acceptodds
Under review as a conference paper at ICLR 2027

TANGO: TASK-ADAPTIVE AGENT ORCHESTRATION ACROSS THE MULTI-AGENT LIFECYCLE

Abstract

Multi-agent systems can improve complex reasoning in LLMs, but many frameworks predetermine one or more stages of the orchestration lifecycle: worker construction, communication, or final aggregation. We present TANGO, a taskadaptive framework that manages this orchestration lifecycle from creation to coordination and conclusion. For each task, the meta-agent creates workers with complementary reasoning strategies and tool configurations. Each worker maintains a structured memory state that records its progress, hypotheses, unresolved questions, and supporting evidence. At periodic checkpoints, the meta-agent coordinates the workers by constructing a sparse directed communication topology and selecting the information to route. It then concludes the process by reasoning over terminal answers and structured memories to synthesize the final answer. We evaluate TANGO with Gemini-2.5-Flash across general agentic, scientific, mathematical, and code-generation tasks using GAIA, GPQA-Diamond, OmniMATH, and HumanEval. We report mean accuracy across three runs, Pass@3, and a four-benchmark macro-average for both metrics. TANGO attains the highest observed macro-averages in our comparison. Our method achieves the highest scores on GAIA, Omni-MATH, and GPQA-Diamond, while being competitive on HumanEval. We further investigate how meta-agent control across the multi-agent lifecycle affects performance, how these components interact, and how structured memory enables the meta-agent to orchestrate workers more effectively. The code is available at https://anonymous.4open.science/r/TANGO-27C/README.md.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
multi-agent systems, LLM agents, agent orchestration, task-adaptive agents, agentic reasoning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_GYxVOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes a multi-agent orchestration framework organized around three stages: Create, Coordinate, and Conclude. The framework dynamically creates specialized workers, manages information exchange between them through a meta-agent, and aggregates their outputs into a final response.

The approach is evaluated on GAIA, GPQA Diamond, Omni-MATH, and HumanEval, achieving competitive or superior performance against a broad set of baselines. Additional experiments on GAIA and Omni-MATH examine the contributions of individual components, communication strategies, and key architectural parameters.

Critical strengths
  1. Thoughtful integration of agent specialization, memory, and communication. The framework combines task-specific worker creation with meta-agent-controlled information exchange based on workers' states and memory. The selective communication mechanism is particularly interesting, as it offers a structured alternative to indiscriminate context sharing. Although the individual ideas build on prior work, their integration into a unified orchestration framework is a meaningful contribution.

  2. Strong experimental analysis and clear presentation. The paper is well organized, with Algorithm 1, Figure 1, and equations making the architecture accessible despite its complexity. The evaluation covers multiple benchmarks and numerous baselines. Ablations on GAIA and Omni-MATH examine individual modules, communication strategies, worker counts, and checkpoint intervals. Together with detailed case studies and failure analyses in the appendix, these experiments provide substantial evidence supporting the design choices.

Critical weaknesses
  1. Limited clarity of the contribution beyond existing agent orchestration approaches. The three stages of the framework address problems that have already received considerable attention. While their integration is well executed, it remains unclear which aspects represent fundamentally new mechanisms rather than an effective combination of existing techniques. Similar high-level orchestration capabilities are also present in practical agentic systems such as Claude Code and Codex. A more explicit comparison of the proposed coordination and memory mechanisms with these approaches would help establish the paper's distinct contribution.

  2. Insufficient evidence regarding the efficiency of the proposed orchestration. Repeated meta-agent interventions and inter-agent communication introduce additional computational and coordination overhead. Although the accuracy improvements are promising, the evaluation does not sufficiently clarify their trade-offs in terms of token usage, monetary cost, and end-to-end latency. Without this information, it is difficult to assess whether the additional architectural complexity is justified by the performance gains.

Questions for authors
  1. Cost efficiency. Could the authors report average token consumption and estimated API cost per task for the proposed framework and the evaluated baselines? Since Gemini 2.5 Flash is relatively inexpensive, it would be particularly interesting to determine whether the performance improvements translate into a favorable cost–accuracy trade-off.

