Publications
Our teams aspire to make discoveries that impact everyone, and core to our approach is sharing our research and tools to fuel progress in the field.
Our teams aspire to make discoveries that impact everyone, and core to our approach is sharing our research and tools to fuel progress in the field.
Sort By
1 - 15 of 11420 publications
Preview abstract
Understanding long visual documents, where information is distributed across extensive pages of text and graphics, remains a critical challenge for modern Vision-Language Models (VLMs). This difficulty is rooted in two fundamental obstacles: poor evidence localization and a high tendency for model hallucination. To address these issues, we propose DocLens, a multi-agent framework that decomposes the task into two specialized stages.
First, a Lens Module leverages document parsing tools for fine-grained, hierarchical evidence localization at both the page and element level. Second, a Reasoning Module employs a sampling-adjudication mechanism to systematically analyze the localized evidence, mitigating hallucination and synthesizing reliable answers.
Paired with Gemini-2.5-Pro, DocLens achieves state-of-the-art performance on MMLongBench-Doc and FinRAGBench-V, even surpassing human experts. Furthermore, our framework offers a highly cost-effective variant that delivers comparable performance to strong baselines at a five-fold reduction in cost.
View details
AI-PROPELLER: Warehouse-Scale Inter-Procedural Code Layout Optimization with AlphaEvolve
Chaitanya Mamatha Ananda
Rajiv Gupta
Aiden Grossman
2026
Preview abstract
Post-link optimizers (PLOs) such as Propeller and BOLT have demonstrated that precise, profile-guided code layout can extract significant performance gains from heavily optimized binaries. However, these systems are currently restricted to intra-procedural techniques, leaving the global potential of inter-procedural layout largely untapped. Inter-procedural code layout is historically difficult due to a combinatorially intractable search space and complex call-return semantics that are challenging to model. Consequently, the performance potential of fine-grained inter-procedural layout remains unproven in practice.Ours uses AlphaEvolve, an agentic workflow to evolve the compiler heuristic in Propeller into a fine-grained inter-procedural optimizer. While AlphaEvolve synthesizes novel code layout policies, Vizier fine-tunes the resulting policy hyperparameters.
To ensure high-fidelity, we move away from approximate static cost models and the agentic workflow generates multiple layout variants that are executed on actual hardware to measure real performance counters, providing a precise reward signal for the evolutionary loop. Ours has been evaluated on several benchmarks including large warehouse-scale applications and experiments show performance improvements of 0.23% to 1.6% on these benchmarks optimized with state-of-the-art FDO and PLO.
This is the first time ever that real-world applications have been optimized with fine-grained inter-procedural code layout.
View details
Dynamic Cogeneration of Bug Reproduction Test in Agentic Program Repair
Michele Tufano
José Cambronero
Renyao Wei
Grant Uy
34th ACM International Conference on the Foundations of Software Engineering (FSE) (2026)
Preview abstract
Bug Reproduction Tests (BRTs) have been used in many Automated Program Repair (APR) systems, primarily for validating fixes and aiding fix generation. In practice, when developers submit a patch, they often implement the BRT alongside the fix. Our experience deploying agentic APR reveals that developers desire a BRT within AI-generated patches to increase their confidence. However, canonical APR systems tend to generate BRTs and fixes separately, and focus on producing only the fix in the final patch. In this paper, we study agentic APR in the context of cogeneration, where the APR agent is instructed to generate both a fix and a BRT in the same patch. We evaluate the effectiveness of different cogeneration strategies on 120 human-reported bugs at Google and characterize different cogeneration strategies by their influence on APR agent behavior. We develop and evaluate patch selectors that account for test change to select patches with plausible fixes (and plausible BRTs). Finally, we analyze the root causes of failed cogeneration trajectories. We show that cogeneration allows the APR agent to generate BRTs for at least as many bugs as a dedicated BRT agent, without compromising the generation rate of plausible fixes, thereby reducing engineering effort in maintaining and coordinating separate generation pipelines for fix and BRT at scale.
