Selective and Hybrid-State Communication for Latent Multi-Agent Systems

Miles Liao; Aaron Smieszek; Yuhao Ge

cs.MAOct 27, 2026

Selective and Hybrid-State Communication for Latent Multi-Agent Systems

Authors: Miles Liao, Aaron Smieszek, Yuhao Ge

Organizations: Independent Researcher

Abstract

Efficient communication remains a major bottleneck in large language model multiagent systems. Latent communication reduces repeated natural-language exchange by transferring internal model states between agents, but these transfers remain costly and are typically designed around conventional KV-cache representations. We address this problem from two complementary perspectives. We introduce SelectKV, a boundary-aware method that selectively communicates KV states while protecting inherited context, and HyLaMAS, which extends latent collaboration to hybrid linear-attention models by communicating both KV caches and recurrent states through a shared workspace. Across six benchmarks, SelectKV reduces communicated KV positions and logical KV payload by 9.7-9.9%, with accuracy unchanged on two benchmarks, within one percentage point on three, and six points lower on HumanEval+. Across Qwen3.5 model scales, HyLaMAS reduces decoded output tokens by 53-80% and achieves 1.16-2.78x speedups relative to text-based multi-agent communication. Together, these independently evaluated methods represent two complementary approaches to efficient latent-state communication across conventional Transformer and hybrid model architectures.

Figures & tables

	ARC-E			ARC-C			GSM8K
Metric	LM	SKV-90	$\Delta$	LM	SKV-90	$\Delta$	LM	SKV-90	$\Delta$
Acc.	99.0 [97.0, 100.0]	99.0 [97.0, 100.0]	0.0	90.0 [84.0, 95.0]	90.0 [84.0, 95.0]	0.0	94.0 [89.0, 98.0]	93.0 [88.0, 98.0]	$\downarrow$ 1.0
Token	461.89	429.23	$\downarrow$ 7.1%	561.03	515.62	$\downarrow$ 8.1%	622.16	626.78	$\uparrow$ 0.7%
KV Pos.	1023.04	923.42	$\downarrow$ 9.7%	1099.60	992.28	$\downarrow$ 9.8%	1068.46	964.29	$\downarrow$ 9.7%
Payload	143.87	129.86	$\downarrow$ 9.7%	154.63	139.54	$\downarrow$ 9.8%	150.25	135.60	$\downarrow$ 9.7%
Time	16.31	14.83	$\times$ 1.10	19.33	17.81	$\times$ 1.09	21.31	21.87	$\times$ 0.97

Table 1: Sequential multi-agent results on Qwen3-4B with 10 latent reasoning steps and

n=100

examples per benchmark. LM denotes uncompressed LatentMAS; SKV-90 denotes boundary-aware SelectKV with a 90% retention ratio applied to the newly communicable KV region while preserving inherited context. Metrics report accuracy (%), generated output tokens, communicated KV positions, logical KV payload (MiB), and inference time (s), each averaged per example. Accuracy is reported with 95% percentile bootstrap confidence intervals in brackets, computed using 10,000 resamples over evaluation examples.

\Delta

reports the accuracy difference in percentage points, relative change for Token, KV Pos., and Payload, and the LM/SKV-90 inference-time ratio. Green shading indicates values favorable to SKV-90.

Tasks	Metrics	Qwen3.5-4B				Qwen3.5-9B
Tasks	Metrics	Single	TextMAS	HyLaMAS	Improve	Single	TextMAS	HyLaMAS	Improve
	Acc.	100.0	100.0	99.0	$\downarrow$ 1.0	100.0	98.0	99.0	$\uparrow$ 1.0
	Token	361	1138	361	$\downarrow$ 68.3%	386	1350	356	$\downarrow$ 73.6%
ARC-E	Time (s)	464	1294	846	$\times$ 1.53	426	1269	618	$\times$ 2.05
	Acc.	94.0	92.0	91.0	$\downarrow$ 1.0	94.0	95.0	93.0	$\downarrow$ 2.0
	Token	436	1297	481	$\downarrow$ 62.9%	465	1524	430	$\downarrow$ 71.8%

Table 2: Sequential multi-agent results on Qwen3.5-4B and Qwen3.5-9B (Gated DeltaNet layers interleaved with periodic full-attention layers),

n{=}100

problems per benchmark. Single : one-agent inference, shown for context. TextMAS : agents exchange natural-language messages. HyLaMAS : agents exchange the KV cache of full-attention layers and the recurrent and causal-convolution states of Gated DeltaNet layers via the shared workspace. Metrics are accuracy (%), output tokens per problem, and total inference time (s); lower is better for the latter two. Improve is HyLaMAS vs. TextMAS: accuracy in absolute points, token reduction (%), speed-up (TextMAS time / HyLaMAS time). Shaded Improve cells mark a change in the favorable direction.

