The paper investigates the problem of over‑editing by large language models (LLMs) when repairing code, showing that even state‑of‑the‑art models like GPT‑5.5 often rewrite more code than necessary. Using a benchmark of 400 BigCodeBench problems with controlled AST‑level corruptions, the authors quantify over‑editing and demonstrate that a simple preservation instruction can significantly reduce excess edits and cognitive complexity while improving Pass@1. They further explore post‑training strategies, finding that reinforcement learning yields the best balance between edit fidelity and performance retention, thereby establishing edit fidelity as a distinct, measurable dimension of code‑repair quality.
FiMI Banking presents a controlled Indian retail‑banking environment designed to train conversational agents that can answer product questions, handle account requests, and comply with strict operational and regulatory constraints. The system is built from vetted banking documents, structured ground truth, synthetic customer backgrounds, and banking tools. Evaluation of two post‑training methods—preference optimization and reinforcement learning with verifiable rewards—shows that preference optimization significantly increases safe behavior (out‑of‑scope refusal from 52% to 80%), while reinforcement learning improves edge‑case and order‑sensitive task performance and reduces token usage.
The paper introduces Test Cases Scaling (TCS), a two‑stage reinforcement learning framework that automatically generates sound and discriminative test cases for code‑generation large language models. Stage 1 trains a test generator to produce tests aligned with a reference solution, while Stage 2 refines the generator to produce counterexample tests that target the model’s current failure modes. Experiments on TACO and LiveCodeBench show that TCS improves pass@1 and inference‑time answer selection, and the learned generator also aids in selecting better outputs from other LLMs.
The paper introduces Multi-step Proximal Policy Improvement (MPI), a method that refines offline reinforcement learning policies through sequential re-centered proximal steps. By modeling policies as a probability manifold, MPI interprets a wide range of offline actor objectives as a single proximal policy improvement step and extends this to multiple steps for controlled policy improvement beyond the behavior distribution. Experiments on D4RL benchmarks demonstrate that a few MPI refinements enhance strong offline baselines such as TD3+BC, ReBRAC, and IQL across many tasks, while diagnostics clarify the benefits of re-centered refinement over fixed-objective scheduling and highlight critic error limitations.
SVG-Score introduces a human‑aligned evaluation framework for text‑to‑SVG generation, addressing the inadequacies of existing image‑based metrics like CLIPScore that poorly capture SVG‑specific errors such as color, count, and spatial inaccuracies. The authors first demonstrate that CLIP‑based scores are largely insensitive to these errors and that off‑the‑shelf Vision‑Language Models respond unevenly across error types and styles. They then create a human‑annotated Semantic Alignment dataset and develop two evaluators: a CLIP scorer adapted to vector graphics and a VLM judge trained with supervised fine‑tuning and reinforcement learning, enabling both fast large‑scale and expressive, interpretable assessment of SVG generators.
The paper introduces FactoSR, a factorized reinforcement learning framework designed to improve spatial reasoning in Vision‑Language Models (VLMs). By decomposing the problem into planar correspondence (XY), depth consistency (Z), and temporal reversibility (T), FactoSR addresses the dimensional mismatch between 2D visual inputs and 3D physical reasoning. Experiments on multi‑view and video benchmarks show significant performance gains, with a 5.9% improvement on VSI‑Bench and 4.5% on All‑Angles‑Bench.
The paper introduces Local Updates, Global Learning (LUGL), a framework that separates data collection from model training, allowing non‑incremental learners such as gradient‑boosted trees (LightGBM) to be used in reinforcement learning for games. LUGL alternates between a local phase—where self‑play generates tabular updates—and a global phase—where these updates train a function approximator before resetting the table. Experiments on both perfect‑information and imperfect‑information games show that LightGBM agents perform competitively or better than neural‑network baselines like DQN and DeepCFR.
LeanGRPO eliminates redundant recomputation in diffusion reinforcement learning by reusing the same feed-forward backbone for rollout and policy update, thereby avoiding unnecessary gradient tracking. It introduces two training schedules—LeanGRPO‑Retain, which reuses computation graphs and activations, and LeanGRPO‑Reweight, which backpropagates provisional gradients and corrects them later. These methods achieve up to a 1.83× speedup on FlowGRPO/DanceGRPO with FLUX.1‑dev and Wan while preserving the original optimization objective.
