CS Paper Radar

Many Right Answers, One Training Example: Removing Responder Noise from GRPO for LLM Recommenders

Xin Li, Hao Jiang, Jinghao Guo, Weisi Lin, Chau Yuen

cs.CL ·submitted 2026-10-13 05:57 UTC · v1
positioncode/data

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.

Comment
Project Page: https://lixin.ai/ReMaRk/

Closest papers

Nearest neighbours by meaning across the whole index.

Better Rankers, Worse Retrievers: The Hidden Cost of RL Post-Training in Generative Recommendation

cs.IR cs.AI ·2026-11-17 ·Yuheng Chen, Fan Zhang, Yixuan Xu, Chao Yang, Yang Xiang +1 more
similarity 0.74
Abstract

Generative recommenders decode the semantic ID of a user's next item with a language model, and post-training them with reinforcement learning (RL) after supervised fine-tuning (SFT) has become standard practice. The gains are mostly reported on the first ten results, yet in deployment such models hand hundreds of candidates to a downstream ranker or reward model. We show that RL post-training has a hidden cost: it improves accuracy at the top of the list (hit@K for small K) but, on all five datasets we study, lowers the recall of the candidate list (hit@K for K in the hundreds). The model becomes a better ranker and a worse retriever. Controlled experiments trace the loss to the negative advantages that RL assigns to the candidates it samples, the rollout. The root cause is that only one next item is logged per user, so RL treats every other candidate as wrong. Yet these candidates are the model's most plausible alternatives, and often the next items of other users with similar histories. Penalising them pushes down exactly what similar users choose next, and these items drop out of those users' candidate lists. This diagnosis points directly to a remedy: out-of-rollout anchoring, a single additional loss term that preserves the SFT model's ordering of the items outside the rollout. It keeps the gain at the top and restores recall to the SFT level on all five datasets, and it is effective under other RL configurations as well. To our knowledge, this is the first systematic study of this recall loss: we measure it, identify its cause, and show that, once the cause is known, a simple remedy suffices.

0.14

Are Dense Rewards Worth Their Tokens in GRPO?

cs.CL cs.AI ·2026-11-10 ·Vikash Singh, Weicong Chen, Debargha Ganguly, Yanyan Zhang, Mohsen Hariri +14 more
positioncode/data similarity 0.69
Abstract · 59 pages (10 pages of main text), 18 figures, 27 tables. Code: https://github.com/vicky157/grpo-validation

RLVR trains language models to reason by rewarding correct final answers. A widely used algorithm, group-relative policy optimization (GRPO), gives every token of a solution the same credit, so dense process rewards have been proposed to credit intermediate steps, for example with a Monte-Carlo reward that scores each step by sampling continuations from it. Comparing such a reward with GRPO per training update, or vetting it with a gradient test on a single batch, leaves out what it costs: in our setting, the continuations alone take 91% of all generated tokens. We introduce a matched-compute protocol that gives every training intervention the same budget, measured in generated tokens and training GPU-hours, and tests it in preregistered studies with controls for side effects. Across 1,668 runs on six models, matching compute reverses the ranking. On Qwen2.5-1.5B, the Monte-Carlo reward leads standard GRPO per update, yet ends 14.3 percentage points lower at equal tokens and 7.8 points lower at equal training time, and none of four dense-reward variants outperforms GRPO even with as much training time. A per-token gradient test orders these rewards as training does but overestimates each by one to two points, as much as the gains it is meant to detect. In contrast, training on the prompts whose predicted success is nearest 1/2 improves accuracy at matched compute, by 1.7 points on Qwen3-0.6B and 0.8 on Qwen2.5-1.5B. These gains replicate on fresh seeds and unseen problems, and a simple rule, that the candidate prompts are predicted to fail more often than not, predicted where they appear in five of six preregistered tests. Both gains come early in training, and standard GRPO with twice the initial learning rate comes within half a point of them. We argue that training interventions should be validated at matched compute, which overturns some apparent gains and confirms others.

