English (United Kingdom)

https://curated-unify.zendy.io/wp-json/zendy-region/v1/featured_content/oa?rat=en

https://curated-unify.zendy.io/wp-json/zendy-region/v1/highlighted_journal/

Global: Zendy Plus annual

Presents the access of premium content as premium feature

Premium Content

Presents the keyphrase highlighting as premium feature

Keyphrase Highlighting

Presents the summarisation as premium feature

Summarisation

Insights

Presents the pdf analysis as premium feature

PDF Analysis

Presents the zaia usage as premium feature

ZAIA

Global: Zendy Plus monthly

Global: Zendy Tools annual

Global: Zendy Tools monthly

Global: Zendy Free Plan

Standard practice within Reinforcement Learning from Human Feedback (RLHF)involves optimizing against a Reward Model (RM), which itself is trained toreflect human preferences for desirable generations. A notable subject that isunderstudied is the (in-)consistency of RMs -- whether they can recognize thesemantic changes to different prompts and appropriately adapt their rewardassignments -- and their impact on the downstream RLHF model.  In this paper, we visit a series of research questions relevant to RMinconsistency: (1) How can we measure the consistency of reward models? (2) Howconsistent are the existing RMs and how can we improve them? (3) In what waysdoes reward inconsistency influence the chatbots resulting from the RLHF modeltraining?  We propose Contrast Instructions -- a benchmarking strategy for theconsistency of RM. Each example in Contrast Instructions features a pair oflexically similar instructions with different ground truth responses. Aconsistent RM is expected to rank the corresponding instruction and responsehigher than other combinations. We observe that current RMs trained with thestandard ranking objective fail miserably on Contrast Instructions compared toaverage humans. To show that RM consistency can be improved efficiently withoutusing extra training budget, we propose two techniques ConvexDA andRewardFusion, which enhance reward consistency through extrapolation during theRM training and inference stage, respectively. We show that RLHF models trainedwith a more consistent RM yield more useful responses, suggesting that rewardinconsistency exhibits a trickle-down effect on the downstream RLHF process.

The Trickle-down Impact of Reward (In-)consistency on RLHF