acceptodds
Withdrawn from ICLR 2027

RhyCo: Rhythm-Controlled Coordination for Multi-Agent LLM Reasoning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

In current multi-agent LLM systems, orchestration happens only between turns, so an inactive agent has no structured channel to intervene before being scheduled, even when it could provide decisive evidence that the ongoing reasoning has gone wrong. We name the missing property communication rhythm: the ability to listen, stay silent, and proactively intervene at a self-adaptive pace, as effective teams do. We propose RhyCo (Rhythm-Controlled Coordination), a framework that adds communication rhythm to multi-agent systems (MAS) by moving orchestration inside the turn. RhyCo introduces a dual-channel speaker-listener interface that while the speaker streams reasoning chunks, each listener exposes a probabilistic signal at every chunk and decodes a compact structured whisper only on escalation. A two-tier orchestrator then decides whether to stay silent, retain the speaker, or hand control to a listener, with an MLP gate handling routine chunks and an LLM arbitrator engaged only on escalation. Experiments show that unrestricted listener intervention can degrade MAS performance, while RhyCo turns the same proactive capability into consistent gains on math, code, and multi-hop benchmarks. On MATH-L5, RhyCo-Base uses about the output tokens of a single agent and RhyCo-RL uses about , compared to for Multi-Agent Debate and for Self-Consistency at five samples.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Large Language Models, Multi-Agent Systems, Multi-Agent Reasoning, Agent Communication, Agent Coordination

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_jufLOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces chunk level coordination for multi-agent LLM reasoning. While one agent generates a reasoning chunk, other agents provide probability-based assessments. A trained MLP gate decides whether to invoke an arbitrator, which can retain the speaker or discard the chunk and transfer control. RhyCo-Base uses a prompted arbitrator; RhyCo-RL additionally trains it with GRPO. Experiments on six math, code, and multi-hop QA benchmarks report improved trade offs between accuracy and cost.

Critical strengths

--- Meaningful coordination problem. Separating routine monitoring from selective arbitration is a reasonable way to regulate intervention while keeping the reasoning backbones fixed.

--- Useful empirical coverage. The evaluation involved six benchmarks and includes adaptive methods such as AFlow and MaAS. The ablations and trajectory examples provide useful evidence about the risks of excessive intervention.

Critical weaknesses

--- The contribution appears incremental, and its distinctive value is insufficiently isolated. Learned coordination has precedents in Puppeteer and ECON, while intermediate checking and rollback are explored in GA-Rollback; Meta-R1 also studies selective monitoring during generation. RhyCo’s contribution appears to lie primarily in combining listener signals, learned gating, arbitration, and speaker handoff. However, the “w/o metrics” ablation also removes the gate and invokes the arbitrator at every chunk. Its degradation therefore does not establish the specific benefit of listener-driven timing over simpler selective intervention.

--- The benefit attributed to training is confounded with additional computation. RhyCo-RL receives three times the chunk budget of RhyCo-Base, while their performance difference is attributed to GRPO. Figure 3 provides useful cost comparisons but does not isolate the training effect at matched budgets. Moreover, Appendix A.1 excludes listener prefill from the training cost penalty, and whether Figure 3 fully accounts for this monitoring overhead is unclear.

Questions for authors

--- What is the specific advance over learned orchestration and online intervention methods, including HANDRAISER? I think the setting is very simlart with Rhyco. Can its benefit be isolated against a simpler controller with comparable intervention frequency and inference cost?

--- How much of the RL improvement remains when both policies use identical chunk budgets?

--- Does Figure 3 include all listener prefill, discarded chunks, whisper generation, and arbitrator computation? Does the reported efficiency advantage persist under this complete accounting?

5 comments
AuthorsOpenReview ↗

Response to Reviewer jufL (1/2)

We thank Reviewer jufL for a constructive review, whose questions on HANDRAISER, simpler controllers, and matched budgets led us to add direct comparisons. We answer in four parts.