0.28

Scaling Laws for Generative Reward Models

cs.AI ·2026-10-22 ·Angel Raychev, Nathan Lile, Rafael Rafailov
code/data similarity 0.67
Abstract · 28 pages. Code and research artifacts: https://github.com/RaychevAngel/generative-reward-models

We study generative reward models (GenRMs) as online preference providers for policy optimization in creative writing. Across Qwen3 reward-model sizes from 0.6B to 14B, Thinking judges trained with GRPO have higher static Gold-label accuracy than Answer-Only judges trained with SFT. This ranking often reverses when the judges train policies through online DPO: Answer-Only rewarders yield higher peak and top-five-average policy Elo in seven of eight matched settings under a fixed, learned 32B Gold evaluator. Intermediate Thinking-judge checkpoints can also outperform later, more statically accurate checkpoints as rewarders. Sweeps over rewarder size, policy size, training duration and DPO regularization show substantial benefits from reward-model capacity, including judges larger than the policy. We analyze trajectories with evaluator-specific Elo arenas, conditional match-resampling intervals and descriptive scaling fits. These results demonstrate an evaluator-rewarder gap in the studied pipelines: fixed-distribution judging accuracy does not determine downstream optimization quality. The comparison changes judge-training objective as well as response format, and the main ranking is conditional on its learned reference rather than a universal human-quality result. A limited follow-up using a different learned judge is inconclusive overall.

0.52

Rubric Dropout: A Simple Way to Mitigate Reward Hacking in Rubric-Based RL

cs.LG cs.AIcs.CL ·2026-08-12 ·Minglai Yang, Xinyu Guo, Zhengliang Shi, Utkarsh Tyagi, Mian Zhang +6 more
highly cited similarity 0.67
Abstract · 24 pages, 14 figures, 5 tables

Rubric-based reinforcement learning, in which an LLM judge scores responses against prompt-specific criteria, has become a standard way to post-train language models on tasks without verifiable answers. The rubric, however, is only a proxy for response quality, and a policy trained against it long enough learns to exploit the difference. We train models with GRPO on medical and science rubrics and grade their responses on out-of-distribution (OOD) benchmarks throughout training with two judges: the training judge and a stronger OOD judge. The two scores first rise together and then diverge. The policy is learning to satisfy the training judge without writing better responses, which is a form of reward hacking. To mitigate it, we propose Rubric Dropout, which applies dropout to rubric criteria: at each training step, a random subset of each rubric's criteria is removed from the reward, so the rewarded subset of criteria is resampled at each step. All responses sampled for the same prompt are scored on the same subset, so their rewards remain comparable, and evaluation always uses the full rubric. We also introduce Block Dropout, which drops semantically similar criteria together so that a near duplicate of a dropped criterion cannot keep rewarding the same behavior. Across both benchmarks and models from 1.7B to 8B, dropout improves the gold judge's score on OOD benchmarks by up to 3 percent on HealthBench-Hard and 7 percent on ResearchQA, and it reduces both the gap between the two judges and over-crediting. It keeps in-domain reward close to the baseline, and dropout rates of 30-50 percent work best. Block Dropout improves the gold judge's OOD score further at the same dropout rate. Changing which criteria are rewarded at each step is a simple, inexpensive way to make rubric rewards harder to exploit.

0.50

Reinforcement Learning with Verifiable Rewards for Small Search Agents

cs.AI cs.IR ·2026-09-23 ·Gaurisankar Jayadas, Aske Plaat, Álvaro Serra-Gómez, Sandheep P
benchmark similarity 0.67
Abstract

Reinforcement Learning with Verifiable Rewards (RLVR) performs well on problems with clear rewards, such as mathematics and coding, but whether it also works where the reward is less clear remains open. The reason-over-search recipe applies RLVR to open-domain question answering, where retrieval grounds the answer and a match against the reference supplies the reward. So far it has been demonstrated on large models, and below one billion parameters only with distillation from a larger teacher. We test the recipe on a small model. We train Qwen3.5-0.8B with Group Relative Policy Optimization (GRPO) and an interleaved Wikipedia-search tool on MuSiQue, varying only the reward across three shapes over three seeds each, and we evaluate every checkpoint held-out on a seven-benchmark question-answering suite. The recipe works: the best run reaches 0.352 average exact match against a 0.092 untrained floor, a 3.8-fold gain, with no distillation step in the training loop. The reward shape also matters. The Search-R1-faithful exact-match-only reward is the worst of the three at every seed at the matched training horizon, and it is worst even on exact match, the metric it directly optimises. We conclude that the sparse exact-match reward, RLVR's default in mathematics and code, is the wrong starting point for models of this size. The reason-over-search setting can supply a suitable reward for RLVR on small models, but small-model RLVR needs its own reward-design study rather than a scaled-down copy of a large-model recipe.

