Search papers, labs, and topics across Lattice.
This paper investigates the potential for heterogeneous agents in multi-agent systems to achieve effective communication through dense latent alignment, addressing the limitations of traditional text-based communication and existing homogeneous models. By employing a novel two-phase training approach that includes reconstruction and generation, the authors demonstrate that their method significantly enhances KV-cache communication across diverse model sizes and benchmarks. The results indicate that their approach not only outperforms existing heterogeneous baselines but also matches or exceeds the performance of text communication while reducing computational costs by 2 to 3 times.
Heterogeneous agents can achieve "mind reading" capabilities through dense alignment, outperforming traditional methods with significantly lower compute costs.
Multi-agent systems communicate mostly through text, paying a lossy and expensive decode and re-encode cost. KV-cache communication is a promising alternative, yet most prior work is homogeneous, using duplicate copies of the same model, and avoids the central challenge of cross-model latent alignment; existing heterogeneous methods are also restrictive, typically assuming shared input and using transferred caches mainly for steering. We study a more fundamental question: can heterogeneous agents be aligned well enough to perform real"mind reading"and transfer both what one agent sees and how it thinks? Our information-structure analysis reveals a duality: context-aware transfer is driven by sparse reasoning signals, while context-unaware transfer, where the receiver sees no input, requires dense contextual knowledge preservation. Motivated by this, we propose dense alignment for heterogeneous KV-cache communication via a lightweight cross-model cache transformation and two-phase training: reconstruction followed by generation. Across all six directions of {Qwen3-4B, 8B, 14B} and six in-domain and out-of-domain benchmarks, our method outperforms prior heterogeneous baselines, matches or exceeds text communication in context-aware settings at roughly 2 to 3 times lower compute, and remains effective in context-unaware transfer where prior methods collapse.
