AutoOR is a scalable synthetic data generation and reinforcement learning pipeline that trains large language models to autoformalize operations research problems expressed in natural language across linear, mixed‑integer, and non‑linear categories. By generating verified training data from standard optimization forms and using solver execution feedback as a reward signal, AutoOR enables post‑training of an 8B model to achieve state‑of‑the‑art or competitive results on six established OR benchmarks, matching significantly larger frontier models. For non‑linear problems involving physical dynamics, a curriculum RL strategy bootstraps from limited initial data, making this class tractable for post‑training.
By Sumeet Ramesh Motwani, Chuan Du, Aleksander Petrov, Christopher Davis, Philip Torr, Antonio Papania-Davis, Weishi Yan
The paper studies optimal placement of millimeter-wave base stations in a realistic, non-convex campus layout using deep reinforcement learning. It compares four DRL methods—single-agent DQN, multi-agent partitioned DQN, single-agent DDPG, and multi-agent partitioned DDPG—and finds that the multi-agent DDPG approach achieves full coverage, a Jain's fairness index of 0.94, and superior performance in dense scenarios with 400 users. The multi-agent DDPG also converges more efficiently than single-agent methods.
By Omar Rady, Mohamed Ayman, Ali Arafa, Mohamed Shalma
The paper introduces Continual Reasoning Gym, a continual reinforcement learning with verifiable rewards (RLVR) environment that sequences text and visual reasoning tasks. It finds that sequential RLVR suffers modest forgetting and underperforms multitask RLVR, but that shared reasoning structures can be leveraged. The authors propose Continual Prompt Replay (CPR), which replays previous-task prompts and regenerates responses, achieving performance comparable to multitask RLVR.
By Lirui Luo, Guoxi Zhang, Hongming Xu, Rongqing Li, Cong Fang, Lifeng Fan
Vector Symbolic Policy Gradient (VSPG) is a discrete-action actor that encodes each action as a unit‑norm hypervector and evaluates it by similarity to the encoded state. Its policy‑gradient update reduces to advantage‑weighted hypervector bundling followed by normalization, enabling the use of standard advantage estimators. The learned action hypervectors act as fixed‑size compressed kernel memories that store advantage‑weighted expansions over visited states, allowing evidence transfer via encoder‑induced similarity and providing a robustness guarantee for greedy action selection under random bit flips.
By Ryozo Masukawa, Sanggeon Yun, SungHeon Jeong, Hyunwoo Oh, Raheeb Hassan, Pietro Mercati, Nathaniel D. Bastian, Mahdi Imani, Mohsen Imani
ADEPT is a reinforcement‑learning framework that first pre‑trains a dexterous policy on a generic object reposing task and then post‑trains downstream policies using this pretrained behavior as a prior. The approach avoids relearning basic skills for each new task, and employs a stable post‑training recipe—behavior‑cloning distillation, critic warm‑up, and conservative on‑policy updates—to preserve the pretrained capabilities. ADEPT’s joint‑space Geometric Fabric mediates between the policy and the robot, enabling zero‑shot sim‑to‑real transfer on a 23‑DoF Kuka‑Allegro and a 29‑DoF Flexiv‑Sharpa, where the robots solve long‑horizon tasks from challenging initial states at human‑level speed.
By Jayjun Lee, Jessica Yin, Asif Rana, Nicholas Blauch, Sam Mady, Mohak Bhardwaj, Nima Fazeli, Nathan Ratliff, Karl Van Wyk, Ankur Handa
MLREF introduces a Module Level Reward Evolution Framework that uses a persistent pool of reusable reward components to improve reward design in reinforcement learning. The framework constructs reward functions as linear combinations of modules from this pool and evolves the pool through reflection-based refinement, hybrid credit assignment, and a merge strategy with rollback. Experiments on 17 tasks demonstrate that MLREF outperforms strong baselines, achieving a 25.2% improvement in locomotion and 6.6% in manipulation with more stable optimization dynamics.
