Source-linked AI summary
GRASP: Generating, Revising, and Assessing for Strategic Planning with Agentic AI
Arunabh Srivastava, Mohammad A., Khojastepour, Srimat Chakradhar, Sennur Ulukus
TL;DR
Complexity can degrade LLM reliability on long-horizon planning tasks. GRASP separates planning into context-isolated generation, strategy refinement, and verification; the authors report that it neutralizes multi-task degradation and improves accuracy over direct planners.
Problem
As task complexity increases, LLMs can suffer constraint collisions, attention fatigue, hallucinations, and multi-step execution failures when given long-horizon instructions.
Source: §1 · p. 1Method
Results
GRASP neutralizes multi-task contextual degradation and achieves up to 16.7% absolute accuracy gain over direct baselines in interleaved dual-task environments.
Source: §1 · p. 2Takeaways & Limitations
The evaluations support the conclusion that explicit structural guardrails can help standard LLMs outperform unguided internal reasoning chains of frontier reasoning models.
Source: §5 · p. 8The optimal number of strategy plans is task-dependent, so deployment across heterogeneous domains currently requires manual tuning.
Source: Limitations · p. 9
Abstract
from arXiv · showhide
Large Language Models (LLMs) typically exhibit a performance profile where reliability degrades as task complexity increases. We address the challenge of generating high-quality natural language executable plans for complex tasks by introducing $\textbf{GRASP}$, a strategy-aware, multi-stage planning framework. GRASP decouples the planning pipeline across specialized, context-isolated modules: it pre-compiles global macro-guidelines (GenPlan), explores alternative localized strategies within isolated context windows (RevPlan), and independently evaluates trajectories using a multi-criteria discriminator (VerPlan). Empirical evaluations show that GRASP consistently establishes a new state-of-the-art frontier across diverse datasets, yielding substantial accuracy gains over direct LLM planners on Natural Plan Calendar Scheduling ($\sim$12.4$\%$$\uparrow$), ZebraLogic ($\sim$30.8$\%$$\uparrow$), and SciBench Math. Crucially, under multi-task scaling-where standard planners suffer immediate performance collapse-GRASP completely flattens the multi-task degradation penalty. In interleaved dual-task environments, GRASP achieves an absolute accuracy gain of up to 16.7$\%$ over direct LLM planners. Furthermore, by isolating context and enforcing strict macro-regularization, GRASP outperforms frontier reasoning models (such as GPT-5-mini) by a margin of 14.5$\%$.
1 Introduction
Complexity scaling can make long-horizon instructions unreliable, motivating explicit plan generation. GRASP separates plan generation, strategic refinement, and verification, and reports gains under multi-task scaling.
1 Introduction
- As long-horizon instructions grow more complex, constraint collisions and attention fatigue can trigger hallucinations and multi-step execution failures.The paper calls this scaling-related vulnerability the “Curse of Instructions”.Source: §1 · p. 1
- GRASP compiles global constraints and guidelines, explores localized strategies in separate context windows, and independently evaluates trajectories.Its modules are GenPlan, RevPlan, and VerPlan.Source: §1 · p. 1
- The authors interface GRASP with RunAgent, a multi-agent execution platform, because monolithic execution can undermine plan decomposition through context pollution.They identify robust execution as important for translating plans into correct downstream results.Source: §1 · p. 1
- GRASP neutralizes multi-task contextual degradation; interleaved dual-task accuracy gains reach 16.7% over direct baselines and 14.5% over GPT-5-mini.The paper also reports that GRASP establishes a new state-of-the-art frontier.Source: §1 · p. 2
2 Related Work
Related work spans autonomous agents, inference-time reasoning methods, and automated planning. The paper positions GRASP as combining context-isolated generation, revision, and verification with autonomous constraint-guided planning.
