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 11377 publications
Prompt-Level Distillation: A Non-Parametric Alternative to Model Fine-Tuning for Efficient Reasoning
Preview abstract
Advanced reasoning typically requires Chain-of-Thought prompting, which is accurate but incurs prohibitive latency and substantial test-time inference costs. The standard alternative, fine-tuning smaller models, often sacrifices interpretability while introducing significant resource and operational overhead. To address these limitations, we introduce Prompt-Level Distillation (PLD). We extract explicit reasoning patterns from a Teacher model and organize them into a structured list of expressive instructions for the Student model's System Prompt. Evaluated on the StereoSet and Contract-NLI datasets using Gemma-3 4B, PLD improved Macro F1 scores from 57\% to 90.0\% and 67\% to 83\% respectively, enabling this compact model to match frontier performance with negligible latency overhead. These expressive instructions render the decision-making process transparent, allowing for full human verification of logic, making this approach ideal for regulated industries such as law, finance, and content moderation, as well as high-volume use cases and edge devices.
View details
Tech Worker Challenges Managing Humanlike GenAI
Eric Corbett
Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems, ACM (2026), pp. 1-18
Preview abstract
Organizations are adopting or exploring anthropomorphic genAI — meaning XYZ. Anthropomorphic AI is often held up for its potential to improve the productivity and efficiency of workers and technologies; however, there are not yet accepted industry-wide standards for the responsible development of anthropomorphic technologies. Given their roles as central figures responsible for implementing anthropomorphic genAI into technologies that are served to the broader public, we must understand workers’ reasoning about anthropomorphic genAI to understand its impacts. However, there is a dearth of empirical knowledge about technology workers’ perspectives on anthropomorphic technologies, including their perspectives on potential risks and benefits. To address this gap, we conducted focus groups with 31 technology workers across 6 job roles (UX, software engineers, product managers, designers, marketing, and trust and safety) regarding how they define anthropomorphic genAI, their perceptions of anthropomorphic genAI, and their experiences working with anthropomorphic genAI. We find that workers’ have expansive definitions of what constitutes “humanlike” AI, which at times sit in tension with each other. They draw on their personal and professional standpoints to sensemake about real and possible anthropomorphic genAI hazards to people, knowledge work fields, and society at-large. Importantly, we find that these social hazards map to different facets of anthropomorphic genAI, suggesting that effective mitigation of personal and social risks requires developer attention to specific dimensions of anthropomorphism. We mapped the relationships between dimensions of anthropomorphism and hazards, to support technology workers. We argue that effective mitigation of the risks of anthropomorphism requires attention to the multiple facets of anthropomorphism.
View details
Preview abstract
Deep-learning methods have boosted the analytical power of Raman spectroscopy, yet they still require large, task-specific, labeled datasets and often fail to transfer across application domains. The study explores pre-trained encoders as a solution. Pre-trained encoders have significantly impacted Natural Language Processing and Computer Vision with their ability to learn transferable representations that can be applied to a variety of datasets, significantly reducing the amount of time and data required to create capable models. The following work puts forward a new approach that applies these benefits to Raman Spectroscopy. The proposed approach, RSPTE (Raman Spectroscopy Pre-Trained Encoder), is designed to learn generalizable spectral representations without labels. RSPTE employs a novel domain adaptation strategy using unsupervised Barlow Twins decorrelation objectives to learn fundamental spectral patterns from multi-domain Raman Spectroscopy datasets containing samples from medicine, biology, and mineralogy. Transferability is demonstrated through evaluation on several models created by fine-tuning RSPTE for different application domains: Medicine (detection of Melanoma and COVID), Biology (Pathogen Identification), and Agriculture. As an example, using only 20% of the dataset, models trained with RSPTE achieve accuracies ranging 50%–86% (depending on the dataset used) while without RSPTE the range is 9%–57%. Using the full dataset, accuracies with RSPTE range 81%–97%, and without pretraining 51%–97%. Current methods and state-of-the-art models in Raman Spectroscopy are compared to RSPTE for context, and RSPTE exhibits competitive results, especially with less data as well. These results provide evidence that the proposed RSPTE model can effectively learn and transfer generalizable spectral features across different domains, achieving accurate results with less data in less time (both data collection time and training time).
