Learning to Interrupt in Language-based Multi-agent Communication
Danqing Wang, Da Yin, Ruta Desai, Lei Li, Asli Celikyilmaz, Ansong Ni
CoLM 2026
cs.CL
2026-04-08
Listeners interrupt mid-generation. Prompting cuts in too early; HandRaiser SFT cuts cost 32.2% vs a non-interruptible baseline, with equal or better success.
In typical LLM multi-agent setups, a listener waits until the speaker finishes a full message. Models are verbose, so context balloons, quality drops, and latency rises. Prior work mostly compresses on the speaker side: prompt the model to be concise, or train it to emit fewer tokens.
That misses a listener-specific fact. In their text Pictionary example, a 16-word description is already enough for some guessers and still useless for others. The speaker cannot know which parts a given listener still needs. Humans interrupt when they have heard enough. Current multi-agent protocols do not.
The protocol is streaming with a veto. The speaker emits fixed-size chunks (16 tokens by default). After each chunk, the listener looks at the prefix so far and decides whether to raise a hand. If it does, the speaker stops and the listener replies immediately. Round-robin and broadcast both decompose into this one-to-one pattern. If two listeners raise at once, first-come, first-served.
Prompting "interrupt when you have enough" fails. Models are overconfident. On meeting scheduling, 87.6% of prompt-based interrupts fire on the first chunk, while a planner message is typically about 3.8 chunks. The current turn gets cheaper, but the conversation grows from 11.5 messages to 15.0, so total tokens go up.
HandRaiser turns "interrupt at this chunk?" into supervised classification. For each candidate cut, they tree-sample the rest of the conversation, estimate later token cost and task success, and compare against the uncut message. A point is a positive label only if cost falls and success does not. Llama-3.1-8B and 70B-Instruct are then SFT'd at 1e-7 for 500 steps. At inference, each chunk costs one extra Yes/No token.
The label is conservative on purpose: beat "never interrupt," do not chase the globally best cut. That makes the estimate less sensitive to the rollout policy. They sample with random interruption, cap branches at 3, and take 10 rollouts per node. Chunk size during sampling is 8 for Pictionary and 16 for the other two tasks.
Three tasks: 2-agent text Pictionary (100 entities), 3-agent meeting scheduling (50 cases, one traveler and two planners with private calendars), and 3-agent MMLU-Pro debate (100 questions, pro/con plus a moderator). In the main setup only the listener may interrupt. Speakers are Llama-3.1-70B, 405B, and Gemini-2.0-flash; numbers average over speakers.
Against a non-interruptible Generic baseline, HandRaiser cuts communication cost 32.2% on average across the three tasks, with equal or higher success. Cost is total generated tokens, including the interrupt decisions.
| Task | Generic SR / cost | Concise | Prompting | HandRaiser |
| Text Pictionary | 0.758 / 373.6 | 0.777 / 301.7 | 0.582 / 492.1 | 0.767 / 278.1 |
| Meeting scheduling | 0.347 / 1294.2 | 0.347 / 1090.1 | 0.310 / 1781.6 | 0.377 / 1026.5 |
| MMLU-Pro debate | 0.583 / 1598.2 | 0.590 / 838.1 | 0.595 / 1952.8 | 0.620 / 794.4 |
Speaker-side Concise already drops debate cost from 1598.2 to 838.1. HandRaiser lands at 794.4 with success 0.620 versus Generic 0.583. Prompting and random interruption inflate round count, so they often cost more than doing nothing. For an 8B listener versus Generic, the cuts are 24.3% on Pictionary, 23.4% on scheduling, 48.9% on debate.
A listener trained on Llama-70B, 405B, and Gemini speakers still saves tokens on an unseen GPT-4o speaker, with a slight success bump. Cross-task transfer works in some cells, but same-task SFT is best: a Pictionary-trained model on scheduling raises cost from 1361 to 1508.
Stacking HandRaiser with a concise speaker trims cost a bit more. On Llama-8B, scheduling success falls from 0.307 to 0.273 and debate from 0.583 to 0.563. The two knobs are orthogonal; stacking them carelessly hurts the task.
When all three scheduling agents can interrupt (70B, no extra finetune), per-round tokens drop from 103 to 64, rounds rise from 13.39 to 20.29, total cost from 1339 to 1209, success stuck at 0.45. It runs. The saving is much thinner than the single-listener setup.
Most multi-agent token-saving work either drops edges or shortens what the speaker says. HandRaiser shows the listener can be the compressor, and can cut at a point that matches how much it already understands. Speaker-side compression cannot do that. For latency-sensitive debate or negotiation with private state, it is a path that leaves the speaker alone and adds a one-token Yes/No head on the listener.
It can stack with speaker compression. It is not a drop-in product: the setup still needs tree-sampled labels and SFT, and the main experiments only train 8B and 70B listeners.
There is no standalone Limitations section. The caveats sit in the body and appendix.
The main experiments allow only one role to interrupt. Symmetric interruption inflates rounds and thins the net saving. Tree sampling is expensive; they randomly subsample nodes down to about O(TNM). Positive labels only beat "never interrupt," so the policy is "do not lose to the full message," not "find the best cut."
Manual inspection of about 50 failures finds four patterns: cutting too early and missing a later correction; a chunk boundary splitting start time from end time, so the listener grabs half a constraint; interrupting then asking the speaker to continue, which gives the tokens back; cutting too late when chunks are large and messages are short. At chunk size 4 the model barely notices the extra words and interrupts even earlier.
The tasks are synthetic: filtered Pictionary words, scripted meeting constraints, MMLU-Pro recast as pro/con debate. There are no end-to-end numbers on software-engineering or tool-using agents. Communication cost is generated tokens as a latency proxy, assuming a local network and KV cache; they mostly ignore cross-machine network delay. Training sets are modest: 3010, 1973, and 4431 trajectories.