acceptodds
Under review as a conference paper at ICLR 2027

BDI Architecture for Efficient Orchestration of LLM-based Multi-Agent Systems

Abstract

LLM-based multi-agent systems (MAS) are shifting from manually designed workflows toward automated orchestration, which often relies on repeated workflow search and optimization or additional training. We, human-beings, do not reconsider all possibilities whenever a course of action fails, but maintain an intention and revise the plan only when it is no longer appropriate for reasoning. Inspired by this phenomenon, we propose a BDI (Belief–Desire–Intention)-based orchestration framework that dynamically constructs and adapts MAS workflows at inference time without explicit workflow search or additional training. A single orchestrator maintains beliefs about agent states and execution outcomes, represents candidate sub-goals as desires, and commits to a selected sub-goal as an intention. It then constructs and adapts an execution plan based on the current system state and accumulated experience. When plan adaptation is insufficient to achieve the intended objective, the orchestrator abandons its current intention and selects a new sub-goal from its desires. This intention commitment mechanism reduces repeated deliberation while preserving adaptability to changing execution conditions. Experiments across agentic and general-domain reasoning benchmarks show that our framework improves GAIA accuracy by 4.24 percentage points and ALFWorld success rate by 19.40%p over the strongest baselines, while achieving the highest general-domain average of 80.00 with a GAIA execution cost of $0.0127 per task.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Large Language Models, Multi-Agent Systems, LLM Agents, Agent Orchestration

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_bfJoOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper tries to study how LLM agents can recover from mid-task failures without redoing everything, which could be good for efficiency. It proposes to keep the verified intermediate results and fixing only the failed step when possible, and only revising the plan when local repairs do not work to maintain the performance. Experiments show that, with GPT-4o-mini, the full system reports more task successes on several benchmarks than the tested baselines.

Critical strengths
  1. The problem this paper is trying to address is a practical problem: how much of an agent workflow should change after a local failure to avoid repeating works? Preserving useful intermediate results while repairing the failed part could be a natural way to go

  2. The results seem to be positive. ALFWorld success increases from 23.88% to 43.28%, and GAIA accuracy increases from 17.58% to 21.82% over the strongest evaluated baselines. The paper also reports token usage and call counts, which is helpful

Critical weaknesses
  1. Some ablations do not clearly ablate the components they are meant to isolate. In particular, Appendix A.8 describes removing Desire filtering as removing answer-format constraints. The resulting drop could reflect less complete task instructions, rather than the value of filtering candidate subgoals.

  2. It is unclear what the BDI formulation adds more than retaining intermediate results, verifying outputs, and applying local repairs. A simpler controller with the same tools, verifier, available information, and execution budget would help to understand the advantage of the proposed method.

Questions for authors
  1. What verifier criteria and repair budgets are used to determine that plan adaptation is insufficient? Algorithm 1 describes an escalation rule, but how do you assess whether it abandons an intention too early or persists for too long? A sensitivity analysis of these choices could be useful.

  2. A bit more clarification on implemented definitions of Desire and Intention would be helpful.

  3. How does the method compare with a simple local-repair controller using the same verifier, tools, retained results, and execution budget?

2 comments

Thank you for recognizing the practical importance of deciding how much of an agent workflow should change after a local failure. Below, we address your questions one by one.


W1. Clarifying the Desire Ablation

The Desire ablation is not solely attributable to missing format instructions. Removing either the task goal or instructions independently reduces ALFWorld success by over 20pp.

Variant (ALFWorld, N=134)Task goalInstructionsSuccess Rate (%)Δ vs. Ours (pp)
OursOnOn44.03-
w/o task goalOffOn23.13−20.90
w/o instructionsOnOff19.40−24.63

The w/o Desire setting in Table 4 removes the environment description (rooms and receptacles) used to guide sub-goal selection, rather than answer-format constraints. We will clarify these settings and add the new ablations in Appendix A.8.