View details
Reinforcement Learning with Discrete Diffusion Policies for Combinatorial Action-Spaces
Haitong Ma
Ofir Nabati
Bo Dai
Na Li
Shie Mannor
Guy Tennenholtz
Proceedings of the 43rd International Conference on Machine Learning (ICML-26), Seoul, South Korea (2026)
Preview abstract
Reinforcement learning (RL) algorithms have achieved superhuman performance
on many sequential decision-making tasks, but often struggle in domains with
large, combinatorial action spaces. To address this, we introduce a practical and
stable algorithm for training discrete diffusion models to represent policies in
such environments. We formulate a policy mirror descent algorithm that enhances
training stability by reframing policy optimization as an inference problem, which
naturally aligns with the learning objective of discrete diffusion models. Through
extensive experiments on a suite of challenging benchmark tasks, we demonstrate
that our approach achieves significant improvements over existing methods in both
performance and sample efficiency. This work opens a promising new direction
for applying discrete diffusion models in RL to tackle long-standing challenges in
large-scale combinatorial action spaces.
View details
Preview abstract
Optimizing large-language model (LLM) training and serving on large-sacle distributed systems with hundreds and thousands of accelerators is always a challenging task due to the fast evloving LLMs, strong domain expertise required, and various optimization goals from different worklaods. Existing methods rely on either handcrafted optimization performed by human experts, which is tedious and time-consuming or resource-intensive black-box searches, which lack the extensibility to keep pace with evolving models and hardware. To address this, we introduce PROMPTS, a novel multi-agent framework that complements traditional search methods with expert-informed reasoning. It automates the diagnosis of performance bottlenecks by synthesizing profiler data and leverages a knowledge base to propose optimized sharding configurations with detailed justifications.
Across eight real-world production workloads, PROMPTS demonstrated remarkable efficiency and accuracy, delivering performance improvements of up to 434%. These workloads spanned diverse model architectures, hardware platforms, computational scales, and various stages of the machine learning lifecycle (pre-training, serving, and post-training). In every case, the configuration adopted by human engineers was identified within the agent's top three proposals from a single invocation. Furthermore, the agent's top-ranked recommendation was the one ultimately adopted in 87.5% of cases, showcasing its ability to not only find optimized solutions, but also to correctly prioritize them. Our work establishes PROMPTS as a scalable, extensible, and explainable methodology for AI-assisted performance engineering in large-scale ML systems.
View details
Preview abstract
Online video platforms face an exponential challenge
in detecting and mitigating the flood of AI-generated “slop”
and synthetic spam perpetuated by coordinated malicious actors.
This content is increasingly designed to exploit the limitations
of traditional media forensics, often utilizing generative AI to
produce unique, localized variations of harmful or low-quality
material at scale. Traditional content-centric moderation fails
against this coordinated, adversarial generation strategy.
This paper presents a novel, scalable detection and classifi-
cation framework designed for online video platforms (OVP) to
identify and triage clusters of coordinated accounts exhibiting
a prevalence of adversarial synthetic content. The approach
leverages a multi-faceted architecture incorporating two core
machine learning components: a robust Coordinated Bot-Net
Detector (via Account Relatedness) and a Synthetic Pattern Clas-
sifier. Crucially, we introduce an advanced AI enhancement layer
utilizing Large Language Models (LLMs), specialized via Low-
Rank Adaptation (LoRA) and Automatic Prompt Optimization
(APO), to achieve rapid, high-precision semantic understanding
of emerging synthetic spam trends.
Evaluated across a representative evaluation dataset (N =
16, 250 weekly candidate channels across six major synthetic
abuse verticals), the system demonstrates high precision (FPR <
0.05%) in identifying coordinated synthetic spam networks.
Furthermore, the LLM-driven classification achieves a 74%
automated triage routing rate, saving over 1, 100 operational
review hours per week while reducing investigation turnaround
times by up to 50% (p < 0.001). This work details a critical
system design that provides essential scalability and adversarial
resilience against sophisticated generative attacks.
View details
Preview abstract
Quantization methods have significantly improved the compute and memory efficiency of Large Language Model (LLM) training. However, existing approaches still rely on accumulating their updates into high precision: concretely, gradient updates must be applied to a high-precision weight buffer, known as \textit{master weights}. This buffer introduces substantial memory overhead, particularly for Sparse Mixture of Experts (SMoE) models, where model parameters and optimizer states dominate memory usage. In this work, we introduce the Error-Compensating Optimizer (ECO), which \textit{for the first time} enables the complete elimination of master weights by directly accumulating updates into quantized parameters, by leveraging existing optimizer states. ECO quantizes the weights after every gradient step and injects the resulting quantization error into the optimizer's momentum buffer, creating an error-feedback loop with zero additional memory overhead for quantization. Beyond its practical efficiency, ECO comes with theoretical guarantees. Specifically, under standard assumptions, naive master weight removal can lead to unbounded drift from the ideal parameter trajectory, whereas ECO provably bounds this drift, ensuring stable convergence. We validate ECO across a range of models, including small transformers (30M--800M), Gemma-3 1B, and an SMoE 2.1B model, using FP8 quantization. In all cases, ECO achieves near-lossless accuracy compared to high-precision baselines. For large SMoE models, ECO reduces memory usage by up to 25\%, establishing a new Pareto frontier for the trade-off between static memory and training loss.