  2. Execution latency. How does the end-to-end execution time compare with the baselines? The repeated checkpoints and meta-agent interventions may introduce substantial overhead, even when workers operate in parallel. Reporting average or median wall-clock latency would help clarify the practical implications of this design.

  3. Communication patterns. Could the authors report the average number of inter-agent communication events per checkpoint, including the frequency of checkpoints with no communication and those with extensive communication? This would help assess how selectively the meta-agent establishes communication and how these decisions relate to computational overhead.

  4. Scope of the orchestration framework. The current architecture is restricted to parallel workers independently solving the complete task. Have authors considered the meta-agent that also construct workflows involving dependent subtasks with sequentially working agents? Similarly, can the framework support selecting different backbone models for different workers?

4 comments
AuthorsOpenReview ↗

Response to Reviewer GYxV Part 1

We thank the reviewer for the positive assessment of the integration and the analysis, and for the four precise questions. Questions 1–3 are answered quantitatively in the common response and we further discuss them here, and answer W1 and Q4 in full.

W1. What is new relative to existing orchestration

We do not claim to have invented worker specialization, agent memory or sparse communication and each has precedents that we cite. The contribution is (a) the specific coordination mechanism and (b) the evidence that the three lifecycle stages interact and must be adapted together.

Comparison with Claude Code, Codex and the open manager-led systems. These systems are manager-led: a single orchestrator decomposes the task, delegates sub-tasks to subagents. Communication is vertical and happens when a subagent finishes and subagents do not see each other's intermediate state. TANGO differs in three mechanisms. First, all workers attempt the whole task, so the system's robustness comes from strategy and tool diversity rather than from a correct decomposition. Second, communication is lateral and mid-trajectory: the router reads structured states of all workers at checkpoints and copies specific evidence from one worker to another while both are still active. Third, the final decision is made over terminal structured memories and tool evidence, not over the workers' answers or a manager's running summary, and Table 6 shows this matters (final answers alone: −11.1 GAIA A@3 points). Magentic-One, AOrchestra and Flash Searcher are the open analogues of the manager-led pattern and are in Table 1 with their own tool stacks. TANGO obtains higher A@3 than all three on GAIA, GPQA-Diamond and Omni-MATH with a smaller tool catalogue.

What the evidence adds beyond "a good combination". Table 2 shows the stages are not additive as routing lowers accuracy when added to fixed workers with a synthesizer but raises it with authored workers. Similarly, structured memory does nothing under dense sharing but is the best configuration under selective routing (Table 3). A system that fixes any one stage cannot exhibit, or exploit, these interactions.

W2 / Q1. Cost efficiency

Table A gives per-task tokens and cost for TANGO and the sampling baselines with the paper's backbone, and Table B does the same for Gemini-3.5-Flash. With Gemini-2.5-Flash, TANGO costs more than Majority Vote (15) on GAIA for +11.9 A@3 points and 47% less on Omni-MATH for +0.9 points. With Gemini-3.5-Flash, TANGO is cheaper than Majority Vote (15) on both benchmarks. The favourable trade-off comes from using five workers instead of fifteen and from keeping the controller's output small, also TANGO uses fewer model steps per task.

Q2. Execution latency

Table C reports mean, median and P90 wall-clock per task. On GAIA, TANGO's median is 3.97 minutes versus 2.43 for Majority Vote (15) and 0.85 for ReAct and 4.59 versus 2.04 and 0.97 on Omni-MATH. TANGO makes far fewer model calls than Majority Vote (15) but more of them are on the critical path. The author call precedes all workers, each checkpoint is a barrier where the workers wait for one router call, and synthesis follows the last worker. Majority Vote's latency is just its slowest trajectory. Two mitigations follow directly from Table D. First, routing demand is task-dependent: 86% of GAIA checkpoints return no route and the empty fraction rises from 67% at the first checkpoint to 87% by the fifth, whereas on Omni-MATH 77% of checkpoints carry routes, so an adaptive routing schedule would remove most barrier waits on GAIA-like tasks with little effect on what is routed. Second, routing can run asynchronously, delivering packets at the recipient's next turn without pausing the others. We have added the latency table and this discussion to Appendix E and to the Limitations (Section 6).

