Neural Interactive Proofs
Authors: Lewis Hammond, Sam Adam-Day
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Finally, we support this theory with experiments in two domains: a toy graph isomorphism problem that illustrates the key ideas, and a code validation task using large language models. In so doing, we aim to create a foundation for future work on neural interactive proofs and their application in building safer AI systems. |
| Researcher Affiliation | Academia | Lewis Hammond EMAIL Sam Adam-Day EMAIL Department of Computer Science, University of Oxford, Oxford, United Kingdom |
| Pseudocode | No | The paper describes theoretical frameworks and experimental procedures but does not include any clearly labeled pseudocode or algorithm blocks. Procedural descriptions are provided in narrative text. |
| Open Source Code | Yes | (iv) a well-documented codebase for testing different protocols in different domains, available at https://github.com/Sam Adam Day/neural-interactive-proofs. |
| Open Datasets | Yes | Our second experiment involves a much more complex problem: checking that a given Python program satisfies a natural language specification. In particular, we make use of the Automated Programming Progress Standard (APPS) dataset (Hendrycks et al., 2021) |
| Dataset Splits | Yes | The train-test split is 80:20. The train-test split of the eventual dataset is 90:10. |
| Hardware Specification | No | The paper mentions using GPT-4o and GPT-4o-mini models and the Open AI fine-tuning API, but it does not specify the underlying hardware (e.g., GPU/CPU models, memory) used for running the experiments or training these models. |
| Software Dependencies | No | The paper mentions using algorithms like independent PPO and expert iteration, and models like GPT-4o and GPT-4o-mini via the OpenAI API. However, it does not specify version numbers for any key software components, libraries, or APIs used for implementation. |
| Experiment Setup | Yes | We use a clipped objective with value ϵ = 0.2, with hyperparameters γ = 0.95 and λ = 0.95. We additionally use advantage normalisation and entropy regularisation with coefficient 0.001. The learning rate is 0.003. For each protocol we train across 10 seeds for 5,000 steps. We train both provers and verifiers via the Open AI fine-tuning API using expert iteration for eight rounds (Anthony et al., 2017). This works by fine-tuning the models in each round on the rollouts on which they received positive reward. We use 10% of the underlying dataset at a time, iteratively adding positive examples to the fine-tuning dataset. Following Kirchner et al. (2024), we fine-tune each model from scratch in each iteration. |