acceptodds
Under review as a conference paper at ICLR 2027

PERSEUS: Non-Blocking Parallel Exploration with a Speculative Swarm

Abstract

Tool-using agents face a breadth–coordination trade-off: a single agent follows a coherent task trajectory but discovers evidence mainly along its chosen call path, while multi-agent exploration broadens discovery through additional coordination. PERSEUS couples one authoritative main agent, the Actor, with a non-blocking speculative swarm. Single-round Speculators use constrained variants of native tools to explore complementary information needs in independent work copies, while the Actor retains task decisions and state changes. Persistent Futures carry unfinished acquisitions across Actor requests. A source-aware ledger admits completed observations with their provenance, allowing the Actor to continue while pending exploration proceeds asynchronously. Cooperation occurs through evidence: a Speculator's inspection can inform a different later action. On selected tasks from four benchmarks, PERSEUS achieves the highest overall end-to-end task success among 13 baselines, improving success from ReAct's 56.3% to 67.8% and Execution Speed from 2.43 to 3.20 (). Temporal analysis, ablations, and scaling relate these outcomes to evidence flow, exploration structure, and persistence. Harness comparisons show fewer Actor turns with maintained or improved quality relative to single-agent controls. The results support broad exploration through asynchronous cooperation under one authority.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLM agents, multi-agent systems, agent swarm, parallel exploration, asynchronous speculative execution

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

PERSEUS is a training-free runtime that couples one authoritative Actor with a speculative swarm. Speculators receive constrained variants of native tools and independently acquire observations in isolated work copies; only the Actor changes authoritative task state. Persistent Futures keep unfinished acquisitions alive across Actor requests, and a source-aware ledger admits completed evidence without requiring the Actor to wait for the swarm. The intended benefit is broader discovery with fewer sequential information-gathering steps. On 183 selected tasks from AutomationBench, τ²-Bench, Terminal-Bench, and GAIA, with three attempts per task, PERSEUS reports 67.8% task success versus 56.3% for ReAct.

Critical strengths
  1. A concrete architecture for asynchronous evidence acquisition under one task authority. The runtime combines constrained exploration domains, isolated executions, persistent futures, and provenance-preserving evidence admission. Speculative results can inform a later action even when it differs from the explored call. This is a useful systems contribution because it broadens information discovery without requiring multiple agents to reconcile authoritative task updates, and the implementation protocol is described in enough detail to make the intended behavior understandable.

  2. The evaluation supports a meaningful system-level improvement and reports important trade-offs. PERSEUS achieves higher task success than ReAct on all four evaluated subsets, with an 11.5-percentage-point pooled improvement. The evaluation additionally compares oracle best-of-n sampling and reports a pooled price-normalized usage ratio of 1.122 relative to ReAct. The paper also acknowledges settings where PERSEUS is weaker: on BFCL it reaches 75.64% accuracy at 14.38 s versus ReAct's 78.85% at 8.34 s. These qualifications make the practical contribution more informative than an unqualified acceleration claim.

  3. The structured-exploration ablation provides a useful design insight. Without Meta-Tool derivation, Speculators use unmodified tool schemas, consuming 14.56M equivalent tokens while reaching 47.1% task success on the diagnostic subset. Full PERSEUS reaches 56.9% at 9.57M equivalent tokens. This comparison supports organizing complementary acquisition directions rather than simply launching more unconstrained exploration, and helps explain why the runtime can use auxiliary computation productively.

Critical weaknesses
  1. The evaluation establishes benefits on a selected task regime, but does not establish how broadly that regime occurs. The task-selection procedure excludes short workflows and screens tasks partly for being within the Actor's practical capability range. The 85-task diagnostic subset is further selected for recurring speculative-evidence injection in the full system. These are reasonable populations for mechanism diagnosis, but they favor settings where there is both time and opportunity to use speculative evidence. The negative BFCL result shows that this boundary matters. The paper should more clearly characterize which observable task properties predict a benefit, and distinguish performance on these selected long-horizon subsets from performance across the original benchmark distributions.

