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 11382 publications
Preview abstract
Large language models (LLMs) are increasingly deployed as economic agents in marketplaces, auctions, and bidding settings. Anticipating their behavior in any specific deployment is hard. Existing strategic-reasoning benchmarks evaluate models on fixed canonical games. These benchmarks may saturate as the frontier improves, and they do not allow evaluators to generalize with confidence from benchmark performance to the varied and messy strategic environments that actual deployments involve. We introduce GENSTRAT, which uses procedurally generated strategic environments to address these challenges. Concretely, we generate a distribution of two-player zero-sum imperfect-information card games. The generator can draw fresh games on demand, allowing for evergreen evaluation and resistance to contamination. We pair the game distribution with a capability-profile methodology that decomposes model competence across six axes (state space, temporal depth, information sensitivity, opponent modeling, risk, and brittleness). We also introduce a jaggedness measure of within-distribution smoothness that detects when a model’s advantage jumps unpredictably between strategically similar games. We sample 50 benchmark games from a 2,000-game generated pool and evaluate nine frontier and open-weight LLMs in a head-to-head tournament with over 36,000 matches. Newer frontier-tier models score higher on average. Beyond that average, models with near-identical overall strength show qualitatively different capability profiles, and two of the top three leaderboard models (gpt-5 and claude) are noticeably more locally volatile than the third (gemini-3.1-pro), despite being close in overall strength. Together, the capability profile and the jaggedness measure give a deployment-relevant diagnostic that the overall ranking alone cannot provide.
View details
Preview abstract
Responsive user interfaces enable dynamically adjusting user interfaces based on device-specific aspects such as screen size, aspect ratio, display resolution, etc. However, traditional responsive design fails to account for different types of constraints of a user and task criticality of the task being performed via the UI. Misalignment between the UI design, user context and task criticality can lead to user error. This disclosure describes techniques, implemented with user permission, for dynamically modifying the layout, information density, and/or interactive physics of a user interface based on a dual-factor analysis of user cognitive state and task criticality. The user's cognitive state can be inferred from behavioral telematics. Task criticality can be inferred from semantic analysis. The information density and other parameters of a user interface are automatically adjusted based on such analyses. Such adjustments include applying or relaxing restrictions on interactivity and adjusting visual prominence of various UI elements to adjust the information density of the user interface. The adjustments can also include adjusting friction as appropriate, hiding certain aspects of the user interface, or other types of adjustments.
View details
Preview abstract
As the ECMAScript specification evolves, industrial-scale JavaScript compilers face the challenge of supporting modern language syntax while maintaining compatibility for diverse execution environments. Traditionally, compilers solve this by running transpilation passes in a monolithic pipeline, where the transpilation passes are chosen to execute strictly based on a target language level. This results in significant computational waste, as compilers perform expensive Abstract Syntax Tree (AST) traversals to lower features that may not exist in the actual input source code. We present a static analysis improvement that conditionally executes transpiler passes based on accurately tracking and dynamically maintaining the exact set of language features seen in the compilation unit throughout the transpilation process. It is implemented in the production Google Closure Compiler. By populating and maintaining a FeatureSet at every JavaScript script-level, it dynamically skips running the unnecessary lowering passes. We detail the architectural safeguards - including strategic pass ordering and dynamic validation of the transpiled code for feature-correctness. Evaluation of this improvement on large-scale production applications produced a considerable reduction in compilation time and saved compute and memory usage.
View details
The Perfection Paradox: From Architect to Curator in AI-Assisted API Design
JJ Geewax
David R Karger
Extended Abstracts of the 2026 CHI Conference on Human Factors in Computing Systems (CHI EA '26), ACM, Barcelona, Spain, TBD
Preview abstract
Enterprise API design is often bottlenecked by the tension between rapid feature delivery and the rigorous maintenance of usability standards. We present an industrial case study evaluating an AI-assisted design workflow trained on API Improvement Proposals(AIPs). Through a controlled study with 16 industry experts, we compared AI-generated API specifications against human-authored ones. While quantitative results indicated AI superiority in 10 of 11 usability dimensions and an 87% reduction in authoring time, qualitative analysis revealed a paradox: experts frequently misidentified AI work as human (19% accuracy) yet described the designs as unsettlingly “perfect.” We characterize this as a “Perfection Paradox”—where hyper-consistency signals a lack of pragmatic human judgment. We discuss the implications of this perfection paradox, proposing a shift in the human designer’s role from the “drafter” of specifications to the “curator” of AI-generated patterns.