By Chenglin Liu, Xun Wang, Ruishuo Chen, Zhuoran Li, Longbo Huang
SkillGate is a method that trains agents to select the correct skill from a large slate during an episode by separating credit signals for skill selection and execution. It addresses the problem of selector credit starvation, where traditional outcome-rewarded RL fails to give sufficient credit to the skill-naming tokens, especially in long-horizon tasks. Experiments on five benchmarks show that SkillGate improves a 9B policy’s success rate from 40.8% to 53.2%, reduces exposure to misleading candidates, and requires fewer skill reads.
By Qingyao Li, Wenxiang Jiao, Shuai Shao, Kangning Zhang, Yuan Lu, Yi Guo, Weiwen Liu, Weinan Zhang, Yong Yu
The paper introduces CurriPO, a tree‑structured curriculum that automatically adapts to diverse user reward models in AI alignment tasks. By exploiting the natural hierarchy between easy‑ and hard‑to‑optimize reward models, CurriPO covers a broad user population in a single traversal, reusing previously incorporated reward models. Experiments on personalized continuous control show that CurriPO improves population satisfaction by 1.2–2.1× over the strongest baseline while cutting training time and better serving users traditionally underserved by conventional optimization.
By Taehyung Kim, Jongeun Choi
PGFS++ is a synthesis‑aware reinforcement learning framework that improves molecular properties while ensuring the resulting molecules can be synthesized and remain structurally similar to the input. It builds on PGFS+ by using trainable embedding lookup tables for reaction templates and second reactants, a more effective scoring function, and a refined RL algorithm. Experiments demonstrate that PGFS++ enhances target properties and preserves high output diversity, overcoming the reward‑hacking failure mode seen in earlier versions.
By Boqiao Zhang, Godbless James, Sai Krishna Gottipati, Andrew Fitzgibbon
The article introduces the Transition Complexity Profile (TCP), a set of reproducible metrics designed to quantify the difficulty of transition prediction in game world modeling and reinforcement learning. TCP evaluates environments based on one-step branching, interaction-induced uncertainty, and temporal/spatial dependency span, and it includes reference distributions, protocol stochasticity, and a versioned measurement budget for comparability. The authors argue that TCP should become standard benchmark metadata and a required statistic in GWM and RL research.
By Lele Cao
The paper introduces Reverse‑Turn Policy Optimization (RTPO), a method that restructures multi‑turn agentic reinforcement learning rollouts into sparse reverse trees and updates policies in temporal reverse order. This approach addresses three key instability sources—context mismatch, weak turn‑level credit assignment, and asynchronous policy drift—by aligning each decision with its downstream continuation. Theoretical analysis shows RTPO eliminates context mismatch and drift, reduces credit bias, and converges to recursive optimality, while experiments demonstrate performance gains of 21.50% over trajectory‑level and 10.76% over turn‑level baselines on multi‑turn agentic RL benchmarks.
By Yugu Li, Jimmy Cao, Jianglin Qiao, Siyi Hu
SPADE (Self-Play in Adaptive Synthetic Executable Environments) is a reinforcement‑learning framework where a single large language model acts as both an Environment Designer—creating executable, long‑horizon training environments—and a Reasoning Agent—learning to act within those environments. The framework uses a regret signal based on the difference between rewarded performance with and without privileged hints to guide the Designer toward environments that are challenging yet solvable. Experiments show that, when scaled to 30‑billion‑parameter models, SPADE outperforms fixed‑environment baselines by significant margins across math, science, code, and reasoning benchmarks, and improves tool‑use performance on BFCL‑v4 and ACEBench‑Agent.
whyItMatters":"By making environment design a learnable component, SPADE enables continuous self‑improvement and demonstrates that adaptive, self‑generated training environments can substantially boost language‑model performance across diverse tasks."
By Bo Liu, Simon Yu, Yiding Jiang, Ao Qu, Andrew Zhao, Zichen Liu, Junsu Kim, Zijian Zhou, Seungone Kim, Tongzheng Ren, Mickel Liu, Hanfei Yu, Zhaorun Chen, Weiyan Shi, Paul Pu Liang, Luke Zettlemoyer, Yejin Choi, Natasha Jaques
The paper introduces SingularClip, a method that periodically clips the singular values of all weight matrices to prevent spectral collapse, a newly identified cause of plasticity loss in neural networks trained on nonstationary tasks. The authors empirically and theoretically analyze how growing anisotropy of singular values degrades the ability to fit new targets, and demonstrate that SingularClip outperforms baseline approaches in both continual supervised learning and deep reinforcement learning settings.
