Cai et al.: InternLM2 conditional online preference training

What changed

Documented a long-context language-model training pipeline and conditional online RLHF designed to handle conflicting preferences and reward hacking, releasing intermediate model stages.

Scientific scope

Documented a long-context language-model training pipeline and conditional online RLHF designed to handle conflicting preferences and reward hacking, releasing intermediate model stages. The source reports empirical mitigation, not elimination of reward hacking or universal preference compatibility. Long-context retrieval success is not broad long-document reasoning reliability. Model sizes and training stages remain one contribution.

Evidence

The linked original sources describe the stated contribution and its evidence limits. WRS has not independently rerun the experiments or certified a proof. The source reports empirical mitigation, not elimination of reward hacking or universal preference compatibility. Long-context retrieval success is not broad long-document reasoning reliability. Model sizes and training stages remain one contribution.

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.

• InternLM2 Technical Report
• Cai et al.: InternLM2 conditional online preference training — author/company record

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