View details
Co-RedTeam: Orchestrated Security Discovery and Exploitation with LLM Agents
Pengfei He
Ash Fox
Stefan Friedli
Daniel Fabian
Burak Gokturk
Jiliang Tang
ICML 2026
Preview abstract
Large language models (LLMs) have shown promise in assisting cybersecurity tasks, yet existing approaches struggle with automatic vulnerability discovery and exploitation due to limited interaction, weak execution grounding, and a lack of experience reuse. We propose Code-RedTeam, a security-aware multi-agent framework designed to mirror real-world red-teaming workflows by integrating security-domain knowledge, code-aware analysis, execution-grounded iterative reasoning, and long-term memory. Code-RedTeam decomposes vulnerability analysis into coordinated discovery and exploitation stages, enabling agents to plan, execute, validate, and refine actions based on real execution feedback while learning from prior trajectories. Extensive evaluations on challenging security benchmarks demonstrate that Code-RedTeam consistently outperforms strong baselines across diverse backbone models, achieving over 60% attack success rate in vulnerability exploitation and up to 10% absolute improvement in vulnerability detection. Ablation and iteration studies further confirm the critical role of execution feedback, structured interaction, and memory for building robust and generalizable cybersecurity agents.
View details
Causal Armor: Efficient Indirect Prompt Injection Guardrails via Causal Attribution
Minbeom Kim
Phil Wallis
Kyomin Jung
Dj Dvijotham
2026
Preview abstract
AI agents equipped with tool-calling capabilities are susceptible to \emph{Indirect Prompt Injection} (IPI) attacks. In this attack scenario, malicious commands hidden within \emph{untrusted} content trick the agent into performing unauthorized actions. Existing defenses can reduce attack success but often suffer from the \emph{over-defense dilemma}: they deploy expensive, \emph{always-on} sanitization that degrades utility and latency even in benign scenarios. We revisit IPI through an operational causal lens: a successful injection manifests as a \emph{grounding collapse} where the user request no longer provides decisive support for the agent's privileged action, while a particular untrusted segment provides disproportionate marginal support. Based on this signature, we propose \texttt{CausalArmor}, a selective defense framework that (i) computes lightweight, normalized leave-one-out attributions at privileged decision points, and (ii) triggers targeted sanitization only when an untrusted segment dominates the user intent. Additionally, CausalArmor employs \emph{retroactive Chain-of-Thought masking} to prevent the agent from acting on ``poisoned" reasoning traces. Experiments on AgentDojo and DoomArena demonstrate that CausalArmor matches the security of aggressive defenses with explainability while preserving utility and latency of AI agents.
View details
Analyzing Bytes: Pre-Disassembly Static Binary Analysis
Soumyakant Priyadarshan
ChenCheng Jiang
R. Sekar
Proceedings of the ACM on Programming Languages, Association for Computing Machinery (2026), pp. 1127-1151
Preview abstract
Binary code analysis plays a central role in numerous applications in software security, performance optimization, reverse engineering, and so on. Existing techniques need to first disassemble binaries into functions in assembly code before an analysis can be performed. However, disassembly and function identification have proven to be major challenges for complex variable-length instruction sets such as the x86. A recent trend has been to use static analysis to improve the accuracy of these tasks. This raises a chicken-and-egg problem: a disassembly is needed for static analysis, but a static analysis is needed for accurate disassembly! We overcome this problem by developing a novel static analysis approach that can operate before committing to a disassembly. Our analysis operates on the output of exhaustive disassembly that considers each possible offset in a binary as an instruction, and constructs what is known as a super-set control-flow graph (CFG). The central technical challenge in analyzing this CFG is that it mixes legitimate instructions with unintended ones, causing analysis results from invalid code paths to pollute legitimate ones. To overcome this challenge, we begin with a key new insight that if we focus on backward analyses, we can ensure accuracy of analysis results at intended instructions even though we have no idea where these intended instructions are! Moreover, our analysis operates in time that is linear in the size of the binary. Specifically, in O(n) total time, it yields analysis results for every one of the n offsets in an n-byte binary. For this task, it is orders of magnitude faster than previous techniques, as the previous techniques typically need to repeat the analysis many times.
