Inferring Latent Intentions: Attributional Natural Language Inference in LLM Agents

TL;DR

This paper introduces Attributional NLI (Att-NLI), a two-stage abductive-deductive framework enabling LLMs to infer latent intentions of other agents in multi-agent settings and to draw logically coherent conclusions. It operationalizes Att-NLI in a textual game, Undercover-V, to empirically evaluate and compare three agent types: Standard NLI, Standard Att-NLI, and Neuro-Symbolic Att-NLI (the latter integrating an external theorem prover). Across four LLMs, results show a clear hierarchy in attributional and game performance, with neuro-symbolic Att-NLI delivering the strongest results (e.g., average spy win rate of 17.08%) and benefiting from external proof feedback and refinement of strategy. The work demonstrates the value of combining abductive inference with symbolic reasoning to improve rational, intention-aware behavior in LLM agents operating in multi-agent environments and suggests substantial gains from neuro-symbolic integration for complex social deduction tasks.

Abstract

Attributional inference, the ability to predict latent intentions behind observed actions, is a critical yet underexplored capability for large language models (LLMs) operating in multi-agent environments. Traditional natural language inference (NLI), in fact, fails to capture the nuanced, intention-driven reasoning essential for complex interactive systems. To address this gap, we introduce Attributional NLI (Att-NLI), a framework that extends NLI with principles from social psychology to assess an agent's capacity for abductive intentional inference (generating hypotheses about latent intentions), and subsequent deductive verification (drawing valid logical conclusions). We instantiate Att-NLI via a textual game, Undercover-V, experimenting with three types of LLM agents with varying reasoning capabilities and access to external tools: a standard NLI agent using only deductive inference, an Att-NLI agent employing abductive-deductive inference, and a neuro-symbolic Att-NLI agent performing abductive-deductive inference with external theorem provers. Extensive experiments demonstrate a clear hierarchy of attributional inference capabilities, with neuro-symbolic agents consistently outperforming others, achieving an average win rate of 17.08%. Our results underscore the role that Att-NLI can play in developing agents with sophisticated reasoning capabilities, highlighting, at the same time, the potential impact of neuro-symbolic AI in building rational LLM agents acting in multi-agent environments.

Figures (11)

• Figure 1: Multi-agent LLM interactions often require inferring latent intentions beyond surface-level propositions, exposing a gap in traditional NLI evaluation. We propose Attributional NLI, a two-stage abductive-deductive framework grounded in attribution theory, consisting of intention selection followed by conclusion verification. We operationalise Att-NLI with Undercover-V, a verifiable social-deduction game that makes latent-intent attribution empirically testable.
• Figure 2: Illustration of the three agent types tested for attributional natural language inference (Att-NLI) on the Undercover-V textual game. During the description phase, Standard NLI uses deduction only; Standard Att‑NLI performs abduction followed by deduction to infer it is not the spy and describes the word card based on the selected intention; Neuro‑Symbolic Att‑NLI further integrates TP to build a logical record that guides intention selection and identifies player 2 as the spy. After all descriptions, players vote simultaneously, and only the neuro‑symbolic agent correctly finds the spy through the intention selection stage.
• Figure 3: The spy performance comparison between GPT-4o-mini, Mixtral-Medium, Mixtral-8x22b, and GPT-4o across different player types.
• Figure 4: Comparison between Standard Att-NLI and Standard NLI Player (1 Standard Att-NLI (Spy) vs. 5 NLI (Cit.) and 1 NLI (Spy) vs. 5 Att-NLI (Cit.)).
• Figure 5: Comparison between Standard NLI and Neuro-Symbolic Att-NLI Player (1 Standard NLI (Spy) vs. 5 Neuro-Symbolic (Cit.) and 1 Neuro-Symbolic (Spy) vs. 5 Standard NLI (Cit.)).

Theorems & Definitions (1)

• Definition 1: Attributional NLI (Att-NLI)