LLMs learn to interrupt mid-generation to cut multi-agent chatter in half

Humans naturally interrupt conversations when they've grasped enough information—a social norm that large language models have so far ignored. Current multi-agent systems built on LLMs tackle communication efficiency by making speakers more concise, but researchers have flipped the problem: what if listeners could simply tell speakers to stop?

The problem with verbose AI#

LLM-based multi-agent systems tend toward excessive communication. Existing efficiency approaches focus on the speaker's side, compressing outgoing messages before transmission. This leaves the listener in a passive role, forced to process every word a partner generates. The result is bloated dialogue that mirrors the natural verbosity of language models rather than the efficiency of human teams.

How interruption works#

The team's core insight is to grant listeners a new power: the ability to halt a speaker mid-generation. Testing this capability revealed an immediate challenge—LLMs, given freedom to interrupt, become overconfident and cut speakers off prematurely, before hearing information necessary to complete tasks.

HANDRAISER: learning when to stop listening#

To solve this, researchers developed HANDRAISER, a learning-based method that trains agents to recognize the optimal interruption point. The approach estimates two competing factors: the expected value of hearing more (future reward) against the cost of continued listening (communication overhead). This balance allows agents to interrupt at exactly the moment they have sufficient information.

Evaluation and results#

The framework was tested across three collaborative tasks:

• Text pictionary (2-agent scenario)
• Meeting scheduling (3-agent task)
• Debate (3-agent task)

HANDRAISER achieved a 32.2% reduction in communication cost compared to baseline systems where interruption was impossible, while maintaining comparable or even better task performance. Notably, the learned interruption behavior transferred across different agent architectures and tasks, suggesting the method captures genuine principles about when to stop listening.