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 11415 publications
Preview abstract
We introduce AASE (Activation-based AI Safety Enforcement), a framework for post-perception safety monitoring in large language models. Unlike pre-perception approaches that analyze input or output text, AASE monitors the model's internal activation patterns—what the model "understands" rather than what text it processes or generates—enabling detection of safety-relevant states before harmful outputs are produced. The framework comprises three techniques: Activation Fingerprinting (AF) for harmful content detection, Agent Action Gating (AAG) for prompt injection defense, and Activation Policy Compliance (APC) for enterprise policy enforcement. We introduce paired contrastive training to isolate safety-relevant signals from confounding factors such as topic and style, addressing signal entanglement in polysemantic activations. Validation across 7 models from 3 architecture families shows strong class separation: Gemma-2-9B achieves AUC 1.00 with 7.2σ separation across all probes; AAG achieves AUC ≥0.88 across all models on the InjecAgent benchmark; APC achieves 0.97-1.00 AUC across three enterprise policies. Model size correlates with probe quality—Gemma-2-9B (7.2σ separation) outperforms Gemma-2-2B (4.3σ). All techniques survive INT4 quantization with minimal separation degradation. AASE is 9× faster than Llama Guard 3 (33ms vs 306ms) with higher TPR (88% vs 50%) at a tunable threshold that trades FPR for detection sensitivity, adding only 0.002ms probe overhead to existing inference.
View details
Preview abstract
While the Latin script is used informally by speakers of many languages with more complex native scripts, high quality Latin script corpora for such languages that reflect actual natural romanizations are scarce and often difficult to collect. In this work, we propose a method for mining romanized language corpora in languages for which we do not have any pre-existing samples of naturally romanized text, focusing on Tigrinya as a test case. First we examine the efficacy of learning romanizations for a language based on observed romanizations in other languages that use the same native script. We then extrinsically assess such methods by using a romanization model trained on Amharic data to bootstrap coverage of romanized Tigrinya in a language identification system. Manual evaluation by two L1 and one L2 Tigrinya speakers suggests our method extracts romanized Tigrinya text with acceptably high precision. We release code to run our mining pipeline on public web corpora, such as MADLAD-400.
View details
Preview abstract
Generative AI (GenAI) is evolving from standalone tools to interconnected ecosystems that integrate chatbots, cloud platforms, and third-party services. While this ecosystem model enables personalization and extended services, it also introduces complex information flows and amplifies privacy risks. Existing solutions focus on system-level protections, offering little support for users to make meaningful privacy choices. To address this gap, we conducted two vignette-based survey studies with 486 participants and a followup interview study with 16 participants. We also explored users’ needs and preferences for privacy choice design across both GenAI personalization and data-sharing. Our results reveal paradoxical patterns: participants sometimes trusted third-party ecosystems more for personalization but perceived greater control in first-party ecosystems when data was shared externally. We discuss design implications for privacy choice interfaces that enhance transparency, control, and trust in GenAI ecosystems.
View details
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
Preview abstract
Modern user interfaces are complex composites, with elements originating from various sources, such as the operating system, apps, a web browser, or websites. Many security and privacy models implicitly depend on users correctly identifying an element's source, a concept we term ''surface attribution.'' Through two large-scale vignette-based surveys (N=4,400 and N=3,057), we present the first empirical measurement of this ability.
We find that users struggle, correctly attributing UI source only 55% of the time on desktop and 53% on mobile. Familiarity and strong brand cues significantly improve accuracy, whereas UI positioning, a long-held security design concept especially for browsers, has minimal impact. Furthermore, simply adding a ''Security & Privacy'' brand cue to Android permission prompts failed to improve attribution. These findings demonstrate a fundamental gap in users' mental models, indicating that relying on them to distinguish trusted UI is a fragile security paradigm.
View details
ToolGrad: Efficient Tool-use Dataset Generation with Textual "Gradients"
Kohei Uehara
Haoyu Zhang
Lin Gu
Zheng Xu
Tatsuya Harada
ACL 2026 (2026)
Preview abstract
Prior work synthesizes tool-use LLM datasets by first generating a user query, followed by complex tool-use annotations like depth-first search (DFS). This leads to inevitable annotation failures and low efficiency in data generation. We introduce ToolGrad, an agentic framework that inverts this paradigm. ToolGrad first constructs valid tool-use chains through an iterative process guided by textual "gradients", and then synthesizes corresponding user queries. This "answer-first" approach led to ToolGrad-500, a dataset generated with more complex tool use, lower cost, and almost 100% pass rate. Experiments show that ToolGrad models outperform those trained on expensive baseline datasets and proprietary LLMs.
