GritHopper-7B

<p align="center"> <img src="static/GritHopperLogo.jpeg" alt="GritHopper Logo" height="250px" align="left" style="position: relative; z-index: 1;"> <div align="center"> <h1> <h1>GritHopper: Decomposition-Free<br> Multi-Hop Dense Retrieval</h1> </h1> <p align="center"> 🌐 <a href="https://ukplab.github.io/eacl2026-GritHopper/" target="_blank">Website</a> | 🏗️ <a href="https://github.com/UKPLab/GritHopper" target="_blank">GitHub Repo</a> | 📃 <a href="https://aclanthology.org/2026.eacl-long.5.pdf" target="_blank">Paper</a> </p> </div> </p> <br clear="left"/> <!--- BADGES: START, copied from sentence transformers, will be replaced with the actual once (removed for anonymity)--->

---

GritHopper-7B is the first decoder-based multi-hop dense retrieval model and achieves state-of-the-art performance on both in-distribution and out-of-distribution benchmarks for decomposition-free multi-hop dense retrieval. Built on GRITLM, it is trained across diverse datasets spanning question-answering and fact-checking. Unlike traditional decomposition-based approaches, GritHopper iteratively retrieves passages without explicit sub-question decomposition, concatenating retrieved evidence with the query at each step.

Using the decoder model in an encoder-only approach (like MDR), it performs each retrieval step in a single forward pass. In contrast to previous SOTA BERT-based approaches (like BeamRetriever or MDR), GritHopper generalizes significantly better to out-of-distribution data.

Key Strengths of GritHopper

• Encoder-Only Efficiency: Each retrieval iteration requires only a single forward pass (rather than multiple autoregressive steps).
• Out-of-Distribution Robustness: Achieves state-of-the-art performance compared to other decomposition-free methods on multiple OOD benchmarks.
• Unified Training: Combines dense retrieval with generative objectives, exploring how post-retrieval information on the generation loss improves dense retrieval performance.
• Stopping: GritHopper utilizes its generative capabilities via ReAct to control its own state. This way, it can stop itself through causal next-token prediction.

---

Staring with GritHopper

GritHopper is trained on MuSiQue, 2WikiMultiHopQA, HotPotQA, EX-Fever and HoVer.

GritHopper Models

1. Installation

pip install grithopper

2. Initialization

from grithopper import GritHopper

# Initialize GritHopper with your GRITLM model checkpoint or huggingface path
hopper = GritHopper(
    model_name_or_path="UKPLab/GritHopper-7B",  
    device="cuda"  # or "cpu"
)

3. Load Document Candidates

You can either load from a list of (title, passage) pairs and optionally dump them to a file:

documents = [
    ("Title A", "Passage text for document A."),
    ("Title B", "Passage text for document B."),
    # ...
]

hopper.load_document_candidates(
    document_candidates=documents,
    device="cuda",
    output_directory_candidates_dump="my_candidates.pkl"  # optional
)

Or load them from a pre-encoded dump:

hopper.load_candidates_from_file(
    dump_filepath="my_candidates.pkl",
    device="cuda"
)

4. Encode a Query

question = "Who wrote the novel that was adapted into the film Blade Runner?"
previous_evidences = [("Blade Runner (Movie)", " The Movie....")] # optional


query_vector = hopper.encode_query(
    multi_hop_question=question,
    previous_evidences=previous_evidences, # optional
    instruction_type="multi-hop"  # or "fact-check" alternatively you can provide a custom instruction with insruction="your_instruction"
)

5. Single-Step Retrieval

result = hopper.retrieve_(
    query=query_vector,
    top_k=1,
    get_stopping_probability=True
)

# {
#   "retrieved": [
#       {
#         "title": "Title B",
#         "passage": "Passage text for document B.",
#         "score": 0.873
#       }
#   ],
#   "continue_probability": 0.65,  # present if get_stopping_probability=True
#   "stop_probability": 0.35
# }

If you prefer to pass the question string directly:

result = hopper.retrieve_(
    query="Who is the mother of the writer who wrote the novel that was adapted into the film Blade Runner?",
    # optional previous_evidences=[("Blade Runner (Movie)", " The Movie....")],
    top_k=1,
    get_stopping_probability=True,
)

# {
#   "retrieved": [
#       { "title": "Blade Runner (Movie)", "passage": "...", "score": 0.92 }
#   ],
#   "continue_probability": 0.75,
#   "stop_probability": 0.25
# }

6. Iterative (Multi-Hop) Retrieval

chain_of_retrieval = hopper.iterative_retrieve(
    multi_hop_question="Who wrote the novel that was adapted into the film Blade Runner?",
    instruction_type="multi-hop",
    automatic_stopping=True,
    max_hops=4
)

# [
#   {
#     "retrieved": [
#       { "title": "Blade Runner (Movie)", "passage": "...", "score": 0.92 }
#     ],
#     "continue_probability": 0.75,
#     "stop_probability": 0.25
#   },
#   {
#     "retrieved": [
#       { "title": "Philip K.", "passage": "...", "score": 0.88 }
#     ],
#     "continue_probability": 0.65,
#     "stop_probability": 0.35
# },
#   ...
# ]

This process continues until either:

1.	The model determines it should stop (if automatic_stopping=True and stop_probability > continue_probability).
2.	It hits max_hops.
3.	Or no documents can be retrieved at a given step.

---

Citation

If you use GritHopper in your research, please cite the following paper:

@inproceedings{erker-etal-2026-grithopper,
    title = "{GRITH}opper: Decomposition-Free Multi-Hop Dense Retrieval",
    author = "Erker, Justus-Jonas  and
      Reimers, Nils  and
      Gurevych, Iryna",
    editor = "Demberg, Vera  and
      Inui, Kentaro  and
      Marquez, Llu{\'i}s",
    booktitle = "Proceedings of the 19th Conference of the {E}uropean Chapter of the {A}ssociation for {C}omputational {L}inguistics (Volume 1: Long Papers)",
    month = mar,
    year = "2026",
    address = "Rabat, Morocco",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2026.eacl-long.5/",
    doi = "10.18653/v1/2026.eacl-long.5",
    pages = "77--94",
    ISBN = "979-8-89176-380-7",
    abstract = "Decomposition-based multi-hop retrieval methods rely on many autoregressive steps to break down complex queries, which breaks end-to-end differentiability and is computationally expensive. Decomposition-free methods tackle this, but current approaches struggle with longer multi-hop problems and generalization to out-of-distribution data. To address these challenges, we introduce GRITHopper-7B, a novel multi-hop dense retrieval model that achieves state-of-the-art performance on both in-distribution and out-of-distribution benchmarks. GRITHopper-7B combines generative and representational instruction tuning by integrating causal language modeling with dense retrieval training. Through controlled studies, we find that incorporating additional context after the retrieval process, referred to as post-retrieval language modeling, enhances dense retrieval performance. By including elements such as final answers during training, the model learns to better contextualize and retrieve relevant information. GRITHopper-7B offers a robust, scalable, and generalizable solution for multi-hop dense retrieval, and we release it to the community for future research and applications requiring complex reasoning and retrieval capabilities."
}

Contact

Contact person: Justus-Jonas Erker, [email protected]

https://www.ukp.tu-darmstadt.de/

https://www.tu-darmstadt.de/

Don't hesitate to send us an e-mail or report an issue, if something is broken (and it shouldn't be) or if you have further questions. This repository contains experimental software and is published for the sole purpose of giving additional background details on the respective publication.

License

GritHopper is licensed under the Apache License, Version 2.0. See LICENSE for the full license text.

Acknowledgement

this Model is based upon the GRITLM.