Researchers from MIT Computer Science and Artificial Intelligence Laboratory and Harvard University School of Engineering and Applied Sciences have demonstrated a method to significantly enhance artificial intelligence agents ability to generate informative questions in uncertain environments. The findings, presented at the International Conference on Learning Representations in April, address a critical limitation in current large language models which are optimized to answer queries but struggle to formulate effective inquiries. To isolate and study this capability, the team developed Collaborative Battleship, a variant of the classic guessing game that requires participants to exchange natural language questions and responses. Over forty human players generated a baseline dataset, which was used to evaluate state-of-the-art and smaller language models. Initial testing revealed that while top-tier models could outperform average humans with minimal turns, they frequently generated redundant or low-information questions. Smaller models performed poorly without structural intervention. The researchers introduced two key technical modifications to bridge this gap. First, they implemented a Monte Carlo inference strategy for the questioning AI, allowing it to treat potential guesses as probabilistic particles that are dynamically weighted based on real-time feedback. This adaptive reasoning enables the system to prioritize questions that maximize information gain. Second, the answering AI was instructed to convert each query into executable Python code. This auto-formalization approach forced the model to explicitly verify spatial data and constraints before responding, effectively reducing hallucinations and misaligned answers. The results demonstrated substantial performance improvements across model scales. Llama 4 Scout, a smaller language model, increased its win rate against human opponents from eight percent to eighty-two percent. By leveraging the refined inference and verification pipeline, the system also outperformed the frontier model GPT-5 while operating at approximately one percent of its computational cost. Answering models similarly improved, with GPT-4o-mini showing a nearly thirty percent accuracy boost and Claude 4 Opus gaining eight points. Validation tests on Guess Who yielded comparable gains, confirming that structured information seeking is broadly applicable. The study underscores that effective AI inquiry depends on robust world modeling and formal verification mechanisms. Researchers note that while these techniques dramatically improve exploratory search, agents still lag behind expert human players in highly complex scenarios, highlighting ongoing challenges in pragmatic reasoning and collaborative dynamics. The work suggests that embedding explicit reasoning and code-based verification into AI agents will be essential for high-stakes applications in scientific discovery, software development, and autonomous research. As agentic systems advance, the ability to efficiently navigate vast solution spaces and adapt to human partners remains a central bottleneck. Future research will focus on scaling these methods to complex, multi-variable environments and evaluating sustained human-AI collaboration.