The paper introduces Privileged Self-Practice (PSP), a method that retains privileged information (PI) in the prompt rather than the loss during on‑policy self‑distillation for multi‑turn agents. PSP injects short per‑task instructions from an analyzer model when rollouts fail, sampling again with the instruction in context and training with the unchanged GRPO objective. Experiments on AppWorld and SWE‑bench Verified show PSP consistently outperforms plain GRPO, boosting task‑goal completion by up to 65% and resolved rate by up to 61% across three student models.
By Xingyu Su, Abhishek Kumar, Qing Ping, Youzhi Luo, Jonathan Buck, Zach Zhang, Subramanian Chidambaram, Vinayak Arannil
Distributed learning in embodied reinforcement-learning agents offers a degree of privacy by retaining raw sensor data on-device and transmitting only policy gradients to the server. Yet temporal stru...
The paper presents a graph-based framework for large-scale railway network management that combines a hierarchical Bayesian model with a Gaussian Process on a graph kernel to model spatially correlated maintenance environments, and a topology-aware Multi-Agent Reinforcement Learning system using graph neural networks and Transformers to optimize network-level policies. It demonstrates scalability by training agents on small network segments and deploying them zero-shot on larger, unseen networks, achieving superior performance over heuristics and standard MARL baselines while reducing training time. The approach addresses the computational challenges of centralized methods and the coordination gaps of decentralized methods in complex, long-horizon infrastructure asset management.
The paper introduces Evo-Rec, a three‑stage framework that improves generative recommendation by learning better reasoning traces for Semantic ID (SID) generation. First, it aligns SIDs with textual and behavioral contexts; second, it samples and selects reasoning traces that improve ground‑truth item prediction via supervised fine‑tuning; third, it refines the reasoning policy with reinforcement learning using catalog‑constrained generation and ranking‑aware feedback. Experiments on three Amazon Review datasets show Evo‑Rec consistently outperforms discriminative, generative, and other reasoning‑enhanced recommenders across all metrics.
The paper examines on‑policy self‑distillation (OPSD) for multi‑turn agents, showing that using privileged information (PI) in the loss can make agents appear confident yet underperform plain RL, sometimes worse than the untrained base model. To address this, the authors propose Privileged Self‑Practice (PSP), which keeps PI in the prompt and uses it only during sampling, not in the loss. PSP consistently outperforms plain GRPO across AppWorld and SWE‑bench Verified, improving task‑goal completion by up to 65% and resolved rate by up to 61%.
FairTest is a search-based testing framework designed to uncover fairness failures in Multi-Agent Reinforcement Learning (MARL) systems. It guides candidate generation using three fitness functions—measuring observed fairness, predicting fairness from abstract states, and assessing policy decision uncertainty—and prioritizes tests based on predicted fairness and uncertainty. Evaluations on three environments and two MARL algorithms show FairTest detects significantly more fairness failures than three baselines, with a 221% increase in failure count and 23% better coverage on average.
By Xiaotong Wang, Xuan Xie
The paper introduces LEAP-CBF, a safety filter that uses Least‑Effort Adversarial Potentials to quantify how much disturbance effort is needed to cause failure in nonlinear dynamical systems. LEAP serves as a control barrier function for the undisturbed system and can be combined with a robust safety filter that tolerates disturbances with bounded cumulative effort. The authors develop a deep reinforcement learning method to construct LEAPs and validate their effectiveness through simulations of multi‑agent systems and hardware experiments on a quadruped and quadrotors.
By Oswin So, Eric Yu, Chuchu Fan
The paper introduces a fully differentiable message‑passing neural network (MPNN) designed to approximate the Uniform Facility Location (UniFL) problem. Unlike many learning‑based approaches that require supervision or reinforcement learning, this model incorporates principles from classical approximation algorithms, providing provable approximation guarantees. Empirical results show that it outperforms standard approximation algorithms and reduces the performance gap to integer linear programming solutions.
By Chendi Qian, Christopher Morris, Stefanie Jegelka, Christian Sohler
arXiv:2604. 06039v2 Announce Type: replace-cross Abstract: Value iteration-type methods have been extensively studied for computing a nearly optimal value function in reinforcement learning (RL).
By Zhichao Jia, Guanghui Lan
The paper introduces Verifier-Certified Rule Transport (VCRT), a method that uses native verifiers to replay adjacent operation pairs and identify commutation certificates or anti-diamonds, thereby distinguishing true logical dependencies from mere serialization choices in reinforcement learning with verifiable rewards. VCRT assigns policy credit based on the total probability mass of each certified orbit and imposes constraints on post-swap consistency, source retention, and policy drift. In leave-one-environment-out transfer experiments across ProofWriter, CLRS, and Lean, VCRT achieves a 77.60% macro pass rate, outperforming the strongest baseline by 13.06 points, with the largest gains observed in Lean.
By Bang Xie, Hao Liu, Zhiyuan Peng, Xin Yin, Chenhao Ying, Yuan Luo, Senjian Zhang, Wei Chen
The paper introduces a stability‑aware autoscaling framework for edge serverless workloads that combines an Attention‑Enhanced Double‑Stacked LSTM with Proximal Policy Optimization to address temporal blindness in deep reinforcement learning. By weighting recent historical states non‑uniformly, the method suppresses high‑frequency jitter while preserving demand trends, outperforming single‑layer LSTM, static HPA, and KEDA baselines in latency reduction and stability. Experiments on two Kubernetes clusters with real Azure Functions traces show a ~67% reduction in P90 latency and improved adherence to a 50 ms hard SLO.