  2. The experiments do not isolate the contribution of non-blocking evidence delivery from broader exploration and preservation of completed work. Removing the swarm removes auxiliary evidence altogether, while removing persistent Futures cancels unfinished acquisitions at the next request boundary. The latter comparison tests the loss of slower evidence as well as its scheduling. Neither comparison retains the same acquisition process while changing whether the Actor waits for it. The width sweeps and best-of-n comparison are useful, but do not resolve this attribution. Consequently, the results establish that the complete runtime is effective more strongly than they establish that its particular asynchronous cooperation protocol is responsible for the gains. A comparison controlling the acquisition policy and resources while varying delivery/synchronization would strengthen that central claim.

  3. The practical consequences of isolated, potentially stale evidence are not sufficiently evaluated. The execution protocol requires the host to isolate all mutable resources, including remote services, and notes that a work copy can be created after the Speculator's context was recorded. Persistent acquisitions can then finish after the Actor has changed authoritative state. The ledger preserves provenance and filters recognized conflicts, but leaves applicability judgments to the Actor, particularly for opaque outputs. The paper does not quantify how often evidence is stale or copy-specific, how often the Actor must revalidate it, or what those checks cost.

Questions for authors
  1. Were the main task subset and selection criteria fixed before comparing PERSEUS with the baselines, and did screening use any task-success results?
  2. Is there an existing comparison that keeps the same Speculators and preserves their work but changes when the Actor waits for or receives their results? If not, which conclusions do the authors attribute specifically to non-blocking delivery, rather than to additional exploration and preserving slower acquisitions?
  3. Do reported execution times include provisioning isolated work copies and external-service replicas, evidence admission, and cancellation/cleanup of pending work? What request/tool concurrency or rate limits were shared across methods?
1 comment

W1/Q1: Task regime and selection

Yes. The subset and the criteria were fixed before any agent experiment, for PERSEUS and the baselines alike. Screening ran the bare Actor model in its native tool-call loop, with no agent harness, and kept tasks on which it made effective progress; success was not required. Tasks also had to be compatible with the evaluation host and within the time limits (Appendix B.1). No result of PERSEUS or of any baseline was used.

Original distributions. Our results and claims concern the selected long-horizon tasks, not the original benchmark distributions; we will state this scope explicitly in Section 3.1 and the conclusion. Within that scope, the 85-case subset served only for mechanism diagnosis: we have rerun every ablation on all 183 tasks, where removing any one component lowers TS by 18.5–21.5 points.

Which task properties predict a benefit. Two observable properties predict it: task horizon and exploration width, i.e. how many independent conditions a task exposes. On horizon, speculative evidence is injected mainly at Actor turns 2–4 (Figure 2), so a task that finishes within one or two Actor turns has no later decision that can use it; BFCL's single-turn calls are this case (Appendix B.6). On width, the evidence comes from parallel exploration, so tasks with more independent conditions give the swarm more to acquire ahead of demand: raising the Speculator cap from 0 to 8 increases PCS by 0.131 and ES by 41% (Figure 3). Both match the utility condition in Appendix A.5: evidence helps only when the downstream work it saves exceeds the acquisition overhead. We will state this boundary in the main text.

W2/Q2: Isolating non-blocking delivery

The submission did not include one, so we added PERSEUS-Blocking. It keeps the same algorithm and Speculators and preserves all their work; the only change is delivery: each acquisition direction blocks until its result returns, and the result is admitted in the same turn instead of asynchronously at a later one. The Actor's interface is unchanged and it is not told about the wait. While one direction is pending it continues with other planned work; once nothing is left to issue, the runtime holds its next request. All rows use 183 tasks and three runs.

ConfigurationTS (%)PCSTime (s)ES
Block on each result (PERSEUS-Blocking)44.5±1.80.444471.4±92.00.94
Cancel unfinished work (w/o Persistent Futures)46.3±3.00.530282.8±11.01.63
No Speculators (w/o Swarm)49.3±2.90.564277.1±17.01.78
Admit asynchronously, keep unfinished work (PERSEUS)67.8±2.80.746211.5±11.13.20

What non-blocking delivery contributes. With the same Speculators and preserved work, admitting results asynchronously instead of blocking on them raises TS from 44.5% to 67.8%, cuts execution time from 471 s to 212 s, and cuts equivalent tokens from 79.3M to 42.2M. Timeouts do not explain the blocking result (12 of 549 runs). Under blocking, the Actor does more of the exploration itself, in fewer but larger turns (2.7× the equivalent tokens per turn), and its own observations bypass the ledger. Non-blocking delivery keeps exploration on the inexpensive Speculators and behind the ledger.

