HANDRAISER: Interruptible Communication Framework

• HANDRAISER is an interruptible communication framework where listeners decide when to cut in mid-message, reducing token usage and latency.
• It segments speaker messages into fixed-size chunks, allowing listeners to determine sufficiency and optimize the cost-performance trade-off.
• Empirical results demonstrate up to 32.2% token savings and enhanced task success in scenarios like Text Pictionary, Meeting Scheduling, and Debate.

HANDRAISER is an interruptible communication framework for language-based multi-agent systems in which the listener, rather than the speaker, decides whether to cut into an ongoing message. It is motivated by three failures of standard multi-agent LLM communication: context overload from verbose messages, higher compute cost from unnecessary tokens, and longer latency because listeners ordinarily wait for a full message before responding. The framework treats communication as streaming generation over fixed-sized chunks, so that a listener can interrupt after receiving only a prefix of the current message if that prefix is already sufficient for an effective response. In the paper introducing the method, the name “HandRaiser” is used consistently, but the paper does not explicitly expand it as an acronym (Wang et al., 7 Apr 2026).

1. Motivation, scope, and interaction protocol

HANDRAISER is explicitly framed as a listener-oriented alternative to speaker-side compression. Earlier approaches in the paper’s discussion attempt to make speakers more concise, but the central objection is that a speaker does not know exactly which parts of a message are indispensable to a specific listener, and different listeners may need different amounts of information. The motivating example is Text Pictionary: a describer may produce the same 16-word description for multiple guessers, yet some guessers can already infer the answer while others cannot. This motivates a protocol in which the listener decides when “enough information” has arrived (Wang et al., 7 Apr 2026).

Operationally, the framework assumes that the current speaker generates a message in fixed-sized chunks and transmits each chunk as soon as it is produced. After each chunk, the listener decides whether to interrupt. If the listener interrupts, it sends an interruption signal, immediately starts generating its own response, and the speaker halts generation and waits. The paper emphasizes that interruptions are therefore neither full-turn nor sentence-level by necessity; they are chunk-level decisions. It also notes that chunk size is a hyperparameter and can range from a single token to a full message, although the experimental default is chunk size 16 unless otherwise specified (Wang et al., 7 Apr 2026).

This protocol is intended to save tokens by truncating messages early and to reduce latency by allowing the next agent to act before the current speaker has finished. A plausible implication is that HANDRAISER treats turn-taking itself as a learned control problem, rather than as a fixed conversational convention.

2. Formal communication model

The framework models a conversation among agents $A=\{a_1,\cdots,a_m\}$ across $T$ rounds as

$$C_A^{1:T} = \{X_{a_1}^1, X_{a_2}^1, \cdots, X_{a_m}^1, \cdots, X_{a_m}^T\},$$

where $X_a^t$ is the message generated by agent $a$ at round $t$ (Wang et al., 7 Apr 2026).

For a single speaking turn, the speaker’s message at round $t$ is segmented into chunks,

$X^t=\{c_0^t, \cdots, c_n^t\},$

and after the listener has received chunks up to index $i$, it has observed the prefix

$\hat{X}^t=\{c_0^t, \cdots, c_i^t\}.$

The interruption decision is evaluated after each such prefix (Wang et al., 7 Apr 2026).

Communication cost is defined as total generated tokens over the conversation,

$T$0

while task quality is measured by a normalized task-specific performance score,

$T$1

The paper then defines the effect of interrupting speaker $T$2 at round $T$3 after chunk $T$4 through changes in expected future cost and expected final performance relative to not interrupting (Wang et al., 7 Apr 2026).

The positive-payoff condition is the central decision criterion:

$T$5

An interruption point is therefore desirable only if it reduces communication cost while preserving or improving task performance. The formulation is sequential and decision-theoretic, but the paper does not formalize the problem as a full MDP with explicit per-step rewards; instead, it estimates downstream outcomes and turns the problem into supervised classification (Wang et al., 7 Apr 2026).

3. Why prompting fails and how HANDRAISER is learned

Before introducing the learned policy, the paper evaluates a direct prompt-based interruption baseline. In that baseline, after each received chunk, the listener is prompted to output whether it should interrupt. The reported failure mode is overconfidence: current LLMs tend to interrupt before they have received enough information. To probe this, the paper asks a GPT-4o listener to self-report one of five understanding levels—“not at all,” “minimal,” “partial,” “good,” and “fully”—after each chunk, and finds that even when the listener’s guess is incorrect, it still rates its understanding as “good” or “fully” in most cases (Wang et al., 7 Apr 2026).

The quantitative illustration is clearest in meeting scheduling. There, the prompt baseline interrupts in the first chunk 87.6% of the time, while a typical planning-agent message spans about 3.8 chunks. This premature interruption increases the average message count from 11.5 under the generic baseline to 15.0 under prompting. The paper’s interpretation is that local token savings are more than offset by the extra clarification rounds triggered by acting on incomplete information (Wang et al., 7 Apr 2026).

