Learning to Interrupt in Language-based Multi-agent Communication

Danqing Wang; Da Yin; Ruta Desai; Lei Li; Asli Celikyilmaz; Ansong Ni

cs.CLApr 7, 2026

Learning to Interrupt in Language-based Multi-agent Communication

Authors: Danqing Wang, Da Yin, Ruta Desai, Lei Li, Asli Celikyilmaz, Ansong Ni

Organizations: CMU · Meta FAIR

Abstract

When a colleague starts explaining something you already understand, you interrupt them. This simple act, a listener taking control of the conversation, is natural in human communication but absent in current verbose LLM multi-agent systems. Current approaches address communication efficiency only from the speaker side, compressing messages before they are sent. We flip the perspective: rather than making speakers more concise, we let listeners decide when they have heard enough. We propose a new communication paradigm in which the listener can interrupt the speaker mid-generation. We find that LLMs, given this ability, are overconfident and interrupt too early before receiving sufficient information. This finding motivates HANDRAISER, a learning method that predicts the right moment to interrupt based on estimated future reward and communication cost. We evaluate our framework on three multi-agent tasks: 2-agent text pictionary, 3-agent meeting scheduling, and 3-agent debate. HANDRAISER reduces communication cost by 32.2% over the non-interruptible baseline while achieving comparable or superior task performance, with interruption behavior that generalizes across different agents and tasks.

Figures & tables

Panel of Figure 1 — Figure 1: Text Pictionary: the speaker describes words for the listener to guess.

Figure 2: In non-interruptible communication (left), one agent (Bob) must wait for the completion of a full generation from another agent (Alice) to start responding. However, in the interruptible case, the speaker Alice sends messages to the listener Bob in chunks, and Bob will decide whether to interrupt. Upon interruption, Alice halts, and Bob responds.

Figure 3: Sampling to estimate the expectation of communication cost and task performance.

		Text Pictionary		Meeting Scheduling		MMLU-Pro-Debate
Listener	Methods	SR	Cost	SR	Cost	SR	Cost
Llama-8B	Generic	$0.743_{\pm 0.033}$	$346_{\pm 37}$	$0.273_{\pm 0.077}$	$1361_{\pm 86}$	$0.520_{\pm 0.050}$	$1531_{\pm 126}$
	Concise	$0.753_{\pm 0.023}$	$277_{\pm 24}$	$0.257_{\pm 0.053}$	$1155_{\pm 80}$	$0.523_{\pm 0.090}$	$829_{\pm 176}$
	Random	$0.710_{\pm 0.033}$	$360_{\pm 22}$	$0.197_{\pm 0.040}$	$1760_{\pm 114}$	$0.537_{\pm 0.050}$	$1531_{\pm 85}$
	Prompting	$0.500_{\pm 0.033}$	$461_{\pm 25}$	$0.200_{\pm 0.020}$	$1778_{\pm 129}$	$0.537_{\pm 0.063}$	$1690_{\pm 130}$
	HandRaiser	$0.743_{\pm 0.023}$	$262_{\pm 27}$	$0.307_{\pm 0.049}$	$1042_{\pm 83}$	$0.583_{\pm 0.080}$	$782_{\pm 181}$

Table 1: Different communication methods. The “Generic” and “Concise” baselines are non-interruptible, and the others are of different interruption strategies. Results are averaged over three speaker agents, and the full results can be found in Tab. 7 . “SR”

=

Success Rate.

