Researchers have developed a novel method called EvoDuet that helps large language models (LLMs) improve their problem-solving abilities by evolving both their search strategies and solutions simultaneously. This approach tackles a common challenge in AI-driven scientific discovery: when models hit a knowledge wall because they lack access to the latest or most relevant external information. EvoDuet’s innovation lies in dynamically refining web search queries and solution candidates together, enabling AI systems to find better information and make more progress on complex tasks.
Key Takeaways
- EvoDuet introduces a bi-level optimization process that co-evolves search queries and candidate solutions, allowing AI models to adapt their information retrieval dynamically.
- The method uses a “retrieval gate” that lets the AI decide when to fetch new documents, reuse existing ones, or proceed without external data, based on its current knowledge gaps.
- Testing EvoDuet on 21 scientific optimization tasks showed significant improvements in discovery performance with advanced models like GPT-5.6-Luna and Gemini-3.8-Flash.
- EvoDuet also demonstrated flexibility by enhancing other evolutionary search frameworks, suggesting broad applicability in AI-driven research and problem solving.
At its core, EvoDuet addresses a key limitation in evolutionary search methods that rely on LLMs: these models often stall if they cannot access fresh or relevant external knowledge. Simply adding a web search tool isn’t enough because search queries might return the same documents repeatedly, even as the AI’s proposed solutions evolve. EvoDuet overcomes this by pairing two loops of optimization. The “inner loop” focuses on improving search queries and ranking retrieved documents based on how promising they seem to be for generating better solutions. Meanwhile, the “outer loop” generates multiple solution candidates in parallel, tests them, and feeds back their performance to refine future searches.
One of EvoDuet’s unique features is the retrieval gate, a mechanism that allows the AI to self-assess its understanding and decide whether to pull in new information, reuse what it already has, or continue working without additional data. This flexibility helps avoid redundant searches and ensures the AI’s knowledge base evolves in sync with its problem-solving attempts.
The researchers evaluated EvoDuet using a set of 21 optimization tasks related to scientific discovery, such as reducing swaps in computational problems or improving symbolic reasoning benchmarks. Using state-of-the-art models like GPT-5.6-Luna and Gemini-3.8-Flash, EvoDuet consistently outperformed previous methods, raising the normalized discovery gain—a measure of how effectively the AI finds better solutions. Notably, the method improved results on eight tasks beyond the best scores reported before and matched top performance on three others. However, some models, like Qwen3.5-9B, did not see benefits, highlighting that EvoDuet’s gains depend on the underlying AI’s capabilities.
Looking ahead, EvoDuet’s approach of tightly coupling search and solution evolution could be a valuable tool for enhancing AI-driven scientific research, where accessing and integrating fresh knowledge is crucial. By enabling AI models to adaptively seek out the most relevant information as they work through complex problems, EvoDuet paves the way for more efficient and effective discovery processes. Future exploration could involve applying this method to even broader domains or integrating it with other AI systems to accelerate innovation in science and technology.
Based on research published on arXiv by Young-Jun Lee, Jinheon Baek, Soyeong Jeong et al..