storage, runtime, admission, API, and network isolation layers; and (4) a complete open-source implementation based on Kubernetes 1.32, Intel TDX, and Sysbox. We evaluate the security properties, attestation correctness, and performance characteristics of dstack-capsule, demonstrating that it achieves Pod-granularity verification without the resource overhead of per-VM isolation.

Kevin Wang, Junbo Li, Neel P. Bhatt et al.

Recent advancements in Large Language Models (LLMs) have showcased their ability to perform complex reasoning tasks, but their effectiveness in planning remains underexplored. In this study, we evaluate the planning capabilities of OpenAI's o1 models across a variety of benchmark tasks, focusing on three key aspects: feasibility, optimality, and generalizability. Through empirical evaluations on constraint-heavy tasks (e.g., $\textit{Barman}$, $\textit{Tyreworld}$) and spatially complex environments (e.g., $\textit{Termes}$, $\textit{Floortile}$), we highlight o1-preview's strengths in self-evaluation and constraint-following, while also identifying bottlenecks in decision-making and memory management, particularly in tasks requiring robust spatial reasoning. Our results reveal that o1-preview outperforms GPT-4 in adhering to task constraints and managing state transitions in structured environments. However, the model often generates suboptimal solutions with redundant actions and struggles to generalize effectively in spatially complex tasks. This pilot study provides foundational insights into the planning limitations of LLMs, offering key directions for future research on improving memory management, decision-making, and generalization in LLM-based planning. Code available at https://github.com/VITA-Group/o1-planning.

Kevin Wang, Alexandre Variengien, Arthur Conmy et al.

Research in mechanistic interpretability seeks to explain behaviors of machine learning models in terms of their internal components. However, most previous work either focuses on simple behaviors in small models, or describes complicated behaviors in larger models with broad strokes. In this work, we bridge this gap by presenting an explanation for how GPT-2 small performs a natural language task called indirect object identification (IOI). Our explanation encompasses 26 attention heads grouped into 7 main classes, which we discovered using a combination of interpretability approaches relying on causal interventions. To our knowledge, this investigation is the largest end-to-end attempt at reverse-engineering a natural behavior "in the wild" in a language model. We evaluate the reliability of our explanation using three quantitative criteria--faithfulness, completeness and minimality. Though these criteria support our explanation, they also point to remaining gaps in our understanding. Our work provides evidence that a mechanistic understanding of large ML models is feasible, opening opportunities to scale our understanding to both larger models and more complex tasks.

Wenqing Zheng, S P Sharan, Ajay Kumar Jaiswal et al.

For a complicated algorithm, its implementation by a human programmer usually starts with outlining a rough control flow followed by iterative enrichments, eventually yielding carefully generated syntactic structures and variables in a hierarchy. However, state-of-the-art large language models generate codes in a single pass, without intermediate warm-ups to reflect the structured thought process of "outline-then-detail". Inspired by the recent success of chain-of-thought prompting, we propose ChainCoder, a program synthesis language model that generates Python code progressively, i.e. from coarse to fine in multiple passes. We first decompose source code into layout frame components and accessory components via abstract syntax tree parsing to construct a hierarchical representation. We then reform our prediction target into a multi-pass objective, each pass generates a subsequence, which is concatenated in the hierarchy. Finally, a tailored transformer architecture is leveraged to jointly encode the natural language descriptions and syntactically aligned I/O data samples. Extensive evaluations show that ChainCoder outperforms state-of-the-arts, demonstrating that our progressive generation eases the reasoning procedure and guides the language model to generate higher-quality solutions. Our codes are available at: https://github.com/VITA-Group/ChainCoder.

Yunfei Xie, Kevin Wang, Bobby Cheng et al.

Multi-turn, multi-agent LLM game evaluations often exhibit substantial run-to-run variance. In long-horizon interactions, small early deviations compound across turns and are amplified by multi-agent coupling. This biases win rate estimates and makes rankings unreliable across repeated tournaments. Prompt choice worsens this further by producing different effective policies. We address both instability and underperformance with MEMO (Memory-augmented MOdel context optimization), a self-play framework that optimizes inference-time context by coupling retention and exploration. Retention maintains a persistent memory bank that stores structured insights from self-play trajectories and injects them as priors during later play. Exploration runs tournament-style prompt evolution with uncertainty-aware selection via TrueSkill, and uses prioritized replay to revisit rare and decisive states. Across five text-based games, MEMO raises mean win rate from 25.1% to 49.5% for GPT-4o-mini and from 20.9% to 44.3% for Qwen-2.5-7B-Instruct, using $2,000$ self-play games per task. Run-to-run variance also drops, giving more stable rankings across prompt variations. These results suggest that multi-agent LLM game performance and robustness have substantial room for improvement through context optimization. MEMO achieves the largest gains in negotiation and imperfect-information games, while RL remains more effective in perfect-information settings. All code is open-source and available here: https://github.com/openverse-ai/MEMO

Yunhao Yang, Neel P. Bhatt, Kevin Wang et al.

Reusable robot skills are becoming the basic units through which embodied agents turn open-ended instructions into long-horizon physical behavior. We argue that, while foundation models have collapsed the cost of creating these skills, the cost of trusting them has not. Existing skill-evolution loops refine skills through execution feedback, unit tests, environment reward, or LLM self-critique, but these signals provide only trace-level evidence: they show that a skill worked on sampled executions, not that skill-induced plans satisfy temporal safety contracts under untested conditions. We introduce VASO, a framework for verification-guided self-evolution of LLM-generated robot skill contracts. In VASO, each skill is represented as a semantic contract with two coupled interfaces: a formal interface that aligns robot states, observations, and control commands with logical propositions for model checking, and a planner-facing interface that guides executable behavior generation. A model checker first filters logically inconsistent skill contracts, then verifies plans induced by the skill against global and local temporal specifications. When verification fails, VASO translates the counterexample trace into a textual gradient that updates the reusable skill contract while keeping foundation-model weights frozen. On Clearpath Jackal and PX4 quadcopter tasks, VASO reaches 97.2% formal-specification compliance using fewer than 100 optimization samples, outperforming execution-feedback, prompt-optimization, and fine-tuning baselines. To our knowledge, VASO is the first framework that closes the loop between formal verification and self-evolving LLM-generated skills for physical AI agents: formal counterexamples become optimization feedback for reusable robot skill contracts, rather than merely verifying one-off plans, tuning planner prompts, or fine-tuning model weights.

