acceptodds
Under review as a conference paper at ICLR 2027

Topology-Consistent Task Planning over Cellular Workflow Complexes for LLM-based Agents

Abstract

Task planning for LLM agents requires workflows that satisfy both user intent and complex sub-task dependencies. While existing planners excel at sequential or directed acyclic graph (DAG)-like structures, they struggle with high-dimensional dependencies—such as verification-correction loops and convergent branch merging—inherent in real-world tool orchestration. We present TopoPlanner, a framework that lifts tool dependency graphs into cellular workflow complexes and uses them as topology-aware context for LLM tool planning. Specifically, TopoPlanner attaches 2-cells to fundamental cycles, retrieves a request-relevant closed subcomplex with cosheaf-consistency regularization, performs multi-dimensional structural reasoning over 0-, 1-, and 2-cells, and interfaces the resulting cellular representation with the planner LLM through a topology-guided conditioning layer. Evaluated on four tool-planning benchmarks, including the three TaskBench domains and ToolBench, all extended with curated cyclic-dependency cases, TopoPlanner consistently outperforms prior baselines, showing significant gains on non-linear workflows. Code and datasets are available at: https://anonymous.4open.science/r/TopoPlaner-AE05.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLM Agents, Tool Planning, Topological Deep Learning

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_3XfcOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The authors propose TopoPlanner, a framework designed to lift tool-dependency graphs into two-dimensional cellular workflow complexes, aiming to address task planning challenges faced by large language model (LLM) agents when dealing with non-linear workflows. By leveraging cosheaf-consistency-based subcomplex retrieval and multidimensional structural reasoning, the paper extracts topology-aware graph tokens to guide the model in generating tool sequences. The authors evaluate their framework on an augmented benchmark dataset containing non-linear structures, demonstrating performance improvements across key metrics.

Critical strengths

Introducing high-order topological tools into the domain of LLM agent task planning offers a novel research perspective.

Demonstrating good empirical results on datasets containing non-linear structures such as loops, branches, and merges.

Critical weaknesses

The mathematical formulation appears unrigorous to me: Equation (8) merely performs pairwise cosine alignment rather than establishing true categorical cosheaf consistency. In Equation (9), because the retrieval objective relies on a simple arithmetic mean without size or connectivity constraints, how do the authors prevent the algorithm from severely degrading into selecting a single high-scoring node? Could the authors also provide the missing discrete optimization pseudo-code used to solve this retrieval objective?

I am concerned about the stability of cellular lifting, as it heavily depends on the arbitrary choice of the underlying spanning tree. How sensitive is the framework when different spanning trees fundamentally alter 2-cell neighborhoods and message-passing paths? Additionally, how do the authors justify using an unoriented cycle topology, given its evident semantic misalignment with actual directed control-flow (such as DAG merges versus feedback loops)?

The ablation studies fail to isolate the true causal contribution of the topological representation. Have the authors considered evaluating against control baselines with randomized 2-cell associations or equal-capacity hypergraph models to prove that the gains stem from correct topology rather than added parameter capacity?

I found the evaluation protocol quite opaque regarding how flat text sequences generated by LLMs are parsed back into directed graphs with loops and merges. Without reporting standard deviations across multiple runs, how can we rule out overfitting to synthetic data motifs? Moreover, why is quantitative evaluation of computational complexity and inference latency missing, given that introducing high-order cellular retrieval and multidimensional reasoning undeniably increases overhead?

I question the practical utility of relying so excessively on pre-constructed static topologies in real-world open environments. How does the framework maintain robustness when facing dynamic API schema updates, parameter changes, or temporary service disruptions? Does the framework offer any online incremental update mechanism, and what is the computational cost of rebuilding these cellular complexes from scratch?

Questions for authors

please refer to Weaknesses

5 comments
AuthorsOpenReview ↗

Response to Reviewer 3Xfc

We thank the reviewer for recognizing the value of representing higher-order workflow dependencies and for pressing on the mathematical formulation, lifting stability, structural controls, evaluation protocol, and open-environment cost. Our central claim is specific: a 2-cell makes a region jointly formed by valid tool dependencies available to request-conditioned retrieval and finite-depth structural reasoning. A directed graph can encode the same dependencies; an unoriented region is not itself an executable feedback policy. We answer the five weaknesses in their original order and distinguish structural plan prediction from runtime task completion throughout.

W1. What do Eqs. 8–9 and the retrieval algorithm actually establish?

Eq. 8 supplies the incidence-level compatibility signal used in cellular retrieval. In its displayed form, each coface-to-boundary term is nonnegative and, for nonzero vectors, 1 − cos(a,b) = ½||â − b̂||². Zero residual on every incidence of a selected, boundary-closed subcomplex therefore gives agreement of the normalized representations along its included boundaries. This is useful specifically for the proposed regional representation: selecting a 2-cell also retains the edges and nodes against which that region is assessed. The result concerns compatibility of the selected representations. Whether learned incidence maps satisfy path-composition identities is a separate, stronger question that Eq. 8 alone does not test. We will state the local zero-residual result with its assumptions and distinguish it from categorical cosheaf consistency.

