The paper introduces Iris-mini and Iris-pro, two search agents trained at 35B and 397B parameter scales. They use a novel data pipeline that constructs reverse‑engineered multi‑hop queries from web hyperlinks, filters trajectories, and alternates supervised fine‑tuning with reinforcement learning in a process called SFT‑RL climbing. Evaluations on several benchmarks show that, with inference‑time context management, the agents achieve the best open‑source results in their parameter ranges.
By Ziyuan Liu, Hengqi Liu, Zichuan Wang, Yang Qin, Jiachen Liang, Xu Chu, Shaowei Chen, Yuantao Gu, Mu Chuan
The paper introduces a reinforcement learning method to improve the quality of feedback generated by large language models for creative writers. By training with group relative policy optimization and a multi‑component reward that emphasizes tailored, actionable, and critical‑issue‑focused feedback, the authors demonstrate that their approach outperforms existing LLMs and baselines on three story corpora. The study shows that actionable suggestions are the key factor driving constructive feedback.
By Maja Stahl, Timon Ziegenbein, Henning Wachsmuth
HLS-Seek is a natural‑language‑to‑High‑Level‑Synthesis framework that optimizes Quality of Results (QoR) such as latency and resource usage by using a comparative proxy reward model instead of full synthesis‑in‑the‑loop reinforcement learning. The system achieves high syntax and functional correctness (84.7% and 81.4% respectively) with only 7 B parameters, surpasses GPT‑5.1 on functional pass@5, and trains 8.5× faster than real‑reward RL. In QoR evaluation, HLS‑Seek attains the lowest latency on 19 of 30 kernels and Pareto‑dominates baseline HLS tools on 9 kernels.
By Qingyun Zou, Feng Yu, Hongshi Tan, Yao Chen, Bingsheng He, WengFai Wong
The paper explores how large language models (LLMs) can be trained for small-molecule drug design by using synthetic tasks that are cheaper to evaluate. By employing a curriculum that gradually increases task difficulty, the authors demonstrate that LLMs can learn design strategies that outperform larger models on structure-based lead optimization. This approach shows that scaling post‑training with synthetic tasks can effectively adapt LLMs to high‑cost experimental scenarios that are otherwise infeasible to train on directly.
By Frank Hu, Shriram Chennakesavalu, Zichen Wang, Patricia Suriana, Bodhi Vani, Kirill Shmilovich, Kangway Chuang, Colin Grambow
SCRIPT is a scalable diffusion policy that uses a Joint Action-State-Text Diffusion Transformer (JAST‑DiT) to jointly encode actions, physical states, and natural‑language instructions, enabling direct interaction between language semantics and control dynamics. The method employs a multi‑stage training framework, including supervised imitation pre‑training, a nonlinear history conditioning mechanism for stable autoregressive control, and a post‑training stage with Reinforcement Learning with Hybrid Rewards (RLHR) that injects learnable noise to improve motion quality and instruction following. Experiments on the 1200‑hour MotionMillion dataset show that SCRIPT outperforms prior state‑of‑the‑art methods across text alignment, motion quality, and physical realism, and its performance scales consistently with model size.
By Jingyan Zhang, Han Liang, Ruichi Zhang, Bin Li, Juze Zhang, Xin Chen, Jingya Wang, Lan Xu, Jingyi Yu
Artificial Intelligence now underpins investment workflows from data and prediction to execution and tool use, yet its technical prowess does not automatically translate into profitability. A comprehensive review of public research up to 31 August 2026 across equities, ETFs, crypto spot, perpetual futures, and on‑chain markets shows real progress in prediction, text processing, portfolio design, and workflow integration, but evidence for durable net performance remains thin. The study highlights that factors such as temporal contamination, survivorship bias, weak benchmarks, implementation costs, and venue mechanics can erode alpha, and no single AI architecture has proven to deliver persistent, cross‑regime, capacity‑aware net alpha.
"whyItMatters":"The findings underscore that while AI advances are evident, investors must rigorously test and govern AI systems to avoid overestimating their profitability potential."