View details
Preview abstract
Source-to-source compilers may perform inefficiently by executing transpilation passes on scripts that do not contain the specific language features a pass is designed to transform, potentially leading to redundant processing. A compiler can analyze a script to generate a per-script feature map, for example, by identifying language features in its abstract syntax tree (AST). Before executing a transpilation pass, the compiler can check this map and may bypass the pass for that script if the specific feature targeted by the pass is not present. This feature map can also be dynamically updated throughout the compilation process as other passes transform the code. This method of conditional pass execution based on content-aware analysis may reduce redundant AST traversals, which could decrease overall compilation time and computational resource consumption.
View details
The Synthetic Gap: Automating Forensic Investigation of "AI Slop" with the Scaled Abuse Forensics Examiner (SAFE)
Vahid Jalali
Longling Wang
Geethik Narayana Kamineni
Utkarsh Chaudhary
Crystal Zhao
Lucas Liu
2026
Preview abstract
Generative AI capabilities have enabled malicious actors to flood online platforms with "AI slop"—mass-produced, low-quality synthetic media designed to overwhelm traditional integrity systems. These adversarial campaigns often utilize coordinated networks to distribute unique, localized variations of synthetic content, rendering static detection methods ineffective. The signals to detect coordination often have recall gaps. The content is not exactly duplicative to be in the same repetitive video cluster. The abusers however show similar patterns of behavior which need forensics. Manual forensic investigations cannot scale to match the velocity of these generative attacks.
To address this, we present SAFE (Scaled Abuse Forensics Examiner), an automated multi-agent architecture designed for the scalable forensics of adversarial synthetic media. The system decomposes the investigation process into specialized agents: a Cluster Understanding Agent specialized in analyzing the relations between channels in a cluster, a Behavior Understanding Agent that identifies inorganic spatiotemporal patterns, and a Content Understanding Agent that utilizes LoRA-adapted Large Language Models (LLMs) and few-shot learning to detect existing policy violations and spirit of the policy violations respectively . A Root Agent synthesizes these multimodal signals to render a final verdict. Early deployment results indicate that SAFE significantly accelerates the identification of novel synthetic threats, reducing forensic investigation time compared to human-in-the-loop workflows.
View details
Gaze Target Estimation Anywhere with Concepts
Xu Cao
Houze Yang
Vipin Gunda
Inki Kim
Jim Rehg
Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2026)
Preview abstract
Estimating human gaze targets in-the-wild is a formidable challenge. Existing computer vision algorithms rely on brittle, multi-stage pipelines that require explicit inputs like head bounding boxes and human pose, causing initial detection errors to cascade and lead to system failure. To overcome this, we introduce the \textbf{Promptable Gaze Target Estimation (PGE)} task, a new end-to-end, concept-driven paradigm. PGE conditions gaze prediction on flexible user text or visual prompts (e.g., "the boy in the red shirt" or "person in point [0.52, 0.48]") to identify a specific subject's target, which eliminates the rigid dependency on intermediate localization cues. We develop a scalable data engine to generate \textbf{Gaze-Co}, a dataset and benchmark of 120K high-quality, prompt-annotated image pairs. We also propose \textbf{AnyGaze}, the first model designed for PGE. AnyGaze uses a Transformer-based detector to fuse features from frozen encoders and simultaneously solves subject localization, in/out-of-frame presence, and gaze target heatmap estimation. AnyGaze achieves state-of-the-art performance on standard gaze target estimation benchmarks, setting a strong baseline for this new problem even on a difficult out-of-domain, real-world clinical dataset. We will open-source the AnyGaze model and the Gaze-Co benchmark.