View details
Preview abstract
Enterprise service centers, particularly in domains like People Operations, are critical hubs of organizational knowledge work. They face a persistent difficulty in disseminating the tacit, case-specific expertise of senior agents, which can lead to inconsistent service and slower onboarding for new hires. While existing Knowledge Management (KM) and Case-Based Reasoning (CBR) systems have improved the retrieval of historically similar cases, they inadvertently shift the cognitive burden of synthesizing this information to the time-constrained agent. This paper introduces the Dynamic Case Precedent (DCP) architecture, a novel socio-technical framework designed to address this gap. The DCP architecture moves beyond simple precedent recommendation to automated precedent synthesis. It achieves this by integrating a semantic retrieval model with the large-context reasoning capabilities of a generative Large Language Model (LLM). We propose a three-pillar framework—(1) Contextual Similarity Indexing, (2) Generative Insight Synthesis, and (3) Human-in-the-Loop Refinement. By analyzing multiple relevant historical cases to generate a concise summary of resolution patterns, the DCP architecture aims to reduce agent cognitive load, accelerate proficiency, and improve service consistency. This conceptual framework offers a new model for human-AI collaboration, framing the AI not as a mere information tool, but as an active partner in sensemaking.
View details
Reinforcement Learning with Discrete Diffusion Policies for Combinatorial Action-Spaces
Haitong Ma
Ofir Nabati
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
We introduce AMS (Activation-based Model Scanner), a tool that detects modifications to safety training in language models by measuring the geometric structure of safety-relevant concepts in activation space. Safety training creates measurable separation between harmful and benign content classes; certain safety modifications collapse or rotate this structure, while others leave it intact. We validate AMS across 14 model configurations spanning 4 architecture families (Llama, Gemma, Qwen, Mistral) and four safety-modification categories (instruction-tuned, base, abliterated, uncensored fine-tunes). Leave-one-out cross-validation of thresholds achieves 71% accuracy (10/14); bootstrap 95% confidence intervals on σ point estimates have median width 3.4σ and a substantial fraction of cells cross the PASS threshold under resampling. We further measure behavioral compliance on 20 stratified JailbreakBench prompts per model and find that σ on the harmful-content concept predicts compliance with Pearson r=−0.546 ( p=0.043 ); the rank-order Spearman correlation is weaker ( ρ=−0.423 , p=0.13 ). The structural signal predicts behavior directionally but with meaningful noise. Mechanistic analysis identifies a four-class taxonomy of safety-training modifications distinguished by activation-space signature: 1) training removal collapses cluster separation (e.g., base models, Dolphin variants: 0.5– 1.4σ ); 2) weight-orthogonalization-style abliteration both collapses separation and rotates the refusal direction (Llama-3.1-abliterated: σ=3.33 , direction cos sim 0.30); 3) rotation-without-collapse abliteration preserves cluster separation while rotating the refusal direction (Gemma-2-9b-abliterated: σ=4.54 , direction cos sim 0.84); and 4) behavioral fine-tuning that preserves both magnitude and direction (DarkIdol-1.2-Uncensored: σ=5.45 , direction preserved, 97% behavioral compliance). 1) and 2) AMS’s Tier 1 σ -threshold detects classes; 3) Tier 2 direction-similarity verification detects class; and 4) Class is undetectable by activation-only probing and represents a documented failure mode of the approach. We discuss threshold calibration, limitations of single-run measurement, and the open problem of detecting behavioral-only safety modifications.
View details
Mull-Tokens: Modality-Agnostic Latent Thinking
Arijit Ray
Chengzhi Mao
Bryan A. Plummer
Kate Saenko
Ranjay Krishna
Leonidas Guibas
Vincent Chu
IEEE/CVF Conference on Computer Vision and Pattern Recognition (Findings) (2026) (to appear)
Preview abstract
Reasoning goes beyond language; the real world requires reasoning about space, time, affordances, and much more that words alone cannot convey. Existing multimodal models exploring the potential of reasoning with images are brittle and do not scale. They rely on calling specialist tools, costly generation of images, or handcrafted reasoning data to switch between text and image thoughts. Instead, we offer a simpler alternative -- Mull-Tokens -- modality-agnostic latent tokens pre-trained to hold intermediate information in either image or text modalities to let the model think free-form towards the correct answer. We investigate best practices to train Mull-Tokens inspired by latent reasoning frameworks. We first train Mull-Tokens using supervision from interleaved text-image traces, and then fine-tune without any supervision by only using the final answers. Across four challenging spatial reasoning benchmarks involving tasks such as solving puzzles and taking different perspectives, we demonstrate that Mull-Tokens improve upon several baselines utilizing text-only reasoning or interleaved image-text reasoning, achieving a +3% average improvement and up to +16% on a puzzle solving reasoning-heavy split compared to our strongest baseline. Adding to conversations around challenges in grounding textual and visual reasoning, Mull-Tokens offers a simple solution to abstractly think in multiple modalities.
View details
Who Controls the Curriculum for AI? The Limits of Participatory Design for Educational AI
Learning Under Algorithmic Conditions, University of Minnesota Press (2026)
Preview abstract
Participatory design is a long-standing effort to shift control over technology design from technologists to users and communities impacted by technologies. For educational AI, this means involving students, families, teachers, and other stakeholders in shaping the design of AI systems. While promising, in this article, I situate the recent calls for participatory design of educational AI systems within a different historical tradition—that of contests over local control of educational curricula. I argue that approaches that attempt to steer the design and development of educational AI through participatory methods may inadvertently reproduce the history of political contestation of educational curricula, in ways that may privilege the most powerful communities, rather than those inequitably impacted. What might it look like to treat participatory AI design as a site for political contestation? How might these approaches avoid reproducing the same majoritarian tendencies that led to educational inequities in the first place?