W2, Q3. Comparison with a Simple Local-Repair Controller

Our commitment rule improves ALFWorld success by 9.70pp over local repair under identical conditions. Unlike a controller that repeatedly repairs blocked sub-goals, our method reconsiders them after exhausting the repair ladder. Frequent reconsideration also reduces success, although the Table 4 comparison additionally changes the repair policy.

Controller (ALFWorld, N=134)RunSuccess Rate (%)Δ vs. Ours (pp)
Local repair (no reconsideration)Same-time run34.33−9.70
Ours (Single-minded)Same-time run44.03-
Open-minded (reconsider at every observation)Table 426.87−16.42
Ours (Single-minded)Table 443.28-

Q1. Reconsideration Criterion and Sensitivity

Intention reconsideration is triggered by repair-budget exhaustion, with performance sensitive to the budget but robust to the stall-detection window. GAIA allows 5 repairs per node and 20 per run, while ALFWorld allows 8 per sub-goal. The verifier accepts steps with completeness ≥ 0.8.

Repair budget (ALFWorld, N=134)Success Rate (%)Δ vs. Ours (pp)Int. switch / ep
238.06−5.971.54
8 (default)44.03-3.91
1646.27+2.244.51
LLM feasibility judgment41.79−2.245.31

A smaller budget reduces success by 5.97pp, while a larger budget improves it by 2.24pp. LLM-based feasibility judgment underperforms the default by 2.24pp.

Stall window (ALFWorld, N=134)Success Rate (%)Δ vs. Ours (pp)
2 steps42.54−1.49
3 steps (default)44.03-
4 steps44.030.00

Stall-window changes affect success by at most 1.49pp. We will report these results in Table 5 of the revision.


Q2. Definitions of Desire and Intention

Desire defines task objectives and constraints, Intention represents committed sub-goals, and Plan specifies their execution. Re-planning preserves committed objectives, whereas intention switching revises them after repair exhaustion. We will align Eq. (1), Section 3.1.2, and Appendix A.1 accordingly.

ComponentGAIA / QAALFWorld
BeliefVerified results, evidence, execution historyInteraction history, visited and discovered locations
DesireQuery and required answer format (source of candidate sub-goals)Task goal and environment description
IntentionCommitted sub-goal DAGCommitted ordered sub-goal list
PlanAgent, tool, and input per sub-goalNext action selection
Re-planRetry, augment, reassign, or locally re-decompose a sub-goal; its target result stays committedRegenerate the sub-goal list
Intention switchDrop a sub-goal from the committed set after its repairs are exhaustedRegenerated list differs in content
AuthorsOpenReview ↗

Brief Summary of Our Response and Gentle Reminder

Thank you again for your thoughtful and constructive feedback, and for recognizing the practical importance of determining how much an agent workflow should change after a local failure.


Regarding your concerns:

Clarification of the Desire ablation (W1): We clarified that the original Desire ablation removes environment descriptions rather than format constraints, and provided additional ablations showing that removing either the task goal or instructions independently reduces ALFWorld success by over 20 percentage points.

Comparison with a simple local-repair controller (W2, Q3): We conducted a controlled comparison demonstrating that our commitment rule improves ALFWorld success by 9.70pp over local repair without intention reconsideration, while also examining the limitations of frequent reconsideration.

Reconsideration criterion and sensitivity (Q1): We clarified that intention reconsideration is triggered by repair-budget exhaustion and conducted sensitivity analyses showing that performance depends on the repair budget but remains relatively stable across stall-detection windows (≤1.49pp variation).

Definitions of Desire and Intention (Q2): We clarified the definitions of Desire, Intention, and Plan across GAIA/QA and ALFWorld, distinguishing re-planning, which preserves committed objectives, from intention switching, which revises them after repair exhaustion.


As today is the final day of the rebuttal period, we hope that these additional experiments and clarifications address your concerns, and would sincerely appreciate it if you could reconsider your assessment of our manuscript in light of these updates.

