Research map / Multi-agent systems and games / LLM agents
Large Language Models Are Neurosymbolic Reasoners
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Shows that a zero-shot prompted GPT-3.5-turbo agent, given symbolic modules (calculator, navigator, sorter, knowledge base) as extra actions, solves four text-based symbolic games at 88% average score, beating DRRN and a behavior-cloned transformer baseline.
Problem
Whether LLMs can act as neurosymbolic reasoners in text-based games that require symbolic tasks such as arithmetic, map navigation, sorting, and common-sense knowledge lookup, without costly training on expert data.
Approach
Wrap an off-the-shelf LLM in a prompting agent that receives observations and a valid action set (game actions plus symbolic-module actions), with role initialization and task-specific constraint prompts, and chooses actions zero-shot without any weight updates.
Key results
- The LLM agent (GPT-3.5-turbo) with symbolic modules achieves an average normalized score of 0.88 on the test set across four games, with 7 average steps.
- Behavior Cloned Transformer with symbolic modules averages 0.99 score, slightly above the LLM agent's 0.88, but requires expert demonstration training.
- Removing constrained prompts drops the LLM agent's average score from 0.88 to 0.67; GPT-4 reaches 0.91 average versus 0.88 for GPT-3.5, with the largest gains on MapReader and Sorting.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add constrained prompts (ablation) | LLM Agent (GPT-3.5-turbo) · Arithmetic (Test) | score: 0.96 -> 1.00 | Helped |
| Add constrained prompts (ablation) | LLM Agent (GPT-3.5-turbo) · MapReader (Test) | score: 0.64 -> 0.86 | Helped |
| Add constrained prompts (ablation) | LLM Agent (GPT-3.5-turbo) · Sorting (Test) | score: 0.35 -> 0.71 | Helped |
| Add constrained prompts (ablation) | LLM Agent (GPT-3.5-turbo) · TWC (Test) | score: 0.73 -> 0.94 | Helped |
| Add constrained prompts (ablation) | LLM Agent (GPT-3.5-turbo) · Average (Test) | score: 0.67 -> 0.88 | Helped |
| Replace GPT-3.5 with GPT-4 | LLM Agent · MapReader (Test) | score: 0.86 -> 0.99 | Helped |
| Replace GPT-3.5 with GPT-4 | LLM Agent · Sorting (Test) | score: 0.71 -> 0.93 | Helped |
| Replace GPT-3.5 with GPT-4 | LLM Agent · TWC (Test) | score: 0.94 -> 0.71 | Hurt |
| Replace GPT-3.5 with GPT-4 | LLM Agent · Average (Test) | score: 0.88 -> 0.91 | Helped |
Surprising findings
- The LLM agent's Sorting score (0.71) is low, attributed to limited memory of the ascending order of objects.
- GPT-4 performs worse than GPT-3.5 on TWC (0.71 vs 0.94) despite overall gains.
Limitations
- Sorting performance limited by agent memory of ordering; MapReader requires many steps due to forgetting routes.
- Prompts require hand-crafted task-specific constraints; the authors suggest more detailed prompts may help.
- Evaluated only on four simple TextWorldExpress games; no evaluation on broader domains.
Experimental setup
Datasets: TextWorldExpress game environments (Arithmetic, MapReader, Sorting, TWC)
Benchmarks: TextWorldExpress Arithmetic TextWorldExpress MapReader TextWorldExpress Sorting TextWorldExpress TWC
Metrics: normalized score average steps per episode
Baselines: DRRN DRRN + symbolic module Behavior Cloned Transformer Behavior Cloned Transformer + symbolic module LLM agent without constrained prompts
Builds on: ReAct Chain-of-Thought DRRN Decision Transformer TextWorldExpress
Closest papers by meaning
- GTBench: Uncovering the Strategic Reasoning Capabilities of LLMs via Game-Theoretic Evaluations (NeurIPS 2024 · similarity 0.89)
- When LLMs Develop Languages: Symbolic Communication for Efficient Multi-Agent Reasoning (ICML 2026 · similarity 0.89)
- LLM+AL: Bridging Large Language Models and Action Languages for Complex Reasoning About Actions (AAAI 2025 · similarity 0.88)
- Distilling LLM Agent into Small Models with Retrieval and Code Tools (NeurIPS 2025 · similarity 0.88)
- GSM-Symbolic: Understanding the Limitations of Mathematical Reasoning in Large Language Models (ICLR 2025 · similarity 0.88)
- Improving Rationality in the Reasoning Process of Language Models through Self-playing Game (ICML 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.