1: Novelty (W1, Q1)

We appreciate why the reviewer sees connections between RhyCo and the cited works. Recent work has made substantial progress in adaptive orchestration, online monitoring, interruption, recovery, and efficient coordination, opening a broader question: can multi-agent systems communicate with the flexibility of a human team? Human coordination is not rigid turn-taking: people listen while others speak, usually remain silent, intervene when useful, and flexibly hand control to whoever is best positioned to contribute. Our goal is the same adaptive freedom with bounded coordination cost. We call this balance communication rhythm.

The cited works also reveal why this broader communication problem requires joint system design. Puppeteer places dynamic agent selection in a centralized orchestrator, while HANDRAISER gives listeners learned control over when to interrupt; Meta-R1 introduces explicit online meta-level regulation, whereas ECON reduces costly direct exchanges through belief-driven coordination; and GA-Rollback responds to detected errors by rewinding the trajectory. These are effective choices in their respective settings, but they give different answers to three coupled questions: where control resides, how information is exposed during reasoning, and what happens once intervention is triggered. Once these choices operate within the same communication process, changing one also changes the state, cost, and decision context faced by the others. The problem is therefore one of joint system design: these coupled choices must be resolved under a common quality–cost objective.

RhyCo addresses these coupled choices through a system-level design of communication rhythm that jointly structures information flow, intervention timing, and floor control. Listeners can raise concerns and recommend intervention, but they cannot directly take over the reasoning floor; the system decides whether a handoff is needed and which specialist should take over.

W1: Incremental vs Puppeteer, ECON, GA-Rollback, Meta-R1

Our ablations already show that each part of RhyCo matters, since rule-based gating, removing whispers, or an untrained arbitrator all lower performance. We test each cited mechanism that maps onto a RhyCo decision inside RhyCo. Puppeteer's centralized selection maps onto an every-chunk controller, which performs poorly: Table 2's Unrestricted Intervention averages 19.2 against 61.3 for RhyCo-Base, and the gate-free arbitrator below collapses at small sizes. ECON, GA-Rollback and Meta-R1 act on other decisions, described above, with no direct counterpart in RhyCo.

Q1: Advance over HANDRAISER?

HANDRAISER and RhyCo share the premise that intervention timing should be learned from rollout outcomes; they differ in who decides. In HANDRAISER, a listener decides on its own to cut off a message in conversation, and in its main experiments only one agent per task may interrupt. In RhyCo, several listeners monitor one reasoning chain and may want to take the floor at once.

To test whether a central decision matters, we remove the gate and arbitrator and let each listener, prompted after every chunk, interrupt and take the floor on its own. It collapses to 17.9% on MATH-L5 because nearly every listener finds a problem and interrupts, as HANDRAISER also found for untrained listeners; Table 2's Unrestricted Intervention, which also intervenes at every chunk but centrally, reaches 31.1%, so central control does less damage. A trigger trained on rollout payoffs would instead fall silent, since competing interruptions mostly lower the payoff, reducing it to the single-agent baseline: 60.9% against 66.9% for RhyCo-Base. Decentralized listeners thus either interrupt too often or fall silent; RhyCo avoids both by separating reporting from deciding.


2: Gate and timing (W1, Q1)

W1: "w/o metrics" also removes the gate; isolate listener-driven timing?

Gate vs. metrics (W1). Removing only the gate, a prompted arbitrator reads the metrics but decides at every chunk:

ConfigBigCodeBench (100)MATH-L5 (100)
Single44.056.0
SC52.063.0
Gate-free, 0.6B52.0—
Gate-free, 1.7B60.020.0
Gate-free, 4B62.060.0
Gate-free, 8B62.064.0

On BigCodeBench, the metrics carry the gate-free arbitrator, which beats SC by 8–10 points from 1.7B up, whereas the same 1.7B arbitrator without the metrics switches constantly and scores 0.0% (Table 2, 570 problems). On MATH-L5, the metrics overwhelm the 1.7B arbitrator, which interrupts 8 or more times per query and falls to 20%; only the 8B arbitrator regulates itself, at 64%. The gate is what lets a small arbitrator interrupt with discipline.

AuthorsOpenReview ↗

Response to Reviewer jufL (2/2)

Q1: Beat a simpler controller at matched frequency and cost?

