README.md

metadata

license: mit
task_categories:
  - question-answering
  - text-retrieval
  - text-generation
language:
  - en
tags:
  - rag
  - llm
  - reasoning
  - search
  - multi-hop-reasoning
  - fact-verification
  - reinforced-self-play

AceSearcher: Bootstrapping Reasoning and Search for LLMs via Reinforced Self-Play

This repository contains the datasets and resources for AceSearcher: Bootstrapping Reasoning and Search for LLMs via Reinforced Self-Play. AceSearcher is a cooperative self-play framework that trains a single large language model (LLM) to alternate between a decomposer that breaks down complex queries and a solver that integrates retrieved contexts for answer generation, eliminating the need for intermediate annotations. AceSearcher significantly enhances LLMs' ability to tackle complex reasoning tasks by coupling supervised fine-tuning with reinforcement fine-tuning optimized for final answer accuracy.

• Paper: AceSearcher: Bootstrapping Reasoning and Search for LLMs via Reinforced Self-Play
• Code: https://github.com/ritaranx/AceSearcher/

Data Download

The AceSearcher project comprises several datasets available on Hugging Face:

Resource	Link
SFT Data	AceSearcher/Search-SFT
RFT Data	AceSearcher/Search-RFT-Pairs
RFT Prompts	AceSearcher/Search-RFT-Prompts
Evaluation Data	AceSearcher/evaluation_datasets

Data Generation

Most of the data generation used in AceSearcher is in the rollout folder of the code repository. The description for files are listed as belows:

• rs_mhqa.py | rs_cot.py | rs_pot.py: [Step 1] the rollout pipeline for multi-hop QA, chain-of-thought, and program-of-thought datasets.
• create_training_pairs.py: [Step 2] the process for filtering & selecting preference pairs in mDPO iterations.
• create_dpo_pairs.py: [Step 3] the process of curating the final preference pairs for reinforcement finetuning

Evaluation

For detailed evaluation scripts, please refer to the code repository:

• For QA / Fact Verification Datasets:
  • Use decompose_vllm.py to first decompose the data.
  • Use main_qa.py to generate the final answer.
• For Document-level Financial Reasoning Datasets:
  • Use main_reasoning.py for evaluation.

Sample Usage

Below are examples demonstrating how to use the models for various tasks, as provided in the Github repository.

For question decomposition on QA tasks:

prompt_plan_qa = """Please break down the question "{question}" into multiple specific sub-questions that address individual components of the original question. 
Mark each sub-question with ### at the beginning.  If you need to refer to answers from earlier sub-questions, use #1, #2, etc., to indicate the corresponding answers.
Decomposed Question:"""

prompt_qa = prompt_plan_qa.replace("{question}", question)

prompt = [
    {"role": "user", "content": prompt_qa.strip()}
] 

text = tokenizer.apply_chat_template(
    prompt,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=False
)

outputs = llm.generate([text], sampling_params)
generated_text = outputs[0].outputs[0].text

For question decomposition on fact verification tasks:

prompt_plan_claim = """Please break down the claim "{claim}" into multiple smaller sub-claims that each focus on a specific component of the original statement, making it easier for a model to verify.
Begin each sub-claim with ###. If needed, refer to answers from earlier sub-claims using #1, #2, etc.
Decomposed claim:"""

prompt_plan_claim = prompt_plan_claim.replace("{question}", question)

prompt = [
    {"role": "user", "content": prompt_plan_claim.strip()}
] 

text = tokenizer.apply_chat_template(
    prompt,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=False
)

outputs = llm.generate([text], sampling_params)
generated_text = outputs[0].outputs[0].text

For question answering for subquestions:

prompt = f"""You have the following context passages:
{context_text}

Please answer the question '{sub_q}' with a short span using the context as reference.
If no answer is found in the context, use your own knowledge. Your answer needs to be as short as possible."""

For fact verification tasks for subquestions:

prompt = f"""You have the following context passages:
{context_text}

Please verify whether the claim '{sub_q}' is correct using the context as reference. 
If no answer is found in the context, use your own knowledge.
Please only output Yes or No and do not give any explanation."""