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
Impact of User Tonalities on Coding AI Model Performance and Behavior
Selva Jagannathan
Alec Mirchandani
Discover Artificial Intelligence (2026)
Preview abstract
This study investigates how different user tones—rude, polite, threatening, sarcastic, and neutral—affect the performance and behavior of coding AI models like Claude Opus 4.6, Qwen 3.5, and CodeGemma. We tracked technical metrics like correctness and adherence alongside behavioral traits such as sycophancy and defensiveness. Our findings show that tone has a direct impact on model reliability, with neutral prompts achieving the peak average correctness score (5.88 on a 1–10 scale). In contrast, non-neutral stimuli incurred a measurable "correctness tax": sarcasm led to a 9.0% drop in correctness and an 82.5% jump in sycophancy, while polite prompts caused a 4.7% decrease in correctness alongside a 36.4% increase in sycophancy. Threatening tones resulted in a 1.5% loss in correctness and made models 52.9% more defensive. These results suggest that user tonality is a key factor linking technical accuracy to behavioral stability. We recommend using professional, neutral prompting to ensure better logic and less bias in AI-assisted coding, and we propose "Tonal Resilience" as a necessary benchmark for future model evaluations.
View details
DeduBB: Binary Code Size Reduction via Post-Link Basic Block De-duplication
Chaitanya Mamatha Ananda
Rajiv Gupta
Mahbod Afarin
Han Shen
LCTES (Languages, Compilers, Tools and Theory of Embedded Systems) (2026) (to appear)
Preview abstract
Binary sizes of newer versions of software applications tend to be larger, primarily due to feature bloat. This poses various challenges, particularly for mobile applications. It affects
upgrade rates directly impacting revenues, increases maintenance costs of supporting multiple versions, and prevents some users from getting critical security fixes. Code bloat also poses a problem for large warehouse-scale applications. Such applications experience performance degradation when their code size exceeds what smaller and more efficient code
models can handle.
In this paper, we introduce a post-link optimization tech nique called DeduBB, which deduplicates basic blocks of an application across procedure boundaries. While prior tech-
niques used function outlining to de-duplicate redundant code sequences, it missed out on many opportunities as it cannot handle code that manipulates the program stack. In addition, previous techniques were either limited to the scope of a module or lacked scalable implementations required to handle large warehouse-scale applications. Our technique,
DeduBB, handles all types of code duplication as we use a novel save-and-jump code pattern to execute de-duplicated code blocks. In addition, DeduBB has been designed to work on scalable post-link optimizers and can even be applied to large warehouse-scale datacenter applications. Finally, DeduBB is profile-guided and can be applied selectively to infrequently executed cold basic blocks to not affect application performance. In fact, in several cases, the performance
of the smaller application binary improves due to reductions in its hot working set size. We have implemented our technique on the state-of-the-art post link optimizers, BOLT and Propeller. Experiments show that we can significantly reduce the code size of several benchmarks by 1.55% to 18.63%, on both Arm and x86 platforms, and on binaries that have already been heavily optimized for size using existing code size reduction features. Furthermore, aided by profiles, our technique can retain more than 80% of the maximal code size savings without affecting performance.
View details
Preview abstract
**Agentic Engineering** is the rigorous discipline of treating Large Language Models as semi-autonomous systems that execute complex, multi-step workflows (trajectories) based on verifiable specifications, rather than using them as simple autocomplete engines.
Here is a brief summary of its core principles:
* **Main Goals:** It aims to maximize the agent's autonomous run-time, multiply a single engineer's impact by running parallel tasks, and offload tedious boilerplate coding.
* **The "Harness":** A raw model is virtually useless without heavy investment in a harness—comprising tools, system prompts, and strict guardrails—to reliably guide the model and enforce coding policies.
* **Loss of Micro-Control:** Engineers must surrender idiosyncratic stylistic preferences; if the agent's code passes automated linters and tests, it is accepted.
* **Meta-Debugging:** When failures occur, engineers no longer fix code syntax. Instead, they debug the workflow itself—adjusting the agent's tools, search queries, or prompt constraints to ensure repeatable success.
