EcoLANG: Efficient and Effective Agent Communication Language Induction for Social Simulation

TL;DR

EcoLANG tackles the high inference costs of LLM-driven social simulations by introducing a two-stage framework: vocabulary compression to build a compact, expressive lexicon and evolutionary rule induction to craft concise, effective communication prompts. The induced language is then deployed in simulations, achieving token reductions of over $20\%$ without compromising simulation accuracy, and transferring across different LLMs. The approach is framework-agnostic and demonstrates robustness across datasets (PHEME and HiSim) and models (including Qwen and Mistral). By coupling semantic clustering with evolutionary language rules, EcoLANG enables scalable, efficient social simulations while preserving the fidelity of social dynamics.

Abstract

Large language models (LLMs) have demonstrated an impressive ability to role-play humans and replicate complex social dynamics. While large-scale social simulations are gaining increasing attention, they still face significant challenges, particularly regarding high time and computation costs. Existing solutions, such as distributed mechanisms or hybrid agent-based model (ABM) integrations, either fail to address inference costs or compromise accuracy and generalizability. To this end, we propose EcoLANG: Efficient and Effective Agent Communication Language Induction for Social Simulation. EcoLANG operates in two stages: (1) language evolution, where we filter synonymous words and optimize sentence-level rules through natural selection, and (2) language utilization, where agents in social simulations communicate using the evolved language. Experimental results demonstrate that EcoLANG reduces token consumption by over 20%, enhancing efficiency without sacrificing simulation accuracy.

Figures (5)

• Figure 1: Responses generated by humans and agents when discussing the MeToo movement. There is information redundancy in the vanilla setting, i.e., Agents w/ natural language (NL), such as long but unnecessary sentences and advanced but uncommon words (left). The average sentence length (Len(sent)) and word length (Len(word)) of responses further highlight this issue in current agent communication with NL, and this can be mitigated by our proposed EcoLANG (right).
• Figure 2: Overview of the EcoLANG framework. We get the language through vocabulary compression and rule evolution in dialogue-intensive scenarios. Then, we enable agents to use this language in social simulations.
• Figure 3: (a) Average fitness score change and (b) language drift change on synthetic-persona-chat simulated dialogues across iterations; (c) Performance and token consumption in HiSim using the best language rules acquired across iterations.
• Figure 4: Case study: responses of agents without any communication optimization and with the best evolved rule at iteration 1 and 5. In most cases, agents express more concisely while sometimes fail to follow instructions.
• Figure 5: Alignment and expressiveness score distribution in the first iteration.