AceSearcher/AceSearcher-32B

Hugging Face
TEXT GENERATIONConcurrency Cost:2Model Size:32.8BQuant:FP8Ctx Length:32kPublished:Jun 16, 2025License:mitArchitecture:Transformer Open Weights Warm

AceSearcher/AceSearcher-32B is a 32.8 billion parameter language model developed by AceSearcher, built upon the Qwen-2.5-Instruct-32B backbone. This model is specifically fine-tuned for complex question answering and fact verification tasks by bootstrapping reasoning and search through reinforced self-play. It excels at decomposing intricate questions and claims into manageable sub-components, facilitating more accurate and grounded responses. The model is optimized for applications requiring structured reasoning over provided contexts, such as document-level financial analysis.

Loading preview...

AceSearcher-32B: Enhanced Reasoning and Search for LLMs

AceSearcher/AceSearcher-32B is a 32.8 billion parameter model, leveraging the Qwen-2.5-Instruct-32B architecture, specifically designed to improve reasoning and search capabilities in large language models. Its development is detailed in the paper "AceSearcher: Bootstrapping Reasoning and Search for LLMs via Reinforced Self-Play".

Key Capabilities

  • Question Decomposition: Breaks down complex questions into multiple specific sub-questions, marking each with ### for structured processing.
  • Claim Decomposition: Decomposes intricate claims into smaller, verifiable sub-claims, aiding in fact verification tasks.
  • Contextual Question Answering: Answers sub-questions using provided context passages, with a fallback to general knowledge if context is insufficient.
  • Fact Verification: Verifies sub-claims against given contexts, outputting a concise 'Yes' or 'No'.
  • Final Answer Generation: Synthesizes information from passages and sub-question answers to generate comprehensive final answers for original questions or claims, ensuring responses are grounded and provide clear reasoning.
  • Document-Level Financial Reasoning: Demonstrates specialized decomposition and Python-based problem-solving for financial analysis questions based on tabular and textual data.

Good For

  • Applications requiring robust, multi-step reasoning over complex information.
  • Systems needing to verify facts or claims with high accuracy.
  • Tasks involving structured decomposition of user queries for improved answer generation.
  • Financial analysis and other domain-specific tasks where precise, context-grounded answers are critical.

Finetunes
1 models