By Linsen Zhu, Mengqing Cai
MultihopSpatial is a new benchmark for Vision‑Language Models that focuses on multi‑hop, compositional spatial reasoning with queries ranging from 1 to 3 hops across varied spatial perspectives. It introduces the Acc@50IoU metric, which jointly evaluates answer selection and precise bounding‑box prediction, and provides a large‑scale training corpus, MultihopSpatial‑Train, to improve spatial intelligence. Evaluation of 37 state‑of‑the‑art VLMs shows that compositional spatial reasoning remains a significant challenge, and reinforcement learning fine‑tuning on the corpus boosts both intrinsic spatial reasoning and downstream embodied manipulation performance.
By Youngwan Lee, Soojin Jang, Yoorhim Cho, Seunghwan Lee, Yong-Ju Lee, Sung Ju Hwang
REFINE is a framework that refines medical concept representations by creating patient‑specific temporal graphs from a global text‑attributed knowledge graph. It uses a reinforcement learning policy to allocate a personalized graph expansion budget for each observed code, then processes the resulting graph with a heterogeneous GNN and a frozen LLM that refines representations via graph‑aware soft prompts. Experiments on MIMIC‑III and MIMIC‑IV demonstrate that REFINE consistently improves various EHR prediction backbones, surpasses strong baselines, and shows robust gains across ablation studies, KG selection, and data insufficiency scenarios.
By Mohsen Nayebi Kerdabadi, Arya Hadizadeh Moghaddam, Dongjie Wang, Zijun Yao
The paper introduces the problem of teaching multimodal large language models (MLLMs) to prioritize must‑have requirements over nice‑to‑have ones in autonomous agent tasks. It evaluates current MLLMs on 3,649 realistic service‑scenario problems and finds widespread failures, then proposes First Things First Reinforcement Learning (FTF‑rl) to explicitly optimize reasoning over multi‑priority user requirements, achieving significant performance gains and generalizing to logical and mathematical reasoning benchmarks.
By Tianjie Ju, Xinyue Xu, Wanxuan Sun, Lingxiao Diao, Gongshen Liu, Zhuosheng Zhang, Cheng Yang
The paper introduces ROBORMBENCH, a benchmark comprising 2,390 real‑robot trajectories, 21,673 verified paraphrases, and ground‑truth progress labels, to evaluate paraphrase robustness in vision‑language reward models (VLMs). It demonstrates that current VLMs often give different rewards for semantically equivalent goal descriptions, sometimes flipping a robot’s outcome from failure to success. The study finds that this instability is widespread, worsens with more divergent rewrites, and is not mitigated by model scale or explicit reasoning, though dedicated reward models trained with trajectory‑grounded supervision show greater stability.
By Wonje Jeung, Sangyeon Yoon, Hyesoo Hong, Yoonjun Cho, Dongjae Jeon, Bumjun Kim, Jean Oh, Youngjae Yu, Albert No
The paper introduces DEX-Comp, a two‑stage training method for soft context compression in Retrieval‑Augmented Generation (RAG). First, a pure distillation warm‑start trains the compression model on correct responses from an uncompressed RAG. Then, hard exploration uses reinforcement learning on queries where the uncompressed RAG fails, encouraging better computation patterns for compressed representations. Experiments on five open‑domain QA benchmarks show that DEX‑Comp compresses retrieved contexts 16×, speeds inference 4×–24×, and matches or surpasses the uncompressed RAG baseline across various retrieval depths.
By Shuyu Guo, Shuo Zhang, Zhaochun Ren
The paper introduces PRISM, a Compositional Reward Model framework that decomposes image quality into multiple verifier‑grounded stages for conditional medical image generation. By assigning distinct rewards for fine‑to‑coarse properties—such as intensity, texture, structural alignment, and semantic fidelity—and combining them via a Hierarchical Constrained Propagation mechanism, PRISM addresses shortcomings of single‑scalar reward approaches. Experiments on PanNuke, CeDeM, and ISIC datasets show that data generated with PRISM improves downstream model performance, achieving higher mDice, lower MRE, and increased F1 scores compared to baseline methods.