By Tyler Kastner, Nimrod De La Vega, Amir-massoud Farahmand
Open-MOPD addresses the capability imbalance problem in multi-teacher on-policy distillation (M-OPD) by isolating capability integration from routing ambiguity and revealing a 35.6% headroom gap compared to a domain-routed oracle ensemble. The study identifies three key factors—sequence-length disparities, convergence drift, and reward staleness—that misallocate token-level optimization budgets, leading to severe degradation in concise tasks. The proposed Open-MOPD framework introduces token-share balancing, gap-aware dynamic budget allocation, and student reward refresh, boosting headroom recovery to 83.4% and providing an open-source, reproducible post‑training recipe and evaluation suite.
By Huan-ang Gao, Haohan Chi, Yong Yan, Shiyuan Feng, Hanlin Wu, Zheng Jiang, Bingxiang He, Wei-Ying Ma, Ya-Qin Zhang, Hao Zhou
While entropy regularization is widely used to stabilize and accelerate Natural Policy Gradient methods, its ability to yield faster convergence rates for the unregularized objective remains underexplored. Existing analyses often rely on double-loop architectures and invoke a linear entropy penalty.
PGFS++ is a synthesis‑aware reinforcement learning framework that improves molecular properties such as drug‑likeness or binding affinity while ensuring the resulting molecules can be synthesized and remain structurally similar to the input. It builds on PGFS+ by using trainable embedding lookup tables for reaction templates and second reactants, a more effective scoring function, and a refined RL algorithm. The method addresses a reward‑hacking failure mode by treating each input molecule as the start of a forward‑synthesis trajectory, applying learned reaction templates with in‑stock building blocks, and producing diverse, high‑quality outputs with explicit synthesis routes.
Falcon Perception-HD applies reinforcement learning (GRPO) to autoregressive perception models, aligning them directly with precision and recall metrics rather than relying on maximum‑likelihood fine‑tuning. The RL framework introduces reward design for set‑structured outputs and multi‑head sampling control, enabling state‑of‑the‑art performance in very dense scenes (up to 500 objects) and eliminating common issues such as mask repetitions, NMS, and coordinate deduplication. Hybrid self‑annotation pipelines tailored for difficult referring expressions and dense scenes further boost RL training, with improvements observed across all difficulty levels on PBench and SACO‑Gold, and the model preserves object existence knowledge without negative samples.
The paper introduces CurriPO, a tree‑structured curriculum that adapts to diverse user reward models in AI alignment. By automatically building a curriculum that branches and reuses reward models, it addresses the problem of users whose reward models are hard to optimize, a group often underserved by conventional methods. Experiments on personalized continuous control demonstrate that CurriPO improves population satisfaction by 1.2–2.1× over the best baseline while cutting training time.
The paper presents a cost‑effective approach for industrial explainable‑recommendation systems by decoupling explanation generation from selection. Explanations are pre‑generated using six prompt styles and two commodity LLMs, then a lightweight CPU‑resident selector (e.g., LambdaRank) chooses the best one at request time, achieving sub‑100 ms latency without GPUs. Experiments on a 2,958‑pair Google Local subset and a 300‑pair MovieLens‑1M split show that pairwise ranking outperforms single‑action RL methods, while KG‑path selectors achieve near‑perfect unique‑output rates, and the overall end‑to‑end build cost is around $15 on commodity hardware.
Data-DPO is a target model‑oriented supervised fine‑tuning data selection method that uses one‑step probing of the target model to generate pairwise data preferences, trains a lightweight reward model to capture these preferences, and then selects a training subset by combining target‑model preference, external quality scores, and marginal diversity. Experiments on Vision‑Flan and LLaVA‑CoT demonstrate that Data‑DPO consistently outperforms existing data selection baselines across multiple data budgets and even surpasses full data training performance.
By Peng Sun, Yi Yang, Antong Zhang, Chunxiao Li, Yanbo Wang, Dianbo Liu, xin chen, Kai Yu, Lu Chen, Tianfan Fu