MaxMin-RLHF on TL;DR
MaxMin-RLHF (Chakraborty et al., ICML 2024), Algorithm 1 with the panel's objectives as the known groups: three alternating rounds of 100 updates, each on the objective with the lowest reference-standardised utility. This revision replaces an earlier upload: the checkpoint was rebuilt after a fix (scale-free t=0 tie-break for MaxMin-RLHF; a precompute artifact that the current integrity check rejects, for NBPO).
- Panel: TL;DR
- Objectives: coverage, faithfulness, conciseness, helpfulness
- Backbone / reference policy:
meta-llama/Llama-3.1-8B-Instruct - Training budget: 300 optimiser updates, global batch 16
- Reported in: Nash Bargaining Preference Optimization (NBPO), Table 2 (primary cross-method evaluation)
Evaluation protocol: independent objective-wise win rate against the common reference,
judged by Llama-3.3-70B-Instruct on prompt-disjoint held-out prompts, both presentation orders.