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

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

In the field of natural language processing, lexical-based retrieval scoring functions like BM25 have become increasingly popular. These functions allow for efficient retrieval of relevant information by using inverted indexes and dynamic pruning algorithms like WAND. By avoiding the exhaustive scoring of all documents, retrieval can be done in sub-linear time. However, there are other methods that can further accelerate dense retrieval, such as approximate nearest neighbor search using HNSW graph indexing.

While lexical-based retrieval scoring functions have proven to be effective, there is ongoing debate about the capabilities of pretrained Transformer language models (LLMs) when it comes to reasoning and planning. LLMs have demonstrated exceptional performance in generating ideas for various tasks, including those involving reasoning. This ability can be leveraged to support reasoning and planning processes effectively.

It is important to note that questioning the claims about LLMs' capabilities in planning and reasoning does not imply that LLMs are devoid of any useful roles in solving such tasks. In fact, their aptitude for generating potential candidate solutions can be valuable in conjunction with model-based planners, external solvers, or expert humans in the loop. However, it is crucial to acknowledge that LLMs are primarily generating ideas to be checked and refined by external solvers, rather than ascribing autonomous reasoning capabilities to them.

One way to test the planning abilities of LLMs is by obfuscating the names of actions and objects in the planning problem. When this was done for test domains, the empirical performance of GPT4, a popular LLM, significantly dropped. This was in contrast to standard AI planners, which had no trouble with the obfuscation. In the Blocks World domain, GPT4's empirical accuracy reached 30%, although it was lower in other domains. To ensure the correctness of the final solution, an external model-based plan verifier can be employed for back prompting and certification.

It is worth noting that LLMs are prone to the Clever Hans effect, where the LLM generates guesses while the human in the loop guides the LLM towards the correct solution, even if unintentionally. This highlights the importance of understanding that LLMs excel at extracting planning knowledge, which can be leveraged effectively.

In conclusion, pretrained Transformer language models have proven to be powerful tools in search and retrieval tasks. While there are ongoing discussions about their reasoning and planning capabilities, their ability to generate ideas and potential candidate solutions can be harnessed effectively. By incorporating external solvers, model-based planners, or expert humans in the loop, LLMs can contribute to solving complex planning and reasoning tasks. However, it is important to avoid ascribing autonomous reasoning capabilities to LLMs and instead view them as valuable aids in the overall problem-solving process.

Actionable advice:

1. When using lexical-based retrieval scoring functions, consider implementing dynamic pruning algorithms like WAND to accelerate the retrieval process.
2. Explore the use of approximate nearest neighbor search, such as HNSW graph indexing, to further improve the efficiency of dense retrieval.
3. When leveraging LLMs for planning and reasoning tasks, incorporate external solvers or model-based plan verifiers to check and refine the generated solutions. Avoid relying solely on the LLM's guesses without external validation.

By following these actionable advice and understanding the role of LLMs in planning and reasoning tasks, researchers and practitioners can harness the full potential of pretrained Transformer language models in search and retrieval processes.