Q3. Communication patterns

Table D answers this over the GAIA and Omni-MATH runs. On GAIA the router is consulted often but rarely acts. It is called about seven times per task, yet 86% of those calls return no route, and the traffic it does send is front-loaded with 1.3 routes per call at the first checkpoint falling to 0.3 by the fifth. On Omni-MATH the router is consulted less than three times per task but acts on three out of four calls, sending two routes per call on average. The router is therefore not uniformly sparse. It is sparse where workers’ intermediate states are of little use to each other and dense where they are, which depends on the task itself and is the behaviour a task-adaptive policy should have. Its overhead is the reading of five structured states at every checkpoint, and that reading is wasted mainly on GAIA-like tasks where most checkpoints are empty. This is why we point to an adaptive routing schedule, rather than to further payload reduction, as the next step.

AuthorsOpenReview ↗

Response to Reviewer GYxV Part 2

Q4. Dependent sub-tasks and heterogeneous backbones

Dependent workflows. The current implementation does not construct dependent workflows as the five workers each attempt the full query in parallel, and the router selects transfers between active workers. However, the lifecycle abstraction extends naturally to this setting. The meta-author could emit sub-tasks with a dependency graph, each node being launched when its prerequisites finish, and the router would then route within the permitted edges of that graph, with the evidence buffer carrying upstream results as provenance-tagged packets exactly as it carries lateral evidence now. Flash Searcher (DAG execution) and ROMA (recursive dependency-aware trees) operate on similar themes. TANGO's contribution would be to add selective, verifiable evidence transfer along the edges and evidence-based synthesis at the join points.

Heterogeneous backbones. The worker specification already carries a model identifier, so assigning different models to different workers or roles is supported by the implementation, however we leave that for future work and have not explored it in our manuscript.

Thank you for the detailed responses and additional experiments. I appreciate the effort the authors have put into addressing my concerns. The new results clarify several important aspects of the method, particularly its computational costs, latency, and communication patterns. However, my main concern regarding the novelty of the proposed approach relative to existing agent orchestration methods remains only partially addressed. While the rebuttal strengthens the paper, I do not believe it fully resolves this concern. I therefore maintain my positive assessment and original score.

Thank you for the thoughtful follow-up and for maintaining your positive assessment. Your questions have given us a clearer view of what to pursue next. The latency and communication analyses showed how task-dependent routing demand is, which motivates an adaptive routing schedule, and asynchronous communication between workers to remove the checkpoint barriers and reduce latency. Beyond that, we will explore dependency-aware workflows with sequential sub-tasks and to support heterogeneous backbones per worker. We also take the point on novelty seriously and will sharpen the contrast with manager-led systems in the final version. We appreciate the insight these questions provided.

Reviewer_D1TNOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper presents TANGO, a training-free framework that adapts three stages of multi-agent orchestration: worker creation, communication, and final synthesis. A meta-agent configures five workers with task-specific reasoning strategies and tools. The workers attempt the same problem, maintain structured memories, and exchange selected information through a meta-router at periodic checkpoints. A synthesizer then combines their candidate answers and supporting evidence. Evaluation uses Gemini-2.5-Flash across GAIA, GPQA-Diamond, Omni-MATH, and HumanEval. Component and communication ablations examine how the stages interact, with the clearest mean-accuracy improvement over independent sampling and voting observed on GAIA.

Critical strengths
  1. With Gemini-2.5-Flash on GAIA, TANGO achieves 55.15% mean accuracy, compared with 43.23% for Majority Vote (15) and 49.90% for Magentic-One (Table 1). This is a substantial improvement in the reported answer accuracy.

  2. The component ablations document interactions under this backbone: routing is more beneficial with authored workers than with a fixed portfolio (Table 2), and structured memory with selective routing yields the highest observed mean accuracy among the communication variants (Table 3). This analysis helps characterize the components' behavior in the tested configurations.

Critical weaknesses
  1. An end-to-end resource accounting table for full TANGO is missing.

  2. With only one model evaluated, I cannot assess whether TANGO consistently delivers practically meaningful gains on other models.

Questions for authors
  1. Clarification of existing resource accounting: From the existing evaluation logs, what are the mean total per-task token usage and API/tool charges of full TANGO, Majority Vote (5)/(15), and Author + Synthesizer on GAIA and Omni-MATH? Please include all worker and controller calls, final synthesis, and paid tool use as applicable, and distinguish input/output usage and any separately billed reasoning tokens.

