Guanren Zhou

I am a Ph.D. candidate at UC Berkeley, advised by Prof. Khalid M. Mosalam in the STAIR Lab. My research focuses on grounded multimodal reasoning and generation across vision-language and NLP/LLM systems.

I am seeking a Summer 2027 internship in AI/ML research, machine learning engineering, or software engineering for AI systems.

Email  /  Scholar  /  GitHub  /  LinkedIn

Research

My work asks how models can reason over complex visual and textual inputs while grounding their outputs in identifiable image regions and source passages.

We combine hierarchical SAM 3 grounding with context-aware filtering and DINOv3-based specialist models to structure visual inputs for VLM reasoning. Citation-aligned masks and relational cues connect generated claims to the image regions that support them, without new end-to-end diagnosis annotations or VLM fine-tuning.

GeoScout combines a 3D occupancy belief with refined VLM geometry captions to condition a PPO next-best-view policy. The captions provide semantic priors over unobserved shape, helping the policy reconstruct held-out object instances with fewer views than an otherwise matched no-caption policy.

This review contrasts case-specific optimization with reusable RL policies, highlighting cross-instance generalization, constraint satisfaction, and verifiable evaluation.

Fine-tuned BERTweet estimates state-level stress prevalence from social-media posts to calibrate an interpretable network model, revealing that social influence outweighed direct exposure in over 80% of U.S. states.

We build an event-centric multimodal system that uses fine-tuned RoBERTa models and VLMs to convert web text and event graphics into provenance-preserving records. The shared records support source-grounded LLM synthesis and longitudinal analysis as new information arrives.

A hierarchical LLM pipeline combines CNN-extracted visual attributes with structured metadata to generate per-instance reports and cross-instance summaries from multimodal observations.

Ongoing Research

We formulate target-conditioned event grounding at the proposition level, distinguishing statements about the target event, target-related context, other events, or no event instance. We are developing a two-pass annotation and evaluation framework that separates grounding from proposition-segmentation errors.

Website design borrows from Jon Barron