Simpler controllers (Q1). We compare RhyCo with three simpler controllers: the gate-free arbitrator above, which decides at every chunk and is compared with RhyCo-Base; a random-timing control, which keeps the number of escalations that RhyCo-RL's learned gate makes on each problem but places them at random chunks; and Table 2's rule gate, which replaces RhyCo-Base's learned gate with hand-crafted rules over the same rhythm state. All three fall clearly below the corresponding learned gate, the first two at a comparable number of interventions, so RhyCo's advantage comes from the gate's choice of when to intervene.

On MATH-L5, the gate-free arbitrator interrupts about 1.8 times per query at 4B and 8B, close to RhyCo-Base's roughly 1.6, but on the same 100 problems it reaches only 60% and 64% against 69% for RhyCo-Base with a 1.7B arbitrator, and it costs more because it consults the arbitrator at every chunk. The random-timing control, with the arbitrator still receiving the listener metrics and whispers, lowers RhyCo-RL's accuracy from 74.1% to 68.1%; because the trained arbitrator can still keep the current speaker on an unneeded escalation, this 6.0-point gap is a conservative estimate of what learned timing adds at the same cost. The rule gate falls 5.8 to 8.9 points below RhyCo-Base across MATH-L5, BigCodeBench and HotpotQA. In every comparison the learned gate leads by 5 to 9 points, which is the benefit of listener-driven timing that the reviewer asked us to isolate.


3: RL at matched budgets (W2, Q2)

W2, Q2: RL gain at identical chunk budgets?

Matched budgets. We run the control in both directions: RhyCo-Base with RL's 90-chunk budget, and RhyCo-RL trained and evaluated with Base's 30 chunks. Gains are relative to RhyCo-Base at 30 chunks:

ConfigBudgetMATH-L5 (%)HotpotQA (EM, %)
RhyCo-Base3066.8762.00
RhyCo-Base9068.52 (+1.65)62.00 (+0.00)
RhyCo-RL3067.28 (+0.41)62.60 (+0.60)
RhyCo-RL9074.07 (+7.20)64.00 (+2.00)

Most of the RL improvement remains under matched inference budgets. On MATH-L5, 5.6 of the original 7.2 points (78%) remain when RhyCo-Base and RhyCo-RL are both evaluated with 90 chunks; on HotpotQA, the matched-budget gain is 2.0 points. At the tighter 30-chunk budget, the corresponding gains are 0.4 and 0.6 points, respectively. The smaller gains at 30 chunks reflect the limited room for the trained arbitrator to act before trajectories reach the budget: it often continues checking an answer, after which truncated trajectories fall back to an answer aggregated from the committed history. RhyCo-Base itself benefits little from increasing the budget from 30 to 90 chunks on MATH-L5, only 8 of 486 problems change, all of which were truncated at 30 chunks, yielding a 1.6-point gain; on HotpotQA, no trajectory reaches 30 chunks, so accuracy is unchanged.


4: Cost accounting (W2, Q3)

W2, Q3: A.1 excludes prefill; is all monitoring cost in Figure 3?

We thank the reviewer for this question. Figure 3 already includes all listener prefill, discarded speaker chunks, whisper generation, and arbitrator computation in the reported inference cost. Appendix A.1 describes a different quantity, the token-cost term of the GRPO training reward, which excludes listener prefill and affects only RL training; we will state this distinction explicitly in the revision.

Complete inference accounting. Figure 3(left) counts output tokens only, as labeled, and Figure 3(right) converts output and input tokens into output-token equivalents at 1:8. The table itemizes the per-query cost behind Figure 3(right) on MATH-L5; speaker decoding includes chunks later discarded by a switch.

MATH-L5, per queryRhyCo-BaseRhyCo-RL
Speaker decoding840 out1,985 out
Whisper decoding96 out274 out
Arbitrator output240 out1,100 out
Listener prefill (3 × 64 per chunk)2,520 in5,955 in
Arbitrator input1,176 in3,359 in
History at handoff1,252 in3,610 in
Initial prompts5,100 in5,100 in
Total1,176 out / 10,048 in → 2,4323,359 out / 18,024 in → 5,612

Efficiency under complete accounting. At every input/output price ratio from 1:4 to 1:100, RhyCo-Base costs less than every multi-agent baseline; at the 1:8 ratio of Figure 3, it uses 2,432 output-token units against 5,221 to 9,490 for AFlow, SC, MaAS and MAD. RhyCo-RL, the most accurate method at 74.1%, uses 5,612 units at 1:8, below SC, MaAS and MAD. The efficiency advantage therefore persists under complete accounting, and the revision will make the training-vs.-inference distinction explicit and add the component breakdown and price-ratio sensitivity.