  2. Focused validation of applicability: Is there any evidence that the gains persist on a second representative backbone? A limited comparison on GAIA, using the same new backbone for TANGO and independent sampling/voting controls and reporting total costs, would materially affect my assessment. This is a targeted check of the existing method, not a request to rerun the entire benchmark suite. If such evidence is unavailable, please clarify which conclusions are intended to be specific to Gemini-2.5-Flash and the evaluated tasks.

3 comments
AuthorsOpenReview ↗

Response to Reviewer D1TN

We thank the reviewer for the precise and constructive questions. We have added the requested resource accounting and a second-backbone comparison to the paper (Appendices E and F); the full tables are in our common response above.

W1 / Q1. End-to-end resource accounting. Table A reports per-task input/output tokens and cost for every method, covering all worker, meta-author, meta-router and meta-synthesizer calls plus search charges; reasoning tokens are counted as output. On GAIA, TANGO costs 0.437 USD per task, 0.064 USD more than Majority Vote (15) and 0.043 USD more than Author + Synthesizer, for +11.9/+18.2 and +3.2/+3.0 A@3/P@3 points. The extra cost is almost entirely input tokens (2.4x those of Majority Vote (15)); output tokens, which cost 8x more, rise by only 6%. On Omni-MATH, TANGO is cheaper than Majority Vote (15) (0.443 vs 0.833 USD) and Author + Synthesizer (0.443 vs 0.456 USD) while improving both A@3 and P@3.

W2 / Q2. Second backbone. Yes. Table B repeats the comparison with Gemini-3.5-Flash as the only change. On GAIA, TANGO reaches 85.5% against 60.6% (ReAct) and 66.7% (Majority Vote (15)), margins of +24.8 and +18.8 points; on Omni-MATH the margins are +16.2 and +8.6. TANGO is also cheaper per task than Majority Vote (15) on both benchmarks (0.68 vs 0.74 USD on GAIA; 0.55 vs 0.81 USD on Omni-MATH). These are single runs, so we report them as a targeted check rather than a full replication of our validation sweep.

Thank you for your response. The clarifications have addressed my main concerns, and I will raise my score accordingly.

Thank you for the quick response and for raising your score. Your questions on resource accounting and backbone generalization led to additions that we believe have made the paper stronger, and we appreciate the insight.

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

This work proposes a novel multi-agent system for problem solving. First, the authors design mechanisms for LLM agents to manage their memories and maintain their current states throughout the problem-solving process. Second, they introduce a meta-agent that dynamically updates the multi-agent system's configuration during execution. This is a key distinction from existing multi-agent systems, which typically determine the workflow at the beginning. Finally, the system synthesizes the final answer by aggregating and summarizing the outputs of individual agents.

Critical strengths
  • The paper is generally well written and easy to understand.
  • The evaluation is a plus as authors evaluate the proposed method on proper and relatively more challenging benchmarks.
  • The idea of dynamically adapting the routing policy among agents during the problem solving appears to be a novel contribution.
Critical weaknesses
  • Some implementation details remain unclear, particularly the definitions and distinctions between the various notations. For example, what do the hypotheses PitP_i^t and open questions UitU_i^t represent? While these concepts can be intuitive in the context of math reasoning, it is unclear how they generalize to other problem-solving tasks. Additionally, what specific information is included in the information packets passed between agents Rj→iR_{j \rightarrow i}, and how is this packet generated?

  • The work lacks scientific insights for its design choices. The proposed framework appears to rely on prompting LLMs to generate structured textual representations for memory management and workflow design, but it remains unclear why these particular representations are necessary or whether they are incomplete for effective problem solving. More specifically, the paper does not establish whether the proposed memory and communication paradigm capture all the information needed or introduce redundant information. Therefore, most designs look ad-hoc.

  • There seem to be more baselines to be compared such as AFlow, OneFlow, etc.

Zhang et al., AFlow: Automating Agentic Workflow Generation

Xu et al., Rethinking the Value of Multi-Agent Workflow: A Strong Single Agent Baseline

Questions for authors

See weaknesses.