Exploration alone is not the source of the gain. With Speculators present, neither blocking (44.5%) nor cancelling unfinished work (46.3%) improves on running no Speculators at all (49.3%). The 18.5-point gain over the no-Speculator configuration appears only when results are delivered without blocking and slower acquisitions are kept. This is the evidence that the asynchronous cooperation protocol, and not additional exploration by itself, is responsible for the gains.

W3/Q3: Stale or copy-specific evidence, and what execution time includes

Yes. Reported times are end-to-end and include provisioning of work copies and external-service replicas, evidence admission, and cancellation and cleanup of pending work. All methods shared the same conditions: no rate limit or concurrency cap, 50–100 ms tool handling per call, and the same official images and pre-downloaded resources for environment setup.

Frequency. Of the 26,630 evidence records admitted in the main evaluation, the ledger flags 310 (1.2%) as conflicting or older. A further 704 (2.6%) were admitted without a flag and later misapplied, for example one customer's record used for another customer's request. Against these, 12,655 records (47.5%) had an effective downstream effect.

Revalidation. The Actor re-checks only the 310 flagged records; all others enter its context directly. The cost of these checks is included in the reported end-to-end time and token usage.

Reviewer_9TBAOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

PERSEUS is a training-free framework in which a main agent receives observations from parallel speculative workers without waiting for them. Workers execute calls in isolated copies, and completed results are filtered before entering the main agent’s context. On 183 selected tasks across four benchmarks, the method improves success over ReAct from 56.3% to 67.8%, with additional ablation and cost analyses.

Critical strengths

S1: The design is practical and requires no training. Keeping decisions with one main agent while using isolated workers for exploration makes the control flow easy to manage.

S2: The experiments include repeated runs, ablations, and a cost-aware best-of-n comparison. The paper also reports the negative BFCL result, which helps clarify when speculation may not be useful.

Critical weaknesses

W1: The novelty over asynchronous orchestrator–worker systems is unclear. A main agent receiving findings from parallel background workers is already a familiar design. The paper should explain what schema-based partitioning, persistent Futures, and the ledger add beyond this architecture, and which contributions go beyond engineering integration. It should also discuss closely related agent-swarm systems, such as Kimi Agent Swarm, and clearly distinguish their designs from PERSEUS.

W2: The experiments do not sufficiently isolate the benefits of the proposed design. The most relevant baseline is missing: an Actor that launches multi-round background explorer agents and receives their findings asynchronously, under matched compute. The component evidence also needs stronger analysis: removing the swarm changes TS from 56.9 ± 9.0 to 52.9 ± 5.9, while removing persistent Futures lowers it to 49.0 ± 9.0, below the no-swarm variant (Table 2). Paired statistical comparisons would help establish the reliability of these differences and clarify component interactions. Moreover, only 162 of 493 runs with evidence injection show identified downstream use (Section 3.2). Please quantify unused calls, misleading or stale evidence, and Actor context growth, while distinguishing undetected use from confirmed waste.

W3: The presentation introduces unnecessary difficulty. Terms such as ready-only admission, source-aware ledger, and logical folding obscure a relatively straightforward execution flow. A complete worked example early in the paper, showing worker launch, tool execution, evidence filtering, and the Actor's subsequent decision, would make both the mechanism and its contribution easier to understand.

Questions for authors

Q1: What are the substantive differences between PERSEUS and asynchronous orchestrator–worker or agent-swarm systems such as Kimi Agent Swarm? What does each of schema-based partitioning, persistent Futures, and the ledger contribute beyond that architecture?

Q2: Does PERSEUS retain its advantage over a matched-compute baseline with multi-round asynchronous background agents? Can you provide paired statistical comparisons for the ablations and quantify evidence use, unused calls, misleading or stale evidence, and Actor context growth?

