C.24:10 - Rationale and current practice
Qualification window. This comparison uses the source set as of 2026-08-21. Reopen it when a later result changes the relative value of explicit planning, route grounding, active information gathering, checkpoint use and replanning, multidimensional evaluation, or long-horizon budget and dependency handling for the declared use.
| Contribution | Adopted, adapted, or rejected move | Boundary and trade-off |
|---|---|---|
| ToolPlanner, EMNLP 2024, ToolPlanner: A Tool Augmented LLM for Multi Granularity Instructions with Path Planning and Feedback | Adopt: keep path planning, feedback, and replanning explicit instead of hiding them inside one call loop. | The gain is replayable route revision; the cost is a plan object. The LLM benchmark does not define universal FPF objects. |
| PlanningArena, ACL 2025, PlanningArena: A Modular Benchmark for Multidimensional Evaluation of Planning and Tool Learning | Adapt: check tool selection, reasoning, user-input interpretation, and execution-relevant constraints separately instead of treating one aggregate score as plan quality. | Its scenarios do not set universal weights or safety limits. C.24 keeps only the dimensions that change this plan or checkpoint. |
| IBM Research, ECAI 2025, From Grounding to Planning: Benchmarking Bottlenecks in Web Agents | Retain with a rejected overread: keep route grounding distinct from plan quality, but reject the claim that planning is always the dominant bottleneck. | This preserves a cheap diagnostic split without hard-coding a web-agent bottleneck order. |
| Aghzal et al., 2026 preprint, Why Do LLM-based Web Agents Fail? A Hierarchical Planning Perspective | Adopt: separate high-level planning, low-level execution, and replanning; a sound plan does not excuse failed grounding or adaptive control. | The result makes the IBM split conditional. It does not make every web-agent layer mandatory in a known fixed route. |
| DeepPlanning, ACL 2026, DeepPlanning: Benchmarking Long-Horizon Agentic Planning with Verifiable Constraints | Adapt: retain global budgets, dependencies, or safe parallelism when material, and use the bounded scout and checkpoint cycle for information needed before commitment. | Long-horizon benchmarks expose degradation and efficiency trade-offs, but their task schemas do not belong in every ordinary call plan. |
C.19.1 current scale-comparison sources and method | Adopt conditionally: use Bitter-Lesson pressure only with the actual probe and a named bearer, scale window, evidence, cost, safety, and uncertainty. | Generality is not a winner by label; local policy and waiver stay separate from empirical comparison. |
This set is non-dominated for C.24’s declared use because it keeps the smallest common planning contract while exposing the failure dimensions that later work shows can move independently. Remove a field when it changes no route, stop, reliance, or replay; reopen when a new contribution changes that trade-off rather than merely adding another benchmark.