View details
Preview abstract
Audio Description ( AD) provides essential access to visual media for blind and low vision ( BLV) audiences. Yet current AD production tools remain largely inaccessible to BLV video creators, who possess valuable expertise but face barriers due to visually- driven interfaces. We present ADCanvas, a multimodal authoring system that supports non- visual control
over audio description ( AD) creation. ADCanvas combines conversational interaction with keyboard- based playback control and a plain- text, screen reader–
accessible editor to support end- to- end AD authoring and visual question answering ( VQA). Combining screen- reader- friendly controls with a multimodal
LLM agent, ADCanvas supports live VQA, script generation, and AD modification. Through a user study with 12 BLV video creators, we find that users adopt
the conversational agent as an informational aide and drafting assistant, while maintaining agency through verification and editing. For example, participants
saw themselves as curators who received information from the model and filtered it down for their audience. Our findings offer design implications for
accessible media tools, including precise editing controls, accessibility support for creative ideation, and configurable rules for human- AI collaboration.
View details
Preview abstract
The fundamental theorem of statistical learning establishes that binary PAC learning is governed by a single parameter---the Vapnik-Chervonenkis ($\mathtt{VC}$) dimension---which controls both learnability and sample complexity. Extending this characterization to multiclass classification has long been challenging, since the early work of Natarajan in the late 80's that proposed the Natarajan dimension ($\mathtt{Nat}$) as a natural analogue of the VC dimension.
Daniely and Shalev-Shwartz (2014) introduced the $\mathtt{DS}$ dimension, later shown by Brukhim et al.\ (2022) to characterize multiclass \emph{learnability}.
Brukhim et al.\ (2022) also demonstrated that the Natarajan and $\mathtt{DS}$ dimensions can diverge arbitrarily, so that multiclass learning appears to be governed by $\mathtt{DS}$ rather than $\mathtt{Nat}$.
We show that the agnostic multiclass PAC sample complexity is in fact governed by \emph{two distinct dimensions}. Specifically, we prove nearly tight agnostic sample complexity bounds that, up to logarithmic factors, take the form
$$
\frac{\mathtt{DS}^{1.5}}{\epsilon} + \frac{\mathtt{Nat}}{\epsilon^2}
$$
where $\epsilon$ is the excess risk. This bound is tight up to a $\sqrt{\mathtt{DS}}$ factor in the first lower-order term, nearly matching known $\mathtt{Nat}/\epsilon^2$ and $\mathtt{DS}/\epsilon$ lower bounds.
The first term reflects the DS-controlled regime, while the second reveals that the Natarajan dimension still dictates asymptotic behavior for small $\epsilon$. Thus, unlike in binary or online classification---where a single dimension (VC or Littlestone) controls both phenomena---multiclass learning inherently involves \emph{two structural parameters}.
Our technical approach departs significantly from traditional agnostic learning methods based on uniform convergence or reductions-to-realizable techniques. A key ingredient is a novel online procedure, based on a self-adaptive multiplicative-weights algorithm which performs a label-space reduction. This approach may be of independent interest and find further applications.
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
How many T gates are needed to approximate an arbitrary n-qubit quantum state to within
a given precision ϵ? Improving prior work of Low, Kliuchnikov and Schaeffer, we show that the
optimal asymptotic scaling is Θ(sqrt{2^n log(1/ε)} + log(1/ε)) if we allow an unlimited number of ancilla qubits. We also show that this is the optimal T-count for implementing an arbitrary
diagonal n-qubit unitary to within error ϵ. We describe an application to batched synthesis of
single-qubit unitaries: we can approximate a tensor product of m = O(log log(1/ϵ)) arbitrary
single-qubit unitaries to within error ϵ with the same asymptotic T-count as is required to
approximate just one single-qubit unitary.
