Shoot First, Ask Questions Later? Building Rational Agents that Explore and Act Like People

TL;DR

The paper investigates whether language-model–based agents can reason and act rationally under resource constraints by introducing a Collaborative Battleship setting and a Bayesian Experimental Design (BED) framework. It pairs a Captain, who questions to gather information, with a Spotter, who answers Yes/No under an information bottleneck, and uses Monte Carlo inference to ground questions, moves, and decisions. Across human and LM agents, the study shows significant gaps in grounding and information-seeking quality, but demonstrates that inference-time BED strategies dramatically improve performance: Spotter QA accuracy improves up to 14.7 percentage points; Captain QA gains in expected information gain up to 0.227 bits; weaker LMs reach super-human performance on targets with large cost savings relative to GPT-5. The authors validate generality with Guess Who? and provide a reusable evaluation harness and a new dataset, BattleshipQA, highlighting practical paths to scalable, rational information-seeking agents that collaborate with people. These findings suggest that combining Bayes-rational querying, information-guided action, and scalable inference can empower smaller models to perform competitively at a fraction of the cost in complex hypothesis spaces.

Abstract

Many high-stakes applications of AI require forming data-driven hypotheses and making targeted guesses; e.g., in scientific and diagnostic settings. Given limited resources, to what extent do agents based on language models (LMs) act rationally? We develop methods to benchmark and enhance agentic information-seeking, drawing on insights from human behavior. First, we introduce a strategic decision-oriented dialogue task called Collaborative Battleship, in which a partially-informed Captain must balance exploration (asking questions) and action (taking shots), while a fully-informed Spotter must provide accurate answers under an information bottleneck. Compared to human players (N=42), we find that LM agents struggle to ground answers in context, generate informative questions, and select high-value actions. Next, to address these gaps, we develop novel Monte Carlo inference strategies for LMs based on principles from Bayesian Experimental Design (BED). For Spotter agents, our approach boosts accuracy by up to 14.7% absolute over LM-only baselines; for Captain agents, it raises expected information gain (EIG) by up to 0.227 bits (94.2% of the achievable noise ceiling). Combined, these components yield sharper targeting (+0.303-0.374 F1), and enable weaker LMs, such as Llama-4-Scout, to outperform both humans (8% -> 82% win rate) and frontier models (0% -> 67% win rate vs. GPT-5) at ~1% of GPT-5's cost. We replicate these findings on Guess Who? where our methods significantly boost accuracy (+28.3-42.4 p.p.), demonstrating their general applicability for building rational information-seeking agents.

Figures (21)

• Figure 1: Illustration of our Collaborative Battleship game. On each turn, the Captain must choose whether to gather information (ask a question) or take action (shoot at a tile). The Spotter sees the full board, but can only provide yes/no answers. Each role requires well-defined forms of internal reasoning (thought bubbles), which we implement as Monte Carlo inference over an approximate hypothesis space. This framework allows us to compare both humans and LM agents against idealized Bayesian strategies in a controlled setting.
• Figure 2: Human results highlights. (a) Asking questions correlates with performance ($\rho=0.684, p<0.002$); (b) Different players demonstrate widely varying explore/exploit strategies; some alternate between asking questions and making moves, while others focus on information-gathering before taking any actions.
• Figure 3: SpotterQA key results. (a) Accuracy varies significantly across the 15 LMs tested. (b) Across LMs, code generation improves accuracy over direct answering and CoT baselines. (c) Whereas human accuracy remains roughly consistent, LMs degrade significantly on "complex" questions that require context. Dashed lines indicate mean human accuracy ($\mu = 92.5\%$) w/r/t gold answer labels.
• Figure 4: CaptainQA key results. (a) Incorporating Bayesian strategies for questions ($Q_{\text{Bayes}}$), moves ($M_{\text{Bayes}}$), and decisions ($D_{\text{Bayes}}$) brings weaker LMs from near-random performance to super-human levels. (b) Sampling up to 10 questions with $Q_{\text{Bayes}}$ yields higher EIG. (c) $D_{\text{Bayes}}$ helps Llama-4-Scout and GPT-4o to spread out questions over time, more closely matching the behavior of humans and GPT-5.
• Figure 5: Guess Who?Our Bayesian strategies yield consistent improvements, replicating the core findings from Battleship.