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

Abstract: Chain-of-Thought (CoT) improves LLMs on difficult reasoning tasks, but it often incurs long natural-language rationales that are poorly aligned with efficient machine reasoning. We propose Communicative Language Symbolism Routing (CLSR), a test-time framework in which multiple LLM agents autonomously invent, evolve, and share compact Language Symbolism Frameworks (LSFs), while a latent-free router adaptively selects and composes these languages per query to optimize the accuracy-token trade-off. Unlike prompt optimization that refines surface instructions, CLSR treats each LSF as a reusable symbolic protocol with compact symbols, usage rules, and a message-passing contract, and improves it through an evolutionary loop driven by correctness and token cost. At inference time, the router may invoke a single low-cost LSF call, ensemble multiple LSFs, or execute a multi-round LSF composition protocol on harder queries. Across challenging benchmarks, CLSR reduces latency-oriented generated token completion by $3\sim 6\times$ compared to standard CoT while maintaining accuracy. We further derive an information-theoretic lower bound on token cost under arbitrary symbolism and show that, under an interpreter-realizability premise, multi-round LSF protocols conditionally subsume program-execution pipelines. Code is publicly available (https://github.com/pzqpzq/LSF_MDia).

• The paper introduces CLSR, a framework where LLMs evolve compact symbolic protocols to enhance multi-agent reasoning efficiency.
• It leverages multi-agent evolutionary bootstrapping and adaptive protocol routing to achieve up to 3–6× token reduction while preserving accuracy.
• Empirical validation across benchmarks demonstrates robust performance gains over traditional chain-of-thought methods under compute constraints.

Symbolic Communication for Adaptive Multi-Agent Reasoning: Analysis of "When LLMs Develop Languages"

Introduction and Motivation

The paper "When LLMs Develop Languages: Symbolic Communication for Efficient Multi-Agent Reasoning" (2606.29354) presents Communicative Language Symbolism Routing (CLSR), a test-time inference framework enabling LLMs to autonomously invent, evolve, and route among compact symbolic reasoning protocols called Language Symbolism Frameworks (LSFs). This paradigm addresses the inefficiency of Chain-of-Thought (CoT) reasoning, which—although beneficial for complex tasks—yields verbose, natural-language rationales with high completion costs, suboptimal for compute-constrained settings and not reflective of internal structural representations.

Unlike prompt optimization or human-designed program-based reasoning, CLSR operationalizes a socio-linguistic evolutionary process in which multiple LLM agents generate, critique, mutate, and retain modular LSFs, driving selection toward information-dense symbolic protocols. These LSFs function as reusable, compositional communication contracts. CLSR employs a latent-free router, implemented via LLM, to adaptively select, ensemble, or compose LSFs per query, optimizing the accuracy--token Pareto frontier at inference without access to model weights or external executors.

Figure 1: Visualization of the CLSR pipeline, illustrating LSF self-evolution, multi-agent bootstrapping, and query-adaptive protocol routing.

CLSR Framework: LSF Synthesis, Evolution, and Routing

LSF Specification and Evolutionary Bootstrapping

Each LSF is defined as a tuple $$S_k = (\mathcal{V}_k, \mathcal{G}_k, \mathcal{R}_k, \psi_k, \rho_k)$$, encompassing a compact symbol inventory, compositional grammar, usage constraints, operation mapping, and empirical profiles. LSF synthesis proceeds by prompting LLMs to extract symbolic conventions from small pools of annotated exemplars (ranging from strict program-like to softened structured natural language), yielding high surface diversity essential for downstream selection.

Subsequent LSF evolution implements a generational loop: independent LLM agents propose LSFs, validate on fresh partitioned data, select high-leverage (accurate, token-efficient) traces, and perform mutation/refinement steps. Population diversity is further bolstered by parallel agent sampling and high-temperature decoding. Empirical ablation (see Figures 3, 4) demonstrates that both increased agent populations and deeper evolutionary schedules yield LSFs with superior accuracy--token efficiency characteristics, though with diminishing returns beyond moderate evolution depths.

Figure 2: Increasing the population of LLM agents in evolution enhances accuracy and token efficiency of the final LSF pool.

Figure 3: Deeper evolutionary refinement iterations improve both test accuracy and token efficiency of evolved LSFs, with gains saturating past a certain threshold.

Token Efficiency and Adaptive Protocol Routing

Token consumption is directly measured as the aggregated length of all LLM-generated outputs during routing, LSF execution, and aggregation. CLSR’s core inference objective is to maximize $\mathbb{E}[\mathbb{I}\{\hat{y}=y\}]$ under $\mathbb{E}[C] \leq B$, where $C$ is the total number of completion tokens, and $B$ is a budget.

