Darshan Deshpande

दर्शन देशपांडे

Darshan Deshpande

Research Scientist, Patronus AI

Google ScholarGitHubXLinkedIn

Darshan Deshpande

San Franciscoref. DD/2026/09

Re: reinforcement learning; text-based world models; evaluation; dialogue agents

I’m a research scientist at Patronus AI. My current primary focus is text-based world modeling: language models that simulate an environment’s dynamics, so agents can train on diverse, on-demand experience instead of hand-built tasks. My new NeurIPS 2026 paper shows that masked diffusion language models make stronger, more steerable world models than autoregressive ones, and transfer zero-shot to unseen environments.

Many of my ideas come from cognitive psychology: what drives LLM behaviors and nuances, how people hold mental states, how and when information retrieval is required, and revision of beliefs as evidence arrives. That runs through my work on agent memory (MemTrack), tip-of-the-tongue retrieval (BLUR, DETOUR), finding where agents fail in long traces (TRAIL), and TRACE-ing reward hacks in code environments.

Selected Recent Work

  1. Masked Diffusion Language Models are Strong and Steerable Text-Based World Models for Agentic RL

    NeurIPS 2026 arXiv code data

  2. FigmaTrace: Capturing Creative Nuances in Human Figma Design Workflows

    arXiv preprint, 2026 arXiv

  3. Benchmarking Reward Hack Detection in Code Environments via Contrastive Analysis

    ICML 2026 arXiv

  4. MemTrack: Evaluating Long-Term Memory and State Tracking in Multi-Platform Dynamic Agent Environments

    Scaling Environments for Agents Workshop, NeurIPS 2025 arXiv blog

  5. TRAIL: Trace Reasoning and Agentic Issue Localization

    arXiv preprint, 2025 arXiv code data blog

Before this

At USC’s Information Sciences Institute I was advised by Filip Ilievski and Fred Morstatter on argumentation: identifying logical fallacies robustly and explainably, and grounding argument-quality judgments in background knowledge. With NEC Laboratories Europe I studied whether prototype-based networks make text classifiers more robust. I was later advised by Jonathan May with who I worked on automated generation of open-domain negotiations.

With regards,

Darshan