View details
What’s on My Network? Using Large Language Models to Identify Real-World IoT Devices at Scale
Rameen Mahmood
Danny Yuxing Huang
Proceedings of ACM International Conference on Emerging Networking Experiments and Technologies (CoNEXT), Association for Computing Machinery (2026)
Preview abstract
The Abkhaz-Adyghe and Nakh-Daghestanian language families encompass 35 living languages that possess arguably the most complex modern Cyrillic orthographies due to their very sophisticated phonology. The relevant online data displays idiosyncratic patterns among which the use of confusable characters in input methods is the most prevalent. This work studies one such character---letter \emph{palochka}---that is shared by most writing systems in question. We investigate whether the patterns including variants of this character alone can act as language data markers when mining these languages in a large-scale web-crawled data. Using a wide-coverage off-the-shelf LID model (GlotLID) we further investigate the data mined using such patterns and estimate the effects of confusable character normalization on quality of paragraph-level LID predictions in 14 supported languages. According to GlotLID, the normalization significantly increases the recall (discovery of new language data) for some languages while degrading it for others. However, manual evaluation reveals that only 41\% of wins and 46\% of losses are accurate due to GlotLID prediction errors. We argue that despite finding useful signal higher precision LID approaches tailored to these long-tail languages are needed to improve the quality of mined data.
View details
Mind the Gap: Structure-Aware Consistency in Preference Learning
Proceedings of the 43rd International Conference on Machine Learning (ICML 2026)
Preview abstract
Aligning Large Language Models (LLMs) with human intent, whether through explicit reward modeling or direct methods such as DPO, fundamentally relies on minimizing a surrogate loss as a proxy for the true pairwise ranking objective. We prove that this reliance is flawed for the standard surrogate losses used: for the equicontinuous hypothesis sets characteristic of neural networks, no standard surrogate provides a meaningful consistency guarantee. Minimizing the surrogate loss to zero can leave the true ranking error arbitrarily high. To resolve this, we formulate LLM alignment within a margin-shifted ranking framework and derive $H$-consistency bounds showing that enforcing a confidence margin $\gamma$ is not merely beneficial but necessary for consistency. We further introduce Structure-Aware $H$-consistency and a corresponding objective (SA-DPO) that adapts the margin to the semantic distance between responses, preventing instability on near-synonymous pairs. Finally, we analyze the trade-off between the margin required for consistency and the model's finite capacity to satisfy it, revealing a strict hierarchy of surrogate losses: heavy-tailed surrogates (e.g., the Polynomial Hinge family) offer strictly superior consistency guarantees for capacity-bounded models compared to the logistic loss used in DPO. Experiments on UltraFeedback and Argilla DPO-Mix-7k confirm that SA-DPO consistently outperforms DPO and SimPO, with a 58.5% head-to-head win-rate in downstream generation quality.
View details
In Search of "Weird Corners": Diagnosing the Limits of Convergent AI in Professional Creative Practice
Mark Williams
2026
Preview abstract
Generative AI assistants typically employ convergent interaction paradigms to resolve ambiguity. While effective for technical tasks, this risks premature convergence in creative domains, constraining output variance. Evaluating a convergent AI probe with expert creatives (N=9) indicates an interactional paradox: structural linearity provides "ignition" utility for early ideation, but misaligns with organic workflows, often inducing "aesthetic sanitization" that standardizes individualized nuance. Prioritizing constructive friction over default agreement, the experts requested active, lateral collaborators. In response, we reframe output convergence as a "full-stack" UI challenge, advocating for Generative frameworks that operationalize the Double Diamond via fluid role-shifting and productive tension.
View details
Mining Attribute Subspaces for Efficient Fine-tuning of 3D Foundation Models
Yu Jiang
Hanwen Jiang
Vincent Chu
Brandon Y. Feng
Zhangyang Wang
Qixing Huang
IEEE/CVF Conference on Computer Vision and Pattern Recognition (2026)
Preview abstract
With the emergence of 3D foundation models, there is growing interest in fine-tuning them for downstream tasks, where LoRA is the dominant fine-tuning paradigm. As 3D datasets exhibit distinct variations in texture, geometry, camera motion, and lighting, there are interesting fundamental questions: 1) Are there LoRA subspaces associated with each type of variation? 2) Are these subspaces disentangled (i.e., orthogonal to each other)? 3) How do we compute them effectively? This paper provides answers to all these questions. We introduce a robust approach that generates synthetic datasets with controlled variations, fine-tunes a LoRA adapter on each dataset, and extracts a LoRA sub-space associated with each type of variation. We show that these subspaces are approximately disentangled. Integrating them leads to a reduced LoRA subspace that enables efficient LoRA fine-tuning with improved prediction accuracy for downstream tasks. In particular, we show that such a reduced LoRA subspace, despite being derived entirely from synthetic data, generalizes to real datasets. An ablation study validates the effectiveness of the choices in our approach.
View details
