Beyond Natural Language: LLMs Leveraging Alternative Formats for Enhanced Reasoning and Communication (2402.18439v3)

Abstract: Natural language (NL) has long been the predominant format for human cognition and communication, and by extension, has been similarly pivotal in the development and application of LLMs. Yet, besides NL, LLMs have seen various non-NL formats during pre-training, such as code and logical expression. NL's status as the optimal format for LLMs, particularly in single-LLM reasoning and multi-agent communication, has not been thoroughly examined. In this work, we challenge the default use of NL by exploring the utility of non-NL formats in these contexts. We show that allowing LLMs to autonomously select the most suitable format before reasoning or communicating leads to a 3.3 to 5.7\% improvement in reasoning efficiency for different LLMs, and up to a 72.7\% reduction in token usage in multi-agent communication, all while maintaining communicative effectiveness. Our comprehensive analysis further reveals that LLMs can devise a format from limited task instructions and that the devised format is effectively transferable across different LLMs. Intriguingly, the structured communication format decided by LLMs exhibits notable parallels with established agent communication languages, suggesting a natural evolution towards efficient, structured communication in agent communication. Our code is released at \url{https://github.com/thunlp/AutoForm}.

"Beyond Natural Language: LLMs Leveraging Alternative Formats for Enhanced Reasoning and Communication" investigates the application of non-natural language formats in enhancing the reasoning and communicative efficiency of LLMs. This paper questions the conventional belief that natural language (NL) is the optimal format for cognitive processes in LLMs, particularly in reasoning and agent communication contexts.

Key findings and methodologies include:

1. Alternative Formats: The paper explores the utilization of non-NL formats like code, logical expressions, and structured data for improving LLM reasoning and communication. The authors argue that these formats can lead to improvements in computational efficiency and clarity of communication among agents.
2. Autonomous Format Selection: LLMs were allowed to autonomously select the most appropriate format for each task, resulting in efficiency gains. This approach yielded a 3.3% to 5.7% improvement in reasoning tasks and reduced token usage by up to 72.7% in multi-agent communication while maintaining effectiveness.
3. Methodology and Experiments:
   • The authors implemented a mechanism prompting LLMs to explore non-NL formats for task inputs. Through this, the capability of LLMs to autonomously decide the thought and communicative format was analyzed.
   • Performance improvements were quantified across both single-LLM reasoning tasks and multi-agent communication scenarios.
4. Thought and Communication Transferability: The paper demonstrates that formats devised for reasoning and communication are transferable across different LLM architectures. The selected formats exhibited structural similarities to traditional agent communication languages, reflecting a spontaneous alignment with established communication protocols.
5. Structural Formats in Communication: Interestingly, the structured formats selected by LLMs resemble traditional Agent Communication Languages (ACLs), such as KQML, highlighting the structured and efficient nature of these non-NL formats.
6. Implications for ACLs: The resultant structured formats proposed by LLMs indicate a natural evolution toward efficient structured communication, underscoring the potential for developing new ACLs that leverage the strengths of LLM-autonomously selected formats.

The research underscores the feasibility and potential benefits of adopting non-NL formats for improving LLM functionalities in reasoning and inter-agent communication. This suggests a reevaluation of traditional NL reliance, highlighting the adaptability and efficiency of structured formats that could enrich agent communications and enhance computational efficiency. The code used in the experimentation is available online, facilitating further exploration and validation by the research community.