🚀 LLM-Guided Embedding Refinement

Boost zero-shot classification and retrieval with test-time query optimization

---

📖 Overview

This repository contains code to improve zero-shot classification and retrieval using embedding models through test-time optimization of query embedding representations.

At test time, given a user query, the query embedding is optimized with gradient descent based on targeted feedback from a stronger model. The method uses scores from an LLM or reranker over a small sampled set of candidate documents, then updates the query representation so the embedding space better reflects the task-specific intent of the query.

🎯 Key Features

• ✨ Test-time optimization — No retraining required
• 🔄 Flexible architecture — Works with various text embedding models and LLM rerankers
• 📊 Proven results — Consistent gains across multiple benchmarks

📄 Paper

Task-Adaptive Embedding Refinement via Test-time LLM Guidance
Ariel Gera, Shir Ashury-Tahan, Gal Bloch, Ohad Eytan, Assaf Toledo

---

💡 How It Works

Embedding models are efficient and scalable, but in challenging zero-shot settings they may miss nuanced task constraints. This library explores a test-time refinement procedure that adapts the query representation using external guidance from a generative LLM, without retraining the embedding model.

This approach improves ranking quality across multiple search and classification benchmarks, with consistent gains on tasks such as:

• 📚 Literature search
• 🎯 Intent detection
• 🔑 Key-point matching
• 📋 Instruction-following retrieval

🔄 Workflow

Step-by-step process:

1. 📝 Embed the original query and candidate documents
2. 🔍 Retrieve top candidates by embedding similarity
3. 🎯 Score a sampled subset using an LLM, cross-encoder reranker, or gold labels
4. ⚡ Optimize the query embedding to better align with supervision
5. 🔄 Re-score the corpus using the refined query embedding

---

🚀 Quick Start

📦 Installation

1️⃣ Create and activate a Python environment

python -m venv .venv
source .venv/bin/activate

2️⃣ Install dependencies

pip install -r requirements.txt

3️⃣ Configure inference for LiteLLM or OpenAI

The reranker and HyDE components use an OpenAI-compatible chat-completions API.

Option A — LiteLLM gateway or proxy

Set BASE_URL to your LiteLLM endpoint and API_KEY to the corresponding key:

export BASE_URL="http://localhost:4000"export API_KEY="your-litellm-api-key"

Use LiteLLM explicitly by prefixing the model name with LiteLLM/, for example:

python run_experiments.py \
--reranker_model "LiteLLM/mistralai/Mistral-Small-3.2-24B-Instruct-2506" \

Option B — OpenAI API

Set your OpenAI API key:

export OPENAI_API_KEY="your-openai-api-key"

Use OpenAI explicitly by prefixing the model name with OpenAI/, for example:

python run_experiments.py \
--reranker_model "OpenAI/gpt-4.1-mini" \
--hyde_model "OpenAI/gpt-4.1-mini"

Important notes

• If no service prefix is provided, this repository defaults to LiteLLM for LLM inference. This will fail if you did not define a suitable endpoint in your environment.
• --reranker_model controls the LLM used for relevance feedback during optimization.
• --hyde_model controls the LLM used to generate hypothetical documents for HyDE (HYpothetical Document Embeddings, see here), this is optional and is not required for basic query optimization functionality.
• LiteLLM/OpenAI setup is only required when using LLM-based reranking or HyDE. It is not needed for --optimize_with_gold or cross-encoder rerankers such as cross-encoder/ms-marco-MiniLM-L-6-v2.

▶️ Basic Usage

Run all default models on all datasets:

python run_experiments.py

Run a specific model on selected datasets:

python run_experiments.py \
--embedding_models "Qwen/Qwen3-Embedding-0.6B" \
--datasets "Clinc150""NFCorpus"

Run experiments in parallel:

python run_experiments.py --parallel 3

---

📊 Supported Datasets

The repository currently supports the following datasets through dataset_loaders.py:

Dataset	Description	Reference
🎓 RealScholarQuery	Real-world academic search queries over arXiv CS papers	He et al., 2025
🔑 ArgKP-21	Key-point matching from 2021 KPA shared task	Friedman et al., 2021
📋 FollowIR	Information retrieval from TREC relevance narratives	Weller et al., 2025
💬 Clinc150	Intent classification with 150 intents across 10 domains	Larson et al., 2019
🏦 Banking77	Banking domain with 77 fine-grained intent categories	Casanueva et al., 2020
🏥 NFCorpus	Medical literature retrieval with lay queries	Boteva et al., 2016

---

🛠️ Custom Usage

🎮 Main Entry Points

Script	Purpose
embedding_adaptation.py	Core script for single experiment runs
run_experiments.py	Batch runner for multiple experiments