Stephen McAleer, JB Lanier, Kevin Wang et al.

In competitive two-agent environments, deep reinforcement learning (RL) methods based on the \emph{Double Oracle (DO)} algorithm, such as \emph{Policy Space Response Oracles (PSRO)} and \emph{Anytime PSRO (APSRO)}, iteratively add RL best response policies to a population. Eventually, an optimal mixture of these population policies will approximate a Nash equilibrium. However, these methods might need to add all deterministic policies before converging. In this work, we introduce \emph{Self-Play PSRO (SP-PSRO)}, a method that adds an approximately optimal stochastic policy to the population in each iteration. Instead of adding only deterministic best responses to the opponent's least exploitable population mixture, SP-PSRO also learns an approximately optimal stochastic policy and adds it to the population as well. As a result, SP-PSRO empirically tends to converge much faster than APSRO and in many games converges in just a few iterations.

Zhiwen Fan, Kevin Wang, Kairun Wen et al.

Recent advances in real-time neural rendering using point-based techniques have enabled broader adoption of 3D representations. However, foundational approaches like 3D Gaussian Splatting impose substantial storage overhead, as Structure-from-Motion (SfM) points can grow to millions, often requiring gigabyte-level disk space for a single unbounded scene. This growth presents scalability challenges and hinders splatting efficiency. To address this, we introduce LightGaussian, a method for transforming 3D Gaussians into a more compact format. Inspired by Network Pruning, LightGaussian identifies Gaussians with minimal global significance on scene reconstruction, and applies a pruning and recovery process to reduce redundancy while preserving visual quality. Knowledge distillation and pseudo-view augmentation then transfer spherical harmonic coefficients to a lower degree, yielding compact representations. Gaussian Vector Quantization, based on each Gaussian's global significance, further lowers bitwidth with minimal accuracy loss. LightGaussian achieves an average 15x compression rate while boosting FPS from 144 to 237 within the 3D-GS framework, enabling efficient complex scene representation on the Mip-NeRF 360 and Tank & Temple datasets. The proposed Gaussian pruning approach is also adaptable to other 3D representations (e.g., Scaffold-GS), demonstrating strong generalization capabilities.

James Wang, Primo Pu, Zephyr Fung et al.

The acquisition of high-quality, action-aligned demonstration data remains a fundamental bottleneck in scaling foundation models for dexterous robot manipulation. Although robot-free human demonstrations (e.g., the UMI paradigm) offer a scalable alternative to traditional teleoperation, current systems are constrained by sub-optimal hardware ergonomics, open-loop workflows, and a lack of systematic data-mixing strategies. To address these limitations, we present XRZero-G0, a hardware-software co-designed system for embodied data collection and policy learning. The system features an ergonomic, virtual reality interface equipped with a top-view camera and dual specialized grippers to directly improve collection efficiency. To ensure dataset reliability, we propose a closed-loop collection, inspection, training, and evaluation pipeline for non-proprioceptive data. This workflow achieves an 85% data validity rate and establishes a transparent mechanism for quality control. Furthermore, we investigate the empirical scaling behaviors and optimal mixing ratios of robot-free data. Extensive experiments indicate that combining a minimal volume of real-robot data with large-scale robot-free data (e.g., a 10:1 ratio) achieves performance comparable to exclusively real-robot datasets, while reducing acquisition costs by a factor of twenty. Utilizing XRZero-G0, we construct a 2,000-hour robot-free dataset that enables zero-shot cross-embodiment transfer to a target physical robot, demonstrating a highly scalable methodology for generalized real-world manipulation.Our project repository: https://github.com/X-Square-Robot/XRZero-G0

Zhiwen Fan, Jian Zhang, Renjie Li et al.

The rapid advancement of Large Multimodal Models (LMMs) for 2D images and videos has motivated extending these models to understand 3D scenes, aiming for human-like visual-spatial intelligence. Nevertheless, achieving deep spatial understanding comparable to human capabilities poses significant challenges in model encoding and data acquisition. Existing methods frequently depend on external depth sensors for geometry capture or utilize off-the-shelf algorithms for pre-constructing 3D maps, thereby limiting their scalability, especially with prevalent monocular video inputs and for time-sensitive applications. In this work, we introduce VLM-3R, a unified framework for Vision-Language Models (VLMs) that incorporates 3D Reconstructive instruction tuning. VLM-3R processes monocular video frames by employing a geometry encoder to derive implicit 3D tokens that represent spatial understanding. Leveraging our Spatial-Visual-View Fusion and over 200K curated 3D reconstructive instruction tuning question-answer (QA) pairs, VLM-3R effectively aligns real-world spatial context with language instructions. This enables monocular 3D spatial assistance and embodied reasoning. To facilitate the evaluation of temporal reasoning, we introduce the Vision-Spatial-Temporal Intelligence benchmark, featuring over 138.6K QA pairs across five distinct tasks focused on evolving spatial relationships. Extensive experiments demonstrate that our model, VLM-3R, not only facilitates robust visual-spatial reasoning but also enables the understanding of temporal 3D context changes, excelling in both accuracy and scalability.