Method	GSM8K	ARC-E	ARC-C	MedQA	MBPP+	HumanEval+
Accuracy (%) $\uparrow$
Single	73	96	87	61	42	50
TextMAS	67	91	81	61	49	52
HyLaMAS	80	94	81	63	49	58
Improve	+13	+3	+0	+2	+0	+6
Tokens / sample $\downarrow$

Table 3: Latent collaboration at small scale (Qwen3.5-2B),

n{=}100

per benchmark. Single is one-agent inference (context), TextMAS passes natural-language messages, HyLaMAS passes hybrid state. Improve is HyLaMAS vs. TextMAS (accuracy in points, tokens as % reduction, speed as speed-up); shaded where it favors HyLaMAS.

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 1: Boundary-protection ablation for SelectKV on MedQA. Removing protection permits greater compression but produces a larger observed accuracy loss. Boundary-aware SelectKV remains one accuracy point below full-cache LatentMAS while reducing communicated KV payload.

Figure 2: Matched-budget SelectKV selector ablation on MedQA. SelectKV outperforms random and recency-only selection at the same retained-KV budget, indicating that its selection criterion contributes beyond compression alone. The selector ablation was conducted as a separate experimental run from the main SelectKV evaluation in Table 1 .

Figure 3: Hybrid-state channel ablation for HyLaMAS on MedQA. Both retains full-attention KV and the complete GDN state, including recurrent and causal-convolution states; KV only removes GDN state; State only removes KV; and None removes both. The left panel reports accuracy with approximate 95% binomial confidence intervals. The right panel reports mean decoded tokens per completed problem and reports the logical state payload shared across agent boundaries in MiB.

	SelectKV	HyLaMAS
Model	Qwen3-4B	Qwen3.5-{2B, 4B, 9B}
Architecture	dense attention, 36 layers	GDN : attention $3{:}1$
Weights	bfloat16	bfloat16
Attention backend	—	sdpa
GPU	NVIDIA L40S (46,068 MiB)	A100-SXM4-40GB (2B/4B); L40S (9B)
PyTorch / CUDA	2.13.0 / 13.0; 2.8.0 / 12.9	2.13.0 / 13.0 (2B/4B); 2.8.0 / 12.9 (9B)

Appendix

Table 4: Experimental configurations for SelectKV and HyLaMAS. Runtime comparisons use matched model, benchmark, hardware, and batch-size settings.

Multi-Agent Communication

Multi-agent communication studies how autonomous agents exchange information to coordinate decisions, share knowledge, and act together under constraints.

See all 380 papers

LatCom: Cross-Agent Latent Compression for Efficient Multi-Agent Collaboration
Sep 29, 2026Shinan Zhang, Tao Zhang, Qihui Zhu +7Multi-Agent LLM Systems Multi-Agent Collaboration
University of Science and Technology of China · Institute of Artificial Intelligence, Hefei Comprehensive National Science Center
HyLaT: Efficient Multi-Agent Communication via Hybrid Latent-Text Protocol
May 25, 2026Xinyi Mou, Siyuan Wang, Zejun Li +2Agent Communication Protocols LLM Agent Training
Fudan University · The Chinese University of Hong Kong · King’s College London +2
Task-Oriented Key-Layer KV Communication for Efficient Latent Multi-Agent Collaboration
Sep 24, 2026Dongsen Zhang, Peipei Li, Zekun Li +1Multi-Agent Communication Multi-Agent LLM Systems
Beijing University of Posts and Telecommunications · University of California, Santa Barbara

Sep 29, 2026J/K move · Enter open · S save

LatCom: Cross-Agent Latent Compression for Efficient Multi-Agent Collaboration

Shinan Zhang, Tao Zhang, Qihui Zhu, Mengjie Zhang+6

University of Science and Technology of China · Institute of Artificial Intelligence, Hefei Comprehensive National Science Center

Multi-Agent LLM Systems Multi-Agent Collaboration LLM Inference Acceleration Multi-Agent Communication

Open paper AskPDF arXiv

LLM-based multi-agent systems (MAS) increasingly use latent collaboration to avoid the information loss and repeated encoding-decoding overhead of natural-language communication. However, directly forwarding all sender latents makes the receiver-side context scale with both the number of agents and the reasoning length, increasing computation, memory usage, and collaboration latency. A natural solution is latent compression. But we find that cross-agent redundancy remains unresolved in existing latent compression approaches, which typically compress each sender independently and then concatenate the results. We propose LatCom, a cross-agent latent compression framework for efficient multi-agent latent collaboration. LatCom maps multiple sender latents into a fixed number of receiver-readable and task-relevant slots. Rather than reconstructing all sender hidden states, it optimizes the compressed latents for receiver-side task utility. LatCom trains the compressor in two stages: single-sender readability learning first establishes a latent interface interpretable by the frozen receiver, and multi-sender fusion learning then trains the compressor to fuse complementary evidence and remove redundancy across agents. Experiments on multiple benchmarks with Qwen3-4B show that LatCom achieves an average 2.46x inference speed-up over LatentMAS and reduces output token usage by 70.3% while maintaining comparable average accuracy.

Abstract

Multi-Agent Communication

Explore similar work

LatCom: Cross-Agent Latent Compression for Efficient Multi-Agent Collaboration

HyLaT: Efficient Multi-Agent Communication via Hybrid Latent-Text Protocol

Task-Oriented Key-Layer KV Communication for Efficient Latent Multi-Agent Collaboration