LRAT-E5
LRAT-E5 is a dense retriever obtained by fine-tuning intfloat/multilingual-e5-large-instruct with LRAT (Learning to Retrieve from Agent Trajectories), a training framework that adapts retrieval models to the behavior of search agents.
The central idea of LRAT is simple: if search is increasingly consumed by agents rather than humans, retrieval models should be trained from agent interaction data rather than only from human-centric supervision. This checkpoint is therefore optimized using query-document supervision mined from deep research trajectories.
What This Model Is For
This checkpoint is intended for:
- dense retrieval in agentic search and deep research systems
- retrieval for long-horizon information-seeking tasks
- evidence retrieval for multi-step reasoning pipelines
- replacing a general-purpose E5 retriever with one aligned to search-agent behavior
It is less suitable for:
- generic sentence embedding leaderboards without task-specific evaluation
- non-retrieval classification workflows
- fully unrelated domains without adaptation
Model Details
- Base model:
intfloat/multilingual-e5-large-instruct
- Training framework: LRAT
- Retriever type: dense bi-encoder
- Objective: weighted contrastive learning with trajectory-derived supervision
- Intended use: retrieval for search agents operating in multi-step reasoning environments
Training Data
The trajectory-derived supervision used for LRAT is constructed from:
- 10K seed queries from InfoSeekQA
- Tongyi-DeepResearch-30B-A3B as the trajectory generation agent
- Wiki-25-Dump as the retrieval corpus
- multiple retrievers during collection, including BM25 and Qwen3-based retrievers
The paper reports:
- 26,482 valid trajectories
- 91,713 training pairs
For more details, please refer to the
Training Data.
Training Configuration
The paper reports the following fine-tuning setup:
- epochs: 2
- batch size: 32
- learning rate:
1e-6
- maximum sequence length: 512
- group size: 10
- temperature: 0.02
The training is implemented with a FlagEmbedding-based dense retriever recipe.
Evaluation Summary
LRAT-E5 is evaluated inside six different search agents on:
- InfoSeek-Eval
- BrowseComp-Plus
The gains are consistent across both smaller task-optimized agents and larger generalist agentic foundation models.
Representative Results for This Backbone
| Agent Backbone | InfoSeek-Eval SR | BrowseComp-Plus SR | BrowseComp-Plus Recall |
|---|
| AgentCPM-Explore (base) | 47.3 | 15.9 | 26.5 |
| AgentCPM-Explore (+ LRAT) | 49.7 | 15.9 | 32.1 |
| WebExplore (base) | 60.0 | 25.4 | 50.4 |
| WebExplore (+ LRAT) | 63.3 | 29.0 | 56.1 |
| Tongyi-DeepResearch (base) | 56.7 | 20.7 | 54.8 |
| Tongyi-DeepResearch (+ LRAT) | 68.0 | 23.9 | 61.8 |
| GPT-OSS (120B, base) | 41.7 | 10.8 | 50.1 |
| GPT-OSS (120B, + LRAT) | 50.7 | 13.1 | 56.0 |
| MiniMax-M2.1 (base) | 64.0 | 46.4 | 64.9 |
| MiniMax-M2.1 (+ LRAT) | 75.0 | 48.7 | 69.7 |
| GLM-4.7 (base) | 73.7 | 46.4 | 68.7 |
| GLM-4.7 (+ LRAT) | 81.7 | 50.6 | 76.3 |
Across these agents, LRAT improves:
- InfoSeek-Eval success rate by roughly
+5.1% to +21.6% relative
- BrowseComp-Plus success rate by roughly
+0.0% to +21.3% relative
- BrowseComp-Plus evidence recall by roughly
+7.4% to +21.1% relative
Usage
This model should be used with the same preprocessing and embedding extraction recipe as the upstream E5 backbone and the LRAT codebase.
1from transformers import AutoTokenizer, AutoModel
2
3model_id = "Yuqi-Zhou/LRAT-multilingual-e5-large"
4
5tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
6model = AutoModel.from_pretrained(model_id, trust_remote_code=True)
7
8# Apply the same pooling / normalization setup used for
9# multilingual-e5-large-instruct in your retrieval pipeline.
If you use E5-style query instructions or prefixes in your system, keep them consistent with your evaluation and indexing setup.
License and Release Notes
Please confirm the final release license against:
- the upstream base model license
- the released training dataset license
- any additional source corpus obligations
This card can be updated later with a final project-level license statement if needed.
Citation
If you use this checkpoint, please cite the LRAT paper.
1@inproceedings{zhou2026lrat,
2 title={Learning to Retrieve from Agent Trajectories},
3 author={Zhou, Yuqi and Dai, Sunhao and Qu, Changle and Pang, Liang and Xu, Jun and Wen, Ji-Rong},
4 booktitle={Proceedings of the 49th International ACM SIGIR Conference on Research and Development in Information Retrieval},
5 year={2026}
6}
Links
- Paper:
https://arxiv.org/abs/2604.04949
- Project page:
https://yuqi-zhou.github.io/LRAT-homepage/
- Code:
https://github.com/Yuqi-Zhou/LRAT
- Model:
https://huggingface.co/Yuqi-Zhou/LRAT-multilingual-e5-large
- Dataset:
https://huggingface.co/datasets/Yuqi-Zhou/LRAT-Train
- Companion Qwen3 checkpoint:
https://huggingface.co/Yuqi-Zhou/LRAT-Qwen3-Embedding-0.6B