Wenqing Zheng, S P Sharan, Zhiwen Fan et al.

Learning efficient and interpretable policies has been a challenging task in reinforcement learning (RL), particularly in the visual RL setting with complex scenes. While neural networks have achieved competitive performance, the resulting policies are often over-parameterized black boxes that are difficult to interpret and deploy efficiently. More recent symbolic RL frameworks have shown that high-level domain-specific programming logic can be designed to handle both policy learning and symbolic planning. However, these approaches rely on coded primitives with little feature learning, and when applied to high-dimensional visual scenes, they can suffer from scalability issues and perform poorly when images have complex object interactions. To address these challenges, we propose \textit{Differentiable Symbolic Expression Search} (DiffSES), a novel symbolic learning approach that discovers discrete symbolic policies using partially differentiable optimization. By using object-level abstractions instead of raw pixel-level inputs, DiffSES is able to leverage the simplicity and scalability advantages of symbolic expressions, while also incorporating the strengths of neural networks for feature learning and optimization. Our experiments demonstrate that DiffSES is able to generate symbolic policies that are simpler and more and scalable than state-of-the-art symbolic RL methods, with a reduced amount of symbolic prior knowledge.

Utsab Saha, Lola Egherman, Ramiz Rahman et al.

Undergraduates in work-based learning experiences often produce meaningful contributions as viewed by their supervisors, yet report a negative perception of their contributions because they struggled during the process or produced only a few lines of code change. As a result, many omit these contributions from their resumes and job interviews, losing a meaningful signal of technical ability. This study examines how guided blog posts help CS students in work based learning experiences reflect on what they learned and contextualize their experiences. It also evaluates the depth of reflection produced. The study included twenty-five juniors and seniors studying CS at CTCs and other affordable local colleges. All participated in one cohort during Fall 2024. Each student was assigned a simple open source issue to solve from a popular open source project over the course of several weeks with the help of an industry mentor. While working on the project, students drafted a LinkedIn blog post using a five-section outline covering project mission, assigned issue, technical architecture, challenges faced, and submitted solution. We conducted a thematic analysis of the published posts and measured reflection depth using Mejia and Turns's Knowledge Gain instrument. Four themes emerged from the posts: identifying problem solving techniques, growth mindset, the challenges and benefits of collaborative development, and the impacts of their contribution on users. Additionally, students demonstrated deep reflection across all four Knowledge Gain constructs. Structured blog posts offer a low-cost addition to experiential CS learning such as capstones, micro-internships, internships, and apprenticeships. This study is descriptive; future work should compare outcomes against a control group.

Jianing Zhu, Yeonju Ro, John Robertson et al.

Long-lived AI agents are increasingly deployed as persistent operational systems, yet they are still evaluated like freshly initialized models. Day-one benchmarks miss a basic systems question: how long does an agent remain reliable after deployment? Even when model weights are frozen, an agent's effective state keeps changing as it compresses interaction history, retrieves from a growing memory store, revises facts after updates, and undergoes routine maintenance. Reliability therefore becomes a lifespan property of the full agent harness, not only a snapshot property of the base model. We introduce AgingBench, a longitudinal reliability benchmark for agent lifespan engineering: measuring not only whether deployed agents degrade, but what form the degradation takes and where repair should target. AgingBench organizes agent aging into four mechanisms: compression aging, interference aging, revision aging, and maintenance aging. To diagnose these failures, AgingBench uses temporal dependency graphs and paired counterfactual probes that produce diagnostic profiles for the write, retrieval, and utilization stages of the memory pipeline. Across 7 scenarios, 14 models, multiple memory policies, and both runner-controlled and autonomous agents, over ~400 runs spanning 8 - 200 sessions show that agent aging is not one-dimensional: behavioral tests can remain clean while factual precision decays; derived-state tracking can collapse sharply within a single model; and the same wrong answer can require different repairs depending on what the diagnostic profile points to. These results suggest that reliable agent deployment requires lifespan evaluation, mechanism-level diagnosis, and stage-targeted repair, not only stronger day-one models.

Tyrone Serapio, Arjun Prakash, Haoyang Xu et al.

Large Language Models (LLMs) have shown great ability in generating executable code from natural language, opening the possibility of automatically constructing environments for AI agents. Recent work on Code World Models (CWMs) demonstrates that LLMs can translate game rules into Python implementations compatible with solvers like Monte Carlo Tree Search. We study this problem in game settings, where generated environments must implement rules, legal actions, state transitions, observations, and rewards. We refer to these game-specific executable models as Game Code World Models (GameCWMs). However, current approaches to generating code world models rely on frontier models and inference-time refinement loops, limiting accessibility and scalability. This work investigates whether GameCWM generation capabilities can be distilled into smaller models through post-training. We introduce: (1) a curated dataset of 30 games spanning perfect and imperfect information games, (2) a verification framework that evaluates generated code against structural and semantic game properties, and (3) a post-training pipeline combining Supervised Fine-Tuning (SFT) with Reinforcement Learning with Verifiable Rewards (RLVR). We experiment with Qwen2.5-3B-Instruct and find that SFT can increase syntactic correctness, while RLVR can improve execution-level adherence to game rules, thereby improving Qwen's ability to generate valid GameCWMs in both perfect and imperfect information games. Overall, our pipeline makes Qwen2.5-3B-Instruct more capable of generating valid GameCWMs, thereby offering a scalable path toward automatic environment generation from natural language.

Harrison Li, Kevin Wang, Cheol Jun Cho et al.