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

This paper proposes a Belief–Desire–Intention (BDI) framework for adaptive orchestration of LLM-based multi-agent systems. It maintains execution beliefs, commits to selected intentions, and prioritizes local plan repairs over global replanning. Experiments across six benchmarks demonstrate improved task performance and competitive inference costs.

Critical strengths
  1. Addresses the trade-off between rigid workflow execution and excessive replanning.

  2. Training-free orchestration with hierarchical failure recovery and reusable failure memory.

  3. Notable improvements on GAIA (+4.24%p) and ALFWorld (+19.40%p), supported by component ablations.

Critical weaknesses
  1. The repair hierarchy largely resembles existing verification-driven recovery and adaptive replanning methods. The unique contribution of BDI remains unclear.

  2. Desire and Intention are defined differently across sections, obscuring the distinction between plan repair and intention reconsideration.

  3. Intention switching relies on exhausting predefined repairs rather than explicitly assessing goal feasibility.

  4. The contributions of failure memory, hierarchical repairs, and verification are not independently evaluated.

  5. Efficiency and reproducibility concerns: The method requires relatively many LLM calls, while the verifier implementation, failure classification, and memory retrieval lack sufficient detail.

Questions for authors
  1. What distinguishes BDI-based orchestration from conventional verification-guided local replanning?

  2. Why is the fixed repair hierarchy appropriate, and how is intention infeasibility determined?

  3. Can the authors provide ablations isolating failure memory, repair strategies, and verification?

  4. Can the authors report repair success distributions, intention-switching behavior, and end-to-end latency?

3 comments
AuthorsOpenReview ↗

Official Comment by Authors (1/2)

Thank you for recognizing our training-free orchestration and effective balance between persistence and adaptation. We address your questions below.


W1, Q1. Difference from Verification-guided Local Replanning

Our key contribution is the commitment rule that determines when to stop local repair and reconsider the Intention. Unlike conventional verification-guided recovery, this rule enables intention revision after repair exhaustion, improving ALFWorld success by 9.70pp under identical conditions.

Controller (ALFWorld, N=134)RunSuccess Rate (%)Δ vs. Ours (pp)
Local repair (no reconsideration)Same-time run34.33−9.70
Ours (Single-minded)Same-time run44.03-
Open-minded (reconsider at every observation)Table 426.87−16.42
Ours (Single-minded)Table 443.28-

In BDI terms, Beliefs retain verified results, Plans undergo local repair, and Intentions change only after repair exhaustion. The open-minded variant also performs worse, although it additionally changes the repair policy.


W2. Consistent Definitions of Desire and Intention

Desire defines task objectives and constraints, Intention specifies committed sub-goals, and Plan determines their execution. Re-planning preserves committed objectives, whereas intention switching revises them after repair exhaustion.

ComponentGAIA / QAALFWorld
BeliefVerified results, evidence, execution historyInteraction history, visited and discovered locations
DesireQuery and required answer format (source of candidate sub-goals)Task goal and environment description
IntentionCommitted sub-goal DAGCommitted ordered sub-goal list
PlanAgent, tool, and input per sub-goalNext action selection
Re-planRetry, augment, reassign, or locally re-decompose a sub-goal; its target result stays committedRegenerate the sub-goal list
Intention switchDrop a sub-goal from the committed set after its repairs are exhaustedRegenerated list differs in content

Table 4 counts re-plans and intention switches separately (2.50 vs. 2.43 per episode). We will align Eq. (1), Section 3.1.2, and Appendix A.1 accordingly.


W3, Q2. Repair Ladder and Intention Infeasibility

Intention infeasibility is defined by repair-budget exhaustion, triggering reconsideration without additional LLM calls. The repair ladder progresses from smaller to larger changes before switching Intention.