2 Related Work
- Inference-time methods such as Chain-of-Thought (Wei et al., 2022), Self-Consistency (Wang et al., 2022), and In-Context Learning improve step-by-step reasoning but lack long-horizon planning capabilities.Search methods including Tree of Thoughts (Yao et al., 2023), Graph of Thoughts (Besta et al., 2024), and REBASE expand trajectory exploration but lack structured constraint enforcement.Source: §2 · p. 2
- Prior agent frameworks use incremental adjustments or critique loops, but remain vulnerable to conversational noise; GRASP isolates generation, revision, and verification across context windows.Examples include ReAct (Yao et al., 2022), Pre-Act (Rawat et al., 2025), Reflexion (Shinn et al., 2023), and TextGrad (Yuksekgonul et al., 2024).Source: §2 · p. 2
- Automated planning work includes PlanGEN’s constraint, selection, and verification agents, symbolic-LLM methods, and orchestrators that rely on human-in-the-loop validation.The paper contrasts these approaches with GRASP’s autonomous zero-shot plan generation and structural regularization.Source: §2 · p. 2
3 GRASP
GRASP divides planning into sequential, specialized modules that separate global plan structure, instance-specific revision, and independent evaluation. Context isolation keeps each stage focused and reduces overload from handling constraints, variables, and local steps together.
3 GRASP
- GRASP sequentially uses GenPlan to create a blueprint, RevPlan to adapt it to a task instance, and VerPlan to verify constraints and select a plan.Source: §3 · p. 2
3.1 GenPlan
- GenPlan separates task-level structure from instance-specific details, using specialized agents to build a baseline plan from constraints and guidelines.Source: §3.1 · p. 3
Knowledge Base
- The Knowledge Base stores constraints, guidelines, and a tentative plan, which the agents read and update during GenPlan.Source: Knowledge Base · p. 3
Guidelines Agent (GA)
- The Guidelines Agent identifies and merges soft, non-binding heuristics that encourage plan consistency and reduce generative drift.Source: Guidelines Agent (GA) · pp. 3–4
Constraint Agent (CA)
- The Constraint Agent identifies and merges hard conditions; even without user-provided priors, it can infer constraints from the task description.Source: Constraint Agent (CA) · p. 4
Plan Generation Agent (PGA)
- The Plan Generation Agent synthesizes and iteratively revises a tentative plan within the constraints and guidelines until an external judge determines that updates have converged.Source: Plan Generation Agent (PGA) · p. 4
3.2 RevPlan
- RevPlan adapts the GenPlan blueprint to instance-specific runtime directives by exploring parallel strategy tracks and merging their plans with the structural blueprint.Source: §3.2 · pp. 4–5
3.3 VerPlan
- VerPlan scores candidate plans independently for feasibility and constraint adherence, then selects the highest-scoring plan as GRASP’s final output.Source: §3.3 · pp. 5–6
4 Evaluation
GRASP is evaluated on four benchmarks using programmatically executed plans and on calendar tasks with interleaved objectives. It improves over direct planners on several single-task benchmarks and maintains performance under multi-task scaling, while ablations show that its modules work best together.
4.1 Experimental Setup
- The evaluation covers Natural Plan Calendar Scheduling, GPQA, SciBench Math, and ZebraLogic, using exact match for calendar scheduling and accuracy for the other datasets.Equivalence is judged by GPT-5; manual review of 50 random instances per dataset fully agreed with the judge.Source: §4.1 · p. 6, Evaluation Metrics and Baselines · p. 6
The Execution Platform: RunAgent
- RunAgent executes plans with frontier LLMs or Python code, using bare-bones execution so verification protocols do not compensate for poor plan quality.Python execution is disabled for the dual- and triple-task experiment, which uses LLMs alone.Source: The Execution Platform: RunAgent · p. 6
Evaluation Metrics and Baselines
- Experiment 1 compares GRASP with direct LLM execution, baseline LLM planners, and methods including PlanGEN, Tree-of-thoughts, and Best-of-N.GPT-5-mini is added as a direct-planner baseline for the dual-task experiment.Source: Evaluation Metrics and Baselines · p. 6
4.2 Main Results
- GRASP beats direct planners on Natural Plan Calendar Scheduling by 12.4% for GPT-4o and ZebraLogic by 30.8% for GPT-4o; SciBench Math results vary by subset, while GPQA gains are not statistically significant.GRASP also outperforms all compared methods on Natural Plan Calendar Scheduling, SciBench Stat, and ZebraLogic; PlanGEN ties it on SciBench Calc and Diff.Source: Single-Task Planning with Programmatic… · p. 7, Table 1 · p. 7, Table 2 · p. 7
Single-Task Planning with Programmatic Execution
- Removing GenPlan constraints and guidelines lowers calendar accuracy from 75.4% to 58.5%, while three RevPlan strategies yield an 11.3% gain at approximately 0.95x baseline cost for GPT-4o-mini.GRASP with GPT-4o costs approximately 13.5x the GPT-4o planner while improving accuracy by 12.4%.Source: Single-Task Planning with Programmatic… · p. 7
Robustness for Compound Task Scaling
- Under multi-task scaling, GRASP(GPT-4o) scores 45.7% on dual tasks and 46.6% on triple tasks; GRASP(GPT-4o-mini) reaches 41.2% on dual tasks, a 16.7% absolute gain over its direct baseline.Source: Robustness for Compound Task Scaling · p. 7
- In the multi-task ablation, full GRASP reaches 45.7% accuracy; RevPlan-only and RevPlan-plus-VerPlan configurations score 10.9% and 12.5%, respectively.GenPlan-only scores 40.8%, while VerPlan-only scores 30.7%.Source: Robustness for Compound Task Scaling · p. 8, Table 8 · p. 8
5 Conclusion
The conclusion presents GRASP as an autonomous framework for complex tasks that separates planning, strategy exploration, and validation across context-isolated agents.