2 comments
AuthorsOpenReview ↗

Response to Reviewer RNDx

We thank the reviewer for the thoughtful review and for recognising the novelty of adapting the routing policy during problem solving and the strength of the benchmarks. We address the three weaknesses in turn and have revised Section 3.3 to define every working-memory field and Appendix A to make the packet mechanics explicit.

W1. Definitions of memory fields, generality beyond mathematics, and what an information packet contains

Hypotheses and open questions. A hypothesis (PitP_i^t) is a worker's current, revisable claim about the answer or about an intermediate quantity it needs. An open question (UitU_i^t) is something the worker has identified as unresolved and that could change its answer. These are the two fields the router uses to see disagreements and to see what a worker still needs. Observations and evidence references describe what a worker can offer, while open questions and next actions describe what it needs. The fields are task-agnostic as the same schema is produced on every benchmark and only the content changes. The revised Section 3.3 now gives a one-sentence definition of each of the six fields in these terms, and the new Appendix H.1 shows two complete working-memory snapshots produced by this schema. Figure 9 (Appendix C.2) also shows routed items and their effect on the recipient in GAIA runs.

What an information packet is and how it is produced. A packet is an existing record copied by reference. At each checkpoint the system builds an indexed list of every item the workers currently hold: each non-empty working-memory field, each tool result or tool error, and each length-bounded model response. Each item carries a reference id, its source agent and turn, its kind (working_memory, tool, or model), the field name, and its content (truncated to 2,000 characters). The router sees the task, all workers' structured states and this list (Appendix H.3 shows the exact prompt) and returns only the tuple (source_agent, target_agent, [reference ids]), up to 64 routes per checkpoint with 1–8 references each. For all valid references, the system copies the item with its provenance into the recipient's incoming buffer BitB_i^t. In one full GAIA run, 60% were working-memory fields, 36% tool results and 4% model responses.

W2. Scientific grounding of the representation and communication choices

We understand the concern and ground the design in two ways: in prior work that established the value of each ingredient, and in quantitative tests of whether the combination is sufficient and non-redundant. Structured summaries are known to help agents manage information over long tasks: HiAgent (Hu et al., 2024) organizes working memory around subgoals and WebResearcher (Qiao et al., 2025) consolidates findings into evolving reports, while work on sparse communication (Li et al., 2024) shows the value of limiting which agents exchange messages. These establish memory for an agent's own context and selective communication between agents separately.

Tables 3 and 6 test this quantitatively. Table 3 compares raw trajectories with structured memory under all-to-all sharing and under selective routing. Structured memory alone does not improve accuracy under all-to-all sharing and routing over raw trajectories has mixed results, but together they give the highest A@3 and P@3 of all communication configurations, raising A@3 by 4.44 points on GAIA and 2.02 on Omni-MATH while cutting routed payload from 46.95K to 1.57K tokens per GAIA task and from 28.68K to 1.82K on Omni-MATH. Table 6 tests sufficiency on the synthesis side: a synthesizer given final answers alone scores 44.04 A@3 on GAIA against 55.15 with the structured terminal memories, and raw outputs reach 52.32 using 17.08K input tokens against 10.44K, so the fields carry information that neither terminal answers nor unstructured histories expose as compactly. We therefore claim that the two components are effective together rather than that either helps on its own. An upper bound on the information a representation should carry is hard to define.

W3. Additional baselines

We agree that broader baseline comparisons are always valuable, however adding the baselines the reviewer suggested is not feasible. AFlow selects workflows by tuning on a validation set, while our available protocol uses the full benchmark pool for evaluation. Tuning AFlow on those same examples and reporting its score on them would mix workflow selection with evaluation, so it would not be a correct comparision. A faithful comparison would require a separate workflow-search split and a common held-out evaluation, which our current protocol does not provide, and is also not available for other baselines. For OneFlow, we could not locate a public implementation to reproduce it under our benchmark, model, and tool setup. We therefore cite AFlow and OneFlow as relevant prior work in the revised paper and recognize their contributions.

Dear Reviewer RNDx,

We posted our response to your review on 14 November, addressing the definitions of the memory fields and packet contents (now in Section 3.3, Appendix A and Appendix H.1), the scientific grounding of the representation (Tables 3 and 6), and the AFlow/OneFlow baselines.