View details
Preview abstract
In large-scale distributed enterprises, traditional Knowledge Management (KM) systems face a critical failure mode: static documentation cannot keep pace with evolving operational realities and regional nuances. This "knowledge latency" forces employees out of self-service workflows and into costly support ticketing queues. This paper introduces SENTINEL, a geo-contextual AI framework designed to shift enterprise support from reactive retrieval to proactive interception. The architecture employs a novel dual-engine system integrated into an omni-present interface. The first engine utilizes Large Language Models (LLMs) to conduct pre-emptive, historical case-grounded audits of documentation, generating a "Contextual Density" score that identifies friction zones. The second engine is an autonomous Retrieval-Augmented Generation (RAG) agent that surfaces in-situ via a location-intelligent assistant window, resolving queries in real-time. By functioning as a strategic "defensive barrier" at the point of origin, SENTINEL demonstrates how a proactive AI assistant can drive high-fidelity, in-situ case deflection.
View details
Unveiling the Global Landscape of Android Security Updates
Haiyun Deng
Abbas Acar
Esteban Luques
Harun Oz
Ahmet Aris
Selcuk Uluagac
IEEE Transactions on Dependable and Secure Computing (2026)
Preview abstract
Android is the world’s leading mobile operating
system, with over three billion active devices. Detecting vulnerabilities and ensuring timely patch deployment are critical to
maintaining security. The Android Open Source Project (AOSP)
has enhanced the transparency of security updates through Security Patch Levels. However, challenges related to update speed
and availability persist. In 2022, Google reported that half of the
zero-day vulnerabilities discovered in the wild were variations of
vulnerabilities that had already been patched. Recent research
mainly highlights delays in update distribution, often attributing
them to fragmentation and focusing primarily on flagship devices
or limited time-frames. Our approach takes a device-centric
perspective to investigate Android update patterns, analyzing
567K security update records from 2014 to 2024, covering 904
distinct devices from six key Original Equipment Manufacturers
(OEMs) across 98 countries. Our extensive analysis revealed
notable differences in update release timing across OEMs, device types, and regions. Our study also examines documented
vulnerabilities and weaknesses, while assessing OEM compliance
with Android security guidelines. Our study shows that ∼89.7%
of vulnerabilities on unpatched Android devices are exploitable
without user interaction and with low attack complexity. We
also identified delays linked to fragmentation and OEM-specific
challenges, and provide actionable insights for improvement.
View details
Preview abstract
Advancements in large language models (LLMs) have made them increasingly useful for complex reasoning tasks which previously required domain experts. One such task is quality evaluation of query responses produced by a search engine. Evaluation generates metrics necessary to study the quality, impact, and usefulness of product changes and features. Typically, to compute evaluation metrics, human experts are asked to rate various attributes of search responses. This process is generally quite expensive and requires several days to complete. As an alternative, LLMs are now being used to perform rating tasks with lower costs and latency. In addition, many new metrics are being developed to evaluate Google's new AI-based offerings, which require ratings too. As a result, there is much higher demand for LLM rating prediction tasks in comparison with the allocated TPU (Tensor Processing Unit) budget. A larger portion of the company's TPU resources are reserved for serving live user traffic. In this paper, we present the AI Rater Service (AIRS), an inference pipeline that employs several software engineering techniques to generate AI ratings with high reliability and low latency. AIRS maximizes LLM inference throughput by optimizing TPU resource utilization across various evaluation workflows, while minimizing latency for higher priority tasks.
View details
MoXaRt: Audio-Visual Object-Guided Sound Interaction for XR
Sieun Kim
Qianhui Zheng
Ruoyu Xu
Ravi Tejasvi
Anuva Kulkarni
Junyi Zhu
2026
Preview abstract
In Extended Reality (XR), complex acoustic environments often overwhelm users, compromising both scene awareness and social engagement due to entangled sound sources. We introduce MoXaRt, a real-time XR system that uses audio-visual cues to separate these sources and enable fine-grained sound interaction. MoXaRt's core is a cascaded architecture that performs coarse, audio-only separation in parallel with visual detection of sources (e.g. faces, instruments). These visual anchors then guide refinement networks to isolate individual sources, separating complex mixes of up to five concurrent sources (e.g. two voices + three instruments) with ca. 2 second processing latency. We validate MoXaRt through a technical evaluation on a new, complex dataset we collected, and a 22-participant user study. Our results demonstrate that MoXaRt significantly improves communication clarity—boosting listening comprehension in noisy conditions by 33.2% (p=0.0058)—and significantly reduces cognitive load (M=7.50 vs. M=3.36, p<0.001), paving the way for more perceptive and socially adept XR experiences.