5 Conclusion
- GRASP decouples plan generation, strategy exploration, and validation across context-isolated agents to address context degradation and plan hallucinations in complex tasks.The paper reports that its evaluations show explicit structural guardrails enable standard LLMs to outperform unguided reasoning chains of frontier reasoning models.Source: §5 · p. 8
Limitations
The framework trades additional runtime latency for execution reliability and does not improve the underlying model’s intrinsic reasoning capabilities. Its strategy count is task-dependent, and deployment across heterogeneous domains currently requires manual tuning.
Limitations
- Iterative generation, local refinement, and verification of alternative strategies increase runtime latency, potentially limiting real-time interactive use.The paper identifies parallelized generation or early-exit heuristics as future directions for reducing inference time.Source: Limitations · p. 9
- GRASP is an inference-time plan regularizer, not a way to remedy severe domain ignorance or missing foundational logic in the underlying model.The paper notes that GRASP does not outperform direct planners on GPQA or SciBench Diff and Calc, where underlying reasoning capability matters.Source: Limitations · p. 9
- The optimal number of strategy-based plans is task-dependent; heterogeneous-domain deployment currently requires manual tuning.Calendar Scheduling peaks at three strategies, while generating more can mislead VerPlan toward sub-optimal choices.Source: Limitations · p. 9
Ethical Considerations
The section discusses GRASP’s dual-use and bias risks, privacy and computational concerns, and planned interactive and runtime extensions. It also lists model token prices.
Ethical Considerations
- GRASP could be adapted to optimize harmful plans, so the authors advocate verification guardrails and human-in-the-loop validation in production.The framework lacks physical agency, real-world permissions, and independent tool-use capabilities.Source: Ethical Considerations · p. 9
- Because GRASP relies on its base LLM and does not filter embedded biases, biased premises can yield structurally sound but ethically compromised plans.Source: Ethical Considerations · p. 9
- Real-world calendar scheduling may involve sensitive user data, while multi-agent pipelines risk context-bleed or unauthorized data persistence.The authors caution against leaking sensitive parameters into public training loops.Source: Ethical Considerations · p. 9
- The authors characterize GRASP’s environmental and computational overhead as reasonable relative to its performance gains, citing competitive token use.They report that some baselines use fewer tokens, while others use more and perform worse.Source: Ethical Considerations · pp. 9–10
A Future Work
- Planned extensions include interactive clarification and guidance requests, plus runtime verification and modification of plans based on intermediate outputs.The authors note that eliciting high-level plan-construction guidance is more challenging than asking about missing information.Source: A Future Work · pp. 11–12
B Token Prices
- Appendix B lists input and output token prices per million tokens for GPT-4o, GPT-4o-mini, and GPT-5-mini.Source: B Token Prices · p. 12, Table 10 · p. 12
C Token Usage
GRASP’s token use is competitive with selected PlanGEN variants, while token consumption and accuracy vary across baselines and SciBench subsets. Experimental details specify model settings and run counts.