4 comments

W1/Q1: Difference from asynchronous orchestrator–worker and agent-swarm systems

PERSEUS is not an orchestrator–worker system or an agent swarm in that sense. It is an unorchestrated, non-blocking speculative swarm that only assists one authoritative Actor. In Kimi Agent Swarm (Kimi K2.5 report, arXiv:2602.02276), an orchestrator decomposes the task into subtasks, creates specialized subagents that execute them, and schedules them in parallel stages, each lasting as long as its slowest subagent. PERSEUS has no task decomposition, no subtask plan, and no synchronization point. One Actor owns all task decisions and state changes. Each Speculator makes a single model round and issues native tool calls whose only output is observations; Speculators receive no subtasks, produce no deliverables, and do not coordinate with each other.

What each component adds beyond the orchestrator–worker architecture:

  • Schema-based partitioning decides what to explore without a plan. Variants are derived mechanically from declared action values and observed invocation heads (Appendix A.1), so breadth follows tool coverage, not an orchestrator's decomposition.
  • Persistent Futures decouple exploration from the Actor's request cycle. Unfinished acquisitions survive across requests, and each request proceeds on whatever is ready.
  • The ledger decides what enters the decision. Complete observations are admitted once with their provenance, and duplicates and superseded values are suppressed, so evidence from one call can inform a different later action (Section 2.5).

The contribution beyond engineering integration is this cooperation protocol: orchestration systems distribute task ownership and synchronize on results, whereas PERSEUS distributes only information acquisition and keeps one authority. In a new ablation on all 183 tasks, removing any one component lowers TS by 18.5–21.5 points. We will add this discussion, including Kimi Agent Swarm, to Related Work.

W2/Q2: Asynchronous multi-round baseline, paired statistics, evidence accounting

Baseline. We added CAID (Geng & Neubig, 2026, arXiv:2603.21489) to Table 1. A central manager delegates subtasks to multi-round agents that run asynchronously in isolated workspaces and return their results. It uses the same test environment and conditions as all other methods and its official source configuration.

TS (%)Autoτ²TermGAIAOverall
CAID19.4±2.854.6±4.858.5±7.529.8±1.540.8
PERSEUS39.8±1.677.8±2.681.7±4.967.3±6.367.8
ESAutoτ²TermGAIAOverallEq. Tokens (M)
CAID0.834.302.311.171.9682.4
PERSEUS2.064.812.443.223.2042.2

CAID uses 1.95× the equivalent tokens of PERSEUS, so PERSEUS keeps its advantage in both TS and ES against about twice the compute.

Paired comparisons. Pairing by task on one run per configuration (binary TS, 183 tasks), PERSEUS wins/loses on 47/8 tasks against w/o Meta-Tool, 41/9 against w/o Speculative Swarm, and 54/8 against w/o Persistent Futures; the rest tie (sign test, p < 0.001 each). On component interaction, w/o Persistent Futures does not improve on w/o Speculative Swarm (46.3% vs. 49.3% on the 183 tasks): Speculators add no gain unless their unfinished acquisitions are kept, which is why persistence is part of the protocol.

Evidence accounting. The ledger reviewed 34,904 records. 8,274 (23.7%) were known duplicates and were rejected before delivery: this is the confirmed waste, and it never enters the Actor's context. Of the 26,630 admitted records, 12,655 (47.5%) had an effective downstream effect, 704 (2.6%) had a misleading or stale effect, and 13,271 (49.8%) showed no observed downstream interaction; we report this last group separately and do not count it as use or as confirmed waste. The 162 of 493 runs in Section 3.2 are a run-level, block-matched lower bound on use. For context growth, in turns with admission the Actor's own calls supply 21.28% of the per-turn context on average, and admitted evidence supplies the rest.

W3: Presentation

Thanks for your suggestion. The writing is over-compressed, and terms such as "ready-only admission", "source-aware ledger", and "logical folding" obscure a simple flow. In the revision we will describe these steps in plain language and add a complete worked example early in Section 2, based on the finance:4073 episode in Appendix C: worker launch, tool execution, evidence filtering, and the Actor's subsequent decision.

Dear Reviewer 9TBA,

Thank you for your careful review and constructive feedback. Following your suggestions, we have clarified that PERSEUS is a single-authority speculative swarm, in which one authoritative Actor owns all task decisions and Speculators only acquire observations, rather than an asynchronous orchestrator–worker or agent-swarm system such as Kimi Agent Swarm. We have also explained what schema-based partitioning, persistent Futures, and the ledger each contribute, with this discussion now included in Related Work. In addition, we have added CAID as an asynchronous multi-round baseline in Table 1, and provided task-paired comparisons for the ablations together with a full accounting of the speculative evidence. In the uploaded revision, we have replaced the compressed words with plain descriptions and added a complete worked example early in Section 2.

