Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].

Grounding by Trying: LLMs with Reinforcement Learning-Enhanced Retrieval

Authors: Sheryl Hsu, Omar Khattab, Chelsea Finn, Archit Sharma

ICLR 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Le Re T can improve the absolute retrieval accuracy by up to 29% and the downstream generator evaluations by 17%. The simplicity and flexibility of Le Re T allows it to be applied to arbitrary off-the-shelf retrievers and makes it a promising technique for improving general LLM pipelines. Project website: http://sherylhsu.com/Le Re T/.
Researcher Affiliation Collaboration 1Stanford University,2Databricks,3Physical Intelligence,4Google Deep Mind EMAIL
Pseudocode Yes Algorithm 1 Prompt Driven Diverse Sampling + Training
Open Source Code No Project website: http://sherylhsu.com/Le Re T/.
Open Datasets Yes We test Le Re T on Hotpot QA (Yang et al., 2018) and Ho Ver (Jiang et al., 2020).
Dataset Splits Yes We test Le Re T on Hotpot QA (Yang et al., 2018) and Ho Ver (Jiang et al., 2020). Both datasets are based on a Wikipedia knowledge base and are multi-hop, meaning that models must reason across multiple articles to arrive at the correct answer. The datasets provide both the correct answer and supporting articles.
Hardware Specification No The paper does not provide specific hardware details such as GPU/CPU models, memory, or cloud instance types used for running the experiments.
Software Dependencies No We implement our sampling pipeline on top of DSPy (Khattab et al., 2023), specifically defining a single hop as a program and sampling data using the evaluate functions.
Experiment Setup Yes We use a learning rate of 1e-7 for SFT/context distillation in all our experiments, and use a τ = 0.05 and learning rate of 1e-7. We train SFT for 1 epoch, and we only distill the best performing prompt. We train IPO for 2 epochs.