Eq. 9 describes the relevance–compatibility trade-off; its mean-score form alone does not impose a minimum subcomplex size. For example, a node scoring 1 while all other cells score lower is a feasible singleton optimum. The implemented selector instead uses an explicit node budget followed by closure-aware edge/face selection. The existing appendix algorithm describes cycle-basis construction; to make retrieval reproducible, we will add the following selection steps:

score nodes, directed edges, and faces for the request, including gated incidence error
select the top-k relevant nodes; prefilter edge candidates
retain candidate edges whose endpoints are selected and faces whose boundary nodes are selected
score each edge/face with the gain from adding it together with missing boundary cells
sort candidates by this closure-aware gain; add under edge/face budgets with boundary closure

Table X1 shows the resulting selections under the evaluated default budgets. The non-singleton behavior follows the budgeted algorithm; it is not inferred from the unconstrained mean objective or from an exact PCST guarantee.

Table X1. Default-budget retrieval trace; mean edge/face counts.

DomainTest requestsCandidate nodesSelected nodes per requestSelected edgesSelected facesSingleton-onlyClosure violations
HuggingFace6802321174.61537.22600
Multimedia8204034325.30247.07600

The component diagnostic reports lower structural scores when consistency or boundary closure is removed, supporting their operational role in the planner. We will make the mathematical and implementation descriptions consistent: the inspected implementation uses a gated squared incidence residual, while the displayed Eq. 8 uses cosine. The revision will state the appropriate local compatibility result and the actual budgeted retrieval procedure without treating either as a proof of functoriality.

AuthorsOpenReview ↗

2nd Response to Reviewer 3Xfc

W2. Does the spanning forest change the representation, and what does an unoriented face mean?

For a fixed spanning forest of the underlying undirected graph, each non-tree edge closes one fundamental cycle. The unfiltered cycles form a basis of the same cycle space under different forest choices. Our implemented length bound retains only short cycles; the retained 2-cells and incidence neighborhoods can therefore change with the forest. In a fixed-checkpoint test that changes the prepared face incidence consumed by the model, the default forest yielded 54 short faces on HF and 85 on MM; ten alternate forests yielded 42–62 and 53–77.

Table X2. Fixed-checkpoint forest sensitivity, N-F1 / L-F1 (%). Alternative values are mean ± sample SD across ten forests.

DomainInference budgetDefault forestDefault repeatTen alternative forests
HuggingFaceOriginal81.83 / 55.9581.74 / 55.7681.94±0.09 / 56.07±0.17
MultimediaOriginal88.32 / 68.3088.37 / 68.2688.38±0.10 / 68.37±0.16
HuggingFaceHalf-budget stress test82.33 / 56.6982.29 / 56.7181.09±0.70 / 54.93±0.94
MultimediaHalf-budget stress test87.96 / 67.8588.06 / 68.0187.90±0.14 / 67.75±0.17

Under the original budget, aggregate scores vary little across the tested forests, while the more selective HF half-budget diagnostic shows that stability depends on retrieval budget. The half-budget rows change the inference policy and are not additional training-seed runs. A DAG merge may form a cycle in its undirected view without being directed feedback. The 2-cell marks that joint region; directed 1-cells retain prerequisite direction. This keeps the representation result distinct from observation-conditioned loop execution.

W3. Is the gain due to correct regional incidence rather than extra capacity?

The reason for introducing 2-cells is not that directed edges cannot represent prerequisites. Rather, a merge or loop-shaped workflow couples several such edges through a shared boundary; pairwise propagation leaves that joint region implicit. TopoPlanner uses the boundary incidence of each 2-cell both to exchange information across the region and to retrieve the region together with its boundary. This is the distinction the component study is intended to test.

Table X3 shows the full model, the edge-only 1-skeleton, and a hypergraph alternative across both backbones and domains. On Multimedia, Vicuna-7B reaches 66.24 L-F1 with the full model, compared with 50.13 without 2-cells and 59.73 with hypergraph grouping; CodeLlama-7B reaches 69.59, versus 56.35 and 62.79. The same ordering holds on HuggingFace. Removing boundary closure or the consistency term also lowers performance in every configuration. Taken together, these comparisons connect the gain to how regional boundaries are represented and used, rather than merely to the presence of an additional high-order unit.

Table X3. Component diagnostics on HuggingFace and Multimedia; N-F1 / L-F1 / Acc (%).

BackboneVariantHuggingFace N/L/AccMultimedia N/L/Acc
Vicuna-7Bw/o 2-cell75.58/46.61/30.8873.29/50.13/34.15
Vicuna-7Bw/o retrieval76.07/47.88/32.2177.97/55.11/35.37
Vicuna-7Bw/o consistency loss80.69/53.31/38.6881.47/58.29/40.98
Vicuna-7Bno Eq. 9 consistency77.39/49.09/33.6880.95/57.03/41.71
Vicuna-7Bno boundary closure75.66/47.07/32.5081.41/57.71/43.17
Vicuna-7Bhypergraph variant79.47/51.66/37.6582.38/59.73/44.63
Vicuna-7BTopoPlanner82.23/55.72/42.3586.80/66.24/54.76
CodeLlama-7Bw/o 2-cell78.95/51.87/35.2975.90/56.35/43.54
CodeLlama-7Bw/o retrieval79.75/56.96/35.5983.89/62.92/49.27
CodeLlama-7Bw/o consistency loss80.80/53.83/38.8287.12/67.61/53.54
CodeLlama-7Bno Eq. 9 consistency79.44/53.19/38.2486.24/64.34/53.41
CodeLlama-7Bno boundary closure78.74/52.00/35.5985.59/66.14/55.37
CodeLlama-7Bhypergraph variant77.35/50.24/31.3284.20/62.79/49.15
CodeLlama-7BTopoPlanner83.43/59.62/47.3589.08/69.59/62.20

We will make this mechanism and the different ablation settings clearer in the revision: the edge-only and hypergraph variants probe the representation, while the closure and consistency rows probe how the retrieved regions are used. The evidence supports the boundary-aware design; it does not require claiming that no other formalism can encode the same workflow.