🔧 Command Examples

Preview commands without execution

python run_experiments.py --dry_run

Enable HyDE (Hypothetical Document Embeddings)

python run_experiments.py \
--hyde_model "meta-llama/Llama-3.1-8B-Instruct"

Run single experiment with custom parameters

python embedding_adaptation.py \
--embedding_model "Qwen/Qwen3-Embedding-0.6B" \
--dataset "NFCorpus" \
--reranker_model "mistralai/Mistral-Small-3.2-24B-Instruct-2506" \
--lr 1e-4 \
--num_steps 100 \
--total_scores 20 \
--scores_from_top 20

⚙️ Configuration Parameters

🎯 Single Experiment Parameters (embedding_adaptation.py)

These parameters configure individual experiment runs. They can also be passed to run_experiments.py and will be forwarded to each experiment.

Parameter	Description	Default
--embedding_model	Embedding model to use (single model)	Qwen/Qwen3-Embedding-0.6B
--dataset	Dataset to use for experiment	RealScholarQuery
--reranker_model	Reranker model for feedback	mistralai/Mistral-Small-3.2-24B-Instruct-2506
--hyde_model	LLM for hypothetical document generation (optional)	None
--lr	Learning rate for query embedding optimization	1e-4
--num_steps	Number of optimization steps	100
--total_scores	Total documents to sample for reranking signal	20
--scores_from_top	Documents sampled from top results (by embedding similarity)	20
--optimize_with_gold	Use gold labels instead of reranker scores	False
--embedder_batch_size	Batch size for embedding inference	10
--reranker_batch_size	Batch size for reranker inference	10
--random_seed	Random seed for reproducibility	42
--save_tensors	Save query trajectory tensors for analysis	False
--experiment_name	Custom experiment name (auto-generated if not provided)	None

🔄 Batch Runner Parameters (run_experiments.py only)

These parameters are specific to the batch runner and control how multiple experiments are executed.

Parameter	Description	Default
--parallel	Number of concurrent experiments to run	1
--experiment_prefix	Custom prefix for experiment names	None
--embedding_models	List of embedding models to test (space-separated)	See defaults in script
--datasets	List of datasets to test (space-separated)	See defaults in script
--dry_run	Preview commands without executing	False
--continue_on_error	Continue running experiments even if one fails	False

---

📁 Output Structure

Each experiment run creates a directory under output/<experiment_name>/ containing:

output/
└── <experiment_name>/
├── *_results.csv # Per-topic evaluation metrics
├── *_raw_scores.parquet # Raw document-level scores
├── tensors/ # Query trajectory tensors (if enabled)
└── config.json # Experiment configuration metadata

---

⏱️ Runtime Estimates

The total runtime of the experiments consists of two main components:

1. Document embeddings — Embedding all documents in the corpus. In a deployment environment this would typically be done offline.
2. Per-query computations — Query embedding, refinement optimization, and LLM teacher feedback.

Per-query latency with a GPU is well under a second per query, as shown in the paper. This means just a few minutes to run all the queries in a dataset, assuming an efficient model endpoint for obtaining the LLM feedback scores.

Thus, much of the experiment runtime is devoted to the one-time cost of computing the corpus document embeddings. For convenience, it is possible to run this initial step separately using the script embed_all_documents.py.

Runtime estimates on a single A100-80GB GPU:

• Using a small embedding model, like Qwen/Qwen3-Embedding-0.6B: Running the full experiment on all datasets, including embedding all corpus documents and optimizing all the queries, should take about an hour.

• Using larger embedding models, like Qwen/Qwen3-Embedding-8B:

  • Embedding corpus documents with 7B/8B models can range from a few minutes for small datasets (e.g., ArgKP-21) to 3-4 hours for larger datasets like RealScholarQuery or FollowIR.
  • Note that for datasets with long documents (e.g., FollowIR), it may be necessary to use a small --embedder_batch_size to avoid running out of GPU memory.

---

📝 Notes

💾 Caching: Embeddings, reranker outputs, and generated texts are cached under /cache to avoid repeated computation.

🔬 Research Focus: This implementation is optimized for research and experimentation. For production deployment, consider replacing the file-system cache with a scalable vector-store solution.

---

📚 Citation

If you use this code in your research, please cite our paper:

@article{gera2026taskadaptive,
title={Task-Adaptive Embedding Refinement via Test-time LLM Guidance},
author={Gera, Ariel and Ashury-Tahan, Shir and Bloch, Gal and Eytan, Ohad and Toledo, Assaf},
year={2026},
journal={arXiv:2605.12487},
url={https://arxiv.org/abs/2605.12487},
}