Nov 2026· Applied Sciences· Vol 16, pp. 10828· 0 citations· 18 references
TL;DR
The proposed Plan-Then-Verify (PTV), a verifier-first control pattern in which a candidate workflow is checked against a documented knowledge base before execution begins, significantly improves static SOP conformance on several model families, while remaining close to Plan-and-Execute in token cost.
Abstract
Pre-execution feasibility validation can substantially improve LLM-based task planning without proportional inference cost, but only when it is grounded, deterministic, and tightly bounded. We propose Plan-Then-Verify (PTV), a verifier-first control pattern in which a candidate workflow is checked against a documented knowledge base before execution begins. A planner generates a structured task graph, a deterministic verifier validates functions, artifacts, and ordering constraints, and one bounded repair pass applies local edits when possible. Unsupported workflows are blocked through safe abstention rather than hallucinated capability invention. We evaluate PTV on Core-60, a 60-task public benchmark spanning supported, multi-step, repair-sensitive, and unsupported requests. Across local and server-side model panels, PTV substantially improves the reported static conformance score on several model families, including Gemma, Devstral, and DeepSeek, while remaining close to Plan-and-Execute in token cost. This score combines supported-task success with correct abstention on unsupported requests; it is not a real-execution success rate. Paired bootstrap confidence intervals show positive PTV–PAE conformance differences for the positive model families and near-zero effects for model-serving configurations whose structured outputs collapse. Additional schema-constrained reruns show that this boundary condition is not merely a scoring artifact: Qwen and DeepSeek recover usable structured plans and then show clear PTV gains, whereas GPT-OSS and GLM remain fragile under the tested serving stack. The primary benchmark is static and invokes no real tools; an auxiliary sandbox replay executes stored plans against mock tools derived from the same specification. The primary outcomes measure conformance to the frozen knowledge base and feasibility criteria, while the replay provides side-effect-free trace checks. These results suggest that lightweight feasibility validation improves static SOP conformance primarily by amplifying planners capable of producing stable structured workflows, rather than compensating for weak upstream plan representations. Core-60, the knowledge-base specification, implementation, and analysis materials are publicly available.
LLM planning often relies on reactive executors (e.g., ReAct) that generate a single action at each step to remain grounded in the latest observations, incurring substantial inference cost. Batched executors reduce this cost by predicting multiple actions in a single model invocation, yet they are prone to compounding...
Qing-Fu Zhu, Huan Zheng, Ting-Yu Liu et al.· 0 citations
Agentic robots must turn a high-level goal into actions while keeping the physical conditions needed for future steps valid. A completed action does not ensure that its effects will persist: a later disturbance can invalidate support that the remaining plan still assumes. We present LOGIV, a graph engineering framework...
En-Qiao Lu, Yu-Xiang Xie, Xingrui Yu et al.· 0 citations
Graph-Aware Refinement and Verification (GARV), which composes graph-grounded action contracts with a safety monitor to produce a model-relative certificate or counterexample and, when route choice is enforceable, a minimum-cost safe refinement.
Stabak Das, Priyesh Ranjan, Xiang-Fang Li et al.· 0 citations
LLM-based coding agents such as Claude Code can resolve real-world repository issues, but their iterative interaction loop can be expensive. Existing approaches reduce this cost primarily by limiting, compressing, or selectively avoiding information and actions within the repair process. We investigate a complementary...
Melika Sepidband, Jiho Shin, Song Wang et al.· 0 citations
Large language models (LLMs) perform well on many stepwise reasoning tasks, yet their ability to construct reliable plans over long horizons remains difficult to measure. Existing evaluations often use compact symbolic domains or allow execution-time feedback, making it harder to isolate failures in end-to-end plan con...
P. Anokhin, Roman Khalikov, Stefan Rebrikov et al.· IEEE Access· 0 citations
Language agents increasingly tackle long-horizon tasks in interactive environments, where actions often depend on latent preconditions. The core difficulty is not trajectory length but recovery: when a prerequisite is missing or an earlier decomposition fails, trying another plausible action rarely helps. ReAct and lat...
Ming-Yang Liu, Qing-Can Kang, Shi-Xiong Kai et al.· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.