AuthorsOpenReview ↗

3rd Response to Reviewer 3Xfc

W4. What is recovered from a flat sequence, and what do variance and runtime measurements show?

TopoPlanner introduces topology before decoding: retrieved, boundary-closed regions inform which tools the LLM selects and how it orders their dependencies. Its final output remains an ordered tool list, as in the benchmark's common planning interface. We assess the resulting nonlinear-structure predictions directly. Across four domains, merge/reuse edge recall rises from 19.04% with GTool to 39.72% with TopoPlanner on Vicuna-7B (Table X4; 804 examples and 2,325 relevant gold edges per backbone). On Multimedia, the stricter example-level dependency-order score rises from 8.29% to 40.37% (Table X5); missing prerequisites count as errors. These gains show that structural conditioning improves the selection and ordering of tools involved in shared dependencies, despite the common sequence output.

Table X4. Merge/reuse dependency prediction (%).

BackboneMethodExamplesN-F1L-F1Merge/reuse edge recall
Vicuna-7BGTool80451.1822.0619.04
Vicuna-7BTopoPlanner80476.1646.9239.72
CodeLlama-7BGTool80464.5635.2729.06
CodeLlama-7BTopoPlanner80478.9449.9141.01

Table X5. Gold-dependency ordering, Dep. / Pair (%).

BackboneMethodHFMMDaily Life
Vicuna-7BGTool9.71 / 24.188.29 / 23.6939.86 / 57.32
Vicuna-7BTopoPlanner27.94 / 51.7440.37 / 64.2960.14 / 79.50
CodeLlama-7BGTool25.74 / 44.9132.56 / 49.7348.11 / 62.30
CodeLlama-7BTopoPlanner33.24 / 58.6448.78 / 71.8063.57 / 83.28

The benchmark's shared scorer projects adjacent predicted tool pairs onto the available directed tool graph to compute L-F1. We will explain this mapping and define Acc precisely as the fraction of examples with node F1 at least 0.99, so neither measure is mistaken for tool-execution success. The reported 10,000-resample paired bootstrap measures uncertainty over matched test examples for fixed models; we will label it accordingly and distinguish it from variation across training runs.

We also quantify the cost of this structural interface. Closed-subcomplex retrieval takes 22.89–39.45 ms per request on CPU across the four evaluated graphs (Table X6). In same-card, reverse-order H800 profiles, Full adds 5.36–24.35 ms per request relative to Hypergraph in the measured static-inference comparisons (Table X7).

Table X6. Retrieval-pool size and CPU retrieval mean (50 fixed queries).

DomainNodes / directed edges / retained 2-cellsEdge candidatesCPU ms
HuggingFace23 / 225 / 5422523.65
Multimedia40 / 449 / 8544933.11
Daily Life40 / 1,560 / 14651239.45
ToolBench245 / 366 / 1936622.89

Table X7. Static CodeLlama-7B model inference, batch 1, ten warmup batches excluded (ms/request).

DomainSame-card orderFull mean / median / p95Hypergraph mean / median / p95Full − Hyper mean
HuggingFaceFull → Hypergraph540.47 / 512.83 / 822.60525.63 / 507.36 / 774.44+14.84
HuggingFaceHypergraph → Full532.27 / 510.73 / 787.72507.92 / 488.19 / 750.43+24.35
MultimediaFull → Hypergraph622.07 / 601.50 / 925.23616.71 / 595.68 / 931.90+5.36
MultimediaHypergraph → Full618.47 / 599.59 / 923.50605.68 / 586.00 / 914.46+12.79

These measurements characterize the planning stages; downstream tool calls are outside their timed scope.

Reviewer_zGcGOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes TopoPlanner, a topology-aware task planning framework for LLM-based agents. The authors argue that conventional graph-based planning methods do not adequately capture complex workflow structures, such as loops, branch merging, and reuse of intermediate results. To address this, they introduce cellular workflow complexes that represent higher-order dependencies through 2-cells.

The proposed framework consists of four components: cellular workflow lifting, cosheaf-consistent retrieval, multi-dimensional structural reasoning, and topology-guided LLM planning. The authors also extend four existing tool-planning benchmarks with topology-guided workflows and evaluate the framework using several LLM backbones. The results show substantial improvements in node F1, link F1, and exact plan accuracy compared with prompt-based and graph-enhanced baselines.

Critical strengths
  1. The paper addresses an important problem in LLM-based task planning, particularly when workflows involve multiple dependencies, feedback loops, or merging branches. Representing higher-order dependency structures explicitly is an interesting direction beyond conventional node-edge graph representations. The proposed framework provides a systematic way to incorporate these structures into LLM planning, and the overall methodology is technically well developed.

