HERA jointly evolves agent harnesses and executable environments to improve decisions about when to act and when to abstain. Verified pairs of feasible and infeasible tasks expose agent failures, which guide both harness adaptation and the construction of new tasks.

HERA alternates failure diagnosis, environment evolution, and harness adaptation.

Agentic Abstention studies when tool-using LLM agents should stop acting under uncertainty. Instead of treating abstention as a single-turn answer-or-refuse decision, the project frames it as a sequential choice: answer, gather more evidence, or abstain.

We evaluate 13 LLM-as-agent systems and 2 agent scaffolds on more than 28,000 tasks across web shopping, terminal environments, and interactive question answering. The results show that many agents can eventually abstain but often do so too late.

Timely, delayed, and failed abstention trajectories in a web shopping task.

Abstention recall increases with interaction budget, while early abstention remains difficult.

[My role in this project]
Co-first author; problem formulation, benchmark construction, experiments, analysis, project page, and writing.

This project introduces SAGE, a framework for learning embodied navigation policies from physics-grounded semantic abstractions. Instead of depending only on photorealistic simulation, the agent learns from sandbox environments that preserve the structure needed for planning and control.

The framework combines sandbox task generation, reinforcement learning over abstracted experience, and transfer to open-world navigation. The paper was accepted to ICML 2026 on May 1, 2026 and is available as arXiv:2605.10118.

SAGE overview: sandbox generation, policy evolution, open-world navigation, and benchmark performance.

SAGE framework: Genesis, Evolution, and Navigation phases.

SPASM is a stability-first framework for persona-driven LLM-LLM dialogue simulation. The project addresses long-horizon identity failures such as persona drift, role confusion, and echoing when language models are used to generate synthetic multi-turn dialogue data.

The framework decomposes simulation into persona generation, Client-Responder dialogue generation, and termination detection. Its core mechanism, Egocentric Context Projection, stores dialogue history in a perspective-agnostic form and projects it into each agent's view before generation.

SPASM framework overview: persona generation, dialogue simulation, egocentric context projection, and termination detection.

[My role in this project]
First author; framework design, implementation, experiments, analysis, and writing.

DialogGuard is a web interface for auditing LLM responses in sensitive conversational contexts. It wraps arbitrary generative models with reusable LLM-as-a-judge pipelines and presents practitioner-facing risk scores and natural-language rationales.

The system supports manual input and live chat workflows, evaluates responses across psychosocial safety dimensions, and exposes multiple evaluation mechanisms including single-agent scoring, dual-agent correction, majority voting, and multi-agent debate.

DialogGuard interface for entering dialogue context, comparing evaluation mechanisms, and inspecting per-dimension risk scores.

Reasoning panel exposing step-by-step multi-agent safety judgments.

[My role in this project]
First author; system design, interface implementation, evaluation, study analysis, and writing.