We hope these additions address your remaining concerns and appreciate your support for our work. If any questions or concerns remain, please do not hesitate to let us know.

Thank you for the detailed response. The clarification on W1 makes me understand better how PERSEUS differs from Kimi Agent Swarm.

However, my experimental concerns under W2 remain only partially addressed.

Baseline. The new CAID baseline scores much lower than the single-agent ReAct baseline already in the paper (40.8% vs. 56.3% TS). This makes the comparison harder to interpret, not easier. It suggests that CAID, in its default configuration, may not fit these tasks well. If so, outperforming it says little about whether PERSEUS’s specific design is better than having background explorer agents. My original question is still open: under matched compute, does PERSEUS keep its advantage over a controlled baseline in which the same Actor launches multi-round background explorers and receives their findings asynchronously?

Ablation. In the new 183-task ablation, removing any single component drops TS from 67.8% to 46.3–49.3%. That is well below single-agent ReAct (56.3%). So each ablated variant is not just less helpful than the full system; it is clearly worse than having no speculation at all. If the three components are this tightly coupled, the paper should show and explain that coupling, and I do not currently see it there. For example, why does the swarm without persistent Futures actively hurt performance instead of simply adding nothing?

Presentation. I appreciate the commitment to rewrite the compressed terminology and add a worked example early in Section 2. It would be much more convincing to see these changes directly in the revised paper.

Given these remaining concerns, I will maintain my current score for now.

Thanks for your follow-up.

Full life-cycle multi-turn subagent & CAID Baseline

We acknowledge that we may have overcomplicated your question. To address the question directly: the variant we employed utilizes a "short-cycle" speculator—a specific configuration of PERSEUS rather than a full life-cycle multi-turn subagent. We chose this approach because a full, multi-turn subagent lifecycle would lead to excessive time consumption and "goal drift." Furthermore, it would introduce significant latency in the integration of asynchronous information, rendering that information stale—a critical drawback. In this context, the depth provided by multi-turn interactions often proves more detrimental than beneficial. We will include this configuration as an ablation variant in our camera-ready version to provide a comprehensive comparison.

Why CAID. CAID is the latest published method that matches the description. One manager agent delegates subtasks to worker agents; each worker runs its own multi-round loop in an isolated workspace in the background; the manager receives and integrates their results as they finish. It ran with 1.95× the equivalent tokens of PERSEUS.

Its score is not an outlier. Every multi-agent baseline in Table 1 is below ReAct (56.3%) overall: CMAS 54.3, DMAS 53.9, Multi-Persona 46.8, Magentic-One 44.8, DyLAN 44.3, CAID 40.8. Towards to the modern SOTA models on these long-horizon tasks, handing multi-round work to other agents costs more in coordination than it returns. This is the trade-off the paper starts from (Section 1), and PERSEUS is the one design with auxiliary agents that exceeds ReAct.

Ablation and Components Coupling

Reference point. The no-speculation control for the ablations is "w/o Speculative Swarm" (49.3%): the PERSEUS runtime with Speculators switched off. ReAct (56.3%) is a separately implemented baseline, not that control. Against 49.3%, the variant without persistent Futures is at 46.3% and the one without Meta-Tool derivation at 46.8%. Both differences are within 3 points, the size of the run-to-run SD (2.9–3.2). An incomplete swarm therefore adds nothing; it does not clearly fall below no speculation.

The three components are stages of one path: derivation decides what each Speculator explores and how many are launched, the swarm acquires, and persistent Futures with the ledger deliver. The gain requires the whole path. Each configuration below breaks one link (183 tasks).

ConfigurationLink that is brokenTS (%)
No Speculatorsreference49.3
w/o Meta-Tool derivationSpeculators get no assigned domain46.8
Fixed width 8width no longer follows the derived variants52.8 / 55.6
w/o Persistent Futuresunfinished acquisitions are cancelled at the next request46.3
Blocking deliveryresults reach the Actor only by blocking it44.5
Full PERSEUSnone67.8

With any link broken, TS stays within 7 points of the no-Speculator reference. With all links in place, it is 18.5 points above it.

