Neon and Castform have done something that makes people re-evaluate "small models" — they trained an open-source model with 4B parameters using reinforcement learning. On the document search task, the accuracy is not only no less than GPT-5.6Sol, but even higher, while the cost per inference is about one percent of the latter's.

Behind this is a re-understanding of the search paradigm. Currently, the mainstream approach for document retrieval is embedding-based search, where documents are converted into numerical vectors to find similar content. However, with the popularity of AI agents, the search process is shifting towards agent-based search: the model breaks down complex questions into multiple smaller questions, autonomously decides on search queries, reviews results, and initiates the next round of searches, repeating this cycle until all necessary information is gathered. This approach can handle more complex problems, but at the cost of requiring a high-performance model for each search, leading to high time and cost — according to Neon's typical request specifications, a single search request with GPT-5.6Sol takes over 10 seconds and costs about $0.03.

image.png

Neon and Castform each take on their own roles: Neon provides the document storage location and search capabilities, while Castform is responsible for training the model to determine "what to search for". The training uses reinforcement learning, where the model first attempts to complete the task, and the system gives a score based on the results, which is directly fed back into the next round of experiments. Evaluation dimensions are not only about whether the final answer is correct, but also whether the correct document was found and whether appropriate sections were cited. Ultimately, the model trained by Castform achieved an average evaluation score of 1.447 in Neon's validation, exceeding GPT-5.6Sol's 1.369, and also higher than the control group GPT-5.4's 1.377, setting a new record for this validation.