  2. The experimental results show substantial improvements over existing methods across four benchmarks and multiple LLM backbones. The improvements in link F1 and exact plan accuracy are particularly notable, suggesting that TopoPlanner helps recover more accurate dependency structures rather than simply identifying relevant tools.

Critical weaknesses
  1. The necessity of the proposed topological formulation is not sufficiently justified. The paper introduces a relatively complex framework involving cellular complexes, cosheaf-consistent retrieval, and multi-dimensional structural reasoning. While the motivation for explicitly modeling workflow dependencies is reasonable, it remains unclear why these structures require such a sophisticated topological formulation. The workflow patterns considered in the paper, including loops, branch merging, and intermediate-result reuse, can already be represented using directed graphs or workflow state machines. The paper argues that conventional graph representations cannot adequately capture higher-order relationships, but the distinction between representing these relationships explicitly through 2-cells and modeling them through simpler dependency-based approaches is not fully established.

  2. The evaluation's practical relevance is uncertain. A major concern is the construction of the extended evaluation datasets. In Section 4.2, the authors first specify target workflow topologies, including loop, merge, and loop-merge structures, and then generate natural-language requests conditioned on these structures. Since TopoPlanner is specifically designed to capture these patterns, this evaluation setup may favor the proposed structural representation. The evaluation data do not fully establish whether the generated workflows reflect the complexity and distribution of real-world agent tasks.

  3. The evaluation primarily relies on node F1, link F1, and exact plan accuracy. These metrics measure agreement with reference workflow structures rather than actual task completion.

Questions for authors
  1. What specific advantages does the cellular complex representation provide over simpler directed graphs or workflow state machines that explicitly model loops and merging dependencies? Could the authors compare against a simpler structure-aware planner to better isolate the benefits of the proposed higher-order representation?

  2. Since the extended benchmarks are generated by specifying target topologies before producing natural-language requests, how representative are these tasks of naturally occurring agent workflows? Have the authors considered evaluating the framework on independently collected real-world workflows that were not generated under predefined topology constraints?

  3. Can the authors provide evidence that the improvements in structural plan prediction translate into higher task completion rates in interactive tool environments? This would be particularly valuable for verification-correction loops, where successful execution requires dynamic decisions based on intermediate observations.

  4. Appendix E includes experiments with Qwen3-235B and DeepSeek-V3 using an API-compatible variant of TopoPlanner. However, the paper does not evaluate proprietary frontier models such as GPT, Claude, or Gemini. Have the authors considered comparing against these stronger planners using direct or structured prompting? Given the complexity of the proposed framework, it would be useful to understand whether simpler planning approaches with stronger LLMs can achieve comparable results.

7 comments
AuthorsOpenReview ↗

Response to Reviewer zGcG

We thank the reviewer for recognizing the importance of nonlinear dependencies and for asking what the cellular formulation contributes beyond simpler graph representations, how the topology-guided data should be interpreted, and what the offline scores establish. We address the three weaknesses and four questions in their original order, combining each related weakness and question.

1. What does a cellular complex add beyond a directed graph or workflow state machine? (W1/Q1)

We do not claim that directed graphs cannot encode branches, merges, or feedback. The distinction is whether a finite-depth planner can treat the region jointly formed by those dependencies as an explicit object for representation and retrieval. When two elements of one workflow region are more than L graph hops apart, an L-layer local node–edge network cannot exchange information between them; attaching a shared 2-cell creates a boundary-to-face-to-boundary route in two incidence steps. Boundary-closed retrieval then keeps that region's constituent nodes and edges together. This is why a cell can help this planner use a relation that was already encodable by edges. A suitable hypergraph may offer a similar shortcut.

Table Z1 reports the complete component/representation diagnostic (N-F1 / L-F1 / Acc, %). It includes both a node–edge-only variant and a hypergraph alternative:

Table Z1. Component and structural-representation diagnostics.

BackboneVariantHuggingFace N/L/AccMultimedia N/L/Acc
Vicuna-7Bw/o 2-cell75.58/46.61/30.8873.29/50.13/34.15
Vicuna-7Bw/o retrieval76.07/47.88/32.2177.97/55.11/35.37
Vicuna-7Bw/o L_CI80.69/53.31/38.6881.47/58.29/40.98
Vicuna-7Bno Eq. 9 consistency77.39/49.09/33.6880.95/57.03/41.71
Vicuna-7Bno boundary closure75.66/47.07/32.5081.41/57.71/43.17
Vicuna-7Bhypergraph variant79.47/51.66/37.6582.38/59.73/44.63
Vicuna-7BTopoPlanner82.23/55.72/42.3586.80/66.24/54.76
CodeLlama-7Bw/o 2-cell78.95/51.87/35.2975.90/56.35/43.54
CodeLlama-7Bw/o retrieval79.75/56.96/35.5983.89/62.92/49.27
CodeLlama-7Bw/o L_CI80.80/53.83/38.8287.12/67.61/53.54
CodeLlama-7Bno Eq. 9 consistency79.44/53.19/38.2486.24/64.34/53.41
CodeLlama-7Bno boundary closure78.74/52.00/35.5985.59/66.14/55.37
CodeLlama-7Bhypergraph variant77.35/50.24/31.3284.20/62.79/49.15
CodeLlama-7BTopoPlanner83.43/59.62/47.3589.08/69.59/62.20