Repair budget (ALFWorld, N=134)Success Rate (%)Δ vs. Ours (pp)Int. switch / ep
238.06−5.971.54
8 (default)44.03-3.91
1646.27+2.244.51
LLM feasibility judgment41.79−2.245.31

A smaller budget reduces success, while a larger one improves it. LLM-based feasibility judgment underperforms the default and increases intention switches, supporting budget-based reconsideration.


W4, Q3. Component-wise Ablation

Removing intention reconsideration or verification independently reduces ALFWorld success by 9.70pp and 5.97pp, respectively.

Variant (ALFWorld, N=134)RunSuccess Rate (%)Δ vs. Ours (pp)
OursSame-time run44.03-
w/o intention reconsiderationSame-time run34.33−9.70
w/o verification (stall detection off)Same-time run38.06−5.97
OursTable 443.28-
Re-plan at every action, no repair ladderTable 431.34−11.94

Each Δ uses its run-specific baseline. The Table 4 variant changes both reconsideration and repair policy. Failure memory only selects the initial repair rung (Algorithm 1, line 10); we will clarify its auxiliary role and focus our contribution claim on the reconsideration rule.

AuthorsOpenReview ↗

Official Comment by Authors (2/2)


W5. LLM Calls and Implementation Details

Our method achieves higher GAIA accuracy with 29% fewer LLM calls and one-third the cost of OWL/Workforce.

Method (GAIA 165)Acc (%)LLM calls / taskTok/task (K)Cost/task ($)
OWL / Workforce17.5853.731198.20.1840
Ours21.8238.33387.30.0603

Verification combines deterministic checks with LLM-based scoring (accept ≥ 0.8; augment at 0.5–0.8), with repair budgets of 5 per node and 20 per run. Failure memory selects repair strategies based on past outcomes for similar failures. We will clarify these details in Appendices A.3 and A.4.


Q4. Repair Outcomes, Intention Switching, and Latency

On ALFWorld, repair success rates range from 42.4% to 68.1%, with 3.91 intention switches and 48 s latency per episode.

Repair (ALFWorld, N=134)AttemptsPassedPass rate (%)
Reject a repeated failed command31321368.1
Stall repair (rewrite the sub-goal list)48320542.4
Forced exploration582644.8
MetricALFWorld (/episode)GAIA (/task)
LLM calls43.638.3
Re-plans4.041.58
Intention switches3.91-
Latency48 s165 s (median 147 s)

All statistics are from the same evaluation run (44.03% ALFWorld success). We will report them in Appendix Table 6.

AuthorsOpenReview ↗

Brief Summary of Our Response and Gentle Reminder

Thank you again for your thoughtful and constructive feedback, and for recognizing the value of our training-free orchestration and effective balance between persistence and adaptation.


Regarding your concerns:

Difference from verification-guided local replanning (W1, Q1): We clarified that our key contribution is the explicit commitment rule governing when local repair gives way to intention reconsideration, and demonstrated that this rule improves ALFWorld success by 9.70pp under controlled conditions.

Consistent definitions of Desire and Intention (W2): We clarified the definitions of Desire, Intention, and Plan across GAIA/QA and ALFWorld, distinguishing re-planning, which preserves committed objectives, from intention switching, which revises them after repair exhaustion.

Repair ladder and intention infeasibility (W3, Q2): We clarified that intention reconsideration is triggered by repair-budget exhaustion rather than additional LLM judgments, and showed that increasing the repair budget improves ALFWorld success, while LLM-based feasibility judgment underperforms the default.

Component-wise ablation (W4, Q3): We conducted controlled ablations demonstrating that removing intention reconsideration and verification independently reduces ALFWorld success by 9.70pp and 5.97pp, respectively, isolating their contributions.

LLM calls and implementation details (W5): We provided detailed implementation settings and confirmed that our method achieves higher GAIA accuracy (21.82% vs. 17.58%) with 29% fewer LLM calls and approximately one-third the cost of OWL/Workforce.

Repair outcomes, intention switching, and latency (Q4): We provided additional statistics on repair outcomes, intention switching, and computational latency, including repair success rates of 42.4–68.1%, 3.91 intention switches, and 48 seconds per episode on ALFWorld.