For question answering to generate the final answer:

prompt = f"""You have the following passages:
{passages}

You are also given some subquestions and their answers:
{sub_answer_text}

Please answer the question '{original_question}' with {final_prompt} using the documents and subquestions as reference.
Make sure your response is grounded in documents and provides clear reasoning followed by a concise conclusion. If no relevant information is found, use your own knowledge. 
Wrap your answer with <answer> and </answer> tags."""

For fact verification tasks to generate the final answer:

prompt = f"""You have the following passages:
{passages}

You are given some subquestions and their answers:
{sub_answer_text}

Please verify the correctness of the claim: '{original_question}' using the subquestions as reference. Please provide a concise and clear reasoning followed by a concise conclusion. Your answer should be Yes or No only. 
Wrap your answer with <answer> and </answer> tags."""

For Decomposition for document-level financial reasoning tasks:

decompose_prompt = """You have the following passages and table:
Passages:
{passage}
Please break down the question '{question}' into multiple specific sub-questions that address individual components of the original question, with the table and passages as the reference. Use ### to mark the start of each sub-question."""

qa_prompt = """You have the following passages and table:
Passages:
{passage}
For the question '{question}', here is a referenced breakdown:
{decompose}.

Write a Python program to solve the question. Store the final result in the variable ans."""


question = "What would the change in furniture and fixtures between 2018 and 2019 be if furniture and fixtures were $5,000 thousand in 2018 instead? (in thousand)"

context_text = "
|||December 31,||
||Useful Life|2019|2018|
|Computer equipment and software|3 \u2013 5 years|$57,474|$52,055|
|Furniture and fixtures|7 years|6,096|4,367|
|Leasehold improvements|2 \u2013 6 years|22,800|9,987|
|Renovation in progress|n/a|8|1,984|
|Build-to-suit property|25 years|\u2014|51,058|
|Total property and equipment, gross||86,378|119,451|
|Less: accumulated depreciation and amortization||(49,852)|(42,197)|
|Total property and equipment, net||$36,526|$77,254|
 7. OTHER BALANCE SHEET AMOUNTS The components of property and equipment, net is as follows (in thousands): Depreciation expense for the years ended December 31, 2019, 2018, and 2017 was $11.8 million, $10.2 million, and $10.3 million, respectively.
"

decompose_prompt = decompose_prompt.replace("{passage}" , context_text)
decompose_prompt = decompose_prompt.replace("{question}", question)
message = [{"role": "user", "content": decompose_prompt.strip()}]
prompt = tokenizer.apply_chat_template(message, tokenize=False, add_generation_prompt=True)
generated_text = llm.generate(prompt, sampling_params)[0].outputs[0].text

qa_prompt = qa_prompt.replace("{passage}", context_text)
qa_prompt = qa_prompt.replace("{question}", question)
qa_prompt = qa_prompt.replace("{decompose}", generated_text)
message = [{"role": "user", "content": qa_prompt.strip()}]
prompt = tokenizer.apply_chat_template(message, tokenize=False, add_generation_prompt=True)
output = llm.generate(prompt, sampling_params)[0].outputs[0].text

Training

The authors use the Llama-Factory codebase for both SFT and RFT (mDPO) finetuning. Please see the config folder in the code repository for example configurations.

Citation

If you find this work useful, consider citing it. Thank you in advance:

@inproceedings{
xu2025acesearcher,
title={AceSearcher: Bootstrapping Reasoning and Search for LLMs via Reinforced Self-Play},
author={Ran Xu and Yuchen Zhuang and Zihan Dong and Ruiyu Wang and Yue Yu and Joyce C. Ho and Linjun Zhang and Haoyu Wang and Wenqi Shi and Carl Yang},
booktitle={the 39th Annual Conference on Neural Information Processing Systems},
year={2025},
url={https://openreview.net/forum?id=jSgCM0uZn3}
}