C Token Usage
- GRASP’s token use is competitive with PlanGEN(Mix) and PlanGEN(BoN); PlanGEN(BoN) uses 22% more tokens on Natural Plan Calendar Scheduling yet performs worse.PlanGEN and GRASP perform at the same level on SciBench Calc and Diff; the authors state GRASP does not improve its underlying LLM’s reasoning power.Source: C Token Usage · p. 12
- PlanGEN(Rebase) and Best of N use fewer tokens than GRASP but perform worse, while PlanGEN(ToT) and ToT use more tokens and also perform worse.Source: C Token Usage · p. 12
D Experimental Details
- Each dataset and model combination was run once, except GPQA and the Experiment 2 ablation, which were each run three times.The stated reason for the single runs is high cost per run.Source: D Experimental Details · p. 12
D.1 Single-Task Planning with Programmatic Execution
- Direct LLM baselines received each dataset’s problem without the task or other text; ToT and PlanGEN baselines followed their specified search and generation configurations.BoN generated four plans and selected the one scoring highest with VerPlan; the cited setup describes four PlanGEN variants.Source: §D.1 · pp. 12–13
D.2 Robustness for Compound Task Scaling
The compound-task experiment tests calendar scheduling as objectives expand from one to two or three, adding simple consultant-calendar subproblems. It compares GRASP with component ablations and direct planning.
D.2 Robustness for Compound Task Scaling
- Dual and triple tasks combine the original scheduling problem with one or two auxiliary consultant-calendar objectives and require exactly two or three newline-separated answers.Source: §D.2 · p. 13
- Each composite instance pairs the original scheduling instance with fixed Consultant D schedules for dual tasks or Consultant D and E schedules for triple tasks.Source: §D.2 · p. 13
- The auxiliary calendar tasks are designed to be simple, but combining them with scheduling tasks reduces plan accuracy.They test parsing dense calendars, working-hour bounds, 30-minute gaps, and selecting the latest free slot.Source: §D.2 · p. 13
- The experiment uses four GRASP strategies for GPT-4o and GPT-4o-mini, with direct LLM planners run without Python execution.Source: §D.2 · p. 13
- The ablation compares direct planning and individual modules with combinations that generate, revise, or score plans, alongside the full GRASP framework.The described variants include GenPlan-only, RevPlan-only, VerPlan-only, GenPlan+RevPlan, and RevPlan+VerPlan.Source: §D.2 · p. 13
Experiment 1
The reported statistical tests favor GRASP over baselines on ZebraLogic and most Natural Plan comparisons, while the supplied SciBench and GPQA tests do not establish significant advantages. A separate compound-task result reports significant differences for GRASP and its ablation.
Experiment 1
- On ZebraLogic, GRASP outperforms every baseline with high statistical confidence; the McNemar comparison with PlanGEN(BoN) has p < 0.0001.Source: Experiment 1 · p. 15
- On Natural Plan Calendar Scheduling, GRASP’s 2% advantage over PlanGEN(BoN) is not statistically significant (p = 0.33), while its advantage over other baselines is significant (p < 0.01).Source: Experiment 1 · p. 15
- For SciBench Stat, the bootstrapped 95% confidence intervals are 75%–91.67% for GRASP and 68.06%–87.5% for the other two methods (p = 0.3438); Calc and Diff match GRASP.The authors do not conduct statistical analysis for Calc and Diff because PlanGEN performs at the same level as GRASP.Source: Experiment 1 · p. 15
- On GPQA, GRASP does not significantly outperform the direct GPT-4o planner (p = 0.6793).Source: Experiment 1 · p. 15
Experiment 2
- In the compound-task experiment, GRASP GPT-4o and the ablation study models each outperform their competitors with p < 0.01; ablation results are averaged over three runs.Source: Experiment 2 · p. 15
E.2 Plan Efficiency for Experiment 1
Reported plan lengths vary by dataset: GRASP plans are longer than PlanGEN(BoN) on ZebraLogic but shorter than named baselines on Calendar Scheduling and SciBench subsets.
E.2 Plan Efficiency for Experiment 1
- 12.67 steps for GRASP versus 9.74 for PlanGEN(BoN) on ZebraLogic; the authors report accuracy also increases by a large margin.The passage links the longer plans with the accuracy increase.Source: §E.2 · pp. 15–16
- 11.3 steps for GRASP versus 13.09 for PlanGEN(BoN) on Natural Plan Calendar Scheduling.Source: §E.2 · p. 16
- SciBench plan lengths vary by subset, with GRASP shorter than PlanGEN(Mix) on Diff and Calc.For Stat, GRASP averages 9.18 steps, BoN 6.96, and PlanGEN(Rebase) 9.53; for Diff, GRASP averages 11.66 and PlanGEN(Mix) 14.2; for Calc, GRASP averages 11.39 and PlanGEN(Mix) 14.85.Source: §E.2 · p. 16
E.3 Information Leakage in Experiment 1
This appendix section examines answer leakage and documents the prompts, algorithms, and examples used in Experiment 1. It covers the leakage definition and accuracy reporting, GRASP’s guideline and constraint generation and plan revision, and the direct planner’s generation instructions.
