LLMs excel at understanding and generating human-like text, allowing them to understand and generate responses that mimic human language, improving communication between machines and humans. These models are versatile and adaptable to various tasks, including language translation, summarization, question answering, text generation, sentiment analysis, and more. Its flexibility allows its implementation in various industries and applications.
However, LLMs sometimes hallucinate, resulting in plausible incorrect statements. Large language models, such as GPT models, are very advanced in language understanding and generation and can still produce confabulations for various reasons. If the input or message provided to the model is ambiguous, contradictory, or misleading, the model may generate fabricated responses based on its interpretation of the input.
Google DeepMind researchers overcome this limitation by proposing a method called FunSearch. It combines a pre-trained LLM with an evaluator, protecting against confabulations and incorrect ideas. FunSearch turns low-scoring starter programs into high-scoring programs to discover new knowledge by combining multiple essential ingredients. FunSearch produces programs that generate the solutions.
FunSearch operates as an iterative process where, in each cycle, the system selects certain programs from the current group. These selected programs are then processed by an LLM, which extends them in innovative ways, producing new programs that undergo automatic evaluation. The most promising are reintroduced into the set of existing programs, establishing a cycle of self-improvement.
Researchers sample the best-performing programs and enter them back into LLMs as prompts for improvement. They start with an initial program as a skeleton and develop only the critical parts that govern the logic of the program. They establish a greedy program skeleton and make decisions by placing a priority function on each step. They use island-based evolutionary methods to maintain a large set of diverse programs. They scale it asynchronously to expand the scope of their focus and find new results.
FunSearch uses the same general containerization strategy. Instead of packing items into containers with the lowest capacity, assign items to the lowest capacity only if the fit is very tight after the item is placed. This strategy eliminates small gaps in containers that are unlikely to be filled. One of the crucial components of FunSearch is that it operates in program space rather than searching constructs directly. This gives FunSearch the potential for real-world applications.
Without a doubt, this marks only the initial phase. The advancement of FunSearch will naturally align with the broader evolution of LLMs. Researchers are committed to expanding its functionalities to address various critical scientific and engineering challenges prevailing in society.
Review the Paper and Blog. All credit for this research goes to the researchers of this project. Also, don't forget to join. our 34k+ ML SubReddit, 41k+ Facebook community, Discord channel, and Electronic newsletterwhere we share the latest news on ai research, interesting ai projects and more.
If you like our work, you'll love our newsletter.
Arshad is an intern at MarktechPost. He is currently pursuing his international career. Master's degree in Physics from the Indian Institute of technology Kharagpur. Understanding things down to the fundamental level leads to new discoveries that lead to the advancement of technology. He is passionate about understanding nature fundamentally with the help of tools such as mathematical models, machine learning models, and artificial intelligence.
<!– ai CONTENT END 2 –>
