Evaluating planning capabilities in agentic systems
Year: 2026
Type: dissertation
Abstract: LLM-based systems extend a base language model with external scaffolding, enabling behaviour beyond next-token prediction. Such scaffolding is often claimed to provide capabilities like deliberation, robustness, recoverability, and planning. Yet many existing benchmarks primarily measure end-to-end usefulness rather than the underlying capabilities a system exhibits. This gap is especially problematic in constrained or safety-critical settings, where we must verify that systems satisfy required properties. Planning is one such property: it captures whether a system can pursue a goal over long horizons by selecting and sequencing actions while satisfying hard preconditions and constraints at every step. In this thesis, we introduce a planning benchmark in which tasks are specified in natural language but grounded in an underlying symbolic transition system, expressed in PDDL, enabling formal, stepwise verification. The benchmark covers two common deployment regimes for LLM-based systems: offline planning (one-shot plan generation) and online planning-by-acting with feedback. It also includes unsolvable instances that systems must correctly identify as infeasible. We evaluate constraint compliance, goal achievement, efficiency, infeasibility detection, and resource usage (interaction budget, tokens, and cost), and we record complete execution traces to support fine-grained post-hoc analysis of where and why failures occur. We instantiate the benchmark with five domains and ten instances per domain (50 tasks total: 40 solvable, 10 unsolvable). We use this benchmark to evaluate four LLMs from two model families (ChatGPT~4o, ChatGPT~4o-mini, Gemini~2.0~Flash, and Gemini~2.0~Flash-lite). For offline planning, we implement and evaluate six widely used prompting techniques (Direct, Chain-of-Thought, Plan-and-Solve, Step-Back, Self-Refine, and Tree of Thoughts). For online systems, we implement and evaluate three representative agent architectures (ReAct, Reflexion, and Critic). While several techniques improve performance, results fall short of deployment-grade reliability: the best offline configuration achieves 68% success rate, and the best online configuration achieves a 40% success rate, with failures largely driven by constraint violations. Systems also frequently fail to flag infeasibility, and additional structure is not consistently beneficial---sometimes degrading performance while increasing token and cost overhead. Overall, our findings suggest that broad capability claims for LLM-based systems are often overstated, and that benchmarks should prioritise verifiable properties over end-to-end usefulness alone. We also present the benchmark as an A2A-compatible Green Agent in the AgentBeatsX competition. All code is open source and can be found in the project repository: https://github.com/oriolmirolf/agentic-planning-eval/
Author: Miró I López Feliu, Oriol
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7154816989