We hope these analyses address the reviewer's concerns, and we would be glad to clarify any remaining questions.

Thank you for the detailed rebuttal and additional experiments. The token breakdown and gate-timing ablations address my questions regarding cost accounting and the role of the MLP gate. For me, the main takeaway from the paper is that separating continuous listener assessment from selective judgement can turn multi agent feedback into better reasoning.

However, I am still unsure whether they establish a sufficiently substantial contribution beyond existing learned methods such as HANDRAISER. In addition, under the original 30 chunk budget, RhyCo-RL yields only marginal gains over RhyCo-Base (+0.41% on MATH-L5 and +0.60% on HotpotQA), which means that the RL improvement relies largely on the expanded 90 chunk horizon rather than a clear gain in coordination efficiency at tight budget.

Although likely beyond the scope of this rebuttal, establishing a broader contribution would require a study across tasks that systematically varies listener count and disagreement, comparing trained centralized and decentralized controllers under matched budgets to show when and why central judge helps.

Reviewer_xQuaOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

RhyCo proposes dynamic intra-turn interruption for multi-agent LLM reasoning. A learned gate uses model-internal signals to decide when another agent should intervene. The method is evaluated on math, coding, and multi-hop reasoning tasks.

Critical strengths

The intra-turn interruption setting is interesting and goes beyond standard turn-based multi-agent coordination. The evaluation also covers several reasoning domains.

Critical weaknesses

The method depends heavily on token-level logits and confidence signals, which are unavailable for many black-box models.

More importantly, frequent intervention can substantially hurt performance. The collapse on some tasks suggests that the learned gate does not reliably control over-deliberation.

The reported efficiency also excludes the cost of Claude Opus annotation and counterfactual rollouts used to train the gate.

Questions for authors

1.How would RhyCo work when token-level logits are unavailable?

  1. Can the authors explain or mitigate the performance collapse at high intervention frequency?

  2. What is the total offline annotation and rollout cost?

2 comments
AuthorsOpenReview ↗

Response to Reviewer xQua (1/2)

We appreciate Reviewer xQua's constructive review. Frequent intervention, black-box deployment, and offline cost are fair tests for RhyCo, and we take them up in three parts.


1: Over-deliberation and performance collapse (W2, Q2)

W2, Q2: Performance collapse under frequent intervention?

We sincerely thank the reviewer for highlighting this concern. We understand why Figure 4 can give the impression that increasing intervention frequency leads to a collapse in performance: the x-axis is the number of speaker switches, several curves decline in the high-switch region, and connecting the buckets with lines makes the plot visually resemble a controlled intervention sweep. Together with our use of the term “over-deliberation,” this presentation can naturally suggest that RhyCo is intervening too frequently and that the gate may be failing to regulate this behavior. This is an important ambiguity to resolve.

The key distinction is that the switch count in Figure 4 is not a prescribed intervention budget, but the emergent behavior of RhyCo on each query. The figure is therefore observational rather than an intervention sweep: query difficulty and emergent switch count vary together, so the decline in accuracy across high-switch buckets cannot by itself be interpreted as the causal effect of increasing intervention frequency. To separate these effects more directly, we add a difficulty-stratified analysis and a complementary controller ablation below, and in the revision we will replace Figure 4 with this difficulty-stratified view so that it no longer reads as an intervention sweep.

The difficulty-stratified analysis uses the empirical SC-5 solve rate to define Easy (5/5), Medium (3–4/5), Hard (1–2/5), and Extremely Hard (0/5) groups. Across MATH-L5, HotpotQA, and BigCodeBench, the mean emergent number of speaker switches increases with difficulty, confirming that high-switch trajectories are disproportionately concentrated on harder queries. We further report Rescued cases (CoT wrong → RhyCo-RL correct), Lost cases (CoT correct → RhyCo-RL wrong), and Δ=Rescued−Lost\Delta=\mathrm{Rescued}-\mathrm{Lost} within each difficulty group. On MATH-L5 and HotpotQA, Δ\Delta remains positive across the difficulty tiers and is largest on the harder groups. Thus, a high emergent switch count is not itself evidence of harmful over-deliberation: difficult queries both require more coordination and remain intrinsically harder to solve.