E.3 Information Leakage in Experiment 1
- With leakage-violating plans excluded, ZebraLogic accuracy is 55.7% for GRASP(gpt-4o-mini) versus 27.1% for PlanGEN(BoN), which generated 145 leakage plans versus GRASP’s 31.Information leakage means the task-instance pair’s final answer appears anywhere in the plan.Source: §E.3 · p. 16
- With leakage excluded, Natural Plan Calendar Scheduling accuracy is 74.3% for GRASP versus 62% for PlanGEN(BoN); GRASP had no leakage plans, compared with PlanGEN(BoN)’s 103.The passage reports no leakage for GRASP and 103 leakage plans for PlanGEN(BoN).Source: §E.3 · p. 16
- For SciBench Stat, revised accuracy is 80.56% for GRASP, 77.78% for BoN, and 76.39% for PlanGEN(Rebase), with one leakage plan each for GRASP and PlanGEN(Rebase).BoN had no leakage plans.Source: §E.3 · p. 16
- For SciBench Diff, revised accuracy is 42% for both GRASP and PlanGEN(Mix); for SciBench Calc, it is 68.27% for GRASP versus 56.07% for PlanGEN(Mix).Both methods had four leakage plans on Diff; Calc had two for GRASP and seven for PlanGEN(Mix).Source: §E.3 · p. 16
E.4 Discussion
- Despite some statistically nonsignificant results, GRASP generates shorter plans, achieves higher accuracy while using fewer tokens, and maintains similar accuracy under strict constraints as baselines drop by more than 10%.The discussion attributes shorter plans to downstream token savings.Source: §E.4 · p. 16
F GenPlan Algorithm and Prompts
- GenPlan iteratively generates and merges guidelines and constraints, generates a plan from the knowledge base, and stops when the guidelines and constraints stabilize.Its prompts require actionable steps without outputs or the final solution in the plan.Source: F GenPlan Algorithm and Prompts · p. 17
G RevPlan Algorithm and Prompts
- RevPlan generates distinct LLM-solvable strategies, derives strategy-specific constraints, runs ReAct for each, and extracts specialized candidate plans without including step results.The plan-fitting prompt allows additions, removals, or modifications to make a plan more robust.Source: G RevPlan Algorithm and Prompts · pp. 18–19
H VerPlan Algorithm and Prompts
- VerPlan selects a final plan from candidate plans using constraints, strategy-specific constraints, and rubrics, while rating plans from 1–100 on completeness, correctness, clarity, actionability, and requirements.Its prompt requests only a single integer rating.Source: H VerPlan Algorithm and Prompts · p. 19
I Direct Planner Generation Prompt
- The direct planner prompt asks for a numbered plan whose steps reach the answer organically, without doing the calculation itself.It instructs the model to output only the plan.Source: I Direct Planner Generation Prompt · p. 20
J Examples of Guidelines and Constraints
- The appendix provides examples of GRASP guidelines and constraints for Natural Plan Calendar Scheduling and ZebraLogic in Tables 15–18.Tables 15–16 are identified as the calendar-scheduling example, while Tables 17–18 are identified as the ZebraLogic example.Source: J Examples of Guidelines and Constraints · p. 20, Table 15 · p. 20, Table 16 · p. 21
K Plan Revision
- RevPlan generates alternative strategies and revises initial plans, demonstrated on Calendar Scheduling and GPQA; the GPQA example uses four strategies.For the calendar example, two strategies and their revised plans are presented; for GPQA, four strategies are generated and the initial plan is revised for each.Source: K Plan Revision · pp. 20–21, Table 21 · p. 22, Table 22 · p. 23
- The GPQA refinements include an energy-time uncertainty approach and lifetime-ratio analysis using a 10:1 ratio.The described plans compare energy separation with uncertainty and use relative lifetimes to identify the dominant energy spread.Source: K Plan Revision · p. 24, Table 25 · p. 24