As today is the final day of the rebuttal period, we hope that these additional experiments and clarifications address your concerns, and would sincerely appreciate it if you could reconsider your assessment of our manuscript in light of these updates.

Reviewer_6ZYuOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents a BDI-inspired controller for LLM multi-agent systems that retains verified intermediate results and escalates through retrying, augmenting, reassigning and locally replanning before reconsidering an intention. A memory of verifier-approved repairs can skip previously ineffective recovery steps, without additional model training. Experiments with GPT-4o-mini report gains on GAIA and ALFWorld and the highest four-benchmark general-domain average among the evaluated methods. The practical contribution is promising, but the paper does not yet clearly isolate its distinctive recovery policy or establish the claimed BDI interpretation of the results.

Critical strengths

1. A concrete and useful approach to controlling recovery scope. The repair ladder, preservation of accepted evidence, and memory-based choice of starting rung form an interpretable inference-time policy (Section 3.1.2, Algorithm 1). This is a worthwhile systems contribution: it turns the intuition of persisting with a goal while changing its execution into explicit recovery operations that could be reused and studied independently of model training.

2. The evaluation provides encouraging evidence about persistence versus adaptation. Table 1 reports ALFWorld success of 43.28% versus 23.88% for the strongest listed baseline, while Table 4 compares the default policy with blind commitment (38.81%) and frequent reconsideration (26.87%). Together with the heterogeneous tasks in Table 2, these results suggest that the package has practical value beyond a single task type. The authors also acknowledge that it neither improves GAIA Level 3 nor minimizes every resource measure, which appropriately limits the empirical claim.

Critical weaknesses

1. The distinctive mechanism is insufficiently specified and isolated. Eq. (1) identifies an intention with one selected subgoal, whereas Section 3.1.1 and Appendix A.1 identify it with an entire subtask DAG; this ambiguity directly affects what Table 4 counts as an intention switch. More consequentially, Appendix A.8 defines “w/o desire filtering” as removing answer-format constraints, while Section 4.2.4 interprets that ablation as evidence for restricting candidate subgoals. Removing formatting does not isolate goal selection, and the no-commitment variant also changes the repair-selection policy. This matters particularly because uncited prior systems already implement much of the surrounding structure: VMAO (Zhang et al., 2026) preserves completed DAG results during verification-driven retries, and ALAS (Geng and Chang, 2025) expands local repair while preserving unaffected work. Their different settings do not establish equivalence, but they make the specific escalation-and-memory policy the important increment to substantiate. As written, the evidence supports a useful package more clearly than the claimed explanation in terms of BDI components; resolving this requires consistent operational definitions and unconfounded evidence, not merely additional citations.

2. The headline effectiveness–cost comparison cannot be fully reconciled from the reported protocol. With the 53/86/26 GAIA level counts in Section 4.1, Table 1's scores imply (53×32.10 + 86×21.20 + 26×3.85)/165 = 21.97%, rather than the reported 21.82%; ordinary rounding cannot explain the difference. Table 3 also needs explicit token/call and billing definitions: Vanilla has 399.9K tokens but 1.01 LLM calls per task, while MaAS costs 0.0159 for 4.8K tokens versus 0.0127 for the proposed method's 98.1K, despite the stated common backbone. These observations do not prove incorrect execution or billing, but require reconciled aggregates and input/output/cache accounting. Baseline tool access and interaction budgets are also unspecified. Because efficiency is a central contribution, the shared-model statement alone is insufficient to establish a controlled comparison. These issues appear addressable through reporting corrections and protocol evidence, but currently weaken confidence in the claimed advantage.

Questions for authors
  1. What precisely are desire, intention and plan in the evaluated implementation, and which state changes count as replanning or intention switching? How are downstream dependencies handled when a required subgoal is abandoned?

