analiza aiqbench
This paper introduces ReMaRk, a reinforcement learning method for LLM recommenders that eliminates "responder noise" by jointly scoring generated lists against the union of all valid human replies rather than selecting a single example. By training on the request itself instead of individual answers, this approach significantly reduces computational costs while maintaining or improving recall across multiple backbones and datasets. The work establishes a practical guideline for optimizing RL efficiency in scenarios where multiple correct human responses exist for a single query.
Many Right Answers, One Training Example: Removing Responder Noise from GRPO for LLM Recommenders
arXiv:2610.17771v1 Announce Type: new Abstract: A recommendation request on Reddit draws several replies, each naming different titles. When an LLM recommender is trained on such threads with reinforcement learning (RL), which replies should a generated list be rewarded against? The state-of-the-art recipe, Rank-GRPO, picks one reply per training example, so the advantage mostly records which reply was drawn (85% of its variance) rather than how good the list is. We argue that RL should train on the request, not on each answer, and propose ReMaRk: it scores every list once against the union of all replies' titles and samples each request equally. This makes RL cheaper. On seven backbones and three datasets, ReMaRk reaches the validation recall of 2,000 per-reply RL steps in 25 to 53% of the steps and leads at that budget; trained until neither improves, it is never significantly behind Rank-GRPO and in four of seven pairs reaches Rank-GRPO's best with fewer GPU-hours. The same reward helps on ProtoQA, where about 100 people answer each question, which points to a simple rule for RL with several valid human answers: reward them jointly.
Abstract:A recommendation request on Reddit draws several replies, each naming different titles. When an LLM recommender is trained on such threads with reinforcement learning (RL), which replies should a generated list be rewarded against? The state-of-the-art recipe, Rank-GRPO, picks one reply per training example, so the advantage mostly records which reply was drawn (85% of its variance) rather than how good the list is. We argue that RL should train on the request, not on each answer, and propose ReMaRk: it scores every list once against the union of all replies' titles and samples each request equally. This makes RL cheaper. On seven backbones and three datasets, ReMaRk reaches the validation recall of 2,000 per-reply RL steps in 25 to 53% of the steps and leads at that budget; trained until neither improves, it is never significantly behind Rank-GRPO and in four of seven pairs reaches Rank-GRPO's best with fewer GPU-hours. The same reward helps on ProtoQA, where about 100 people answer each question, which points to a simple rule for RL with several valid human answers: reward them jointly.
Computation and Language (cs.CL)
arXiv:2610.17771 [cs.CL]
(or arXiv:2610.17771v1 [cs.CL] for this version)
https://doi.org/10.48550/arXiv.2610.17771 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Text integral sub licența deschisă a publicației · 275 cuvinte
