Sharing is Caring: Efficient LM Post-Training with Collective RL Experience Sharing

• URL: http://arxiv.org/abs/2509.08721v1
• Date: Wed, 10 Sep 2025 16:14:20 GMT
• Title: Sharing is Caring: Efficient LM Post-Training with Collective RL Experience Sharing
• Abstract summary: Post-training language models (LMs) with reinforcement learning (RL) can enhance their reasoning capabilities without supervised fine-tuning.<n>We present Swarm sAmpling Policy Optimization (SAPO), a fully decentralized and asynchronous RL post-training algorithm.<n>We show SAPO achieved cumulative reward gains of up to 94% in controlled experiments.
• Score: 1.5280055693761307
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Post-training language models (LMs) with reinforcement learning (RL) can enhance their complex reasoning capabilities without supervised fine-tuning, as demonstrated by DeepSeek-R1-Zero. However, effectively utilizing RL for LMs requires significant parallelization to scale-up inference, which introduces non-trivial technical challenges (e.g. latency, memory, and reliability) alongside ever-growing financial costs. We present Swarm sAmpling Policy Optimization (SAPO), a fully decentralized and asynchronous RL post-training algorithm. SAPO is designed for decentralized networks of heterogenous compute nodes, where each node manages its own policy model(s) while "sharing" rollouts with others in the network; no explicit assumptions about latency, model homogeneity, or hardware are required and nodes can operate in silo if desired. As a result, the algorithm avoids common bottlenecks in scaling RL post-training while also allowing (and even encouraging) new possibilities. By sampling rollouts "shared" across the network, it enables "Aha moments" to propagate, thereby bootstrapping the learning process. In this paper we show SAPO achieved cumulative reward gains of up to 94% in controlled experiments. We also share insights from tests on a network with thousands of nodes contributed by Gensyn community members running the algorithm on diverse hardware and models during an open-source demo.

