Lin Guan
lguan9 [at] asu [dot] edu

I'm a 5th-year Ph.D. student in Computer Science at Arizona State University. I work at Yochan Lab (AI Lab), supervised by Dr. Subbarao Kambhampati. My research interests lie at the intersection of machine learning and human-AI interaction, especially reinforcement learning from human feedback (a.k.a. RLHF or PbRL) and LLMs/VLMs for decision making or planning (a.k.a. AI agents). Specifically, I am working on:

(a) Development of systematic mechanisms to robustly harness planning knowledge acquired (particularly via large vision and language models) from vast yet noisy sources of collective knowledge, such as the internet.

(b) Creation of intuitive interfaces (e.g., in natural language) that allow users to effortlessly provide guidance (e.g., within a human-in-the-loop RL system), specify task objectives and preferences, and steer or align the outputs of generative models.

email   |   cv   |   google scholar   |   twitter   |   linkedIn   |   github

• [2023.12]

  1 paper accepted to AAAI 2024.

• [2023.09]

  1 paper accepted to NeurIPS 2023.

• [2023.01]

  1 paper accepted to ICLR 2023.

• [2022.05]

  Student researcher at Google.

• [2022.04]

  1 paper accepted to ICML 2022 (also accepted to RLDM 2022 and received the Best Paper Award at PRL@ICAPS 2022).

• [2021.10]

  1 paper accepted to NeurIPS 2021 as a Spotlight presentation (3%).

• [2021.05]

  Machine learning software engineer intern at TikTok and worked on AutoML.

[Older News]

[2021.10]

1 paper accepted to AAAI 2022 Blue Sky Track.

[2021.06]

1 paper accepted to IROS 2021.

[2019.12]

1 paper accepted to AAAI 2020.

[2019.04]

1 paper accepted to IJCAI 2019.

When no sound verifier is available, can we use large vision and language models (VLMs), which are approximately omniscient, as scalable Behavior Critics to catch undesirable robot behaviors in videos? To answer this, we first construct a benchmark that contains diverse cases of goal-reaching yet undesirable robot policies. Then, we comprehensively evaluate VLM critics to gain a deeper understanding of their strengths and failure modes.

To leverage LLMs in planning tasks, we introduce an alternative paradigm that teases an explicit world (domain) model in PDDL out of LLMs and then uses it to plan with sound domain-independent planners. This is motivated by the insight that LLMs, while incapable of the combinatorial search needed to produce correct plans, may be better suited as the source of world models.

This tutorial discusses the fundamental limitations of LLMs in generating plans (especially those that require resolving subgoal interactions), and also presents constructive uses of LLMs for planning tasks.

We introduce the notion of Relative Behavioral Attributes that enables end users to tweak the agent's behavior through nameable concepts even for a tacit-knowledge skill learning task (e.g., decreasing the steering sharpness of an autonomous driving agent, or increasing the softness of the movement of a two-legged "sneaky" agent).

We make Human-in-the-Loop RL more efficient and feasible by allowing human teachers to not only give binary feedback but also to highlight task-relevant features.

Explicit symbolic knowledge is important for solving long-horizon task and motion planning tasks. But a key resistance to leveraging easily available human knowledge (or knowledge acquired from LLMs/VLMs) has been that it might be inexact. In this work, we present a framework to quantify the relationship between the true task model and an inexact STRIPS model, and introduce a novel approach using landmarks and a diversity objective to make up for potential errors in the symbolic knowledge.

This survey provides a high-level overview of five learning frameworks that primarily rely on human guidance other than conventional, step-by-step action demonstrations.

We showcase how explicit characterization of a domain can facilitate reinforcement learning from ambiguous demonstrations.

We advocated an ambitious research program to broaden the basis for human-AI interaction by proposing that AI systems support a symbolic interface – independent of whether their internal operations themselves are done in human-interpretable symbolic means.

We provide a large-scale, high-quality dataset of human actions with simultaneously recorded eye movements (i.e., gaze info) while humans play Atari video games.

• [2023]

  Served as a reviewer for ICML 2023 and NeurIPS 2023.

• [2022]

  Served as a reviewer for NeurIPS 2022.

• [2021]

  Served as program committee for AAAI 2022 and ICRA 2022.

• [2020]

  Served as program committee for AAAI 2021.