Table Z1 tests both parts of this explanation. Removing 2-cells retains the node–edge scaffold but removes explicit regions; a hypergraph alternative retains higher-order grouping without the same graded cell structure. Full has higher N-F1/L-F1 than both reported variants in both domains and backbones. On Multimedia L-F1, the Full/no-2-cell/hypergraph scores are 66.24/50.13/59.73 with Vicuna-7B and 69.59/56.35/62.79 with CodeLlama-7B. Thus the results support a useful regional inductive bias within this architecture, not a theorem that directed graphs are inadequate. Table Z1 mixes training and inference interventions and is not an equal-capacity causal decomposition; matched state-machine, PDDL, and HTN planners were not tested.

AuthorsOpenReview ↗

2nd Response to Reviewer zGcG

2. Do topology-guided requests favor the proposed inductive bias? (W2/Q2)

The extension is a coverage-controlled test of the paper's regional-structure hypothesis, not an estimate of motif frequencies in natural workflows. We first mine candidate tools and directed dependencies from the source tool graph, validate every selected link and topology, and only then generate the request. Thus the region is built from valid source-graph relations rather than arbitrary tool pairings; this does not establish that the resulting task is executable end to end or that its frequency reflects real use. The remaining question is whether the measured benefit depends on the particular structures selected by this construction. Table Z1 probes the regional representation within our architecture; the held-out merge and original-DAG evaluations below address two narrower distributional explanations.

Table Z2. Human assessment of sampled augmented Cycle requests (1–5 scale).

DomainNaturalnessComplexityAlignmentOverall
HuggingFace4.284.344.164.26
Multimedia4.354.294.224.29
Daily Life4.064.213.884.05

Table Z2 checks request plausibility and alignment with the annotated workflow, not whether the motif distribution is natural. The split audit finds 0/300, 0/220, and 0/160 exact normalized request + ordered node/link construction-key repeats across train/test in HF, MM, and Daily Life; graph-only overlap remains in 21, 13, and 0 cases. This rules out that specific exact-repeat leakage, not motif-level alignment.

First, we withhold one merge topology from training in HF/Vicuna-7B. Table Z3 reports the paired test comparison (%, with 95% bootstrap intervals for the gains):

Table Z3. One held-out merge topology, HF/Vicuna-7B.

MetricTopoPlannerGain over GTool95% CI for gain
N-F178.7+26.6[24.0, 29.2]
L-F149.1+30.6[27.3, 33.9]
Precedence accuracy59.1+35.8[32.1, 39.7]

This tests transfer beyond that particular training topology, but only for one withheld merge pattern within the same benchmark construction.

Second, Table Z4 tests the original DAG portion of the source benchmarks (N-F1 / L-F1, %). These evaluation requests were not generated by our motif-conditioned extension:

Table Z4. Original, non-augmented DAG test requests.

DomainBackboneGTool N/LTopoPlanner N/L
HuggingFaceVicuna-7B48.80/10.7884.38/48.00
HuggingFaceCodeLlama-7B66.81/25.8982.49/47.11
MultimediaVicuna-7B57.79/27.2982.31/49.01
MultimediaCodeLlama-7B76.85/43.5284.45/50.00

On these original DAG evaluation requests, TopoPlanner exceeds GTool on N-F1 and L-F1 in all four backbone–domain comparisons. This shows that the gain is not confined to newly generated test requests; the aggregate DAG result is not, by itself, a test of the 2-cell mechanism on naturally occurring regional structures.

Together, Tables Z3–Z4 challenge two narrow explanations: exact dependence on one training merge topology and success only on augmented test requests. Because training still includes other augmented examples, neither isolates the effect of augmented training; nor do these evaluations establish performance on independently collected nonlinear interactive workflows. We therefore claim targeted structural coverage and the observed transfer, not an unbiased real-world workflow distribution.

AuthorsOpenReview ↗

3rd Response to Reviewer zGcG

3. Do structural scores demonstrate interactive task completion? (W3/Q3)

TopoPlanner outputs a tool-name plan, not an observation-conditioned executor. Tables Z5–Z7 test three plan properties: bounded feedback intent in the request, sequence repetition/length, and gold-prerequisite ordering. Merge-only regions need no repeated call; directed feedback denotes bounded revision or finite transformation. In paper_cycle_dailylife_0005, reviews precede possible cancellation and rebooking, illustrating the request-level rather than execution-level setting.

Table Z5. Semantic cues in loop-bearing requests (%).

SubsetRequestsCondition/stopVerify/checkRefine/correctFinite cyclicBounded/finite
All loop-bearing2,26627.8952.0781.161.99 (45)100.00
Directed loop1,13412.4343.0381.663.00 (34)100.00
Loop–merge1,13243.3761.1380.650.97 (11)100.00

“Bounded/finite” describes requests, not observed termination.

For raw tool-name sequences, Table Z6 flags repetition or length above max(8, 2 × gold tool count). This is an offline guard, and its repeat flag may include legitimate reuse.

Table Z6. Offline guarded-stop diagnostic (%).