0.21

EliteReplay: Preserving Diverse Solutions in RL Post-Training via Quality-Diversity Search

cs.AI ·2026-10-15 ·Hanane Nour Moussa, Doeun Lee, Sachin Kumar
similarity 0.67
Abstract

Reinforcement learning has become a common post-training technique to improve the accuracy of LLMs on verifiable tasks. However, such training can narrow the set of solutions the model can produce. Several approaches have attempted to address this behavior-sharpening effect, but they often optimize metrics that do not reflect the diversity of correct solutions. Ideally, we want a model that can produce multiple good solutions, not only a single reliable one. We propose EliteReplay, a post-training recipe that pairs GRPO with an "elite" archive built by quality-diversity search. Before training, we use the frozen base model to collect solutions that span distinct approaches; during training, GRPO is combined with a replay loss on this archive. The result is a post-trained model that improves accuracy while preserving more of the base model's solution diversity. On multiple text-to-SQL tasks, EliteReplay achieves a higher pass@128 than all baselines, producing up to 77% more distinct correct strategies than the strongest baselines. It preserves most of the base model's correct strategies, outperforming existing methods in strategy retention. This diversity translates into practical benefits on ambiguous questions, where EliteReplay covers both interpretations nearly twice as often as the strongest baseline on AmbiQT.

0.21

Flat-RLVR: Sharpness-Aware Reinforcement Learning with Verifiable Rewards for Reasoning Capacity in LLMs

cs.LG cs.AI ·2026-10-21 ·Yeonbum Kim, Hyun Kyu Lee, Sung Whan Yoon
similarity 0.66
Abstract · 9 pages, 8 figures

Reinforcement learning with verifiable rewards (RLVR), particularly through group-relative policy optimization (GRPO), has emerged as a central paradigm for the post-training of large language models (LLMs). While RLVR is highly effective at enhancing the reasoning capabilities of LLMs, it commonly induces \textit{distribution sharpening}, whereby the policy distribution becomes excessively peaked and rare-yet-correct trajectories are progressively suppressed, thus diminishing sample diversity. To this end, we promote flat-reward optima in RLVR to smooth the policy distribution. We theoretically link flat reward RLVR to reduced distribution sharpening, and empirically show improved sample diversity while maintaining single-attempt accuracy. Our method, called \texttt{Flat}-RLVR, searches for flat reward regions via sharpness-aware minimization during LLM post-training, combining this with GRPO variants to achieve better trade-offs between sampling efficiency and reasoning capacity.

0.08

Expert-Guided Rubrics as Dense Reward Signals for Reinforcement Learning in Non-Verifiable Domains

cs.LG cs.AIcs.IR ·2026-10-17 ·Ashish Bansal, Hossam Amer, Keerat Kaur Guliani, Parinita Edke, Lesley Matheson +1 more
similarity 0.65
Abstract · 11 pages, 4 figures, RL4LLM-Agents workshop ICAIF ACM

Large language models (LLMs) are increasingly deployed as autonomous agents in non-verifiable domains such as financial services, where they must interpret user intent, rewrite queries, and retrieve information under domain-specific constraints. Although reinforcement learning (RL) can train such agents from the consequences of their actions, recent successes largely rely on verifiable tasks with automatically computed binary rewards. In non-verifiable retrieval tasks, assessing query quality requires scarce subject-matter expert (SME) judgment that is inherently graded rather than binary. With Group Relative Policy Optimization (GRPO), sparse rewards yield no useful learning signal when all sampled outputs within a group receive the same outcome, causing optimization to stall. We introduce a method that converts a small, fixed investment of SME judgment into a dense reward and combines it with retrieval-based metrics to form a hybrid, GRPO-compatible training signal. Experiments show improvements of up to 6\% in Recall@3 and 6.7\% in MRR@3 compared with sparse-reward baselines.

0.21