At inference, the router digests query-aligned LSF descriptors—summaries of accuracy, cost, and empirical reliability—and adaptively constructs plans choosing between: (i) direct single-LSF decoding, (ii) multi-LSF aggregation (e.g., majority voting), or (iii) staged symbolic compositions involving multiple LSFs and rounds. This fine-grained budget allocation, including early stopping, enables instance-wise optimal compute usage.

Figure 4: CLSR trace consistently dominates CoT on the accuracy--token trade-off across scientific QA, mathematical, and hard math tasks when scaling token budgets.

Theoretical Properties and Empirical Results

Information-Theoretic Foundations

The paper formalizes the cost--accuracy Pareto frontier using a constrained stochastic control framework. It derives an information-theoretic lower bound on the expected number of generated tokens required for a target accuracy:

$$\mathbb{E}_\pi[L] \geq \frac{\max\{I_{\mathrm{req}}(x,\delta), 0\}}{\kappa_\theta(x)}$$

where $I_{\mathrm{req}}(x, \delta)$ quantifies the required information to reach error rate $\delta$; $\kappa_\theta(x)$ captures the per-token information rate given the LSF protocol and model. CLSR empirically increases $\kappa_\theta(x)$ by evolving more information-dense symbolism, as corroborated by aggregate token savings and accuracy retention (see Table 1 and the relevant figures).

Empirical Validation

Experiments span seven reasoning benchmarks: MMLU-pro, GPQA-main, GSM8K, MATH500, AIME21-24, ScienceQA, and HotpotQA. Evaluations across open-source LLMs (Qwen3-8B/32B, LLaMA3-8B, DeepSeek-R1) consistently show CLSR improves or maintains accuracy with a $\mathbb{E}[\mathbb{I}\{\hat{y}=y\}]$0 reduction in completion length compared to standard CoT, outperforming token-reduction (e.g., CoD, CCoT, SoT) and prompt-optimization baselines. CLSR remains competitive with program-execution pipelines (e.g., PoT, PAL), despite not relying on external tools.

Figures 5–8 present fine-grained scaling analyses; increasing the LSF agent pool and deeper evolutionary schedules consistently boost efficiency and accuracy across domains.

Figure 5: More LLM agents in LSF evolution yield superior token efficiency and accuracy for MMLU-pro.

Figure 6: Larger agent populations in LSF evolution enhance both accuracy and efficiency for GPQA-main.

Figure 7: Increased agent count in LSF evolution boosts performance on MATH500.

Figure 8: Science-QA benchmark also benefits from more evolution agents in terms of tokens and accuracy.

Qualitative analyses (Figures 9–14) highlight that CLSR traces—unlike verbose CoT—encode intermediate state and logical progression with minimal symbolic overhead. These traces act as interpretable, tightly specified reasoning skeletons, making error localization and correction more tractable and facilitating robust multi-step reasoning under severe token constraints.

Sociolinguistic and Practical Implications

The emergence and retention of effective LSFs in CLSR are interpretable via socio-linguistic principles (e.g., conventionalization by least-effort, code-switching). CLSR demonstrates that LLMs, given only black-box access and indirect selection pressure for correctness-compression, converge toward reusable symbolic protocols that are not merely the result of instruction-tuning or handcrafted prompts but are bootstrapped via inter-agent evolution and selection within the model’s inductive priors.

Practically, CLSR is applicable in deployment settings where compute budgets and latency constraints are critical. The modular LSF pool, discovered once, can be efficiently amortized across downstream queries, and the framework is compatible with advances in LLM scaling and architectural enhancements.

Limitations and Future Directions

While CLSR robustly improves token efficiency and practical inference efficiency, offline LSF evolution incurs substantial upfront computational cost, and the framework currently relies on high-quality exemplars containing rich reasoning traces. LSF synthesis and transferability across tool-augmented or multi-modal settings remain open problems for practical deployments. Further, interpretability for human review may be hampered as protocols become highly compressed and less semantically transparent.

Conclusion

The paper establishes CLSR as a robust, practical method for test-time reasoning compression in LLMs, enabling adaptive, token-efficient multi-agent reasoning with emergent symbolic protocols. The methodology and theoretical framing lay groundwork for future research on automated language evolution, self-bootstrapping reasoning architectures, and scalable LLM deployments under hard compute constraints. As LLMs evolve toward collective, communication-competent agents, frameworks such as CLSR will be pivotal in understanding and engineering emergent compositional language in artificial systems.

Figure 9: On GPQA, CLSR generates compact symbolic traces that maintain key state and correctness.

Figure 10: CLSR’s symbolic form yields concise, correct multi-step derivations in MATH500.