HANDRAISER addresses this by learning interruption timing from rollout-derived labels. For each potential interruption point in a speaker’s chunked message, the method constructs a tree of possible futures. Each chunk boundary defines a branch; interrupting after chunk $T$6 yields one branch, while allowing the message to continue corresponds to the no-interruption branch. Future trajectories are then sampled under a random interruption policy, and expected downstream communication cost and task performance are estimated for each branch (Wang et al., 7 Apr 2026).

Label construction is strict. A point is labeled positive only when interrupting there yields lower communication cost and comparable or better task performance than not interrupting. These labeled interruption points are converted into instruction-format training data whose target is “Yes” or “No.” The interruption predictor—called HandRaiser in the paper—is then trained by supervised finetuning with cross-entropy loss. The reported training configuration is finetuning Llama-3.1-$T$7-Instruct with learning rate $T$8 for 500 steps (Wang et al., 7 Apr 2026).

The data-generation procedure uses chunk size 8 for Text Pictionary during sampling and chunk size 16 for the other tasks, with maximum branch size 3, rollout number 10, maximum 10 rounds per rollout, at most 10 non-terminating speaking nodes per case, and temperature 1 during sampling. At inference, the paper uses temperature 0.7 and one interruption decision per chunk (Wang et al., 7 Apr 2026).

4. Experimental domains, models, and baselines

The framework is evaluated in three multi-agent scenarios. The first is 2-agent Text Pictionary, in which a describer provides clues and a guesser must identify the secret entity; the paper reports 100 evaluation entities. The second is 3-agent Meeting Scheduling, involving one traveler and two planning agents with private constraints; the paper reports 50 evaluation tasks. The third is 3-agent MMLU-Pro Debate, in which pro and con debaters argue for correct and incorrect answers and a moderator chooses a side; the paper reports 100 evaluation instances (Wang et al., 7 Apr 2026).

The primary metrics are Success Rate and Communication Cost. Communication Cost is total generated tokens; Success Rate is task completion quality, instantiated by task-specific criteria such as correct guessing, valid schedule construction, or correct moderator choice (Wang et al., 7 Apr 2026).

The main listener models are Llama-3.1-8B-Instruct and Llama-3.1-70B-Instruct. Speakers are averaged across Llama-3.1-70B-Instruct, Llama-3.1-405B-Instruct, and Gemini-2.0-flash, and the paper also reports transfer to GPT-4o as an unseen speaker. Baselines include two non-interruptible methods—Generic and Concise—and three interruptible methods—Random, Prompting, and HandRaiser (Wang et al., 7 Apr 2026).

The baseline structure is important for interpretation. Generic is the non-interruptible reference system. Concise is a speaker-side compression baseline. Random interrupts without payoff estimation. Prompting performs interruption through direct prompting. HandRaiser is the learned listener-side interruption policy. This experimental design isolates whether gains come from interruption per se, from speaker compression, or from learned interruption timing (Wang et al., 7 Apr 2026).

5. Empirical performance

The headline result in the abstract is that HANDRAISER reduces communication cost by 32.2% compared to the baseline while maintaining comparable or superior task performance (Wang et al., 7 Apr 2026). The paper also reports task-level reductions against the Generic baseline: 24.3% on Text Pictionary, 23.4% on Meeting Scheduling, and about 46.2% to 48.9% on Debate, depending on the averaging context reported in the text (Wang et al., 7 Apr 2026).

Listener	Task	Generic SR / Cost	HandRaiser SR / Cost
Llama-8B	Text Pictionary	$T$9 / $$C_A^{1:T} = \{X_{a_1}^1, X_{a_2}^1, \cdots, X_{a_m}^1, \cdots, X_{a_m}^T\},$$0	$$C_A^{1:T} = \{X_{a_1}^1, X_{a_2}^1, \cdots, X_{a_m}^1, \cdots, X_{a_m}^T\},$$1 / $$C_A^{1:T} = \{X_{a_1}^1, X_{a_2}^1, \cdots, X_{a_m}^1, \cdots, X_{a_m}^T\},$$2
Llama-8B	Meeting Scheduling	$$C_A^{1:T} = \{X_{a_1}^1, X_{a_2}^1, \cdots, X_{a_m}^1, \cdots, X_{a_m}^T\},$$3 / $$C_A^{1:T} = \{X_{a_1}^1, X_{a_2}^1, \cdots, X_{a_m}^1, \cdots, X_{a_m}^T\},$$4	$$C_A^{1:T} = \{X_{a_1}^1, X_{a_2}^1, \cdots, X_{a_m}^1, \cdots, X_{a_m}^T\},$$5 / $$C_A^{1:T} = \{X_{a_1}^1, X_{a_2}^1, \cdots, X_{a_m}^1, \cdots, X_{a_m}^T\},$$6
Llama-8B	MMLU-Pro Debate	$$C_A^{1:T} = \{X_{a_1}^1, X_{a_2}^1, \cdots, X_{a_m}^1, \cdots, X_{a_m}^T\},$$7 / $$C_A^{1:T} = \{X_{a_1}^1, X_{a_2}^1, \cdots, X_{a_m}^1, \cdots, X_{a_m}^T\},$$8	$$C_A^{1:T} = \{X_{a_1}^1, X_{a_2}^1, \cdots, X_{a_m}^1, \cdots, X_{a_m}^T\},$$9 / $X_a^t$0
Llama-70B	Text Pictionary	$X_a^t$1 / $X_a^t$2	$X_a^t$3 / $X_a^t$4
Llama-70B	Meeting Scheduling	$X_a^t$5 / $X_a^t$6	$X_a^t$7 / $X_a^t$8
Llama-70B	MMLU-Pro Debate	$X_a^t$9 / $a$0	$a$1 / $a$2