By Faraz Shaikh, Gianluca Reali, Mauro Femminella
The paper investigates whether evolutionary stability guarantees that learning agents can achieve cooperative outcomes in a multi‑agent setting. Using a three‑agent governance game, the authors compare the evolutionary basin of attraction with learning basins derived from independent Q‑learning, scaled Boltzmann exploration, and SA–EA BQL. They find that while the evolutionary basin covers the entire sampled grid, only ε‑greedy Q‑learning attains a substantial learning basin, whereas the other methods fail to sustain cooperation, highlighting a disconnect between population‑level stability and finite‑sample learning accessibility.
By Yijie Wang
The paper proposes using category theory to structure reinforcement learning in high‑dimensional, partially observable environments. By partitioning the state space into equivalence classes (symmetry orbits) and treating each class as a groupoid with a canonical representative, the agent can share learning across similar states, reducing redundancy and improving sample efficiency. Experiments on partially observable benchmarks show that this orbit‑based partitioning consistently enhances performance in environments with latent symmetry.
By Ben Opperman, Eduardo Alonso, Esther Mondrag\'on
The paper introduces AGVF, a multi‑agent framework for generating medical‑necessity appeals that must adhere to payer policy and evidence constraints. AGVF treats appeal synthesis as a Constrained Markov Decision Process involving five agents—policy formalization, evidence retrieval, gap analysis, adversarial critique, and gated synthesis—and proves that refining the policy constraint graph monotonically reduces evidence deficiency. A deterministic citation‑grounding gate ensures no unsupported assertions enter the shared state, and validation on 1,000 synthetic cases shows zero citation violations and monotonic deficiency reduction.
By Harshil Lodhiya, Alex McManus, Reese Walker
The paper introduces a finite‑sample probabilistic safety certification framework for black‑box AI decision models used in closed‑loop grid operation. It transforms the AI‑grid evaluation into a binary unsafe outcome under a safety specification and applies exact binomial inference to provide a tight one‑sided upper bound on the unsafe operation probability, using held‑out calibration scenarios. The framework also incorporates physically interpretable sample‑space adversarial attacks to address distribution shifts and is validated through case studies involving 1,000‑agent AI models for grid‑edge flexibility coordination.
By Yihong Zhou, Hanbin Yang, Thomas Morstyn
The paper investigates hindsight relabeling in preference‑conditioned multi‑objective reinforcement learning (MORL). It finds that relabeling transitions with the achieved preference direction often harms performance, causing a phenomenon called Preference Coverage Collapse where the critic’s coverage narrows to a small region of the preference space. The authors propose a simple method, her_mix, that blends achieved and requested preferences, which restores performance across most settings and dramatically reduces abandoned preference mass.
By Baptiste Bonin, Caro Strickland, Audrey Durand
The paper introduces Safety to Competence (S2C), a two‑stage reinforcement learning framework that first learns a safety filter and then trains a competitive task policy while embedding the filter. By separating safety synthesis from task learning, S2C reduces training complexity and prevents the policy from being exploited by adversarial attacks. Experiments on simulated touchdown games show that S2C achieves higher win rates, better Elo ratings, and lower exploitability than eight safe‑RL baselines, and hardware tests confirm its competence against a human opponent.
By Ruihan Wu, Rui Yang, Donggeon David Oh, Duy Nguyen, Haimin Hu
The paper presents a method for creating Open RAN (O‑RAN) slicing xApps that are both adaptive and fully inspectable. By using a large language model (LLM) to evolve compact Python programs, the authors generate slicing controllers whose decision logic remains readable and editable after optimization. In experiments on the NSF POWDER 5G testbed, the evolved controllers improved best‑effort throughput by 44.5% and reduced SLA misses from 79.9% to 2.2%, while evolutionary search outperformed independent prompting in trace‑driven simulations.
By Faezeh Dehghan Tarzjani, Bhaskar Krishnamachari
The paper investigates curriculum learning for graph neural network-based reinforcement learning applied to the job shop scheduling problem. By training policies on progressively larger instances (from 20×20 up to 30×30), the authors demonstrate that this approach reduces training time and improves performance compared to single-size training. Evaluation on unseen instances from 8×8 to 30×30 shows that curriculum learning lowers the mean optimality gap by about 8–9 percentage points and saves roughly 50 hours of training time at the largest target size.
By Jayakrishnan K. Vasudevan (Rosenheim University of Applied Sciences), Jonathan Hoss (Rosenheim University of Applied Sciences), Noah Klarmann (Rosenheim University of Applied Sciences)
The paper introduces UECR-GRPO, a method that unifies on‑policy distillation and verifier‑based reinforcement learning for mathematical reasoning. It combines verifier rewards and teacher‑derived log‑ratios into a single KL‑regularized objective (Path‑Utility Unification) and then redistributes credit at the token level using entropy‑calibrated redistribution, preserving total task credit. Experiments on five benchmarks show that UECR‑GRPO improves average accuracy by up to 0.89 percentage points over the best baseline for both Qwen3‑1.7B and Qwen3‑4B students.
By Jie Zhang, Jingxiao Yang, Zhehao Huang, Yuhang Liu, Xiaolin Huang