  2. Does the desire-filtering ablation change anything beyond answer-format constraints? Which verifier, formatting, repair operators and budgets remain fixed across commitment variants, and what evidence separates the effect of intention persistence from changing the repair-selection policy?

  3. Can you reconcile the GAIA level and overall accuracies and provide the token/call/billing definitions underlying Table 3, together with baseline tool access and interaction budgets? In particular, are the token and call averages computed over the same tasks and executions, and how are cached tokens counted and priced?

3 comments
AuthorsOpenReview ↗

Official Comment by Authors (1/2)

Thank you for recognizing our interpretable recovery policy and effective intention management. We address your concerns below.


W1-1, Q1. Definition of Intention and Intention Switching

Intention refers to the committed sub-goal DAG, not an individual sub-goal. Re-planning modifies execution while preserving the committed objectives, whereas intention switching changes the committed sub-goals after repair options are exhausted. Table 4 counts these separately (2.50 re-plans and 2.43 switches per episode).

ComponentGAIA / QAALFWorld
BeliefVerified results, evidence, execution historyInteraction history, visited and discovered locations
DesireQuery and required answer formatTask goal and environment description
IntentionCommitted sub-goal DAGCommitted ordered sub-goal list
PlanAgent, tool, and input per sub-goalNext action selection
Re-planModify execution while preserving the sub-goal objectiveRegenerate the list without changing its sub-goals
Intention switchDrop a sub-goal after exhausting repairsChange the committed sub-goal list after failed repairs

In GAIA/QA, dropping a sub-goal blocks its dependents, and the final answer uses the remaining verified results. We will align Eq. (1), Section 3.1.2, and Appendix A.1 with these definitions.


W1-2. Clarifying the Desire Ablation

The Desire ablation is not solely attributable to missing format instructions. Removing either the task goal or instructions separately reduces ALFWorld success by over 20 percentage points.

Variant (ALFWorld, N=134)Task GoalInstructionsSuccess Rate (%)Δ (pp)
OursOnOn44.03–
w/o task goalOffOn23.13−20.90
w/o instructionsOnOff19.40−24.63

The w/o Desire setting in Table 4 removes the environment description (rooms and receptacles) used to guide sub-goal selection, not answer-format constraints. The additional ablations distinguish the effects of task goals and instructions. We will correct Appendix A.8 and report these variants explicitly.


W1-3, Q2. Isolating the Commitment Rule

A controlled comparison isolates the benefit of intention reconsideration. With the repair ladder, verifier, prompts, tools, and budget fixed, disabling reconsideration reduces ALFWorld success by 9.70pp.

Controller (ALFWorld, N=134)RunSuccess Rate (%)Δ vs. Ours (pp)
Local repair (no reconsideration)Same-time run34.33−9.70
Ours (Single-minded)Same-time run44.03–
Open-minded (reconsider at every observation)Table 426.87−16.42
Ours (Single-minded)Table 443.28–

Each Δ uses the corresponding run-specific baseline. Unlike the controlled comparison, the w/o Intention commitment variant in Table 4 also removes the repair ladder, so its 16.42pp gap cannot be attributed solely to reconsideration. We will clarify this distinction and use the controlled comparison to support the commitment rule.


W1-4. Relation to VMAO and ALAS

Our distinction from VMAO and ALAS is the explicit rule governing when local repair gives way to intention revision. While these methods also preserve verified results and support local repair, our framework abandons a sub-goal when its repair options are exhausted. The controlled comparison in W1-3 shows a 9.70pp gain from this rule.

Failure memory only selects the initial repair rung and does not govern intention switching. We will clarify its auxiliary role and focus our contribution claim on the commitment rule.


W2-1. GAIA Level-wise Accuracy

The overall accuracy (21.82%, 36/165) is correct; two level-wise values were mistranscribed.

LevelCorrectAccuracy (%)
Level 117/5332.08
Level 218/8620.93
Level 31/263.85
Overall36/16521.82

We will correct Table 1 and report correct/total counts alongside all percentages.

