Lu et al.: feedback-driven discovery of preference-optimization objectives

What changed

Iteratively proposed and evaluated preference-training loss code with an LLM, yielding a logistic/exponential blending objective and testing transfer to held-out tasks.

Scientific scope

Iteratively proposed and evaluated preference-training loss code with an LLM, yielding a logistic/exponential blending objective and testing transfer to held-out tasks. DPO, preference learning and programmatic search are inherited. Acquired contribution is objective-discovery framework plus scoped DiscoPOP loss, not evidence that human preferences or safety are universally optimized. Numerical results and generated-code novelty were not independently audited.

Evidence

The linked original sources describe the stated contribution and its evidence limits. WRS has not independently rerun the experiments or certified a proof. DPO, preference learning and programmatic search are inherited. Acquired contribution is objective-discovery framework plus scoped DiscoPOP loss, not evidence that human preferences or safety are universally optimized. Numerical results and generated-code novelty were not independently audited.

Contributors

• Source-listed authors and technical contributors · SCIENCE: Source-listed authors or explicitly credited technical contributors for the cited versions. Family entries retain source-specific attribution; a combined list is not one original-paper byline or equal individual work. Names alone do not establish resolved identities.
• GPT-4 / GPT-4 · EXECUTION: Generated candidate preference-loss code inside a human-designed search, training and evaluation loop; exact service snapshot is not specified by this record.

• Discovering Preference Optimization Algorithms with and for Large Language Models
• Lu et al.: feedback-driven discovery of preference-optimization objectives — author/institutional project

Provisional WRS assessment, edition 2026-10-16.1. Independent expert calibration is pending. This grade does not certify correctness or assign ownership.

Full scientific record · Browse contribution grades · Request a correction