MethodSafe stopRepeatOver-stepLoop–merge safeDirected-loop safe
GTool90.156.088.4378.0082.00
TopoPlanner99.110.790.1096.0099.00

Table Z7 tests whether gold prerequisites are ordered: Dep. requires every dependency in an example to be satisfied; Pair averages correct gold pairs, counting missing endpoints as errors.

Table Z7. Gold-dependency ordering (%).

BackboneMethodHF Dep./PairMM Dep./PairDaily Life Dep./Pair
Vicuna-7BGTool9.71/24.188.29/23.6939.86/57.32
Vicuna-7BTopoPlanner27.94/51.7440.37/64.2960.14/79.50
CodeLlama-7BGTool25.74/44.9132.56/49.7348.11/62.30
CodeLlama-7BTopoPlanner33.24/58.6448.78/71.8063.57/83.28

These results support plan quality: TopoPlanner improves dependency ordering in all three domains (Z7), and Z6 shows fewer flagged sequence risks. They do not establish interactive task completion: no parameterized tool calls, responses to intermediate observations, or terminal verifier were evaluated. Acc measures node matching, not task success. Our empirical claim is structural plan prediction rather than execution gain.

4. How does a frontier model with simple prompting compare? (Q4)

We also test a frontier model under simple prompting. We paired GPT-6.1 Sol Direct and Structured prompts on every HF/MM test request with the same tool descriptions; Structured additionally sees public dependency links. Neither sees the gold plan, and both produce tool/edge JSON. Table Z8 reports this control:

Table Z8. GPT-6.1 Sol Direct versus public-graph Structured prompting.

DomainPromptTasksN-F1 (%)L-F1 (%)NodeAcc (%)Input tokensOutput tokens
HuggingFaceDirect68083.6360.5138.9712,531,81953,820
HuggingFaceStructured68084.0362.9440.1515,853,78653,561
MultimediaDirect82088.2567.9549.8815,022,86275,578
MultimediaStructured82089.6872.3057.2023,243,69974,290

Table Z8 shows that simple structured prompting already yields high structural scores and raises Sol L-F1 by about 2.4 HF and 4.3 MM points, at 26.5%/54.7% more input tokens. This directly supports the reviewer's suggestion that stronger LLMs are serious planning controls. At the same time, in our Qwen3-235B API-backbone evaluation, Multimedia Direct scores 77.10/57.49 N/L-F1 and TopoPlanner 87.65/71.00; for DeepSeek-V3, the corresponding scores are 77.61/56.09 and 88.08/69.19. Those same-backbone results show that model scale did not erase the structural benefit in the tested API interface. They are not a Sol comparison: a same-model Sol–TopoPlanner arm is unavailable, so we cannot rank TopoPlanner against Sol, Claude, or Gemini from these data. Neither control tests interactive completion.

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

This paper proposes TopoPlanner, which represents tool dependencies using cellular complexes and combines topology-aware retrieval, multi-dimensional message passing, and LLM conditioning for tool planning. Experiments on four benchmarks augmented with loop and merge structures show substantial improvements over graph-based baselines, supported by ablations and additional structural evaluations.

Critical strengths

TopoPlanner provides a useful integration of higher-order cellular representations, request-conditioned retrieval, and LLM tool planning. The model of loop and merge regions is a reasonable inductive bias beyond conventional node-edge representations. Also, the gains over GTool are substantial across multiple datasets and backbones.

Critical weaknesses

The Cycle benchmark combines synthetic feedback loops with DAG-like merges, while evaluation primarily measures offline tool-sequence and dependency prediction. The output interface and repetition diagnostics do not establish whether agents can correctly execute conditional feedback loops with termination. Also, the results show that the gains remain after removing 2-cells, so the overall improvement cannot be attributed specifically to topology.

Questions for authors

How do the reported 2-cell counts relate to the fundamental-cycle basis (e.g., cycle filtering, spanning-tree selection, and whether the learned maps satisfy cosheaf requirements)?

4 comments
AuthorsOpenReview ↗

Response to Reviewer PmzJ

We thank the reviewer for recognizing the motivation for higher-order workflow structure and for raising important questions about cycle semantics, the contribution of 2-cells, and the lifting construction. We address each question through the corresponding definitions, ablations, and structural diagnostics. Throughout, our evaluated task is offline prediction of tool names and dependencies from a request and an available tool graph, not execution of a feedback policy in a live environment.

1. What does “Cycle” represent, and what do the results establish about feedback and stopping?

The benchmark's “Cycle” label names a closed region in the lifted undirected structure, not a single execution semantics. It includes (i) directed feedback, (ii) convergent paths in a DAG that form a merge, and (iii) loop–merge compositions. A DAG merge can bound a 2-cell without requiring any repeated tool call. Directed feedback denotes a bounded revision or reuse pattern in the annotated workflow; generating a tool-name sequence is not the same as selecting an action after observing a runtime check or proving termination. The extension fixes motifs from valid tool/dependency links before request generation, rather than inventing arbitrary tools.

We audited all 2,266 loop-bearing requests. Table P1 reports non-exclusive semantic categories:

Table P1. Semantic cues in loop-bearing requests.

