When LLMs Invent Their Own Language: CLSR Enables Multi‑Agent Reasoning with Fewer Tokens

Overview

CLSR (Communicative Language Symbolism Routing) asks whether large language models must use natural language to organise reasoning. It proposes that, under correctness‑vs‑token‑cost pressure, LLM agents can discover and evolve discrete machine‑language symbol systems called Language Symbolism Frameworks (LSFs).

Problem Background

Chain‑of‑Thought (CoT) improves accuracy by externalising intermediate steps, but its natural‑language format incurs high token cost, especially when the intermediate state is consumed by another LLM rather than a human. In multi‑agent pipelines (solver, router, critic, verifier, etc.) the bandwidth cost of verbose prose becomes a bottleneck.

If a token is bandwidth, reasoning efficiency is the amount of answer‑relevant state each token carries.

LSF: Reusable Machine‑Reasoning Protocol

An LSF is a structured protocol consisting of slots such as [bind] x=…, y=…, [sub] need: eliminate distractor B/C,

[op] evidence(A) > evidence(D); constraint: mechanism match

, [chk] no contradiction with condition‑2, and [ans] A. Unlike a one‑off prompt, an LSF can be stored, evaluated, routed, combined, and evolved across examples.

How CLSR Works

4.1 LSF Synthesis

Given training exemplars from a benchmark, CLSR samples a batch of exemplars as context and asks the LLM to design an LSF that maintains reasoning ability while reducing tokens. No manual editing of symbols or grammar is performed.

4.2 LSF Evolution

Use current LSFs to answer new training/validation problems.

Record answer correctness and completion‑token cost.

Select high‑leverage traces that are both correct and short.

Feed these traces, parent LSFs, and failure information back to the LLM.

Generate the next generation of LSFs and repeat.

The agents are defined by backbone model, random seed, exemplar subset, and generation context, effectively expanding the search space of symbolic protocols.

4.3 Test‑time Routing

At inference, an LLM‑router selects or composes appropriate LSFs based on the problem and LSF profile. Three reasoning modes are supported: low‑cost strict LSFs for easy questions, multi‑round compositional LSFs for hard questions, and hybrid plans that balance compression and verification.

Token‑efficient reasoning is not “saying less” but “packing more answer‑relevant information per token”.

Experimental Setup

Evaluated on seven benchmarks (MMLU‑Pro, GPQA‑main, GSM8K, MATH500, AIME21‑24, ScienceQA, HotpotQA) using open‑source backbones LLaMA3‑8B, DeepSeek‑R1‑Distill‑Qwen3‑8B, Qwen3‑8B, Qwen3‑32B. Baselines include Raw CoT, CoD, CCoT, SoT, PoT, PAL, Plan‑to‑Solve, PromptBreeder.

Metrics: final answer accuracy and average completion tokens.

CLSR pushes the accuracy‑token frontier: near‑Raw‑CoT accuracy with 3–6× fewer tokens.

Main Results

Across models and tasks, CLSR consistently reduces generation tokens while staying on the Pareto frontier. For Qwen3‑8B, CLSR achieves the same accuracy as Raw CoT but with roughly one‑third of the tokens. Ablation on the number of reasoning rounds $T$ shows $T=3$ improves accuracy on hard tasks (GSM8K, MATH500, GPQA) despite higher token use, confirming that allocating tokens to verification yields gains.

Mechanism Analyses

1. Token‑saving comes from “code‑switching” rather than simple compression. LSFs retain high‑density information such as variable bindings, evidence links, and checks.

2. Hard problems benefit from allocating tokens to structured verification. Experiments demonstrate that extra rounds improve accuracy while still using fewer tokens than Raw CoT.

3. Relation to program‑aided reasoning. CLSR does not rely on external executors but can emulate program‑like state updates when interpreter‑realizability holds. Comparisons on Qwen3‑8B show CLSR’s token cost is lower than PoT/PAL when counting only generation tokens, though external execution remains valuable for exact computation.

From CLSR to Machine Dialectology (MDia)

MDia extends CLSR to heterogeneous LLM societies, tracking speaker‑listener‑dialect‑task events. It introduces “receiver‑relative utility”: a dialect’s usefulness depends on the listener, task, and token budget. MDia improves average accuracy by ~3.6 % over the strongest token‑reduction baseline and cuts generation tokens by ~71 %.

Key findings include listener openness asymmetry, public dialect asymmetry, weak‑speaker teaching effects, foreign‑dialect risk, and the importance of profile‑aware routing.

Application: Principia

Principia leverages CLSR/MDia as a symbolic layer for research‑idea discovery. It decomposes idea generation into reusable principle cards, evidence composition, and symbolic derivations, enabling small open‑source models to achieve higher effective reasoning capacity within controlled workflows.

Takeaway Messages

CoT length comprises necessary reasoning state and redundant natural‑language narration; replacing the latter with symbols yields token savings.

Token‑efficient reasoning is a representation‑learning problem, not merely prompt‑style control.

No single machine dialect is optimal for all tasks; adaptive routing and dialect selection are essential.

Small models benefit disproportionately from efficient protocols.

Machine‑language protocols must be auditable and may require dual‑layer traces for human interpretability.

Limitations and Open Questions

LSFs still rely on human‑learned symbols; they are not invented from scratch.

Compact traces are not guaranteed to be human‑readable explanations.

Cross‑model dialect transfer can be risky; profile‑aware routing mitigates but does not eliminate foreign‑dialect hazards.

Program execution and formal verification remain indispensable for exact computation tasks.

Model updates may invalidate existing dialects, necessitating versioning and drift detection.

References

ICML 2026 poster: https://icml.cc/virtual/2026/poster/61557

Code & data: https://github.com/pzqpzq/LSF_MDia

Principia prototype: https://github.com/pzqpzq/Principia