Building a diagnosis model for primary progressive aphasia (PPA) has been challenging due to the data scarcity. Collecting clinical data at scale is limited by the high vulnerability of clinical population and the high cost of expert labeling. To circumvent this, previous studies simulate dysfluent speech to generate training data. However, those approaches are not comprehensive enough to simulate PPA as holistic, multi-level phenotypes, instead relying on isolated dysfluencies. To address this, we propose a novel, clinically grounded simulation framework, Hierarchical Aphasic Speech Simulation (HASS). HASS aims to simulate behaviors of logopenic variant of PPA (lvPPA) with varying degrees of severity. To this end, semantic, phonological, and temporal deficits of lvPPA are systematically identified by clinical experts, and simulated. We demonstrate that our framework enables more accurate and generalizable detection models.

Aditya Narkhede, Erick Rivas, Kevin Wang

The design of structures and vehicles subject to fluid-structure interaction (FSI) often requires high-fidelity coupled analysis. While the design variables pertain to the structure, the computational cost is dominated by the fluid solver, making iterative optimization prohibitively expensive. This paper presents an adaptive multi-fidelity optimization method combining high-fidelity FSI analysis with a lightweight surrogate for fluid-induced loads and a decision model that selects between surrogate and high-fidelity fluid evaluations. During optimization, completed FSI analyses incrementally update a non-intrusive surrogate model based on nearest-neighbor search and radial interpolation. A hybrid Lagrangian-Eulerian mapping function is developed to transfer fluid loads between structural designs. The evolution of surface orientation is handled by decomposing the traction vectors into local orthonormal bases. An adaptive Gaussian process regression model is employed to predict surrogate error and quantify uncertainty, allowing risk-aware selection of when coupled analysis is required. As design evaluations cluster near the optimum, the accuracy of the surrogate model naturally improves, thereby reducing the reliance on the fluid solver. It requires no offline training, preserves the high-fidelity structural model in all design evaluations, and ensures that the final design is evaluated by high-fidelity FSI analysis. The fundamental idea is justified theoretically using a simplified model problem, which shows that the leading-order error is a monotonically increasing, concave, and bounded function of the fluid added mass. The framework is demonstrated on two benchmark problems. For shape optimization of a flexible panel under shock loading, results show an $80\%$ reduction in computational cost while maintaining accuracy within $2.3\%$ of fully high-fidelity FSI optimization.

Kevin Wang, Seth Akins, Abdallah Mohammed et al.

Generative AI systems such as ChatGPT have a disruptive effect on learning and assessment. Computer science requires practice to develop skills in problem solving and programming that are traditionally developed using assignments. Generative AI has the capability of completing these assignments for students with high accuracy, which dramatically increases the potential for academic integrity issues and students not achieving desired learning outcomes. This work investigates the performance of ChatGPT by evaluating it across three courses (CS1,CS2,databases). ChatGPT completes almost all introductory assessments perfectly. Existing detection methods, such as MOSS and JPlag (based on similarity metrics) and GPTzero (AI detection), have mixed success in identifying AI solutions. Evaluating instructors and teaching assistants using heuristics to distinguish between student and AI code shows that their detection is not sufficiently accurate. These observations emphasize the need for adapting assessments and improved detection methods.

Quinn Jacobson, Joe Luo, Jingfei Xu et al.

NeuroHex is a brain-inspired hexagonal coordinate system designed to support highly efficient world models and reference frames for online adaptive AI systems. Inspired by the hexadirectional firing structure of grid cells in the human brain, NeuroHex adopts a cubic isometric hexagonal coordinate formulation that provides full 60° rotational symmetry and low-cost translation, rotation and distance computation. We develop a mathematical framework that incorporates ring indexing, quantized angular encoding, and a hierarchical library of foundational, simple, and complex geometric shape primitives. These constructs allow low-overhead point-in-shape tests and spatial matching operations that are expensive in Cartesian coordinate systems. To support realistic settings, we also develop a novel tool (OSM2Hex) that can process OpenStreetMap (OSM) data sets and convert them into the NeuroHex coordinate system. The OSM2Hex spatial abstraction processing pipeline can achieve a reduction of 90-99% in geometric complexity while maintaining the relevant spatial structure map for navigation. Our initial results, based on actual city and neighborhood scale data sets, demonstrate that NeuroHex offers a highly efficient substrate for building dynamic world models to enable adaptive spatial reasoning in autonomous energy-efficient AI systems with continuous online-adaptive learning (COAL) capability.

Hengyu Liu, Yuehao Wang, Chenxin Li et al.

3D Gaussian splatting (3DGS) has enabled various applications in 3D scene representation and novel view synthesis due to its efficient rendering capabilities. However, 3DGS demands relatively significant GPU memory, limiting its use on devices with restricted computational resources. Previous approaches have focused on pruning less important Gaussians, effectively compressing 3DGS but often requiring a fine-tuning stage and lacking adaptability for the specific memory needs of different devices. In this work, we present an elastic inference method for 3DGS. Given an input for the desired model size, our method selects and transforms a subset of Gaussians, achieving substantial rendering performance without additional fine-tuning. We introduce a tiny learnable module that controls Gaussian selection based on the input percentage, along with a transformation module that adjusts the selected Gaussians to complement the performance of the reduced model. Comprehensive experiments on ZipNeRF, MipNeRF and Tanks\&Temples scenes demonstrate the effectiveness of our approach. Code is available at https://flexgs.github.io.

Stephen McAleer, John Lanier, Kevin Wang et al.

