Research map / Large language models / Reasoning and chain of thought

When LLMs Develop Languages: Symbolic Communication for Efficient Multi-Agent Reasoning

ICML 2026 · Large language models › Reasoning and chain of thought · approach: Search and self-correction in reasoning · 0 citations (OpenAlex)

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.

Proposes CLSR, a test-time framework where LLM agents evolve compact symbolic reasoning protocols (LSFs) that an LLM router selects, ensembles, or composes per query to cut generated tokens versus CoT.

Problem

Chain-of-thought reasoning produces long natural-language rationales that are token-expensive; the paper asks whether LLMs can invent compact symbolic languages that reason more efficiently under an accuracy-token trade-off.

Approach

LLM agents propose, critique, and mutate reusable Language Symbolism Frameworks (symbol inventory, grammar, usage rules) via an evolutionary loop selecting for correctness and low token cost; an LLM-based router (no trained latent router) then plans single-LSF, ensemble, or multi-round LSF composition per query.

Key results

• On Qwen3-8B, CLSR reaches 91.2% on GSM8K with 89 tokens versus 91.6% at 248 tokens for Raw CoT, and 86.8% on MATH500 at 257 tokens versus 87.8% at 905 tokens for CoT.
• With T=3 rounds on Qwen3-8B, CLSR attains 94.8% on GSM8K and 89.7% on MATH500, exceeding PoT (92.1%, 88.1%) and PAL (92.5%, 88.6%) in accuracy.
• On LLaMA3-8B, CLSR reaches 38.9% on MMLU-Pro with 320 tokens versus 38.6% at 960 tokens for Raw CoT.

Surprising findings

• CLSR with T=3 exceeds program-execution baselines PoT and PAL on GSM8K and MATH500 without any external executor.
• Swapping the LSF generator to a weaker model reduces tokens but degrades accuracy, with the largest drop on GPQA.

Limitations

• Offline language evolution is computationally expensive and takes days.
• Synthesis requires training exemplars containing reasoning content.
• Compact symbolic traces may be less interpretable for human auditing.

Experimental setup

Datasets: MMLU-Pro GPQA GSM8K MATH500 AIME 2021-2024 ScienceQA HotpotQA LongBench-v2

Benchmarks: MMLU-Pro GPQA-main GSM8K MATH500 AIME21-24 ScienceQA HotpotQA LongBench-v2

Metrics: accuracy (%) generated completion tokens

Baselines: Raw CoT CoD CCoT SoT PoT PAL Plan-to-Solve PromptBreeder

Builds on: chain-of-thought self-consistency Chain-of-Draft Sketch-of-Thought PromptBreeder

Closest papers by meaning

• Large Language Models Are Neurosymbolic Reasoners (AAAI 2024 · similarity 0.89)
• Distilling LLM Agent into Small Models with Retrieval and Code Tools (NeurIPS 2025 · similarity 0.88)
• Communicating Activations Between Language Model Agents (ICML 2025 · similarity 0.88)
• R2R: Efficiently Navigating Divergent Reasoning Paths with Small-Large Model Token Routing (NeurIPS 2025 · similarity 0.87)
• Eliciting Chain-of-Thought in Base LLMs via Gradient-Based Representation Optimization (AAAI 2026 · similarity 0.87)
• LLM+AL: Bridging Large Language Models and Action Languages for Complex Reasoning About Actions (AAAI 2025 · similarity 0.87)