Request subsetRequestsCondition/stop (%)Verify/check (%)Refine/correct (%)Finite cyclic (%)Bounded/finite (%)
All loop-bearing2,26627.8952.0781.161.99 (45)100.00
Directed loop1,13412.4343.0381.663.00 (34)100.00
Loop–merge1,13243.3761.1380.650.97 (11)100.00

Table P1 characterizes request semantics; the bounded/finite label is not an observed executor termination rate. For the plan output itself, Table P2 evaluates whether the predicted sequence preserves merge/reuse dependencies. The four-domain subgroup has 804 examples and 2,325 gold merge/reuse edges per backbone:

Table P2. Merge/reuse dependency prediction.

BackboneMethodExamplesN-F1 (%)L-F1 (%)Merge/reuse edge recall (%)
Vicuna-7BGTool80451.1822.0619.04
Vicuna-7BTopoPlanner80476.1646.9239.72
CodeLlama-7BGTool80464.5635.2729.06
CodeLlama-7BTopoPlanner80478.9449.9141.01

Table P3 checks the ordering of every gold dependency: Pair is the mean fraction of correctly ordered gold pairs (missing endpoints count as errors), and Dep. requires all gold dependencies of an example to be present and correctly ordered.

Table P3. Gold-dependency ordering.

BackboneMethodHF Dep./Pair (%)MM Dep./Pair (%)Daily Life Dep./Pair (%)
Vicuna-7BGTool9.71 / 24.188.29 / 23.6939.86 / 57.32
Vicuna-7BTopoPlanner27.94 / 51.7440.37 / 64.2960.14 / 79.50
CodeLlama-7BGTool25.74 / 44.9132.56 / 49.7348.11 / 62.30
CodeLlama-7BTopoPlanner33.24 / 58.6448.78 / 71.8063.57 / 83.28

Table P4 checks generated sequences for any repeated tool call or length above max(8, 2 × gold tool count):

Table P4. Offline guarded-stop diagnostic.

MethodSafe stop (%)Repeat (%)Over-step (%)Loop–merge safe (%)Directed-loop safe (%)
GTool90.156.088.4378.0082.00
TopoPlanner99.110.790.1096.0099.00

The guard indicates fewer obvious repetition/over-length patterns in static output; it can also flag legitimate repeated calls and does not test conditional execution or runtime termination. Thus Tables P2–P4 support better structural plan prediction for merge/reuse and dependency order, while the reviewer's execution-level question remains outside this evaluation. We will separate these claims explicitly.

AuthorsOpenReview ↗

2nd Response to Reviewer PmzJ

2. How much of the improvement is supported by the 2-cell ablation?

The full-versus-GTool difference is a system comparison: TopoPlanner also changes request-specific retrieval, consistency supervision, and the graph-to-LLM interface. We therefore do not assign the entire difference to 2-cells. Table P5 compares the full architecture with its component ablations (N-F1 / L-F1 / Acc, %):

Table P5. Component ablation across backbones and domains.

BackboneMethodHuggingFaceMultimediaDaily Life
Vicuna-7Bw/o 2-cell75.58 / 46.61 / 30.8873.29 / 50.13 / 34.1581.25 / 62.42 / 59.28
Vicuna-7Bw/o retrieval76.07 / 47.88 / 32.2177.97 / 55.11 / 35.3782.99 / 65.65 / 64.26
Vicuna-7Bw/o L_CI80.69 / 53.31 / 38.6881.47 / 58.29 / 40.9893.13 / 71.44 / 69.76
Vicuna-7BTopoPlanner82.23 / 55.72 / 42.3586.80 / 66.24 / 54.7697.27 / 78.53 / 86.25
CodeLlama-7Bw/o 2-cell78.95 / 51.87 / 35.2975.90 / 56.35 / 43.5482.86 / 64.35 / 63.75
CodeLlama-7Bw/o retrieval79.75 / 56.96 / 35.5983.89 / 62.92 / 49.2786.85 / 67.05 / 53.66
CodeLlama-7Bw/o L_CI80.80 / 53.83 / 38.8287.12 / 67.61 / 53.5495.41 / 76.20 / 75.95
CodeLlama-7BTopoPlanner83.43 / 59.62 / 47.3589.08 / 69.59 / 62.2097.43 / 77.95 / 87.29

In Table P5, relative to w/o 2-cell, Full's N-F1 increases by 6.65/13.51/16.02 points with Vicuna-7B and 4.48/13.18/14.57 with CodeLlama-7B across HF/MM/Daily Life; L-F1 increases by 9.11/16.11/16.11 and 7.75/13.24/13.60, respectively. Every reported N-F1 and L-F1 comparison favors Full. Removing retrieval or the consistency loss also reduces reported performance. This supports the usefulness of explicit 2-cell regions within the evaluated architecture, while recognizing that the overall GTool gap has several contributors. We do not claim that a graph or another planning formalism cannot represent the same dependencies. Here Acc is the benchmark's node-match criterion, not end-to-end execution success.

3. How do the 2-cell counts relate to fundamental cycles, tree choice, and cosheaf maps?

Consider the underlying undirected graph before filtering and fix a spanning forest T with c components. Every non-tree edge e=(u,v) closes a unique cycle C_e: the edge itself plus the unique u–v path in T. These cycles are independent because C_e contains its own non-tree edge e and no other fundamental cycle contains that edge. They span because subtracting from any cycle the C_e associated with each of its non-tree edges leaves a cycle supported only on T, and a forest has no nonzero cycles. Thus they form a basis with |E_u|−|V|+c members (for a connected graph, c=1). Direction is retained as dependency information for planning but does not enter this undirected basis count. We then retain only cycles with at most six edges. Consequently, the reported 2-cell counts are retained short faces, not the full cycle rank, and the filtered set need not span the entire cycle space. Different spanning forests can yield different retained short cycles; we do not claim invariance to tree selection.