These results show a consistent pattern: HANDRAISER lowers cost in every reported main setting and usually preserves or improves success rate. The gains are especially large in long-context debate, where the moderator can often determine a preference before the debater finishes generating. In meeting scheduling, the method is notable not only for reducing cost but also for improving success rate relative to Generic for both 8B and 70B listeners (Wang et al., 7 Apr 2026).

The comparison with Prompting is equally informative. Prompting often increases cost and can substantially reduce success rate. For example, with the 8B listener in Text Pictionary, Prompting reaches Success Rate $a$3 and Cost $a$4, compared with HandRaiser at Success Rate $a$5 and Cost $a$6. Random interruption is less pathological than Prompting but still inferior because it lacks an estimated cost-performance tradeoff (Wang et al., 7 Apr 2026).

The comparison with Concise is more nuanced. Concise often also reduces cost, but it is a speaker-side method. The paper does not claim that HandRaiser uniformly dominates Concise; instead, it treats listener interruption and speaker compression as orthogonal mechanisms.

6. Generalization, sensitivity, and limitations

The paper reports that HANDRAISER generalizes across speakers. A listener trained on interactions with Llama-70B, Llama-405B, and Gemini transfers to GPT-4o in Text Pictionary, with the qualitative conclusion that it achieves slightly better performance at lower cost and adapts to different speaker styles. This suggests that the learned policy is not tightly coupled to one source model’s verbosity profile (Wang et al., 7 Apr 2026).

Cross-task transfer is weaker than same-task training but still meaningful. For Llama-8B, a model trained on Meeting Scheduling and evaluated on MMLU-Pro Debate obtains Success Rate $a$7 and Cost $a$8, compared with the Generic baseline at Success Rate $a$9 and Cost $t$0. The paper states that same-task training performs best, which is consistent with the claim that optimal interruption timing is task-specific (Wang et al., 7 Apr 2026).

The method also transfers to a more complex communication pattern in which all three agents in Meeting Scheduling are interruptible. Without additional finetuning, the 70B HandRaiser preserves performance at $t$1 Success Rate while lowering cost from $t$2 to $t$3. However, this comes with a notable increase in the number of communication rounds from 13.39 to 20.29. The paper also reports that per-round cost drops from 103 tokens to 64 tokens in this setting. This suggests that interruption can compress individual turns while still fragmenting the conversation into more turns (Wang et al., 7 Apr 2026).

Sensitivity analysis indicates that stricter interruption thresholds make the policy more conservative, while communication cost follows a U-shaped curve: excessively permissive thresholds interrupt too early and increase repair rounds, whereas overly strict thresholds miss useful interruptions. The paper also finds that smaller chunk sizes do not help, because they increase interruption checks while providing too little semantic change between decisions. This suggests that semantically meaningful chunking is preferable to token-level granularity (Wang et al., 7 Apr 2026).

The principal limitation is the cost of label generation. The appendix gives a rough worst-case complexity of $t$4 for exhaustive tree sampling and then describes a practical approximation with random sampling that reduces cost to $t$5. The policy is therefore inexpensive at inference but expensive to train. Additional limitations are explicit: the framework assumes the speaker always halts when interrupted, it optimizes to outperform the no-interruption baseline rather than a globally optimal interruption policy, and performance remains task-specific (Wang et al., 7 Apr 2026).

Taken together, HANDRAISER establishes interruption timing as a learnable component of multi-agent language systems. Its broader significance lies in shifting communication control from universal speaker compression toward listener-contingent stopping rules. A plausible implication is that future multi-agent systems may treat turn boundaries, partial information sufficiency, and conversational floor transfer as first-class optimization targets rather than as fixed interface assumptions.