Policy Space Response Oracles (PSRO) is a reinforcement learning (RL) algorithm for two-player zero-sum games that has been empirically shown to find approximate Nash equilibria in large games. Although PSRO is guaranteed to converge to an approximate Nash equilibrium and can handle continuous actions, it may take an exponential number of iterations as the number of information states (infostates) grows. We propose Extensive-Form Double Oracle (XDO), an extensive-form double oracle algorithm for two-player zero-sum games that is guaranteed to converge to an approximate Nash equilibrium linearly in the number of infostates. Unlike PSRO, which mixes best responses at the root of the game, XDO mixes best responses at every infostate. We also introduce Neural XDO (NXDO), where the best response is learned through deep RL. In tabular experiments on Leduc poker, we find that XDO achieves an approximate Nash equilibrium in a number of iterations an order of magnitude smaller than PSRO. Experiments on a modified Leduc poker game and Oshi-Zumo show that tabular XDO achieves a lower exploitability than CFR with the same amount of computation. We also find that NXDO outperforms PSRO and NFSP on a sequential multidimensional continuous-action game. NXDO is the first deep RL method that can find an approximate Nash equilibrium in high-dimensional continuous-action sequential games. Experiment code is available at https://github.com/indylab/nxdo.

Kevin Wang, Subre Abdoul Moktar, Jia Li et al.

Large Language Models (LLMs) have become increasingly pervasive, finding applications across many industries and disciplines. Ensuring the trustworthiness of LLM outputs is paramount, where Uncertainty Estimation (UE) plays a key role. In this work, a comprehensive empirical study is conducted to examine the robustness and effectiveness of diverse UE measures regarding aleatoric and epistemic uncertainty in LLMs. It involves twelve different UE methods and four generation quality metrics including LLMScore from LLM criticizers to evaluate the uncertainty of LLM-generated answers in Question-Answering (QA) tasks on both in-distribution (ID) and out-of-distribution (OOD) datasets. Our analysis reveals that information-based methods, which leverage token and sequence probabilities, perform exceptionally well in ID settings due to their alignment with the model's understanding of the data. Conversely, density-based methods and the P(True) metric exhibit superior performance in OOD contexts, highlighting their effectiveness in capturing the model's epistemic uncertainty. Semantic consistency methods, which assess variability in generated answers, show reliable performance across different datasets and generation metrics. These methods generally perform well but may not be optimal for every situation.

Zhiwen Fan, Wenyan Cong, Kairun Wen et al.

While neural 3D reconstruction has advanced substantially, its performance significantly degrades with sparse-view data, which limits its broader applicability, since SfM is often unreliable in sparse-view scenarios where feature matches are scarce. In this paper, we introduce InstantSplat, a novel approach for addressing sparse-view 3D scene reconstruction at lightning-fast speed. InstantSplat employs a self-supervised framework that optimizes 3D scene representation and camera poses by unprojecting 2D pixels into 3D space and aligning them using differentiable neural rendering. The optimization process is initialized with a large-scale trained geometric foundation model, which provides dense priors that yield initial points through model inference, after which we further optimize all scene parameters using photometric errors. To mitigate redundancy introduced by the prior model, we propose a co-visibility-based geometry initialization, and a Gaussian-based bundle adjustment is employed to rapidly adapt both the scene representation and camera parameters without relying on a complex adaptive density control process. Overall, InstantSplat is compatible with multiple point-based representations for view synthesis and surface reconstruction. It achieves an acceleration of over 30x in reconstruction and improves visual quality (SSIM) from 0.3755 to 0.7624 compared to traditional SfM with 3D-GS.

Kevin Wang, Jason Ramos, Ramon Lawrence

With the rapid evolution of Natural Language Processing (NLP), Large Language Models (LLMs) like ChatGPT have emerged as powerful tools capable of transforming various sectors. Their vast knowledge base and dynamic interaction capabilities represent significant potential in improving education by operating as a personalized assistant. However, the possibility of generating incorrect, biased, or unhelpful answers are a key challenge to resolve when deploying LLMs in an education context. This work introduces an innovative architecture that combines the strengths of ChatGPT with a traditional information retrieval based chatbot framework to offer enhanced student support in higher education. Our empirical evaluations underscore the high promise of this approach.

Kevin Wang, Ishaan Javali, Michał Bortkiewicz et al.

Scaling up self-supervised learning has driven breakthroughs in language and vision, yet comparable progress has remained elusive in reinforcement learning (RL). In this paper, we study building blocks for self-supervised RL that unlock substantial improvements in scalability, with network depth serving as a critical factor. Whereas most RL papers in recent years have relied on shallow architectures (around 2 - 5 layers), we demonstrate that increasing the depth up to 1024 layers can significantly boost performance. Our experiments are conducted in an unsupervised goal-conditioned setting, where no demonstrations or rewards are provided, so an agent must explore (from scratch) and learn how to maximize the likelihood of reaching commanded goals. Evaluated on simulated locomotion and manipulation tasks, our approach increases performance by $2\times$ - $50\times$. Increasing the model depth not only increases success rates but also qualitatively changes the behaviors learned.

Jianzhu Yao, Kevin Wang, Ryan Hsieh et al.

Reasoning and strategic behavior in social interactions is a hallmark of intelligence. This form of reasoning is significantly more sophisticated than isolated planning or reasoning tasks in static settings (e.g., math problem solving). In this paper, we present Strategic Planning, Interaction, and Negotiation (SPIN-Bench), a new multi-domain evaluation designed to measure the intelligence of strategic planning and social reasoning. While many existing benchmarks focus on narrow planning or single-agent reasoning, SPIN-Bench combines classical PDDL tasks, competitive board games, cooperative card games, and multi-agent negotiation scenarios in one unified framework. The framework includes both a benchmark as well as an arena to simulate and evaluate the variety of social settings to test reasoning and strategic behavior of AI agents. We formulate the benchmark SPIN-Bench by systematically varying action spaces, state complexity, and the number of interacting agents to simulate a variety of social settings where success depends on not only methodical and step-wise decision making, but also conceptual inference of other (adversarial or cooperative) participants. Our experiments reveal that while contemporary LLMs handle basic fact retrieval and short-range planning reasonably well, they encounter significant performance bottlenecks in tasks requiring deep multi-hop reasoning over large state spaces and socially adept coordination under uncertainty. We envision SPIN-Bench as a catalyst for future research on robust multi-agent planning, social reasoning, and human--AI teaming. Project Website: https://spinbench.github.io/

