Yash Maurya
AI Safety & Evaluation
I'm a Research Engineer at Scale AI, building post-training, evaluation, and red-teaming systems for frontier language models, with a focus on alignment, safety, privacy, and dependable behavior in enterprise deployments.
I care about making advanced models trustworthy in production, especially under adversarial and edge-case behavior. My work sits at the intersection of model capability, policy constraints, and operational risk.
I completed my master's at Carnegie Mellon University in Privacy Engineering, where I worked on privacy, LLM safety, and evaluation research across academia and industry collaborations.
Here's my resume in case you need it.
Privacy is not something that I'm merely entitled to, it's an absolute
prerequisite.
Recent News
Research
* indicates equal contribution
2026
HarnessOpt-Bench: Evaluating LLMs at Harness
Optimization
arXiv preprint arXiv:2608.06301 · 2026
TL;DR: Benchmarks whether frontier LLMs can improve an agent's prompts, tools,
memory, and orchestration under a fixed evaluation budget. Results show that optimizer models
matter more than the coding harness they use, with gains varying widely by task and seed.
Insights Generator: Systematic Corpus-Level Trace
Diagnostics for LLM Agents
arXiv preprint arXiv:2605.21347 · 2026
TL;DR: A multi-agent system that turns large corpora of LLM-agent traces into
evidence-backed diagnostic insights. Experts using its reports improved scaffold performance by
30.4 percentage points over the unmodified baseline.
International Conference on Learning Representations (ICLR 2026)
TL;DR: Evaluates model reasoning on morally ambiguous scenarios, where multiple
conclusions can be defensible. Expert rubrics test whether models identify trade-offs, justify
recommendations, and reason across normative ethics frameworks.
LHAW: Controllable Underspecification for Long-Horizon
Tasks
Lifelong Agents: Learning, Aligning, Evolving Workshop at ICLR 2026
TL;DR: Builds controllably underspecified versions of long-horizon agent tasks
and validates their ambiguity through actual agent execution. The 285-task release measures
whether agents seek clarification when missing information is genuinely outcome-critical.
TAIGR Workshop at ICML 2026 · Best Paper Award
TL;DR: A bilingual safety benchmark that separates language effects from
geopolitical grounding through an English-Korean transcreation matrix. It exposes safety and
over-refusal behaviors that translation-only evaluations miss.
2025
IEEE Symposium on Privacy Expectations (ISoPE) 2025 & SUPA 2025 Workshop
on Societal & User-Centered Privacy in AI
TL;DR: Practitioner-oriented framework that organizes concrete ML privacy
mitigations, tools, and design patterns across the ML lifecycle to help teams operationalize
privacy-preserving AI in real-world deployments.
When Privacy Guarantees Meet
Pre-trained LLMs: A Case Study in Synthetic Data
2025 USENIX Conference on Privacy Engineering Practice and Respect (PEPR'25)
TL;DR: Shows how document formatting and contextual patterns can create privacy
leakage in differentially private synthetic data pipelines built on opaque pre-trained LLMs,
even at conservative privacy budgets.
2024
Beyond the Accept Button: How Information and Control Shape Data
Sharing and AI Engagement
Sponsored by Meta - MSIT-PE Capstone Report 2024 (Advisor: Professor Lorrie
Cranor)
TL;DR: Studied how consent flow design impacts AI engagement and data sharing,
examining length and control options in social media data sharing.
Position: LLM Unlearning Benchmarks are Weak Measures of
Progress
IEEE Conference on Secure and Trustworthy Machine Learning (SaTML), 2025
TL;DR: Shows that modest, benign changes to common unlearning benchmarks can
expose recoverable target information or substantially worse retention loss, making reported
progress look more reliable than it is.
Designing a Benefit Assessment Protocol for AI
Systems
Course Project for Responsible AI Course(10-735) at CMU (Advisor: Professor
Hoda Heidari)
TL;DR: A structured protocol for systematically assessing
AI benefits to enable more
comprehensive AI evaluation.
Unified Locational Differential Privacy Framework
arXiv preprint arXiv:2405.03903 (Advisor: Professor Hana Habib)
TL;DR: A privacy framework for aggregating sensitive location-based data while
protecting individual privacy through differential privacy mechanisms.
AI Governance and Accountability: An Analysis of
Anthropic's
Claude
arXiv preprint arXiv:2407.01557 (Advisor: Professor Norman Sadeh)
TL;DR: Case study on Anthropic's Claude using AI governance and accountability
frameworks, examining compliance with NIST and EU AI Act standards.
Guardrail baselines for unlearning in LLMs
ICLR 2024 Workshop on Secure and Trustworthy Large Language Models
TL;DR: Simple prompting and filtering baselines can match fine-tuning on common
unlearning evaluations, motivating stronger metrics that distinguish inference-time safeguards
from genuine parameter-level forgetting.
Twentieth Symposium on Usable Privacy and Security (SOUPS 2024 Posters)
& SOUPS 2024 Societal & User-Centered Privacy in AI Workshop (SUPA 2024)
TL;DR: UsersFirst taxonomy outperforms LINDDUN PRO in detecting privacy notice and
choice threats in user study.
Twentieth Symposium on Usable Privacy and Security (SOUPS 2024 Posters)
& SOUPS 2024 Workshop on Privacy Threat Modeling (WPTM 2024)
TL;DR: UsersFirst: A user-centric framework for identifying and mitigating privacy
notice and choice threats, extending beyond LINDDUN
Through the Lens of LLMs:
Unveiling Differential Privacy Challenges
2024 USENIX Conference on Privacy Engineering Practice and Respect(PEPR'24)
TL;DR: LLMs demonstrate stronger privacy attacks on Google's Topics API, bypassing
differential privacy safeguards.
Is it Worth Storing Historical Gradients?
Course Project for Federated and Collaborative Learning Course(10-719) at
CMU (Advisor: Professor Virginia Smith)
TL;DR: Current weights beat historical gradients for detecting FL attacks, saving
storage and enhancing privacy.
2022
Federated Learning for Colorectal
Cancer Prediction
2022 IEEE 3rd Global Conference for Advancement in Technology (GCAT), 1-5
TL;DR: Federated learning enables privacy-preserving colorectal cancer prediction
across hospitals with centralized-level accuracy
2021
Responsible Computer Vision (RCV) Workshop at CVPR 2021
TL;DR: Slightly modified version of IS-CAM (described below)
2020
IS-CAM: Integrated Score-CAM for axiomatic-based
explanations
arXiv preprint arXiv:2010.03023
TL;DR: Enhanced CNN interpretability through IS-CAM, integrating Score-CAM to
produce sharper attribution maps