As today is the last day of the discussion period, we would very much appreciate hearing your thoughts on our rebuttal and whether it addresses your concerns, or if there is anything further we can clarify.

Best regards, The Authors

Comments on the paper

3 comments
AuthorsOpenReview ↗

Common response to all reviewers Part 1

Common response to all reviewers

We thank all three reviewers for their careful reading. Two requests were shared across reviews: an end-to-end resource accounting for TANGO and some baselines, and evidence that the gains hold on a second backbone. We address both here and refer to these tables from the individual responses. We have added them to the revised paper as Appendix E (resource accounting), Appendix F (second backbone), and an extended Appendix C.1 (communication statistics). We hope this addresses all the concerns raised by the reviewers.

How we measured cost, tokens and latency

Every model call made by the harness (worker turns, meta-author, meta-router, meta-synthesizer, and search calls) is logged with its provider-reported input and output token counts, and every search call is counted. Cost is computed per call from the official Gemini 2.5 Flash list price (0.30 per million input tokens and 2.50 per million output tokens), plus Serper search charges ($1 per 1,000 queries). Gemini 2.5 Flash reports thinking tokens as part of output tokens, so reasoning tokens are included in the output column and are not billed separately. All figures below are per-task means averaged over the three runs. Wall-clock latency is the elapsed time from the first meta-agent call to the final answer for one task, with workers running in parallel.

Table A. Per-task resources, Gemini-2.5-Flash (Table 7 in the revised paper)

GAIAA@3P@3Input tok./taskOutput tok./taskCost/task
ReAct33.5449.0910.0K7.5K$0.023
Majority Vote (5)40.0052.7356.6K40.2K$0.127
Majority Vote (15)43.2350.30165.4K117.5K$0.373
Author + Synthesizer51.9265.45251.4K122.9K$0.394
TANGO55.1568.48390.8K124.7K$0.437
Omni-MATHA@3P@3Input tok./taskOutput tok./taskCost/task
ReAct63.0275.004.1K21.9K$0.055
Majority Vote (5)69.3477.1522.8K108.9K$0.279
Majority Vote (15)70.9675.2064.3K325.4K$0.833
Author + Synthesizer70.9079.88156.7K162.7K$0.456
TANGO71.8881.45257.1K145.8K$0.443

TANGO is the most accurate method in both tables while staying in the same cost band as the strongest baselines: on GAIA it costs slightly more than Majority Vote (15) and Author + Synthesizer for a much larger accuracy gain, and on Omni-MATH it is cheaper than both while still more accurate.

Table B. Resuts on Gemini-3.5-Flash (Table 9 in the revised paper)

All three methods use the identical worker backbone, tool sets, sampling parameters, turn budgets and scoring as in the paper. Only the model identifier changed. Cost uses the Gemini-3.5-Flash API price of 1.50 per million input tokens and 9.00 per million output tokens, plus the same Serper charges. Owing to time and compute constraints these are single runs rather than three-run means, so we report point estimates only.

GAIAAccuracyInput tok./taskOutput tok./taskCost/task
ReAct60.61% (100/165)12.9K3.6K$0.054
Majority Vote (15)66.67% (110/165)180.0K47.3K$0.737
TANGO85.45% (141/165)195.4K42.0K$0.678
Omni-MATHAccuracyInput tok./taskOutput tok./taskCost/task
ReAct75.59% (387/512)9.2K4.2K$0.052
Majority Vote (15)83.20% (426/512)142.3K64.5K$0.808
TANGO91.80% (470/512)86.5K46.3K$0.548

TANGO's margin over the strongest sampling baseline is larger with the stronger backbone with +18.8 points on GAIA and +8.6 on Omni-MATH, versus +11.9 and +0.9 with Gemini-2.5-Flash. With a stronger backbone like Gemini-3.5-flash, TANGO is cheaper per task than Majority Vote (15) on both benchmarks while being more accurate. The sampling baseline does not benefit from the stronger model nearly as much as orchestration does. Majority Vote (15) still fails on the tasks where every sample makes the same mistake, whereas TANGO recovers a large share of them.

