On the Planning Abilities of Large Language Models (A Critical Investigation with a Proposed Benchmark)

Abstract

Intrigued by the claims of emergent reasoning capabilities in LLMs trained on general web corpora, in this paper, we set out to investigate their planning capabilities. We aim to evaluate (1) how good LLMs are by themselves in generating and validating simple plans in commonsense planning tasks (of the type that humans are generally quite good at) and ( 2) how good LLMs are in being a source of heuristic guidance for other agents-either AI planners or human planners-in their planning tasks. To investigate these questions in a systematic rather than anecdotal manner, we start by developing a benchmark suite based on the kinds of domains employed in the International Planning Competition. On this benchmark, we evaluate LLMs in three modes: autonomous, heuristic and human-in-the-loop. Our results show that LLM's ability to autonomously generate executable plans is quite meager, averaging only about 3% success rate. The heuristic and human-in-the-loop modes show slightly more promise. In addition to these results, we also make our benchmark and evaluation tools available to support investigations by research community.

SayCanPay: Heuristic Planning with Large Language Models Using Learnable Domain Knowledge

“…Existing works have demonstrated the planning abilities of both the decoder type (Pallagani et al 2022) and the encoder-decoder type architectures (Valmeekam et al 2023(Valmeekam et al , 2022. Since the generated plan is in free-form language and may contain unrecognizable (for the environment) words or incorrect syntax, it cannot be directly translated into actionable steps in the environment.…”

“…using a controller to regulate the heater where only a knob exists), leading to infeasible plans. Moreover, such models focus greedily on the next actionable step without considering its relevance to the ultimate goal, resulting in longer, cost-inefficient plans (Valmeekam et al 2023). Recent works like SayCan (Ahn et al 2022) have sought to address the affordance problem by using pretrained skills to evaluate the action's executability -Can the action be executed in the current state?…”

Improving Knowledge Extraction from LLMs for Task Learning through Agent Analysis

“…STARS attempts to verify LLM responses before attempting to achieve the goal indicated by a response. There are many approaches to verification of LLM knowledge, including 1) response sampling (Wang et al 2023), 2) use of other sources of knowledge such as planning (Valmeekam et al 2023) or an LLM (Kim, Baldi, and McAleer 2023), and 3) human feedback/annotation (TidyBot). Recursively Criticizes and Improves (RCI; Kim, Baldi, and McAleer 2023) verifies LLM output by prompting the LLM again to identify (potential) issues.…”

Artificial Intelligence: A Kalam and Sufi Perspective

“…Although recent AI systems show some abstraction and reasoning abilities, however, they seem to use pattern matching, shortcuts, and memorization of some aspect of the reasoning process [50][51][52][53][54]. Despite the remarkable ability of Large Language Models (LLMs) [9,[70][71][72][73][74] in learning some patterns of the reasoning process and then apply them in different context, they still lack understanding of the coherent text they produce when they are probed more deeply [68,[75][76][77][78][86][87][88]. Another flaw in the argument that deep learning will lead to human-level intelligence is the assumption that intelligence will somehow emerge through training neural networks without providing any convincing justification for this assumption.…”