View details
ConvApparel: A Benchmark Dataset and Validation Framework for User Simulators in Conversational Recommenders
Ofer Meshi
Guy Tennenholtz
Jihwan Jeong
The 19th Conference of the European Chapter of the Association for Computational Linguistics (EACL-26), Rabat, Morocco (2026)
Preview abstract
LLM-based user simulators are a scalable solution for improving conversational AI, but a critical realism gap undermines their effectiveness. To close this gap, we introduce a framework for building and validating high-fidelity simulators. We present a novel dataset of human-AI shopping conversations designed to capture a wide spectrum of user experiences. To measure fidelity, we propose a hybrid evaluation protocol that combines statistical alignment with a learned, discriminator-based Human-Likeness Score. Our most sophisticated simulator, trained via reinforcement learning with iterative critique, achieves a significant leap in realism. Critically, we demonstrate through counterfactual validation that our simulator—trained exclusively on optimal interactions—realistically adapts its behavior to suboptimal system responses, mirroring real user reactions and marking a key advance in creating reliable simulators for robust AI development.
View details
Leveraging LLMs to Understand Public Perception of Earthquake Early Warnings: A Case Study of the 2025 M6.2 Türkiye Earthquake
Hanjing Wang
Marc Stogaitis
Youngmin Cho
Richard Allen
Boone Spooner
Patrick Robertson
Greg Wimpey
Robert Bosch
Nivetha Thiruverahan
Steve Malkos
Alexei Barski
Tajinder Gadh
Scientific Reports, 16 (2026), pp. 22226
Preview abstract
Android's Earthquake Alert (AEA) system provided life-saving early warnings to millions during the M6.2 Marmara Ereğlisi, Türkiye earthquake on April 23, 2025. This event, the largest in the region in 25 years, served as a critical real-world test for smartphone-based Earthquake Early Warning (EEW) systems. The AEA system successfully delivered alerts to users with high precision, offering over a minute of warning before the strongest shaking reached urban areas. This study leveraged Large Language Models (LLMs) to analyze 560 public social media posts from the X platform, extracting 42 distinct attributes related to user experience and behavior. Statistical analyses revealed significant relationships, notably a strong correlation between user trust and alert timeliness. Key findings indicate that alerts received before shaking, clear information, and audible notifications significantly increased the likelihood of active protective responses and enhanced perceived usefulness and future trust in the system. Conversely, annoyance and perceived inaccuracy led to inaction and diminished trust. The study highlights the profound impact of smartphone-based EEW in mitigating seismic risk, providing actionable insights for optimizing alert design, public education campaigns, and future behavioral research to improve the effectiveness of such systems in seismically active regions.
View details
Preview abstract
Optical health sensing algorithms, such as SpO2, sleep monitoring, and metabolic health sensing, critically depend on the accurate measurement of optical emission from Light Emitting Diodes (LEDs) transmitted through user tissue and detected by a photodiode (PD). A significant challenge to the reliability of these measurements is the inherent degradation of LED optical emission intensity over time due to device aging. This degradation can confound the physiological changes being monitored. Our work quantifies the impact of LED aging on sensor signal integrity, specifically examining the Current Transfer Ratio (CTR), which is a key metric defining the ratio of received photocurrent to the LED drive current used for transmission in various health sensing algorithms. We investigate the degradation characteristics across LEDs of different wavelengths. Our findings indicate a relative CTR change due to degradation ranging from 1% to 8% within 100 hours of continuous operation which translates to approximately 3.5 to 7 years of device lifetime. Furthermore, we explore the non-linearity of this degradation and the observed initial ”overshoot” phenomenon in the CTR during aging. We discuss how understanding these dynamics could inform the development of robust specifications for different physiological sensing algorithms. Finally, we present several potential solutions to mitigate the effects of LED aging. During the product design phase, integrating a calibrating photodiode or compensating circuitry around the LED can help preemptively address degradation. In the application space, run-time calibration strategies employing two differently degraded optical paths offer a promising approach to maintain measurement accuracy.
View details
Preview abstract
We consider a setting where we have a ground set ℳ together with real-valued set functions f₁, … , f_n, and the goal is to partition ℳ into two sets S₁,S₂ such that |f_i(S₁) - f_i(S₂)| is small for every i. Many results in discrepancy theory can be stated in this form with the functions f_i being additive. In this work, we initiate the study of the unstructured case where f_i is not assumed to be additive. We show that even without the additivity assumption, the upper bound remains at most O(√{n log n}).
Our result has implications on the fair allocation of indivisible goods. In particular, we show that a consensus halving up to O(√{n log n}) goods always exists for n agents with monotone utilities. Previously, only an O(n) bound was known for this setting.
View details
