Can LLMs Really Reason and Plan? Pretrained Transformer Language Models for Search - part 1

Can LLMs Really Reason and Plan? Pretrained Transformer Language Models for Search - part 1

In recent years, there has been a surge of interest in language models, particularly Large Language Models (LLMs), and their capabilities in reasoning and planning. LLMs, such as GPT4, have shown remarkable performance in various language-related tasks, leading some to believe that they possess autonomous reasoning capabilities.

However, it is important to critically examine these claims and understand the true nature of LLMs' abilities. While LLMs excel in idea generation for any task, including reasoning, it does not necessarily mean that they can plan or reason autonomously. This distinction is crucial in leveraging LLMs effectively to support reasoning and planning tasks.

In the context of planning tasks, one way to assess the planning capabilities of LLMs is to reduce the effectiveness of approximate retrieval. By obfuscating the names of actions and objects in the planning problem, the empirical performance of GPT4 plummeted significantly. This indicates that the improved performance of GPT4 may not solely come from its ability to plan.

To further validate the role of LLMs in planning tasks, an external model-based plan verifier can be employed. By letting the verifier do the back prompting and certify the correctness of the final solution, we can ensure that the LLM's generated potential answers are checked and refined by external solvers. This approach provides a clearer understanding of the LLM's contribution to planning tasks without ascribing autonomous reasoning capabilities to them.

It is important to acknowledge that LLMs do have constructive roles to play in solving planning and reasoning tasks. Their uncanny ability to generate ideas and potential candidate solutions, although with no guarantees, can be valuable in what we call "LLM-Modulo" setups. In these setups, LLMs work in conjunction with model-based planners, external solvers, or expert humans in the loop. The key is to recognize that LLMs are generating potential answers to be checked and refined, rather than relying solely on their autonomous reasoning capabilities.

In the realm of information retrieval, lexical-based retrieval scoring functions like BM25 have gained popularity. These functions allow for efficient retrieval using inverted indexes and can be further accelerated by dynamic pruning algorithms like WAND. Dynamic pruning algorithms avoid exhaustively scoring all documents that match at least one of the query terms, leading to sub-linear time retrieval.

Dense retrieval, on the other hand, can be accelerated by employing approximate nearest neighbor search techniques. One such technique is indexing the document vector representation using HNSW graph indexing. This approach enables faster retrieval by approximating the nearest neighbors of a given query vector.

In conclusion, LLMs have their strengths in idea generation and can be effectively leveraged in planning and reasoning tasks. However, it is crucial to avoid ascribing autonomous reasoning capabilities to LLMs and instead utilize them in conjunction with external solvers or verifiers. By understanding the true nature of LLMs' abilities and incorporating them into well-designed frameworks, we can harness their potential while avoiding misleading interpretations.

Actionable advice:

1. When utilizing LLMs for planning tasks, incorporate an external model-based plan verifier to ensure the correctness of the final solution.
2. Obfuscate the names of actions and objects in planning problems to assess the true planning capabilities of LLMs.
3. Leverage LLMs' strengths in idea generation by combining them with model-based planners, external solvers, or expert humans in the loop for more effective planning and reasoning.

By following these actionable advice, researchers and practitioners can make the most out of LLMs' capabilities while avoiding overestimating their autonomous reasoning abilities.