Georgetown University

Using 142 arguments from IBM-Rank-30k as a shared information space, the study feeds 11 open-weight instruction-tuned LLMs' judgments of whether a candidate piece of information is worth pursuing (the probability of an affirmative label, treated as an information-seeking score) into one sequential elicitation simulation with a fixed selection rule, fixed respondent behavior, and no natural-language question generation, characterizing how model-specific information-seeking preferences shape the breadth and depth of elicitation trajectories and testing the effects of interaction history, candidate-set size, response labels, and prompt wording.

Using 142 arguments from IBM-Rank-30k as a shared information space, the study feeds 11 open-weight instruction-tuned LLMs' judgments of whether a candidate piece of information is worth pursuing (the probability of an affirmative label, treated as an information-seeking score) into one sequential elicitation simulation with a fixed selection rule, fixed respondent behavior, and no natural-language question generation, characterizing how model-specific information-seeking preferences shape the breadth and depth of elicitation trajectories and testing the effects of interaction history, candidate-set size, response labels, and prompt wording.

The authors introduce the Analytical Memory Unit (AMU), a memory schema that attaches a full derivation (lineage) graph to every cached result and gates retrieval so a hit is served only when the requester is authorized for every column touched; under complete lineage recording they prove a safety guarantee by construction, and six experiments show lineage-gated retrieval removes the cross-department leakage of content-gated memory at a modest reuse cost, with zero leaks over 9 round-trips in a real-agent integration.

The authors introduce the Analytical Memory Unit (AMU), a memory schema that attaches a full derivation (lineage) graph to every cached result and gates retrieval so a hit is served only when the requester is authorized for every column touched; under complete lineage recording they prove a safety guarantee by construction, and six experiments show lineage-gated retrieval removes the cross-department leakage of content-gated memory at a modest reuse cost, with zero leaks over 9 round-trips in a real-agent integration.

This work adapts the board game Clue into a text-based multi-agent environment with six LLM players (three each from GPT-4o-mini and Gemini-2.5-Flash) across 18 games and introduces an external possibility-matrix tool (YES/NO/MAYBE cells plus an accusation_ready signal) that externalizes belief-state tracking, yielding near-perfect accusation accuracy for tool-augmented agents (Gemini-2.5-Flash 1.00, GPT-4o-mini 0.96), raising non-tool agents in mixed games from about 0.81/3 to about 2.71/3, while leaving agents' autonomy over accusation timing unchanged.

This work adapts the board game Clue into a text-based multi-agent environment with six LLM players (three each from GPT-4o-mini and Gemini-2.5-Flash) across 18 games and introduces an external possibility-matrix tool (YES/NO/MAYBE cells plus an accusation_ready signal) that externalizes belief-state tracking, yielding near-perfect accusation accuracy for tool-augmented agents (Gemini-2.5-Flash 1.00, GPT-4o-mini 0.96), raising non-tool agents in mixed games from about 0.81/3 to about 2.71/3, while leaving agents' autonomy over accusation timing unchanged.