TLDR: This research explores improving LLM perspective-taking in collaborative tasks using structured “thought-action” examples derived from a planning system. While these examples help with basic attentional filtering, LLMs still struggle with reasoning about hidden information and evaluating the costs of information-seeking actions, highlighting the need for explicit belief tracking and cost modeling for robust social intelligence.
Large Language Models (LLMs) have made significant strides in various areas, but understanding what another agent sees, knows, or intends—a concept known as perspective-taking—remains a considerable challenge. This ability is crucial for effective collaboration, especially in multi-agent systems involving human-AI interaction. A recent research paper titled “Who Sees What? Structured Thought-Action Sequences for Epistemic Reasoning in LLMs” delves into this very problem, exploring how LLMs can improve their capacity for perspective-taking.
The study, conducted by researchers Luca Annese, Sabrina Patania, Silvia Serino, Tom Foulsham, Silvia Rossi, Azzurra Ruggeri, and Dimitri Ognibene, investigates the potential of providing LLMs with structured examples to enhance their reasoning in complex scenarios. They specifically focus on tasks that involve active perception, where an agent needs to actively gather information, and collaborative reasoning, where agents work together to achieve a goal.
The core of their approach involves using a planning system called Fast Downward to generate detailed “reasoning trees.” From these trees, they extract three distinct types of example sequences: G-type (optimal goal paths), E-type (paths that lead to new information), and L-type (locally optimal decisions at each step). These sequences are then transformed into “thought-action” examples. This means an LLM is prompted to explain the reasoning behind each decision, turning a simple action into a clear, step-by-step thought process.
To test their method, the researchers used a modified version of the “Director Task” in a simulated household environment. In this setup, a “Director” agent gives instructions to a “Matcher” agent, which must retrieve a target object. The environment is partially observable, meaning agents don’t always see the same things, and includes hidden containers and occlusions. This setup mimics real-world situations where information is limited and asymmetric, requiring the Matcher to infer what the Director can or cannot see to resolve ambiguities.
The experiments revealed interesting insights. While the structured examples did offer some benefits, particularly the L-type examples which slightly reduced the number of clarification questions and overall action steps, they did not consistently lead to significant improvements in perspective-taking. The LLM-based agents performed well in tasks that only required basic “common-ground filtering”—essentially, ignoring objects the Director couldn’t see. However, they struggled significantly when the task demanded “imagining Director-privileged space” (reasoning about hidden content) or “metacognitive cost-benefit evaluation” (weighing the costs of asking questions or exploring unseen areas).
The findings suggest that simply providing structured examples might not be enough for robust perspective-taking. The LLMs, like GPT-o3-mini used in this study, tend to treat asking questions as having almost no cost, leading them to “play it safe” rather than inferring information. The linear nature of the action lists also doesn’t explicitly convey the underlying reasons for certain decisions, especially those related to cost-benefit analysis.
Also Read:
- CausalPlan: Enhancing LLM Agent Collaboration Through Causal Reasoning
- Bridging the Language Gap: How AI is Enabling Robots to Understand Human Instructions
The researchers conclude that future work needs to focus on incorporating explicit “belief state tracking” (what an agent knows or believes), “learned cost models” (understanding the true cost of actions like asking questions or exploring), and more sophisticated prompting strategies. These strategies should encourage the LLM to hypothesize about unseen content and reason about uncertainty. Furthermore, testing these agents in richer, more uncertain environments with graded visibility and real-world sensor data will be crucial for advancing their ability to truly understand and collaborate with others. For more details, you can read the full research paper here.