Why the swarm without Futures adds nothing. Section 3.3 gives these mechanisms component by component. Speculators are still launched and paid for, but every call that has not finished by the Actor's next request is cancelled. Only quick lookups reach the Actor, and slower investigations are lost (Section 3.3, "Preserving evidence across decisions"). The Actor carries the swarm's overhead and receives a partial evidence stream. We will add the joint reading above as a paragraph there to enhance this point.

We hope these additions address your remaining concerns and appreciate your feedback.

Reviewer_iwy4OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper introduces PERSEUS, a training-free runtime that combines a single authoritative Actor with a non-blocking swarm of Speculators. Speculators acquire potentially useful information through constrained variants of native tools and execute their calls in independent work copies. The paper reports an increase in overall task success from 56.3% for ReAct to 67.8%, together with improved success-adjusted execution speed.

Critical strengths
  1. The idea is simple and useful: smaller agents gather information in the background while the main agent continues the task.
  2. The results improve on ReAct across all four tested benchmark subsets, with an overall success gain of 11.5 points.
Critical weaknesses
  1. The ablations do not fully isolate the proposed mechanisms. Removing tool partitioning also changes the number of Speculators, and there is no blocking counterpart to isolate the benefit of non-blocking execution.
  2. The cost comparison uses price-normalized equivalent tokens, which do not directly establish actual dollar costs when input/output ratios differ. Reporting dollar costs and explaining how canceled requests are counted would make the comparison clearer.
Questions for authors
  • How does PERSEUS compare with a blocking version that uses the same Speculator configuration but waits for their results before the Actor’s next request?

  • Does Meta-Tool derivation still improve performance when the number of Speculators is held fixed?

  • Do the ablation findings also hold across all 183 tasks in the main evaluation, beyond the subset selected for recurring speculative evidence injection?

  • In its lossy OS-tuning extension, Speculative Actions[1] provides the Actor with a summary of the Speculator's measurements and actions. How does PERSEUS's evidence-sharing mechanism differ from this earlier design, and which experiments show the benefit of those differences?

[1] ICLR 2026: Speculative Actions: A Lossless Framework for Faster Agentic Systems

3 comments

W1/Q1: Blocking counterpart

To provide a comparison for this setup, we implemented a blocking variant. PERSEUS-Blocking keeps the same algorithm and Speculator configuration. The only change is delivery: each acquisition direction blocks until its result returns, and the result is admitted in the same turn instead of asynchronously at a later one. The Actor is not told about the wait: it continues with its other planned work and is then held until delivery.

Config (183 tasks)TS (%)PCSTurnTime (s)ESEq. Tokens (M)
PERSEUS-Blocking44.5±1.80.4448.62471.4±92.00.9479.3
PERSEUS (non-blocking)67.8±2.80.74612.19211.5±11.13.2042.2

Blocking lowers TS by 23.3 points and takes 2.2× the time. Timeouts are not the cause (12 of 549 runs). The waiting Actor keeps working on other goals, so broad exploration shifts from the Luna Speculators to the Terra Actor, whose own observations enter its history without ledger filtering. This raises tokens and lowers task quality.

W1/Q2: Meta-Tool derivation with the number of Speculators held fixed

We agree that in Table 2, "w/o Meta-Tool Derivation" also changes swarm width. We added a control that fixes the number of Speculators at N = 8 in both arms on all 183 tasks: 8 constrained variants versus 8 Speculators receiving the unmodified schema.

Fixed width, N = 8TS (%)PCSTime (s)ES
Unmodified schema55.6±1.90.556256.2±22.22.17
Meta-Tool variants52.8±1.10.528246.1±5.82.15

At matched width, the Meta-Tool arm is 2.7 points lower at similar time. This reflects a coupling, not a weakness of derivation: Meta-Tool derivation works together with dynamic dispatch, which launches one Speculator per derived variant as the variant set changes with Actor progress. Forcing a fixed width of 8 either truncates the derived variants or dispatches redundant Speculators, which disrupts the timing of speculation and lowers success. The Table 2 gain therefore comes from derivation and dispatch acting together, with width following the derived domains. We will state this in Section 2.2 and the ablation discussion.