BenchmarkDifficulty tier# QueriesMean switchesSingleCoTMADRhyCo-RLRescuedLostΔ
MATH-L5Easy (SC 5/5)2121.3490.193.499.599.1120+5.7
Medium (SC 3–4/5)1152.8069.674.887.084.3176+9.6
Hard (SC 1–2/5)635.5136.533.357.161.9213+28.6
Extremely Hard (SC 0/5)969.682.10.07.314.6140+14.6
HotpotQAEasy (SC 5/5)4001.5069.599.098.899.542+0.5
Medium (SC 3–4/5)712.1045.174.674.681.71813+7.0
Hard (SC 1–2/5)432.6020.930.244.253.5166+23.3
Extremely Hard (SC 0/5)2863.184.90.31.711.5320+11.2

When does over-deliberation emerge? It emerges when no gate filters interventions and an LLM controller decides at every chunk. On the same 100 MATH-L5 problems, RhyCo-Base, whose gate passes only selected chunks to a 1.7B arbitrator, reaches 69%. Removing the gate while keeping the listener metrics leaves the same arbitrator deciding at every chunk: it interrupts eight or more times per query, even on some simple problems, the reasoning never converges, and accuracy falls to 20%. Removing the metrics as well does not prevent the collapse, and on BigCodeBench this variant scores 0.0%. The gate is therefore what keeps intervention selective: listeners may monitor continuously, while the gate determines when that evidence warrants expensive LLM-level intervention. In the revision, we will use “over-deliberation” only for this non-convergent regime of repeated intervention.

AuthorsOpenReview ↗

Response to Reviewer xQua (2/2)

2: Dependence on token-level logits (W1, Q1)

W1, Q1: How does RhyCo work without token-level logits?

RhyCo works with a black-box speaker, and limits on black-box listener. Of the twelve gate features, only the speaker's token confidence conf_min requires the speaker's logits. The three other speaker-side features are computed from the text, and the eight listener metrics come from the listeners' own logits over the speaker's tokens, which every closed API returns. A closed-API speaker monitored by open-weight listeners therefore loses only conf_min. We test this setting and the fully closed one on RhyCo-RL by withholding the corresponding features:

“Speaker closed” removes the speaker’s token-confidence metric (conf_min), the only logit-based speaker-side metric; “Both closed” additionally removes the eight logit-derived listener metrics, leaving the text-derived surface features and whisper channel unchanged.

StateAIME 2024 (30)MATH-L5 (100)HotpotQA (100)Avg.
Full 12-D43.373.067.061.1
Speaker closed40.074.062.058.7
Both closed16.746.058.040.2

With the speaker closed, the average drops by only 2.4 points, from 61.1 to 58.7, so RhyCo works with a black-box speaker. When the listeners are closed as well, the gate keeps only the surface features and the whisper channel, and the average falls to 40.2: without the listener metrics, the gate escalates far more often, and the extra interruptions cost accuracy. RhyCo therefore requires open-weight listeners, which need only the speaker's output text; a fully black-box deployment remains a limitation, which we will state in the revision.


3: Offline annotation and rollout cost (W3, Q3)

W3, Q3: Total offline cost of annotation and rollouts?

We agree that the offline cost of building the gate's labels should be reported alongside the inference cost in Figure 3. It is paid once per benchmark, does not grow with the number of test queries, and is small because the training pools are small and only the failed Single trajectories are annotated.

Annotation. For each failed Single trajectory, Claude Opus 4.6 receives the problem and the trajectory split into numbered chunks (64 tokens; 256 for code) and returns only the index of the first wrong chunk. Each trajectory is annotated three times, and we checked all annotated chunks manually. GSM8K's and AIME's own training pools yield few failed trajectories, so their gates are trained on them together with the MATH-L5 labels; RhyCo-RL's GRPO runs on every training query, correct or not, with an outcome-minus-cost reward and no annotation.

