[Paper] Inferring Latent Intentions: Attributional Natural Language Inference in LLM Agents

Source: arXiv - 2601.08742v1

Overview

The paper “Inferring Latent Intentions: Attributional Natural Language Inference in LLM Agents” tackles a missing piece in today’s large language models (LLMs): the ability to reason about why an agent performed an action, not just what the action was. By blending ideas from social psychology with natural language inference (NLI), the authors propose a new framework—Attributional NLI (Att‑NLI)—that lets LLM‑driven agents generate and verify hypotheses about hidden intentions in multi‑agent settings.

Key Contributions

• Att‑NLI framework: Extends classic NLI with abductive (hypothesis‑generation) and deductive (verification) reasoning to model intention attribution.
• Undercover‑V benchmark: A textual game designed to evaluate agents’ ability to infer latent motives in a controlled, multi‑agent environment.
• Three agent archetypes:
  1. Standard NLI agent – pure deductive reasoning.
  2. Att‑NLI agent – abductive‑deductive reasoning without external tools.
  3. Neuro‑symbolic Att‑NLI agent – abductive‑deductive reasoning augmented by theorem provers.
• Empirical hierarchy: Demonstrates that neuro‑symbolic agents consistently outperform the other two, achieving a +17.08 % win‑rate over baseline NLI agents.
• Cross‑disciplinary insight: Shows how concepts from social psychology can be operationalized in LLM reasoning pipelines.

Methodology

1. Problem Formalization

• The authors view intention inference as a two‑step process:
  • Abduction – generate plausible latent intentions that could explain observed actions.
  • Deduction – test each hypothesis against the current knowledge base to see if it holds.

2. Att‑NLI Design

• Prompt templates encode abductive questions (e.g., “What might X be trying to achieve by doing Y?”) and deductive checks (e.g., “Given hypothesis H, does it follow that …?”).
• The framework can be plugged into any LLM that supports chain‑of‑thought prompting.

3. Neuro‑symbolic Augmentation

• After the LLM proposes hypotheses, the system translates them into logical formulas and feeds them to an external theorem prover (e.g., Prolog or Z3).
• The prover returns a boolean verdict, which the agent uses to prune or reinforce its belief state.

4. Evaluation via Undercover‑V

• A turn‑based textual espionage game where multiple agents (players, NPCs, and the LLM) act covertly.
• Success is measured by the LLM’s ability to correctly identify the hidden goal of a target agent before the game ends.

5. Metrics

• Win rate (percentage of games where the LLM correctly attributes the target’s intention).
• Hypothesis quality (precision/recall of generated intentions).
• Computation overhead (extra latency introduced by theorem proving).

Results & Findings

• Abduction matters: Adding a hypothesis‑generation step alone yields a ~13 % boost in win rate.
• Neuro‑symbolic synergy: External theorem proving further refines hypotheses, pushing precision above 80 % and delivering the highest win rate.
• Trade‑off: The richer reasoning pipeline incurs modest latency, but remains within interactive‑game tolerances (< 0.5 s per turn).

The authors conclude that intention attribution is a distinct reasoning skill that can be systematically cultivated in LLM agents, and that neuro‑symbolic integration offers a practical path to more rational behavior.

Practical Implications

• Multi‑agent AI systems: Chat‑bots, virtual assistants, or game NPCs that need to anticipate user goals can adopt Att‑NLI to become more proactive and context‑aware.
• Security & fraud detection: Systems monitoring logs or communications could generate latent‑intent hypotheses (e.g., “Is this user trying to exfiltrate data?”) and verify them automatically.
• Human‑AI collaboration: By exposing its inferred intentions, an LLM can offer clearer explanations, improving trust in collaborative workflows (e.g., code review assistants suggesting why a developer might have introduced a particular pattern).
• Neuro‑symbolic pipelines: The demonstrated benefit of coupling LLMs with theorem provers encourages developers to embed lightweight symbolic engines (e.g., Z3, MiniZinc) into existing LLM‑powered products without a full rewrite.
• Tool‑augmented agents: The paper’s prompt‑engineering recipe shows how to turn any LLM (GPT‑4, Claude, Llama 2) into an attributional reasoner with just a few API calls and an optional symbolic backend.

Limitations & Future Work

• Scalability of theorem proving: While effective on the modest logical structures in Undercover‑V, more complex real‑world domains may cause prover bottlenecks.
• Prompt sensitivity: The quality of abductive hypotheses heavily depends on prompt phrasing; robustness across diverse domains remains an open question.
• Evaluation breadth: The benchmark focuses on a single textual game; broader testing (e.g., dialogue systems, negotiation bots) is needed to validate generality.
• Explainability: Although the system can output its hypotheses, translating formal proof traces into human‑readable explanations is still rudimentary.

Future research directions suggested by the authors include: expanding Att‑NLI to multimodal inputs (vision + language), integrating probabilistic reasoning to handle ambiguous intentions, and building a library of reusable symbolic modules for common intention patterns.

Authors

• Xin Quan
• Jiafeng Xiong
• Marco Valentino
• André Freitas

Paper Information

• arXiv ID: 2601.08742v1
• Categories: cs.CL
• Published: January 13, 2026
• PDF: Download PDF