Q3: Ablations on all 183 tasks

We reran the ablations on all 183 tasks:

Config (183 tasks)TS (%)PCSTurnTime (s)ES
w/o Meta-Tool Derivation46.8±3.20.5299.19301.2±22.91.55
w/o Speculative Swarm49.3±2.90.56413.31277.1±17.01.78
w/o Persistent Futures46.3±3.00.53010.09282.8±11.01.63
Full PERSEUS67.8±2.80.74612.19211.5±11.13.20

All three findings hold: removing any component lowers TS by 18.5–21.5 points and roughly halves ES. Meanwhile, we have included this result into the Table 2 in the revision.

W2: Dollar cost and cancelled requests

Dollar costs are taken directly from the per-model billing records of each episode. Every request issued in the episode window is counted, including Speculator requests cancelled before completion.

Method$ per 183-task run$ per task
ReAct70.780.39
PERSEUS129.160.71

In dollars, PERSEUS costs 1.82× ReAct overall. On the tasks that ReAct solves, the ratio is 1.14×. The remaining difference comes from tasks that PERSEUS solves and ReAct does not, where broader exploration carries PERSEUS further (67.8% vs. 56.3% TS overall). We will add dollar costs and this accounting rule to Appendix B.5.

Q4: Difference from the lossy OS-tuning extension of Speculative Actions [1]

In Speculative Actions [1], a fast Speculator applies provisional parameter updates to the live system while the Actor deliberates; the Actor's decision overwrites them (last-write-wins), and the Actor reads a compressed chronology of the Speculator's recent (measurement, action) pairs. PERSEUS differs in three ways:

  1. Isolation instead of overwrite. Speculator calls run in independent work copies and never change authoritative state; only observations reach the Actor. Actions therefore need not be reversible; sending an email, for example, is not.
  2. Complementary acquisition instead of acting ahead on the same control loop. Speculators gather information outside the Actor's call path, and admitted evidence can inform a different later action.
  3. Complete source evidence that persists. The ledger admits complete observations with provenance and suppresses duplicates, instead of passing a model-compressed summary; unfinished acquisitions survive across Actor requests.

Experiments: SA in Table 1 is Speculative Actions [1], and PERSEUS has higher TS on all four benchmarks (67.8% vs. 55.4% overall).

Thanks for adding the controls and full-set ablations. These substantially address my concerns, and I am inclined to raise my rating from 2 to 3.

One point still needs clarification: does the blocking variant use the same evidence filtering and Actor execution rules, with only delivery timing changed? A short execution timeline would help explain the large differences in Actor exploration and token use.

Please also reflect the mixed findings clearly: the fixed-width experiment does not establish a standalone benefit from Meta-Tool derivation, and the overall dollar cost is 1.82× ReAct’s. My support is based on the complete system’s success and latency improvements.

Thank you for the follow-up.

Blocking variant. Yes. Evidence filtering and Actor execution rules are identical in both variants: Speculator results pass through the same ledger, and Actor-issued calls run through the native path, which the ledger does not filter in either variant. Only delivery timing differs.

PERSEUSPERSEUS-Blocking
Actor turn t startsOne Speculator per meta-tool variant is launched; the Actor issues its own callsSame
A Speculator call is still runningIt stays pending; the Actor's turn t+1 starts without itThat direction blocks; the Actor continues with turn t and wait for all in-turn results
The result returnsAdmitted through the ledger at a later turnAdmitted through the same ledger into turn t

In PERSEUS, the Actor proceeds on what is ready and never has a wait to fill. Under blocking, it fills the wait with its own calls, so part of the broad exploration moves from the Luna Speculators to the Terra Actor. The result is fewer but larger turns (8.62 vs. 12.19 turns, 2.7× the equivalent tokens per turn).

Mixed findings. The revision paper states both points:

  • Section 3.3: the fixed-width control does not establish a standalone benefit of Meta-Tool derivation; the gain depends on dispatching one Speculator per derived variant.
  • Section 3.1 and Appendix B.5: the billed cost of PERSEUS is 1.82× ReAct's ($129.16 vs. $70.78 per 183-task run).

As you note, the paper's claims concern the complete system's task success and execution speed; it does not claim a dollar-cost advantage.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.