Kevin Wang, Hongqian Niu, Didong Li

Generative Artificial Intelligence (AI), such as large language models (LLMs), has become a transformative force across science, industry, and society. As these systems grow in popularity, web data becomes increasingly interwoven with this AI-generated material and it is increasingly difficult to separate them from naturally generated content. As generative models are updated regularly, later models will inevitably be trained on mixtures of human-generated data and AI-generated data from earlier versions, creating a recursive training process with data contamination. Existing theoretical work has examined only highly simplified settings, where both the real data and the generative model are discrete or Gaussian, where it has been shown that such recursive training leads to model collapse. However, real data distributions are far more complex, and modern generative models are far more flexible than Gaussian and linear mechanisms. To fill this gap, we study recursive training in a general framework with minimal assumptions on the real data distribution and allow the underlying generative model to be a general universal approximator. In this framework, we show that contaminated recursive training still converges, with a convergence rate equal to the minimum of the baseline model's convergence rate and the fraction of real data used in each iteration. To the best of our knowledge, this is the first (positive) theoretical result on recursive training without distributional assumptions on the data. We further extend the analysis to settings where sampling bias is present in data collection and support all theoretical results with empirical studies.

Wenyan Cong, Hanqing Zhu, Kevin Wang et al.

Efficiently reconstructing 3D scenes from monocular video remains a core challenge in computer vision, vital for applications in virtual reality, robotics, and scene understanding. Recently, frame-by-frame progressive reconstruction without camera poses is commonly adopted, incurring high computational overhead and compounding errors when scaling to longer videos. To overcome these issues, we introduce VideoLifter, a novel video-to-3D pipeline that leverages a local-to-global strategy on a fragment basis, achieving both extreme efficiency and SOTA quality. Locally, VideoLifter leverages learnable 3D priors to register fragments, extracting essential information for subsequent 3D Gaussian initialization with enforced inter-fragment consistency and optimized efficiency. Globally, it employs a tree-based hierarchical merging method with key frame guidance for inter-fragment alignment, pairwise merging with Gaussian point pruning, and subsequent joint optimization to ensure global consistency while efficiently mitigating cumulative errors. This approach significantly accelerates the reconstruction process, reducing training time by over 82% while holding better visual quality than current SOTA methods.

Wenyan Cong, Hanqing Zhu, Peihao Wang et al.

World foundation models, which simulate the physical world by predicting future states from current observations and inputs, have become central to many applications in physical intelligence, including autonomous driving and robotics. However, these models require substantial computational resources for pretraining and are further constrained by available data during post-training. As such, scaling computation at test time emerges as both a critical and practical alternative to traditional model enlargement or re-training. In this work, we introduce SWIFT, a test-time scaling framework tailored for WFMs. SWIFT integrates our extensible WFM evaluation toolkit with process-level inference strategies, including fast tokenization, probability-based Top-K pruning, and efficient beam search. Empirical results on the COSMOS model demonstrate that test-time scaling exists even in a compute-optimal way. Our findings reveal that test-time scaling laws hold for WFMs and that SWIFT provides a scalable and effective pathway for improving WFM inference without retraining or increasing model size. Project page: https://scalingwfm.github.io/.

Kaitlyn Wang, Jian Ge, Kevin Willis et al.

This paper presents GPFC, a novel Graphics Processing Unit (GPU) Phase Folding and Convolutional Neural Network (CNN) system to detect exoplanets using the transit method. We devise a fast folding algorithm parallelized on a GPU to amplify low signal-to-noise ratio transit signals, allowing a search at high precision and speed. A CNN trained on two million synthetic light curves reports a score indicating the likelihood of a planetary signal at each period. While the GPFC method has broad applicability across period ranges, this research specifically focuses on detecting ultra-short-period planets with orbital periods less than one day. GPFC improves on speed by three orders of magnitude over the predominant Box-fitting Least Squares (BLS) method. Our simulation results show GPFC achieves $97%$ training accuracy, higher true positive rate at the same false positive rate of detection, and higher precision at the same recall rate when compared to BLS. GPFC recovers $100\%$ of known ultra-short-period planets in $\textit{Kepler}$ light curves from a blind search. These results highlight the promise of GPFC as an alternative approach to the traditional BLS algorithm for finding new transiting exoplanets in data taken with $\textit{Kepler}$ and other space transit missions such as K2, TESS and future PLATO and Earth 2.0.

Kevin Wang, Hongqian Niu, Yixin Wang et al.

Generative networks have shown remarkable success in learning complex data distributions, particularly in generating high-dimensional data from lower-dimensional inputs. While this capability is well-documented empirically, its theoretical underpinning remains unclear. One common theoretical explanation appeals to the widely accepted manifold hypothesis, which suggests that many real-world datasets, such as images and signals, often possess intrinsic low-dimensional geometric structures. Under this manifold hypothesis, it is widely believed that to approximate a distribution on a $d$-dimensional Riemannian manifold, the latent dimension needs to be at least $d$ or $d+1$. In this work, we show that this requirement on the latent dimension is not necessary by demonstrating that generative networks can approximate distributions on $d$-dimensional Riemannian manifolds from inputs of any arbitrary dimension, even lower than $d$, taking inspiration from the concept of space-filling curves. This approach, in turn, leads to a super-exponential complexity bound of the deep neural networks through expanded neurons. Our findings thus challenge the conventional belief on the relationship between input dimensionality and the ability of generative networks to model data distributions. This novel insight not only corroborates the practical effectiveness of generative networks in handling complex data structures, but also underscores a critical trade-off between approximation error, dimensionality, and model complexity.