PPO-STGNN is a DAG task‑scheduling algorithm that combines proximal policy optimization with spatio‑temporal graph neural networks to address the NP‑hard scheduling problem in heterogeneous cloud‑edge‑end environments. It extracts features from both the DAG task topology and the physical resource graph, then optimizes the scheduling policy to minimize makespan and schedule length ratio while improving CPU and memory load balancing. A multi‑teacher behavior‑cloning pretraining step accelerates convergence, and experiments demonstrate significant load‑balancing gains with low completion times in dynamic, heterogeneous settings.
CroCo introduces cross‑lingual contrastive preference tuning on self‑generations, extending prior English‑only methods to 14 high‑ and low‑resource languages. A reward model trained solely on English preferences, applied to a multilingual base, yields effective within‑language rankings and improves performance in both monolingual and multilingual settings without catastrophic forgetting. The approach requires on‑policy data; off‑policy responses and online preference optimization offer limited gains, yet on structured tasks CroCo matches or surpasses the base model in most languages, and on open‑ended generation it wins 28/30 judge evaluations across 15 languages.
By Mike Zhang, Ali Basirat, Desmond Elliott
The paper introduces a method for creating query‑specific rubrics for DeepResearch‑style long‑form report generation by training rubric generators with reinforcement learning. It builds a dataset of queries annotated with human preferences, then uses a hybrid reward that includes preference consistency, format validity, and LLM‑based rubric evaluation. The learned rubrics outperform generic or manually constructed alternatives in distinguishing preferred reports and, when used as rewards, improve performance of both single‑agent and multi‑agent DeepResearch systems.
By Changze Lv, Jie Zhou, Wentao Zhao, Jingwen Xu, Shihan Dou, Zisu Huang, Muzhao Tian, Xiaohua Wang, Zhengkang Guo, Yang Liu, Pluto Zhou, Tao Gui, Le Tian, Xiao Zhou, Xiaoqing Zheng, Xuanjing Huang, Jie Zhou
The paper introduces SPACE, a method for enabling large language model agents to emit variable-length action chunks in long-horizon tasks. By distilling chunk-boundary supervision from programmatic skills derived from successful trajectories, SPACE overcomes the tendency of agents to either act one step at a time or commit to overly long sequences. Experiments on ALFWorld and ScienceWorld demonstrate that SPACE raises success rates by 7.0%–31.3% and cuts LLM decision rounds by up to 78.9%.
By Yanting Yang, Can Jin, Jinman Zhao, Jiahao Wu, Yang Zhou, Zhepeng Wang, Zhendong Wang, Mu Zhou, Dimitris N. Metaxas
The paper introduces a reinforcement learning framework that selects among a portfolio of gradient‑based and derivative‑free optimizers during a run. At each decision point a recurrent policy reads the current run state and chooses both the next optimizer and its usage duration, passing the best solution and step size forward. The method is trained with a decoupled actor‑critic using the same runtime distribution metric as evaluation, and on unseen problems it outperforms all individual portfolio optimizers except at the smallest budgets, remaining robust to distribution shift.
By Martin van der Schelling, Deepesh Toshniwal, Miguel A. Bessa
RideSkill is a hierarchical algorithm for generalized ride sharing that uses large language models (LLMs) to automatically design and train a skill repository, a combiner, and a repositioner. The combiner assigns vehicle-specific skills for adaptive dispatch across varying scenarios and objectives, while the repositioner moves idle vehicles to emerging regions to avoid conflicts. By training all components via an LLM-based evolutionary method, RideSkill eliminates the need for real-time LLM calls, enabling high-performance deployment in large-scale systems.
By Zijian Zhao, Sen Li, Xialiang Tong, Mingxuan Yuan
The paper introduces EqGrid, a closed‑loop simulation that uses a low‑frequency, open‑weight LLM policy agent to set price, carbon limits, and subsidies for a community of empirically‑grounded household personas, while high‑frequency multi‑agent RL traders clear a continuous double auction on a physically constrained IEEE‑33‑bus grid. It demonstrates that the LLM can reduce energy‑poverty inequality—lowering the Gini of energy burden from 0.351 to 0.305 and mean burden by 28%—without increasing net grid cost, and that a compressed sub‑1B model retains 92–95% of this benefit at dramatically lower inference energy. The study also establishes a compute‑efficiency frontier and a decoupled‑safety design that eliminates grid violations.
whyItMatters":"By showing that a lightweight LLM can effectively manage energy markets to reduce poverty and inequality while staying energy‑efficient, the work offers a practical, low‑carbon AI solution for humanitarian energy‑poverty interventions."
