Kevin Wang

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.

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.

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.

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.

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.