Task	Metric	Generic	Concise	Random	Prompting	HandRaiser
Text Pictionary	SR	$0.758_{\pm 0.028}$	$\mathbf{0.777}_{\pm 0.025}$	$0.748_{\pm 0.033}$	$0.582_{\pm 0.037}$	$0.767_{\pm 0.022}$
Text Pictionary	Cost	$373.6_{\pm 41.9}$	$301.7_{\pm 30.9}$	$389.9_{\pm 26.7}$	$492.1_{\pm 34.1}$	$\mathbf{278.1}_{\pm 37.8}$
Meeting Scheduling	SR	$0.347_{\pm 0.059}$	$0.347_{\pm 0.048}$	$0.307_{\pm 0.060}$	$0.310_{\pm 0.040}$	$\mathbf{0.377}_{\pm 0.056}$
Meeting Scheduling	Cost	$1294.2_{\pm 69.0}$	$1090.1_{\pm 80.9}$	$1672.5_{\pm 107.7}$	$1781.6_{\pm 133.0}$	$\mathbf{1026.5}_{\pm 85.0}$
MMLU-Pro-Debate	SR	$0.583_{\pm 0.032}$	$0.590_{\pm 0.074}$	$0.580_{\pm 0.047}$	$0.595_{\pm 0.068}$	$\mathbf{0.620}_{\pm 0.075}$
MMLU-Pro-Debate	Cost	$1598.2_{\pm 101.6}$	$838.1_{\pm 144.5}$	$1714.1_{\pm 118.4}$	$1952.8_{\pm 159.7}$	$\mathbf{794.4}_{\pm 150.3}$

Table 2: Fix the speakers and use different listeners. Average results across three speaker agents ( Llama-70B , Llama-405B , and Gemini-2.0-flash ). “SR”

=

Success Rate. “Generic” and “Concise” are non-interruptible baselines; others are interruptible.

	Success Ratio			Communication Cost
	TP	MS	MMLU-D	TP	MS	MMLU-D
Generic	$0.743_{\pm 0.033}$	$0.273_{\pm 0.077}$	$0.520_{\pm 0.050}$	$329_{\pm 37}$	$1361_{\pm 86}$	$1531_{\pm 126}$
Textual Pictionary	$0.743_{\pm 0.023}$	$0.283_{\pm 0.640}$	$0.537_{\pm 0.030}$	$253_{\pm 27}$	$1508_{\pm 84}$	$879_{\pm 127}$
Meeting Schedule	$0.726_{\pm 0.023}$	$0.307_{\pm 0.049}$	$0.543_{\pm 0.033}$	$261_{\pm 31}$	$1042_{\pm 83}$	$905_{\pm 163}$
MMLU-Pro-Debate	$0.726_{\pm 0.023}$	$0.273_{\pm 0.540}$	$0.583_{\pm 0.080}$	$256_{\pm 28}$	$1274_{\pm 83}$	$782_{\pm 181}$

Table 3: Performance when we apply HandRaiser trained on one task (row) to other tasks (column). These experiments are based on Llama-8B . MMLU-D refer to MMLU-Pro-Debate

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

	Train	Test	# Tractories	Inter	No-inter	Avg. #Message	Avg Length
Textual Pictionary	112	100	3010	54.1%	45.9%	5.1	425.4
Meeting Schedule	100	50	1973	62.0%	38.0%	14.7	1182.7
MMLU-Pro-Debate	200	100	4431	53.6%	46.4%	5.1	1448.1

Appendix

Table 5: Dataset Statistics. Inter and No-inter indicate the percentage of interruption and no interruption in the training trajectories. Avg. #Message is the average number of messages in the training trajectories, and Avg Length is the average length of the trajectory in words.

Figure 7: Ablation on different thresholds and chunk sizes. “SR” denotes success ratio.

		Textual Pictionary		Meeting Scheduling		MMLU-Pro-Debate
Listener	Methods	SR	Cost	SR	Cost	SR	Cost
Llama-8B	HandRaiser	$0.743$	$262$	$0.307$	$1042$	$0.583$	$782$
Llama-8B	+ Concise Speaker	$0.760$	$253$	$0.273$	$1021$	$0.563$	$763$
Llama-70B	HandRaiser	$0.790$	$294$	$0.447$	$1010$	$0.657$	$806$
Llama-70B	+ Concise Speaker	$0.780$	$279$	$0.467$	$1010$	$0.653$	$794$

Appendix

Table 6: Combine listener-oriented HandRaiser with speaker-oriented compression.