Related papers

• AstraFlow: Dataflow-Oriented Reinforcement Learning for Agentic LLMs [57.86040075371121]
  Reinforcement learning (RL) is increasingly used to improve the reasoning, coding, and tool-use capabilities of large language models.<n>Agentic RL remains prohibitively expensive.<n>We propose AstraFlow, a dataflow-oriented RL system that replaces conventional trainer-centered control with principled component abstractions.
  arXiv  Detail & Related papers  (2026-05-15T03:13:35Z)
• AcceRL: A Distributed Asynchronous Reinforcement Learning and World Model Framework for Vision-Language-Action Models [21.204182133307047]
  We propose AcceRL, a fully asynchronous and decoupled RL framework.<n>AcceRL is the first to integrate a plug-and-play, trainable world model into a distributed asynchronous RL pipeline.<n> Experiments on the LIBEROciteliu2023libero benchmark demonstrate that AcceRL achieves state-of-the-art (SOTA) performance.
  arXiv  Detail & Related papers  (2026-03-19T03:50:45Z)
• Online Training and Pruning of Deep Reinforcement Learning Networks [0.0]
  Scaling deep neural networks (NN) of reinforcement learning (RL) algorithms has been shown to enhance performance when feature extraction networks are used.<n>We propose an approach to integrate simultaneous training and pruning within advanced RL methods.
  arXiv  Detail & Related papers  (2025-07-16T07:17:41Z)
• RL for Reasoning by Adaptively Revealing Rationales [36.50924054394857]
  Supervised fine-tuning (SFT) relies on dense ground-truth labels, which become increasingly costly as sequence length grows.<n>We address this by adaptive backtracking (AdaBack), a per-sample curriculum learning algorithm that reveals only a partial prefix of the target output during training.<n>We show that our adaptive curriculum over partial answers reliably solves problems that are otherwise intractable.
  arXiv  Detail & Related papers  (2025-06-22T17:46:14Z)
• StreamRL: Scalable, Heterogeneous, and Elastic RL for LLMs with Disaggregated Stream Generation [55.75008325187133]
  Reinforcement learning (RL) has become the core post-training technique for large language models (LLMs)<n>StreamRL is designed with disaggregation from first principles to address two types of performance bottlenecks.<n> Experiments show that StreamRL improves throughput by up to 2.66x compared to existing state-of-the-art systems.
  arXiv  Detail & Related papers  (2025-04-22T14:19:06Z)
• FusionLLM: A Decentralized LLM Training System on Geo-distributed GPUs with Adaptive Compression [55.992528247880685]
  Decentralized training faces significant challenges regarding system design and efficiency. We present FusionLLM, a decentralized training system designed and implemented for training large deep neural networks (DNNs) We show that our system and method can achieve 1.45 - 9.39x speedup compared to baseline methods while ensuring convergence.
  arXiv  Detail & Related papers  (2024-10-16T16:13:19Z)
• Efficient Parallel Split Learning over Resource-constrained Wireless Edge Networks [44.37047471448793]
  In this paper, we advocate the integration of edge computing paradigm and parallel split learning (PSL) We propose an innovative PSL framework, namely, efficient parallel split learning (EPSL) to accelerate model training. We show that the proposed EPSL framework significantly decreases the training latency needed to achieve a target accuracy.
  arXiv  Detail & Related papers  (2023-03-26T16:09:48Z)
• Federated Deep Reinforcement Learning for the Distributed Control of NextG Wireless Networks [16.12495409295754]
  Next Generation (NextG) networks are expected to support demanding internet tactile applications such as augmented reality and connected autonomous vehicles. Data-driven approaches can improve the ability of the network to adapt to the current operating conditions. Deep RL (DRL) has been shown to achieve good performance even in complex environments.
  arXiv  Detail & Related papers  (2021-12-07T03:13:20Z)
• Adaptive Stochastic ADMM for Decentralized Reinforcement Learning in Edge Industrial IoT [106.83952081124195]
  Reinforcement learning (RL) has been widely investigated and shown to be a promising solution for decision-making and optimal control processes. We propose an adaptive ADMM (asI-ADMM) algorithm and apply it to decentralized RL with edge-computing-empowered IIoT networks. Experiment results show that our proposed algorithms outperform the state of the art in terms of communication costs and scalability, and can well adapt to complex IoT environments.
  arXiv  Detail & Related papers  (2021-06-30T16:49:07Z)
• A Heuristically Assisted Deep Reinforcement Learning Approach for Network Slice Placement [0.7885276250519428]
  We introduce a hybrid placement solution based on Deep Reinforcement Learning (DRL) and a dedicated optimization based on the Power of Two Choices principle. The proposed Heuristically-Assisted DRL (HA-DRL) allows to accelerate the learning process and gain in resource usage when compared against other state-of-the-art approaches.
  arXiv  Detail & Related papers  (2021-05-14T10:04:17Z)
• Unsupervised Learning for Asynchronous Resource Allocation in Ad-hoc Wireless Networks [122.42812336946756]
  We design an unsupervised learning method based on Aggregation Graph Neural Networks (Agg-GNNs) We capture the asynchrony by modeling the activation pattern as a characteristic of each node and train a policy-based resource allocation method.
  arXiv  Detail & Related papers  (2020-11-05T03:38:36Z)
• Learning Dexterous Manipulation from Suboptimal Experts [69.8017067648129]
  Relative Entropy Q-Learning (REQ) is a simple policy algorithm that combines ideas from successful offline and conventional RL algorithms. We show how REQ is also effective for general off-policy RL, offline RL, and RL from demonstrations.
  arXiv  Detail & Related papers  (2020-10-16T18:48:49Z)
• A Low Complexity Decentralized Neural Net with Centralized Equivalence using Layer-wise Learning [49.15799302636519]
  We design a low complexity decentralized learning algorithm to train a recently proposed large neural network in distributed processing nodes (workers) In our setup, the training data is distributed among the workers but is not shared in the training process due to privacy and security concerns. We show that it is possible to achieve equivalent learning performance as if the data is available in a single place.
  arXiv  Detail & Related papers  (2020-09-29T13:08:12Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.