pith. sign in

Title resolution pending

characterized this empirically in RLHF as reward over-optimization, showing that as the policy spends KL budget against a learned RM, the gold reward traces a hill-shaped curve tha

1 Pith paper cite this work. Polarity classification is still indexing.

1 Pith paper citing it
browse 1 citing papers

Title metadata for this work has not finished resolving. The hub is built from the citation graph; the title resolver retries DOI and OpenAlex on its next pass.

fields

cs.LG 1

years

2026 1

verdicts

UNVERDICTED 1

representative citing papers

General Preference Reinforcement Learning

cs.LG · 2026-05-18 · unverdicted · novelty 6.0 · 2 refs

GPRL carries a k-dimensional skew-symmetric preference structure into policy updates with per-dimension advantages and a drift monitor, yielding 56.51% length-controlled win rate on AlpacaEval 2.0 from Llama-3-8B-Instruct while outperforming SimPO and SPPO on other benchmarks.

citing papers explorer

Showing 1 of 1 citing paper.

  • General Preference Reinforcement Learning cs.LG · 2026-05-18 · unverdicted · none · ref 54 · 2 links

    GPRL carries a k-dimensional skew-symmetric preference structure into policy updates with per-dimension advantages and a drift monitor, yielding 56.51% length-controlled win rate on AlpacaEval 2.0 from Llama-3-8B-Instruct while outperforming SimPO and SPPO on other benchmarks.