		Success Ratio					Communication Cost
		Noninterruptible		Interruptible			Noninterruptible		Interruptible
Listener	Speaker	Generic	Concise	Rand.	Prompt	HandRaiser	Generic	Concise	Rand.	Prompt	HandRaiser
Textual Pictionary
L8B	L70B	0.670	0.710	0.690	0.400	0.710	443.8	333.8	448.8	574.3	337.0
	L405B	0.740	0.790	0.710	0.500	0.760	361.1	290.6	386.4	513.3	280.5
	Gemini	0.820	0.760	0.730	0.600	0.760	232.6	206.9	245.6	296.5	168.3

Appendix

Table 7: Full Results for Tab. 1 . Notations are the same. L8B refers to llama 3 8B, L70B refers to llama 3 70B, and L405B refers to llama 3 405B.

		Success Ratio					Communication Cost
		Noninterruptible		Interruptible			Noninterruptible		Interruptible
Listener	Speaker	Generic	Concise	Rand.	Prompt	HandRaiser	Generic	Concise	Rand.	Prompt	HandRaiser
L8B	L70B	0.06	0.01	0.03	0.03	0.018	42.9	29.0	37.0	18.8	27.8
	L405B	0	0.05	0.05	0.02	0.024	40.2	28.1	17.5	32.7	32.4
	Gemini-2.0	0.04	0.01	0.02	0.05	0.028	28.8	16.3	12.6	23.6	19.5
	AVG	0.033	0.023	0.033	0.033	0.023	37.3	24.5	22.4	25.0	26.6

Appendix

Table 8: STD for Tab. 7 . Notations are the same. L8B refers to llama 3 8B, L70B refers to llama 3 70B, and L405B refers to llama 3 405B.

Multi-Agent Communication

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

See all 375 papers

What Should Agents Say? Action-state Communication for Efficient Multi-Agent Systems
Jun 3, 2026Chen Huang, Yuhao Wu, Wenxuan ZhangMulti-Agent LLM Systems Multi-Agent System Optimization
Singapore University of Technology and Design
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
Learning to Communicate: Toward End-to-End Optimization of Multi-Agent Language Systems
Apr 23, 2026Ye Yu, Heming Liu, Haibo Jin +3Multi-Agent System Optimization LLM Agent Training
University of Illinois Urbana-Champaign, IL, USA

Jun 3, 2026J/K move · Enter open · S save

What Should Agents Say? Action-state Communication for Efficient Multi-Agent Systems

Chen Huang, Yuhao Wu, Wenxuan Zhang

Singapore University of Technology and Design

Multi-Agent LLM Systems Multi-Agent System Optimization Agent Communication Protocols Multi-Agent Communication

Open paper AskPDF arXiv

Multi-agent systems (MAS) built on large language models are typically organized around roles, pipelines, and turn schedules, while the content that agents pass to one another is often left as unconstrained natural language. However, this free-form communication can rapidly inflate token usage, consume the shared context window, and ultimately affect both system performance and inference cost. We analyze five common inter-agent communication strategies across two MAS topologies, finding that no fixed strategy is universally optimal. Instead, effective inter-agent messages consistently preserve action-centered information needed by downstream agents. Building on this, we propose the PACT (Protocolized Action-state Communication and Transmission), which treats inter-agent communication as a public state-update problem and projects each raw agent output into a compact action-state record before it enters shared history. Across different MAS topologies, PACT consistently improves the performance-cost trade-off, achieving comparable or stronger task performance with substantially fewer tokens. The gains extend to production coding harnesses: PACT lifts OpenHands' resolve rate at -10% tokens-per-resolved, and is resolve-neutral on SWE-agent while halving input tokens. Our code is publicly available at https://github.com/iNLP-Lab/PACT.

Abstract

Multi-Agent Communication

Explore similar work

What Should Agents Say? Action-state Communication for Efficient Multi-Agent Systems

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

Learning to Communicate: Toward End-to-End Optimization of Multi-Agent Language Systems