BenchmarkTraining problemsFailed trajectoriesCalls (×3)Input tokens per callTotal input tokensTotal output tokensList-price cost
MATH-L511930902.0k≈0.18M0.6k$0.9
GSM8K1008240.4k0.01M0.2k$0.05
BigCodeBench143802400.4k0.09M1.5k$0.5
Subtotal0.28M2.3k$1.5

The table details three benchmarks; summed over all six benchmarks, the annotation costs about $2.6 in total. Costs use Claude Opus 4.6's list price of 5 USD per million input tokens and 25USD per million output tokens (Nov. 2026).

Counterfactual rollouts. At the annotated onset of every failed trajectory, the criticizer, the verifier, and the metareviewer each take over and continue to completion. The original trajectory was produced by the generator, so each failed trajectory needs only three counterfactual continuations, for example 90 for the 30 failed MATH-L5 trajectories. They run on the local Qwen3-8B, so they use no API tokens, and collecting them takes a few minutes per benchmark. Amortized over the 3,546 test queries of the six benchmarks, the annotation adds less than $0.001 per query. We will report this offline cost next to Figure 3 in the revision.


We hope these analyses address the reviewer's concerns, and we would be glad to clarify any remaining questions.

Reviewer_wpTxOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes RhyCo, a framework for within-turn coordination in multi-agent LLM reasoning. It introduces a dual-channel speaker-listener interface in which inactive agents continuously assess the active speaker and provide structured feedback when intervention is needed. A two-tier orchestrator combines a lightweight learned gate for intervention timing with an LLM arbitrator for speaker selection. The authors further optimize the arbitrator using GRPO. Experiments across mathematical reasoning, code generation, and multi-hop QA demonstrate competitive performance against existing multi-agent approaches.

Critical strengths
  • RhyCo extends multi-agent orchestration to intermediate reasoning steps, enabling proactive error detection and adaptive speaker switching. The separation of intervention timing and speaker selection is a sensible design that balances coordination flexibility with computational overhead.
  • The ablation studies demonstrate that unrestricted intervention can severely degrade performance and that selective intervention is essential. Comparisons across multiple reasoning domains also suggest that the framework is broadly applicable.
Critical weaknesses
  • The novelty and independent contribution of the listener-driven mechanism remain unclear. Prior work has explored listener-initiated interruption and adaptive agent coordination. RhyCo combines existing techniques into a more structured framework, but the experiments do not clearly establish which improvements arise specifically from its proposed listener assessments. In particular, removing listener metrics also removes the gate, confounding the contributions of these components. The relatively limited improvements from structured whispers further weaken the evidence for the proposed dual-channel interface.
  • The efficiency claims are insufficiently supported by computational cost measurements. The framework requires multiple listeners to evaluate every reasoning chunk, potentially introducing substantial prefill overhead. However, the cost analysis does not clearly account for all listener computations. Consequently, the reported token efficiency may overstate the actual computational advantage, which is central to the paper's motivation.
  • The contribution of RL optimization is confounded by differences in inference budgets. RhyCo-RL uses a substantially larger chunk budget than RhyCo-Base. Without a comparison under matched inference budgets, it is difficult to determine whether the improvements result from training the arbitrator or simply allowing more reasoning and intervention opportunities.
Questions for authors
  • Can the authors isolate the contribution of listener assessments from that of the learned intervention gate, for example by retaining the gate while using only speaker-side features?
  • Does the inference cost analysis include all listener prefill computations? What are the end-to-end latency and total computational costs relative to the strongest baselines?
  • How much improvement does GRPO provide when RhyCo-Base and RhyCo-RL are evaluated under identical inference budgets?
  • How are the role-specific listener features handled when the corresponding agent becomes the active speaker?
2 comments
AuthorsOpenReview ↗

Response to Reviewer wpTx (1/2)

Thank you for a careful and detailed review. Your suggestions to isolate the listener metrics, count prefill, and match chunk budgets each led to a new analysis, which we present in four parts.


