AceSearcher/AceSearcher-14B: Enhanced Reasoning and Search for LLMs
AceSearcher/AceSearcher-14B is a 14.8 billion parameter model, leveraging the Qwen-2.5-Instruct-14B backbone, specifically developed to improve the reasoning and search abilities of large language models. Its core innovation lies in a reinforced self-play approach, enabling it to effectively decompose complex questions and claims into manageable sub-questions or sub-claims.
Key Capabilities
- Question Decomposition: Breaks down intricate questions for both general Question Answering (QA) and fact verification tasks into specific, interlinked sub-questions.
- Contextual Answering: Provides short, span-based answers to sub-questions using provided context, falling back to its own knowledge if context is insufficient.
- Fact Verification: Verifies sub-claims with a simple 'Yes' or 'No' based on context or internal knowledge.
- Final Answer Synthesis: Integrates answers from sub-questions and provided passages to generate comprehensive, grounded final answers for original complex questions or claims, including clear reasoning.
- Financial Reasoning: Demonstrates specialized capabilities in document-level financial reasoning, including decomposing financial questions and generating Python code to solve them based on tabular and textual data.
Good For
- Applications requiring robust complex question answering and fact-checking by breaking down problems.
- Tasks that benefit from structured reasoning and step-by-step problem-solving in LLMs.
- Financial analysis and other domains where precise, context-aware reasoning and programmatic solutions are beneficial.
This model is the checkpoint used in the paper "AceSearcher: Bootstrapping Reasoning and Search for LLMs via Reinforced Self-Play".