Kyuyoung Kim, Kevin Wang, Yunfei Xie et al.

Training language models to produce both correct answers and sound reasoning remains an open challenge. Reinforcement learning with verifiable rewards typically optimizes only final outcomes, which can lead to a failure mode where task accuracy improves while reasoning becomes less accurate, less complete, or even internally inconsistent. We propose verifiable process supervision (VPS), a post-training framework for verifiable domains that jointly optimizes prediction accuracy and reasoning quality. We first apply supervised fine-tuning to induce a structured reasoning format, enabling syntactic extraction of intermediate claims that are evaluated against ground-truth signals to form process-level rewards. To address the heterogeneous difficulty of reasoning subtasks, we introduce adaptive reward weighting that prioritizes components with the largest remaining errors, creating an implicit curriculum. We evaluate VPS on chess, a controlled testbed where reasoning steps can be deterministically verified against engine signals. While accuracy-only RL improves move accuracy, it sharply degrades reasoning quality, increasing win-rate error by up to 112% and reducing internal consistency by up to 69%. In contrast, VPS preserves accuracy while significantly improving reasoning quality, reducing win-rate error by up to 30% and restoring consistency to near saturation. At matched accuracy, judge evaluation also prefers the process-supervised models. A reasoning-space analysis further shows that, without a structured prior, accuracy-only RL converges to budget-dependent shortcuts rather than sound multi-step reasoning. These results show that VPS enables language models to reason both accurately and reliably in verifiable domains.

Kevin Wang, Neel P. Bhatt, Cong Liu et al.

Large language models (LLMs) can be adapted either through numerical updates that alter model parameters or symbolic manipulations that work on discrete prompts or logical constraints. While numerical fine-tuning excels at injecting new factual knowledge, symbolic updates offer flexible control of style and alignment without retraining. We introduce a neurosymbolic LoRA framework that dynamically combines these two complementary strategies. Specifically, we present a unified monitoring signal and a reward-based classifier to decide when to employ LoRA for deeper factual reconstruction and when to apply TextGrad for token-level edits. Our approach remains memory-efficient by offloading the symbolic transformations to an external LLM only when needed. Additionally, the refined prompts produced during symbolic editing serve as high-quality, reusable training data, an important benefit in data-scarce domains like mathematical reasoning. Extensive experiments across multiple LLM backbones show that neurosymbolic LoRA consistently outperforms purely numerical or purely symbolic baselines, demonstrating superior adaptability and improved performance. Our findings highlight the value of interleaving numerical and symbolic updates to unlock a new level of versatility in language model fine-tuning.

Shunfan Zhou, Kevin Wang, Hang Yin

Web3 applications require execution platforms that maintain confidentiality and integrity without relying on centralized trust authorities. While Trusted Execution Environments (TEEs) offer promising capabilities for confidential computing, current implementations face significant limitations when applied to Web3 contexts, particularly in security reliability, censorship resistance, and vendor independence. This paper presents dstack, a comprehensive framework that transforms raw TEE technology into a true Zero Trust platform. We introduce three key innovations: (1) Portable Confidential Containers that enable seamless workload migration across heterogeneous TEE environments while maintaining security guarantees, (2) Decentralized Code Management that leverages smart contracts for transparent governance of TEE applications, and (3) Verifiable Domain Management that ensures secure and verifiable application identity without centralized authorities. These innovations are implemented through three core components: dstack-OS, dstack-KMS, and dstack-Gateway. Together, they demonstrate how to achieve both the performance advantages of VM-level TEE solutions and the trustless guarantees required by Web3 applications. Our evaluation shows that dstack provides comprehensive security guarantees while maintaining practical usability for real-world applications.

Dharsan Ravindran, Kevin Wang, Zhuoyuan Cao et al.

Recent advances in vision foundation models, such as the Segment Anything Model (SAM) and its successor SAM2, have achieved state-of-the-art performance on general image segmentation benchmarks. However, these models struggle in adverse weather conditions where visual ambiguity is high, largely due to their lack of uncertainty quantification. Inspired by progress in medical imaging, where uncertainty-aware training has improved reliability in ambiguous cases, we investigate two approaches to enhance segmentation robustness for autonomous driving. First, we introduce a multi-step finetuning procedure for SAM2 that incorporates uncertainty metrics directly into the loss function, improving overall scene recognition. Second, we adapt the Uncertainty-Aware Adapter (UAT), originally designed for medical image segmentation, to driving contexts. We evaluate both methods on CamVid, BDD100K, and GTA driving datasets. Experiments show that UAT-SAM outperforms standard SAM in extreme weather, while SAM2 with uncertainty-aware loss achieves improved performance across diverse driving scenes. These findings underscore the value of explicit uncertainty modeling for safety-critical autonomous driving in challenging environments.

David Sychrovský, Christopher Solinas, Revan MacQueen et al.

Nash equilibrium is perhaps the best-known solution concept in game theory. Such a solution assigns a strategy to each player which offers no incentive to unilaterally deviate. While a Nash equilibrium is guaranteed to always exist, the problem of finding one in general-sum games is PPAD-complete, generally considered intractable. Regret minimization is an efficient framework for approximating Nash equilibria in two-player zero-sum games. However, in general-sum games, such algorithms are only guaranteed to converge to a coarse-correlated equilibrium (CCE), a solution concept where players can correlate their strategies. In this work, we use meta-learning to minimize the correlations in strategies produced by a regret minimizer. This encourages the regret minimizer to find strategies that are closer to a Nash equilibrium. The meta-learned regret minimizer is still guaranteed to converge to a CCE, but we give a bound on the distance to Nash equilibrium in terms of our meta-loss. We evaluate our approach in general-sum imperfect information games. Our algorithms provide significantly better approximations of Nash equilibria than state-of-the-art regret minimization techniques.