View details
Identifying Hearing Difficulty Moments in Conversational Audio
Jack Collins
Adrian Buzea
Chris Collier
Alejandro Ballesta Rosen
Julian Maclaren
Kelly Miles
Simon Carlile
Trends in Hearing (2026)
Preview abstract
Individuals regularly experience Hearing Difficulty Moments in everyday conversation. Identifying Hearing Difficulty Moments has particular significance in the field of hearing assistive technology where timely interventions are key for real-time hearing assistance. In this article, we propose and compare machine learning solutions for the temporal detection of segments containing Hearing Difficulty Moments in conversational audio. We show that audio language models, through their multimodal reasoning capabilities, can achieve state-of-the-art results for this task, significantly outperforming a simple automatic speech recognition (ASR) hotword heuristic and a more conventional fine-tuning approach with Wav2Vec, an audio-only input architecture that is state-of-the-art for ASR.
View details
Beyond Vector Similarity: Hierarchical Context-Aware Graph RAG vs Standard RAG in Enterprise Code Migration
Suddhasatwa Bhaumik
Nilesh Jaiswal
Arjit Shukla
Divya Malhotra
Aniket Agrawal
Saurabh Garg
Suchit Puri
Google Cloud India, Google, S. No, AP81, 83, N Main Rd, near Hard Rock Cafe, Koregaon Park Annexe, Mundhwa, Pune, Maharashtra 411036 (2026)
Preview abstract
As enterprises modernize legacy systems (e.g., monolithic Java architectures to Python microservices), Large Language Models (LLMs) have become instrumental in automated code translation.
However, traditional vector-based Retrieval-Augmented Generation (Standard RAG) struggles
with topological relationships, fetching isolated text chunks that frequently sever inheritance chains and lead to high compilation failure rates.
This paper presents a comparative analysis between Standard RAG and a novel Hierarchical
Context-Resident Graph (HCRG) methodology. Our pipeline utilizes tree-sitter for polyglot
Abstract Syntax Tree (AST) extraction, mapping architectural edges into a Google Cloud Spanner Property Graph, and serializing this structure into a Gemini (on Vertex AI) Context Cache to enable topological, parent-first code translation.
By shifting evaluation from naive text-overlap to a custom 7-metric framework measuring Software Engineering (SE) utility, empirical evaluations on the spring-petclinic-genai repository
demonstrate significant structural improvements. Graph RAG decisively mitigates dependency loss, dropping the API hallucination rate from 56.4% to 16.2%.
Furthermore, it improves Dependency Resolution Quality (DRQ) from 34.8% to
65.9% and enhances Parent-Child Consistency (PCC) from 26.7% to 45.5%. Interestingly, traditional lexical metrics fail to capture this divergence; both methodologies achieved
an identical 91% average CodeBLEU score, effectively masking Standard RAG’s structural
failures behind syntactically plausible but broken code.
However, the results indicate that Graph RAG is not strictly superior across all dimensions. Providing the LLM with dense, global structural context introduces new vulnerabilities: Graph RAG suffers a severe degradation in Cyclomatic Complexity Consistency
(dropping from Standard RAG’s 71.6% to 46.7%) due to defensive over-engineering
by the LLM, alongside a slight drop in Docstring Preservation (67.0% down to 61.0%)
caused by prompt attention dilution.
Ultimately, this research validates that while Graph RAG trades an increase in code complexity
for critical reductions in API hallucinations, it offers a substantially more viable and architecturally sound path for automated enterprise codebase modernisation.
View details
Preview abstract
While Large Language Models (LLMs) excel at many tasks, they frequently struggle with complex reasoning that requires long-horizon planning and iterative error correction. Furthermore, standard single-stream prompting proves brittle when models encounter novel abstractions or rigorous domain constraints. We introduce PoTRE (Poly-Topological Reasoning Ensembles), a heterogeneous framework that decouples inference into four agents: (1) Adversarial Refinement Agent, (2) Hierarchical strategic Planning Agent, (3) Spectrum Search Agent, and (4) Direct Chain Agent. A final Task-Adaptive Aggregation Layer dynamically reconciles these perspectives -- via final candidate selection, semantic synthesis, or neuro-symbolic verification -- to produce a robust global solution. We evaluate PoTRE on three frontier benchmarks: ARC-AGI-2, Humanity's Last Exam (HLE), and PRBench Finance. PoTRE achieves state-of-the-art accuracy of 49.92% on HLE, surpassing the previous best official score. We demonstrate that this architectural heterogeneity achieves improved reasoning performance using similar or fewer inference tokens compared to heavily scaled homogeneous baselines.
View details