1: Novelty (W1, Q1)

W1: Unclear novelty; combines existing techniques

We appreciate why the reviewer sees connections between RhyCo and prior work on listener-initiated interruption and adaptive agent coordination. Recent work has made substantial progress in adaptive orchestration, online monitoring, and interruption, opening a broader question: can multi-agent systems communicate with the flexibility of a human team? Human coordination is not rigid turn-taking: people listen while others speak, usually remain silent, intervene when useful, and flexibly hand control to whoever is best positioned to contribute. Our goal is the same adaptive freedom with bounded coordination cost. We call this balance communication rhythm.

Listener-driven interruption makes coordination more responsive, but effective multi-agent coordination also requires global control over when intervention happens and who takes the floor. RhyCo resolves this tension through a system-level design of communication rhythm that jointly structures information flow, intervention timing, and floor control. Non-speaking agents continuously provide lightweight evidence about the ongoing reasoning process, while richer semantic communication and LLM arbitration are activated selectively through escalation. The system then determines when intervention is warranted and whether the current speaker should retain the floor or another specialist should take over.

Q1: Isolate listener assessments from the learned gate

W1: "w/o metrics" also removes the gate

Because RhyCo gives listener-provided information and system-level control distinct roles, the reviewer's concern becomes an attribution question: how much does the listener information itself contribute? Table 2's "w/o metrics" removes the metrics together with the gate that consumes them, so we add two controls that each change only one of the two.

First, we keep the gate–arbitrator controller of RhyCo-RL fixed and close the information sources one at a time: the average over AIME 2024, MATH-L5 and HotpotQA falls from 61.1 to 58.7 without the speaker's token confidence, and to 40.2 once the listener metrics are removed as well. Under the same controller, listener-derived metrics have a substantially larger effect than the speaker-side metric.

“Speaker closed” removes the speaker’s token-confidence metric (conf_min), the only logit-based speaker-side metric; “Both closed” additionally removes the eight logit-derived listener metrics, leaving the text-derived surface features and whisper channel unchanged.

StateAIME 2024 (30)MATH-L5 (100)HotpotQA (100)Avg.
Full 12-D43.373.067.061.1
Speaker closed40.074.062.058.7
Both closed16.746.058.040.2

Second, we keep the listener metrics and remove the gate: a prompted arbitrator reads the metrics but must decide at every chunk, at sizes from 0.6B to 8B:

ConfigBigCodeBench (100)MATH-L5 (100)
Single44.056.0
SC52.063.0
Gate-free, 0.6B52.0—
Gate-free, 1.7B60.020.0
Gate-free, 4B62.060.0
Gate-free, 8B62.064.0

On BigCodeBench, the metrics carry the gate-free arbitrator: from 1.7B up it beats SC by 8–10 points, whereas the same 1.7B arbitrator without the metrics (Table 2, "w/o metrics") switches constantly, fragments the program, and scores 0.0% (on the full 570 problems; Single scores 44% on both sets). On MATH-L5, without the gate, the metrics overwhelm the 1.7B arbitrator: it interrupts 8 or more times per query and falls to 20%, and only the 8B arbitrator regulates itself (1.85 interruptions, 64%). The gate is what lets a small arbitrator interrupt with discipline.

W1: Whisper gains are small

The dual-channel interface addresses a related information–cost trade-off: lightweight metrics provide continuous monitoring, while structured whispers provide richer semantic context only after escalation. In the clean ablation, whispers contribute 0.6 points in average performance. Their communication footprint is sparse: routine chunks generate no whisper tokens, and each escalated whisper is capped at 24 tokens. The two channels therefore occupy different operating points in the information–cost trade-off.

AuthorsOpenReview ↗

Response to Reviewer wpTx (2/2)

2: Matched budgets (W3, Q3)

W3, Q3: GRPO gain at identical chunk budgets?