By Kunal Jadhav, Siddhesh More
The paper reports a specialized training pipeline for large language models to excel in competitive programming, combining problem curation, synthetic reasoning traces, supervised fine‑tuning, and reinforcement learning. Using 22,000 curated problems, the authors train two models—Nemotron‑3‑Nano‑CC (30B) and Nemotron‑3‑Ultra‑CC (550B)—and introduce GenCorrect, a test‑time refinement strategy. On the IOI 2025 benchmark, Nano‑CC scores 468 points with GenCorrect, surpassing the gold‑medal threshold, while Ultra‑CC reaches 502; in IOI 2026, a competition‑specific Ultra‑CC system scores 535.4, exceeding both the gold threshold and the top human score of 498.27, marking the first AI system to outscore the highest‑scoring human contestant on an IOI problem set.
By Aleksander Ficek, Sean Narenthiran, Mehrzad Samadi, Somshubra Majumdar, Boris Ginsburg
DiDrive introduces a risk‑aware hierarchical diffusion framework for offline reinforcement learning in autonomous driving. It combines a low‑level risk‑gated encoder with a high‑level contextual modulator to filter redundant state information, and a 3DICE policy optimization that reduces out‑of‑distribution overestimation and stabilizes gradients. On the CARLA benchmark, DiDrive outperforms baselines such as IQL, CQL, and Diffusion‑QL, achieving an 85% success rate and a 4295.68 average reward in dense traffic with 60 vehicles.
By Qisong Guo, Jingtang Chen, Zhilin Chen, Pei Xu, Mingjian Fu, Wenxi Liu, Yuanlong Yu
The paper introduces a method that integrates action abstraction into policy optimization for reinforcement learning and generative flow networks. By iteratively identifying frequently used action subsequences in high‑reward trajectories and treating them as single high‑level actions, the approach expands the action space and improves sample efficiency. Experiments on synthetic and real‑world tasks show that this technique discovers diverse high‑reward states more effectively, especially on challenging exploration problems, and yields interpretable abstract actions that reflect the underlying reward structure.
By Oussama Boussif, L\'ena N\'ehale Ezzine, Joseph D Viviano, Micha{\l} Koziarski, Moksh Jain, Esmeralda S. Whitammer, Emmanuel Bengio, Rim Assouel, Yoshua Bengio
SafeEvolve is an experience-driven framework that co‑evolves a harness and policy to align large‑language‑model agents with safety goals. It uses completed on‑policy trajectories to update safety prompts and hierarchical skills, then applies a two‑stage SFT‑RL training loop that bootstraps the policy with the evolved harness and refines it through verifier‑augmented rewards. Experiments on agentic safety benchmarks show that SafeEvolve improves the safety‑utility tradeoff, achieving a three‑fold reduction in ASR on AgentDojo for Qwen3.5‑4B while increasing benign utility from 59.79% to 61.86%.
By Qinghua Mao, Wanying Qu, Dadi Guo, Leitao Yuan, Qingyu Liu, Yu Li, Guanxu Chen, Yanwei Fu, Xi Lin, Xia Hu, Dongrui Liu
The paper introduces a two‑step deep reinforcement learning framework for Reach‑Avoid‑Stay (RAS) problems, aiming to compute the maximal robust RAS set and its control policy for general dynamic systems. First, it learns the maximal robust control‑invariant set inside the target and a policy to keep the system within it. Then it uses this invariant set as a target to compute the maximal robust reach‑avoid set, proving equivalence to the maximal robust RAS set and constructing a switching policy that guarantees task completion. Simulation results show the method achieves exact maximal RAS sets without training errors and outperforms baseline approaches in accuracy and performance.
By Gabriel Chenevert, Jingqi Li, Achyuta kannan, Sangjae Bae, Donggun Lee