REVIEW 11 cited by
What is the Alignment Objective of GRPO?
Not yet reviewed by Pith; the record is open.
This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.
SPECIMEN: schema-true, not a live event
T0 review · schema-true
One-sentence machine reading of the paper's core claim.
pith:XXXXXXXX · record.json · timestamp
read the original abstract
In this note, we examine the aggregation of preferences achieved by the Group Policy Optimisation (GRPO) algorithm, a reinforcement learning method used to train advanced artificial intelligence models such as DeepSeek-R1-Zero and DeepSeekMath. The GRPO algorithm trains a policy using a reward preference model, which is computed by sampling a set of outputs for a given context, observing the corresponding rewards, and applying shift-and-scale normalisation to these reward values. Additionally, it incorporates a penalty function to discourage deviations from a reference policy. We present a framework that enables us to characterise the stationary policies of the GRPO algorithm. This analysis reveals that the aggregation of preferences differs fundamentally from standard logarithmic pooling, which is implemented by other approaches such as RLHF. The precise form of preference aggregation arises from the way the reward preference model is defined and from the penalty function, which we show to essentially correspond to the reverse Kullback-Leibler (KL) divergence between the aggregation policy and the reference policy. Interestingly, we demonstrate that for groups of size two, the reward preference model corresponds to pairwise comparison preferences, similar to those in other alignment methods based on pairwise comparison feedback. We provide explicit characterisations of the aggregate preference for binary questions, for groups of size two, and in the limit of large group size. This provides insights into the dependence of the aggregate preference on parameters such as the regularisation constant and the confidence margin of question answers. Finally, we discuss the aggregation of preferences obtained by modifying the GRPO algorithm to use direct KL divergence as the penalty or to use rewards without scale normalisation.
Forward citations
Cited by 11 Pith papers
-
Predictable GRPO: A Closed-Form Model of Training Dynamics
GRPO updates reduce to a damped oscillator whose mass, damping, and stiffness are fixed by optimizer hyperparameters plus one measured curvature scale, subsuming single-exponential saturation while adding inertial slo...
-
Predictable GRPO: A Closed-Form Model of Training Dynamics
A closed-form inertial model of GRPO dynamics that subsumes single-exponential saturation as its overdamped limit and predicts group-size invariance, stability thresholds, and overdamped-to-oscillatory transitions.
-
Listwise Policy Optimization: Group-based RLVR as Target-Projection on the LLM Response Simplex
Listwise Policy Optimization explicitly performs target-projection on the LLM response simplex, unifying and improving group-based RLVR methods with monotonic improvement and flexible divergences.
-
Listwise Policy Optimization: Group-based RLVR as Target-Projection on the LLM Response Simplex
LPO reframes group-based RLVR as explicit target-projection on the LLM response simplex and performs exact divergence minimization to achieve monotonic listwise improvement with bounded gradients.
-
V-tableR1: Process-Supervised Multimodal Table Reasoning with Critic-Guided Policy Optimization
V-tableR1 uses a critic VLM for dense step-level feedback and a new PGPO algorithm to shift multimodal table reasoning from pattern matching to verifiable logical steps, achieving SOTA accuracy with a 4B open-source model.
-
f-GRPO and Beyond: Divergence-Based Reinforcement Learning Algorithms for General LLM Alignment
f-GRPO and f-HAL estimate f-divergences between reward-aligned and reward-unaligned response distributions and prove expected reward improvement for general LLM alignment.
-
Small Agent Group is the Future of Digital Health
A ten-agent group of 3B-4B LLMs with debate, retrieval, and role specialization beats 70B-72B single LLMs on clinical benchmarks while using less memory.
-
Large Language Models Reasoning Abilities Under Non-Ideal Conditions After RL-Fine-Tuning
RL fine-tuning of LLMs improves clean-benchmark accuracy while degrading accuracy under three injected-distractor evaluation scenarios, though one of the three scenarios contradicts the headline claim.
-
Inference-Time Alignment Control for Diffusion Models with Reinforcement Learning Guidance
Blending a base diffusion model with its RL-finetuned version at sampling time lets users dial alignment strength, with the blend weight corresponding to the KL-regularization coefficient beta/w.
-
LaViPlan : Language-Guided Visual Path Planning with RLVR
LaViPlan uses RLVR with GRPO and ADE/FDE rewards to fine-tune a 2B VLM for trajectory prediction, improving ADE/FDE on ROADWork and a normalized safety score on CODA-LM over supervised fine-tuning.
-
Alignment and Safety in Large Language Models: Safety Mechanisms, Training Paradigms, and Emerging Challenges
A broad survey of LLM alignment that catalogs objectives, benchmarks, SFT/RLHF/DPO methods, and safety challenges, without contributing new experimental or theoretical results.
Discussion (0). Sign in to comment.