We run the control in both directions: RhyCo-Base with RL's 90-chunk budget, and RhyCo-RL trained and evaluated with Base's 30 chunks. Gains are relative to RhyCo-Base at 30 chunks:

ConfigBudgetMATH-L5 (%)HotpotQA (EM, %)
RhyCo-Base3066.8762.00
RhyCo-Base9068.52 (+1.65)62.00 (+0.00)
RhyCo-RL3067.28 (+0.41)62.60 (+0.60)
RhyCo-RL9074.07 (+7.20)64.00 (+2.00)

Under identical inference budgets, GRPO improves RhyCo-Base by 5.6 points on MATH-L5 and 2.0 points on HotpotQA at 90 chunks; at 30 chunks, the gains are 0.4 and 0.6 points, respectively. The smaller gain at 30 chunks reflects the limited room for the trained arbitrator to act before the trajectory reaches the budget: it often continues checking an answer, after which truncated trajectories fall back to an answer aggregated from the committed history. The 90-chunk comparison therefore provides a clearer measure of the training gain. RhyCo-Base itself benefits little from the larger budget: increasing its limit from 30 to 90 chunks changes only 8 of 486 MATH-L5 problems, all of which were truncated at 30 chunks, and improves accuracy by 1.6 points; on HotpotQA, no trajectory reaches 30 chunks, so accuracy is unchanged. Thus, of the 7.2-point MATH-L5 gap between the configurations reported in the paper, 5.6 points arise under matched budgets from GRPO training, while 1.6 points come from the larger inference budget.


3: Token cost includes prefill (W2, Q2)

W2, Q2: Prefill counted? Latency and cost vs baselines?

We thank the reviewer for this question. All listener prefill is already included in the cost reported in Figure 3(right). Figure 3(left) counts output tokens only, as labeled, and Figure 3(right) converts output and input tokens into output-token equivalents at 1:8; the input side covers every listener's prefill over each new 64-token chunk, the arbitrator's incremental input, the history passed to an agent that takes over, and all initial prompts.

Complete token accounting. The table itemizes the per-query cost behind Figure 3(right) on MATH-L5; speaker decoding includes chunks later discarded by a switch.

MATH-L5, per queryRhyCo-BaseRhyCo-RL
Speaker decoding840 out1,985 out
Whisper decoding96 out274 out
Arbitrator output240 out1,100 out
Listener prefill (3 × 64 per chunk)2,520 in5,955 in
Arbitrator input1,176 in3,359 in
History at handoff1,252 in3,610 in
Initial prompts5,100 in5,100 in
Total1,176 out / 10,048 in → 2,4323,359 out / 18,024 in → 5,612

The gate itself is a lightweight 12→32→1 MLP and contributes negligible compute relative to the LLM calls.

Measured end-to-end latency. We additionally measure wall-clock performance on identical hardware (one A100 80GB GPU), so the comparison captures the actual listener-forward overhead in addition to token accounting.

MethodTime/queryq/hrOutput tokensPeak mem.Listener %Arbitrator %
RhyCo-Base27.1s1331,17618.4 GB5.013.0
RhyCo-RL83.7s433,35921.4 GB5.114.2
SC@534.8s1035,71218.5 GB——
MAD107s347,87118.4 GB——

Listener monitoring accounts for 5.0–5.1% of end-to-end wall-clock. Each listener processes a 64-token chunk in a single forward pass, while the active speaker generates the same chunk autoregressively token by token; this prefill–decode asymmetry helps explain the measured 5.0–5.1% listener share. On the same hardware, RhyCo-Base is faster than SC@5, while RhyCo-RL remains faster than MAD.

Total cost vs. strongest baselines. At every input/output price ratio from 1:4 to 1:100, RhyCo-Base costs less than every multi-agent baseline; at the 1:8 ratio of Figure 3, it uses 2,432 output-token units against 5,221 to 9,490 for AFlow, SC, MaAS and MAD. RhyCo-RL, the most accurate method at 74.1%, uses 5,612 units at 1:8, below SC, MaAS and MAD. The revision will add the component breakdown, price-ratio sensitivity, per-method counting conventions, and same-hardware wall-clock measurements.


4: Role-specific features (Q4)

Q4: Role-specific listener features when that agent speaks?

The per-role slots are keyed by role. When a role becomes the speaker, its slot is zero-filled; a listener without a slot (the generator) is reflected only through the cross-listener aggregates. The speaker features and the three aggregates are role-agnostic, and the gate is trained on chunks from all four speaker roles (the counterfactual continuations in which each agent takes over).


We hope these analyses address the reviewer's concerns, and we would be glad to clarify any remaining questions.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.