View details
Preview abstract
This paper demonstrates that artificial intelligence can accelerate mathematical discovery by autonomously solving an open problem in theoretical physics. We present a neuro-symbolic system, combining the Gemini Deep Think large language model with a systematic Tree Search (TS) framework and automated numerical feedback, that successfully derived novel, exact analytical solutions for the power spectrum of gravitational radiation emitted by cosmic strings. Specifically, the agent evaluated the core integral for arbitrary loop geometries, directly improving upon recent AI-assisted attempts that only yielded partial asymptotic solutions. To substantiate our methodological claims regarding AI-accelerated discovery and to ensure transparency, we detail system prompts, search constraints, and intermittent feedback loops that guided the model. The agent identified a suite of 6 different analytical methods, the most elegant of which expands the kernel in Gegenbauer polynomials to naturally absorb the integrand's singularities. The methods lead to an asymptotic result for at large that both agrees with numerical results and also connects to the continuous Feynman parameterization of Quantum Field Theory. We detail both the algorithmic methodology that enabled this discovery and the resulting mathematical derivations.
View details
System for a Secure, Outcome-Based Synthetic Labor Market Using Trusted Execution Environments
Patent (2026)
Preview abstract
Some artificial intelligence provisioning models that function as tools for human users or rely on labor arbitrage can present challenges for organizations, such as managing personnel rather than task outcomes and introducing data security risks. An architecture is described for an outcome-based synthetic labor market in which autonomous computational agents can be compensated based on verified task completion. The framework can leverage trusted execution environments to create secure hardware enclaves for processing sensitive data, which can render the data cryptographically inaccessible to a host system or agent provider. This approach can facilitate a secure, transactional market for autonomous professional execution, which may enable a shift from managing labor resources to procuring verified outcomes from a pool of specialized agents.
View details
Preview abstract
Autonomous research agents can now produce competitive solutions and complete manuscripts, yet their papers routinely contain fabricated citations, method descriptions disconnected from the code, and scores on incorrect scales---failures invisible to evaluations that assess fluency rather than evidentiary grounding. The core problem is verifiability: no existing system maintains a traceable chain from each claim in the paper to its grounding evidence, and current evaluation protocols assess output fluency rather than evidentiary grounding. We address this with Chaine-of-Evidence (CoE), a verifiability standard requiring every claim to trace to its grounding evidence, and instantiate it in Scientist One, an end-to-end research system that maintains evidence chains natively, and CoE Audit, an evaluation protocol with four integrity checks targeting the most damaging chain failures. Auditing 60 papers from four systems, we find every baseline exhibits at least one failure: phantom citations at 4--25%, method-code alignment in at most 2/15 papers. Scientist One achieves zero phantom citations (0/830), the highest alignment rate (7/15), and competitive solver scores.
View details
Learning Conditional Averages
Marco Bressan
Nataly Brukhim
Nicolo Cesa-Bianchi
Emmanuel Esposito
Shay Moran
Maximilian Thiessen
COLT (2026)
Preview abstract
We introduce the problem of learning \emph{conditional averages} in the PAC framework.
The learner receives a sample labeled by an unknown target concept from a known concept class, as in standard PAC learning. However, instead of learning the target concept itself, the goal is to predict, for each instance, the average label over its \emph{neighborhood}---an arbitrary subset of points that contains the instance.
In the degenerate case where all neighborhoods are singletons, the problem reduces exactly to classic PAC learning. More generally, it extends PAC learning to a setting that captures learning tasks arising in several domains, including explainability, fairness, and recommendation systems. %including explainability, fairness, and recommendation systems.
Our main contribution is a complete characterization of when conditional averages are learnable, together with sample complexity bounds that are tight up to logarithmic factors. The characterization hinges on the joint finiteness of two novel combinatorial parameters, which depend on both the concept class and the neighborhood system, and are closely related to the independence number of the associated neighborhood graph.
View details
Preview abstract
This study examines the psychological and ethical implications of generative-AI chatbot use among youth, introducing the CTRL framework (Cognitive Trust, Reliance, and Learning Diminution) to explain how repeated use fosters cognitive offloading and reduced verification behavior. Survey data from 420 participants analyzed through factor analysis and structural equation modeling reveal that higher trust predicts greater reliance and diminished critical evaluation, alongside elevated concerns around privacy and academic integrity. Findings highlight the need for AI literacy and responsible design to mitigate unintended cognitive impacts.
View details