AuthorsOpenReview ↗

Official Comment by Authors (2/2)

W2-2, Q3. Token, Call, and Cost Accounting

Recomputed from the corresponding evaluation runs, our method achieves the highest GAIA accuracy at one-third the cost of OWL/Workforce (0.0603 vs. 0.1840 per task).

Method (GAIA 165)Acc (%)Tok/task (K)Cost/task ($)
Vanilla / CoT9.70407.00.0623
Fixed MAS14.55515.00.0793
Reflexion12.12473.30.0730
A-MEM12.12452.20.0693
AFlow12.12195.80.0303
MaAS10.9198.80.0159
OWL / Workforce17.581198.20.1840
Agent-Co-op10.91401.60.0623
Ours21.82387.30.0603

The submitted Ours cost (98.1K tokens, 0.0127) came from a different run with cached-token discounts. The corrected values (387.3K tokens, 38.33 LLM calls, 0.0603) correspond to the reported 21.82% accuracy. Vanilla's previous call count measured sessions rather than LLM responses, while MaAS's token count omitted tool-related content. All methods use GPT-4o-mini (temperature 0). Tokens include input and output, calls count LLM responses, and cached inputs are charged at full rates (0.15/ M input, 0.60 / M output). Costs are averaged over available records from the corresponding runs (165 tasks for Ours/OWL, 164 for Vanilla/A-MEM, and 139 for MaAS). We will correct Table 3.


W2-3, Q3. Baseline Tools and Budgets

We provide the same tool interface and 40-round limit to Vanilla/CoT, Fixed MAS, Reflexion, A-MEM, Memp, Agent-Co-op, and AFlow. The interface includes DuckDuckGo search, web/document extraction, browser simulation, and code execution. Our method uses the same tools, with a 40-round limit and a 900-second timeout per task.

OWL/Workforce, MaAS, and MAS² retain their official configurations. We will document these settings and distinguish shared-budget baselines from those using official configurations in Appendix A.7.

AuthorsOpenReview ↗

Brief Summary of Our Response and Gentle Reminder

Thank you again for your thoughtful and constructive feedback, and for recognizing the value of our interpretable recovery policy and effective intention management.


Regarding your concerns:

Definition of intention and intention switching (W1-1, Q1): We clarified that intention refers to a committed sub-goal DAG rather than an individual sub-goal, distinguishing re-planning, which preserves committed objectives, from intention switching, which revises them after repair options are exhausted.

Clarification of the Desire ablation (W1-2): We clarified that the original Desire ablation removes environment descriptions rather than format constraints, and provided additional ablations showing that removing either the task goal or instructions reduces ALFWorld success by over 20 percentage points.

Isolating the commitment rule (W1-3, Q2): We conducted a controlled comparison with the repair ladder, verifier, prompts, tools, and budget fixed, demonstrating that disabling intention reconsideration reduces ALFWorld success by 9.70pp.

Relation to VMAO and ALAS (W1-4): We clarified that our key distinction lies in the explicit commitment rule governing when local repair gives way to intention revision, supported by the 9.70pp improvement from the controlled comparison.

GAIA level-wise accuracy (W2-1): We identified and corrected two mistranscribed level-wise accuracy values, confirming that the reported overall accuracy of 21.82% (36/165) remains unchanged.

Token, call, and cost accounting (W2-2, Q3): We recomputed resource usage from the corresponding evaluation runs and corrected the accounting, confirming that our method achieves the highest GAIA accuracy (21.82%) at approximately one-third the cost of OWL/Workforce ($0.0603 vs. $0.1840 per task).

Baseline tools and budgets (W2-3, Q3): We clarified the shared tool interface and 40-round limit used by our method and applicable baselines, while distinguishing methods evaluated under their official configurations.


As today is the final day of the rebuttal period, we hope that these additional experiments and clarifications address your concerns, and would sincerely appreciate it if you could reconsider your assessment of our manuscript in light of these updates.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.