Efficient Exploration for LLMs

Authors: Vikranth Dwaracherla, Seyed Mohammad Asghari, Botao Hao, Benjamin Van Roy

ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We present evidence of substantial benefit from efficient exploration in gathering human feedback to improve large language models. In our experiments, an agent sequentially generates queries while fitting a reward model to the feedback received.
Researcher Affiliation Collaboration 1Google DeepMind 2Stanford University.
Pseudocode Yes Algorithm 1 learning interface
Open Source Code No The paper references third-party tools and libraries used (e.g., 'enn library'), but does not explicitly state that its own source code is released or provide a link to it.
Open Datasets Yes Each prompt is sampled uniformly from the Anthropic Helpfulness Base train dataset.
Dataset Splits No The paper mentions using 'Anthropic Helpfulness Base train dataset' and 'Anthropic Helpfulness Base eval dataset' but does not explicitly describe a distinct 'validation' split or provide specific train/validation/test split percentages or sample counts.
Hardware Specification No The paper does not provide specific details about the hardware (e.g., GPU/CPU models, memory) used for running its experiments.
Software Dependencies No The paper mentions using 'ADAM' for optimization and the 'enn library', but it does not specify version numbers for these or other software dependencies.
Experiment Setup Yes at the start of each epoch of interaction, each agents receives a batch of B = 32 prompts... The replay buffers are first-in-first-out (FIFO) buffer, with a maximum capacity of C = 3200 data points. In our experiments, we set N = 100.