Table P6 gives the actual graph pools. Its edge counts are directed and should not be substituted for |E_u| in the rank formula:

Table P6. Tool-graph size and retained 2-cell counts.

DomainNodesDirected edgesRetained 2-cellsEdge candidatesMean CPU retrieval (ms)
HuggingFace232255422523.65
Multimedia404498544933.11
Daily Life401,56014651239.45
ToolBench2453661936622.89

Regarding cosheaf requirements, our construction uses request-conditioned coface-to-boundary incidence operators, boundary-closed retrieval, and a compatibility penalty. Boundary closure and a low compatibility residual do not by themselves prove that the learned operators satisfy all strict cellular-cosheaf axioms. We use the cosheaf structure as an inductive framework for cross-dimensional compatibility, rather than claim a proof that the learned operators form a strict cellular cosheaf. The fundamental-cycle basis result concerns the graph topology and does not depend on this learned-map claim.

Manuscript revision. We will revise the dataset description to define “Cycle” as a topological category and distinguish DAG merges from directed feedback that requires runtime stopping conditions. In the method section, we will separate the complete fundamental-cycle basis from the length-filtered 2-cells, state that the retained faces depend on the spanning forest, and describe the learned incidence operators without implying that strict cosheaf axioms have been proved. In the evaluation section, we will label guarded-stop and dependency-order scores as offline sequence diagnostics, define Acc as the node-match criterion, and state explicitly that conditional execution and runtime termination are not evaluated.

Thank you for the detailed rebuttal. The clarification of the length-filtered 2-cell construction and the Cosheaf formulation addresses several of my technical concerns. I also appreciate the additional discussion of the ablation results. However, my main concern remains that the evaluation focuses on offline structural prediction rather than demonstrating planning capabilities for conditional feedback loops in executable workflows. Therefore, I will maintain my original rating.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to the Reviewers

We thank the reviewers for their careful questions about the role of cellular structure, the topology-guided evaluation, and the interpretation of planning scores. The central motivation of TopoPlanner is not that a directed graph cannot encode branching, merging, or feedback. It is that dependencies forming a shared region remain implicit when a planner only propagates along pairwise edges. By attaching a 2-cell to that region, TopoPlanner makes its boundary available to request-conditioned retrieval and finite-depth structural reasoning while retaining the directions of the underlying dependencies.

What does the regional representation contribute? The component study compares the full planner with its edge-only 1-skeleton and a hypergraph alternative across two backbones. On Multimedia, Vicuna-7B obtains 66.24 L-F1 with the full model, versus 50.13 without 2-cells and 59.73 with hypergraph grouping; CodeLlama-7B obtains 69.59, versus 56.35 and 62.79. The same ordering holds on HuggingFace. Removing retrieval, boundary closure, or consistency supervision also lowers the reported scores. These results support the proposed combination of explicit regions and boundary-aware use of those regions within the evaluated architecture (PmzJ, Table P5; zGcG, Table Z1; 3Xfc, Table X3).

Does the benefit depend on the topology-guided extension? The extension selects valid tools and directed dependencies from the source tool graph before generating requests; it tests the regional structures targeted by the method. We also evaluate beyond those newly generated test requests. With one merge topology withheld from training, HuggingFace/Vicuna-7B gains 30.6 points in L-F1 over GTool (95% paired-bootstrap interval: 27.3–33.9; zGcG, Table Z3). On original, non-augmented DAG test requests, TopoPlanner improves N-F1 and L-F1 in all four HuggingFace/Multimedia backbone comparisons: for example, Vicuna-7B L-F1 rises from 10.78 to 48.00 on HuggingFace and from 27.29 to 49.01 on Multimedia (zGcG, Table Z4). The holdout and original-request results test distinct aspects of transfer; neither treats augmented motif frequencies as estimates of real-world workflow frequencies.

Where does the structural improvement appear? In the four-domain merge/reuse subgroup, Vicuna-7B recall of the relevant gold dependency edges rises from 19.04% with GTool to 39.72% with TopoPlanner; CodeLlama-7B rises from 29.06% to 41.01% (PmzJ, Table P2; 3Xfc, Table X4). For the stricter example-level requirement that every gold prerequisite be present and correctly ordered, Multimedia improves from 8.29% to 40.37% with Vicuna-7B and from 32.56% to 48.78% with CodeLlama-7B (PmzJ, Table P3; 3Xfc, Table X5). These diagnostics connect the representation to tool selection and dependency ordering, the outcomes measured by our offline planning interface.

Manuscript revision. We will distinguish undirected closed regions from directed feedback, clarify the incidence-compatibility signal in Eq. 8 and the budgeted retrieval procedure associated with Eq. 9, and define the sequence-to-link evaluation and node-match Acc precisely. We will also make the forest-choice and computational-cost settings easy to locate, and state the available-tool-graph assumption and the offline planning scope alongside the empirical claims. The reviewer-specific responses give the complete tables and protocols for each point.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.