Pranav Gupta, Rishabh Rengarajan, Viren Bankapur et al.

Combining LiDAR and Camera-view data has become a common approach for 3D Object Detection. However, previous approaches combine the two input streams at a point-level, throwing away semantic information derived from camera features. In this paper we propose Cross-View Center Point-Fusion, a state-of-the-art model to perform 3D object detection by combining camera and LiDAR-derived features in the BEV space to preserve semantic density from the camera stream while incorporating spacial data from the LiDAR stream. Our architecture utilizes aspects from previously established algorithms, Cross-View Transformers and CenterPoint, and runs their backbones in parallel, allowing efficient computation for real-time processing and application. In this paper we find that while an implicitly calculated depth-estimate may be sufficiently accurate in a 2D map-view representation, explicitly calculated geometric and spacial information is needed for precise bounding box prediction in the 3D world-view space.

Kaitlyn Wang, Jian Ge, Kevin Willis et al.

Of over 5,000 exoplanets identified so far, only a few hundred possess sub-Earth radii. The formation processes of these sub-Earths remain elusive, and acquiring additional samples is essential for investigating this unique population. In our study, we employ the GPFC method, a novel GPU Phase Folding algorithm combined with a Convolutional Neural Network, on Kepler photometry data. This method enhances the transit search speed significantly over the traditional Box-fitting Least Squares method, allowing a complete search of the known Kepler KOI data within days using a commercial GPU card. To date, we have identified five new ultra-short-period planets (USPs): Kepler-158d, Kepler-963c, Kepler-879c, Kepler-1489c, and KOI-4978.02. Kepler-879c with a radius of $0.4 R_\oplus$ completes its orbit around a G dwarf in 0.646716 days. Kepler-158d with a radius of $0.43 R_\oplus$ orbits a K dwarf star every 0.645088 days. Kepler-1489c with a radius of $0.51 R_\oplus$ orbits a G dwarf in 0.680741 days. Kepler-963c with a radius of $0.6 R_\oplus$ revolves around a G dwarf in 0.919783 days, and KOI-4978.02 with a radius of $0.7 R_\oplus$ circles a G dwarf in 0.941967 days. Among our findings, Kepler-879c, Kepler-158d and Kepler-963c rank as the first, the third, the fourth smallest USPs identified to date. Notably, Kepler-158d stands as the smallest USP found orbiting K dwarfs while Kepler-963c, Kepler-879c, Kepler-1489c, and KOI-4978.02 are the smallest USPs found orbiting G dwarfs. Kepler-879c, Kepler-158d, Kepler-1489c, and KOI-4978.02 are among the smallest planets that are closest to their host stars, with orbits within 5 stellar radii. In addition, these discoveries highlight GPFC's promising capability in identifying small, new transiting exoplanets within photometry data from Kepler, TESS, and upcoming space transit missions, PLATO and ET.

Stephen McAleer, Kevin Wang, John Lanier et al.

Policy space response oracles (PSRO) is a multi-agent reinforcement learning algorithm that has achieved state-of-the-art performance in very large two-player zero-sum games. PSRO is based on the tabular double oracle (DO) method, an algorithm that is guaranteed to converge to a Nash equilibrium, but may increase exploitability from one iteration to the next. We propose anytime double oracle (ADO), a tabular double oracle algorithm for 2-player zero-sum games that is guaranteed to converge to a Nash equilibrium while decreasing exploitability from one iteration to the next. Unlike DO, in which the restricted distribution is based on the restricted game formed by each player's strategy sets, ADO finds the restricted distribution for each player that minimizes its exploitability against any policy in the full, unrestricted game. We also propose a method of finding this restricted distribution via a no-regret algorithm updated against best responses, called RM-BR DO. Finally, we propose anytime PSRO (APSRO), a version of ADO that calculates best responses via reinforcement learning. In experiments on Leduc poker and random normal form games, we show that our methods achieve far lower exploitability than DO and PSRO and decrease exploitability monotonically.

Kevin Wang, Deva Ramanan, Aayush Bansal

We study video-specific autoencoders that allow a human user to explore, edit, and efficiently transmit videos. Prior work has independently looked at these problems (and sub-problems) and proposed different formulations. In this work, we train a simple autoencoder (from scratch) on multiple frames of a specific video. We observe: (1) latent codes learned by a video-specific autoencoder capture spatial and temporal properties of that video; and (2) autoencoders can project out-of-sample inputs onto the video-specific manifold. These two properties allow us to explore, edit, and efficiently transmit a video using one learned representation. For e.g., linear operations on latent codes allow users to visualize the contents of a video. Associating latent codes of a video and manifold projection enables users to make desired edits. Interpolating latent codes and manifold projection allows the transmission of sparse low-res frames over a network.

Ramon Iglesias, Federico Rossi, Kevin Wang et al.

The goal of this paper is to present an end-to-end, data-driven framework to control Autonomous Mobility-on-Demand systems (AMoD, i.e. fleets of self-driving vehicles). We first model the AMoD system using a time-expanded network, and present a formulation that computes the optimal rebalancing strategy (i.e., preemptive repositioning) and the minimum feasible fleet size for a given travel demand. Then, we adapt this formulation to devise a Model Predictive Control (MPC) algorithm that leverages short-term demand forecasts based on historical data to compute rebalancing strategies. We test the end-to-end performance of this controller with a state-of-the-art LSTM neural network to predict customer demand and real customer data from DiDi Chuxing: we show that this approach scales very well for large systems (indeed, the computational complexity of the MPC algorithm does not depend on the number of customers and of vehicles in the system) and outperforms state-of-the-art rebalancing strategies by reducing the mean customer wait time by up to to 89.6%.