AuthorsOpenReview ↗

Common response to all reviewers Part 2

Table C. End-to-end latency and model calls (Table 8 in the revised paper)

BenchmarkMethodMeanMedianP90Model calls/task
GAIAReAct0.950.852.219.4
GAIAMajority Vote (15)2.972.436.2997.1
GAIATANGO4.643.978.9429.2
Omni-MATHReAct1.440.973.561.5
Omni-MATHMajority Vote (15)2.322.044.5527.5
Omni-MATHTANGO4.464.596.0517.5

TANGO makes 3.3x (GAIA) and 1.6x (Omni-MATH) fewer model calls per task than Majority Vote (15), but its median latency is 1.6x and 2.25x higher, because its calls are partly sequential: the meta-author runs before any worker starts, every checkpoint is a synchronization point at which the five workers wait for one router call, and the synthesizer runs after the last worker finishes. Majority Vote (15) runs 15 fully independent trajectories, so its latency is the slowest single trajectory. How much of this overhead is useful depends on the task type, which Table D quantifies.

Table D. Communication patterns (Table 4 in the revised paper)

Statistics over all three GAIA and all three Omni-MATH runs of the main Gemini-2.5-Flash configuration. A checkpoint is one router call after every k=2 worker turns and a route is one directed source→recipient transfer, which may carry several evidence references.

StatisticGAIAOmni-MATH
Checkpoints per task (mean / median / max)6.99 / 4 / 242.64 / 1 / 24
Routes per checkpoint (mean)0.432.00
Checkpoints with no route86.2%22.7%
Checkpoints with 1–2 routes6.8%48.2%
Checkpoints with 3–5 routes5.3%22.4%
Checkpoints with more than 5 routes (maximum)1.7% (20)6.7% (16)
Routes per non-empty checkpoint (mean)3.122.58
Routes per task / delivered evidence references per task3.00 / 5.495.28 / 7.44

The two benchmarks have very different routing requirements. GAIA tasks are long, tool-heavy searches where workers run many turns, so the router is called 6.99 times per task on average. However, 86% of checkpoints return no route and communication is front-loaded. Omni-MATH problems are shorter and converge quickly and the router is called only 2.64 times per task, but only 22.7% of those routes carry no information. These numbers quantify what the routing-traffic figure in Appendix C.1 also shows. For future work this motivates an adaptive router that can increase or decrease routing for better latency and increased cost without losing out on accuracy.

AuthorsOpenReview ↗

Follow up on rebuttal

Summary of the discussion so far and request for feedback

Dear Reviewers and Area Chair,

Thank you for the constructive reviews. On 14 November we posted a common response and individual replies, and uploaded a revised PDF. In brief:

  • Resource accounting (D1TN W1/Q1, GYxV W2/Q1). Per-task tokens and cost for ReAct, Majority Vote (5/15), Author + Synthesizer and TANGO on GAIA and Omni-MATH (Table A; new Appendix E). TANGO is the most accurate method while staying in the same cost band as the strongest baselines, and is cheaper than Majority Vote (15) on Omni-MATH.
  • Second backbone (D1TN W2/Q2). Single-run results with Gemini-3.5-Flash on all of GAIA and Omni-MATH (Table B; new Appendix F). TANGO's margin over Majority Vote (15) grows to +18.8 and +8.6 points, and it is cheaper per task on both benchmarks.
  • Latency and communication patterns (GYxV Q2/Q3). Wall-clock latency and per-checkpoint routing statistics (Tables C and D; Appendix E and extended Appendix C.1), showing that routing demand is task-dependent and pointing to adaptive routing as the main lever for cost and latency.
  • Definitions, packet mechanics and grounding (RNDx W1/W2). One-sentence definitions of all six working-memory fields (Section 3.3), an explicit description of what a routed item is (Appendix A), two complete working-memory snapshots (Appendix H.1), and the Table 3 and Table 6 evidence for why structured memory and selective routing are effective together.
  • Related work (RNDx W3, GYxV W1). AFlow and OneFlow are now cited and discussed, and we contrast TANGO's lateral, mid-trajectory evidence routing with manager-led systems such as Claude Code, Codex and Magentic-One.

As the discussion period closes on 18 November, we would be grateful if the reviewers could let us know whether these responses address their concerns, and whether anything remains that we should clarify or expand on. We are happy to provide further analyses before the deadline.

Best regards, The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.