DiffMAS: An End-to-End Optimization Framework for Enabling Telepathy in Multi-Agent Systems

The DiffMAS framework innovatively treats latent communication as a learnable component in multi-agent systems, realizing end-to-end joint optimization of communication mechanisms and reasoning capabilities through parameter-efficient supervised training. This framework addresses issues such as information loss, high token overhead, and cumulative latency caused by fixed text communication interfaces in current multi-agent systems, achieving significant performance improvements on benchmarks like mathematical reasoning (AIME24) and scientific question answering (GPQA-Diamond). It represents an important shift in multi-agent systems from manually designed communication protocols to learning-optimized communication mechanisms.

Multi-agent systems based on large language models have demonstrated collective intelligence beyond single agents, but current research has blind spots in communication mechanisms: existing solutions mostly treat communication as a fixed text interface, with three major limitations:

1. Information loss: Compressing complex reasoning states into text loses subtle differences;
2. Token overhead: Long conversations consume context windows;
3. Cumulative latency: Multiple rounds of dialogue increase the number of LLM calls. Latent communication (exchanging internal representations) is an alternative, but existing methods have not been jointly optimized with multi-agent reasoning.

The core innovations of DiffMAS include two points:

Latent Communication as a Learnable Representation

Communication content is dynamically learned based on tasks. Each agent maintains a trainable communication embedding (learning optimal information compression from data), and receiving agents decode the embedding through an attention mechanism to integrate it into reasoning.

Parameter-Efficient Supervised Training

It uses trajectory supervision (expert-demonstrated interaction trajectories), freezes the pre-trained LLM backbone (only trains communication adapters), and hierarchical optimization (first individual strategies then joint communication protocols) to learn efficient collaboration patterns without destroying pre-trained knowledge.

Technical implementation details of DiffMAS:

1. Communication Encoder: A lightweight network compresses LLM hidden states into fixed-dimensional communication vectors, balancing expressive power and computational efficiency;
2. Cross-Agent Attention Mechanism: Receivers dynamically select communication content to focus on through learnable attention weights, enabling selective information integration;
3. Training Stability Techniques: Curriculum learning (from simple to complex tasks), communication dropout (enhancing robustness), and gradient clipping (preventing over-updates) ensure training convergence.

Experimental results show that DiffMAS has significant performance:

Core Indicator Breakthroughs

• AIME24 Math Competition: 26.7% accuracy, outperforming single-agent and text-based multi-agent systems;
• GPQA-Diamond Scientific Question Answering: 20.2% accuracy, demonstrating cross-domain reasoning capabilities;
• Decoding Stability: Significant improvement in output quality consistency.

Comparative Analysis

DiffMAS outperforms single-agent reasoning (proving the value of collaboration), text-based multi-agent systems (proving the superiority of latent communication), and previous latent communication methods (proving the necessity of end-to-end optimization). It also has higher communication efficiency (the token equivalent cost of latent vectors is lower than natural language messages).

Application prospects and challenges of DiffMAS:

• Real-time Collaboration Scenarios: Suitable for latency-sensitive real-time applications (e.g., real-time strategy games, online customer service), reducing communication rounds and token overhead;
• Edge Computing Deployment: Lightweight communication adapters are suitable for resource-constrained environments;
• Interpretability Challenges: Communication vectors have opaque semantics. The team suggests mitigating this through probe technology and visualization, but full interpretability remains an open problem.

Limitations and future directions of DiffMAS:

• Limitations: Relies on expert trajectory supervised training, limiting application in unlabeled data scenarios;
• Future Directions:

1. Reinforcement learning expansion (environment feedback training);
2. Dynamic communication topology (learning optimal communication graph structure);
3. Hierarchical communication (multi-granularity information exchange);
4. Cross-modal expansion (visual, audio, etc. scenarios).

DiffMAS represents a paradigm shift in the communication mechanisms of multi-agent systems: from manually designed communication protocols to learning-optimized communication mechanisms. It proves that direct exchange of internal representations ('telepathy') between agents is not only feasible but also superior. It is recommended that developers explore ways to let systems self-learn efficient communication methods instead of carefully designing dialogue processes.