By Aayush Kumar Tyagi, Prathosh A. P., Mausam
The paper introduces Reflection-Aware GRPO (RA‑GRPO), a reinforcement‑learning framework that aligns diffusion generative models with human preferences. It uses Diffusion Reflection to correct intermediate sampling paths by reversing the diffusion process, and Counterfactual Path Synthesis to embed these corrected trajectories into the policy, avoiding extra inference cost. Experiments on text‑to‑image and text‑to‑video models show RA‑GRPO outperforms existing methods, reducing reward hacking and improving generalization while remaining architecture‑agnostic.
By Junlong Wu, Jiuzhou Lin, Jia Sun, Boheng Zhang, Huaiqing Wang, Dewen Fan, Houde Liu, Qianqian Gan, Fan Yang, Tingting Gao
PerfReasoning is a new benchmark that tests large language models (LLMs) on their ability to reason about hardware performance and generate analytical performance‑model code. The benchmark presents workloads, architectures, and mapping specifications, asking models to compare mappings and predict off‑chip traffic and buffer requirements. While the best closed‑source models achieve over 90% accuracy on reasoning‑based Q&A and the top open‑weight model scores 82.4%, constructing full performance models remains difficult, with most models scoring below 15% and significant variability across runs. Task‑specific reinforcement learning can improve a 4B model’s mapping‑reasoning accuracy by 15.7 points, but feedback‑free self‑revision prompting is not reliably effective.
By Dan Zhao, Karthikeyan Sankaralingam, Christos Kozyrakis, Qijing Huang
SQL-Zero is a self‑playing framework that trains a Text‑to‑SQL agent without any human‑annotated natural‑language/SQL pairs. It uses a challenger and a solver, both based on the same large language model, to generate SQL queries that are only validated by execution against the database. On the BIRD dataset, self‑play improves performance over a zero‑shot baseline by 6.6 to 7.3 points and shows promising transfer to unseen Spider databases, especially at 3B parameter scale.
By Daniel Machado Pedrozo, Julia Soares Dollis, Bryan Lincoln Marques de Oliveira, Vinicius Alboneti Aguiar, S\'avio Salvarino Teles de Oliveira, Telma Woerle de Lima Soares
The paper introduces GUT, a method that uses directed acyclic graphs to represent all possible reasoning branches of Large Language Models (LLMs). It comprises two modules: GUT-Q, which quantifies reasoning uncertainty by approximating graph complexity, and GUT-O, which reduces uncertainty through reinforcement learning that rewards lower uncertainty. Experiments on four LLMs across five datasets demonstrate GUT’s effectiveness in measuring and mitigating reasoning uncertainty.
By Shuang Liang, Xin-Yu Hu, Xiang-Jun Ou, Shao-Qun Zhang
The paper introduces DM-Align, a single-stage optimization framework that jointly performs distribution matching for distillation and aligns video generative models with human preferences. By deriving complementary gradient directions—one minimizing the gap between real and fake models and another guiding the model toward preferred samples—the method eliminates the need for separate reinforcement learning and distillation stages. Experiments on multiple foundational video models show that this sample-guided approach consistently outperforms both standalone variants and traditional two-stage pipelines.
By Jiuzhou Lin, Junlong Wu, Fei Zuo, Huan Ouyang, Dewen Fan, Boheng Zhang, Huaiqing Wang, Jia Sun, Fan Yang, Houde Liu, Kehai Chen, Min Zhang, Tingting Gao, Han Li
Squint is a visual Soft Actor Critic algorithm designed to accelerate reinforcement learning for robotics. It combines parallel simulation, a distributional critic, resolution squinting, layer normalization, a tuned update-to-data ratio, and an optimized implementation to reduce wall‑clock training time. On the SO‑101 Task Set, Squint trains policies in as little as 15 minutes on a single RTX 3090 GPU, with most tasks converging in under 6 minutes and successfully transferring to a real SO‑101 robot.
By Abdulaziz Almuzairee, Henrik I. Christensen
The paper introduces a reinforcement‑learning framework that automatically discovers compact parametrized quantum circuits for modeling power GaN HEMTs and logic nanowire FETs. Using a graph neural network policy trained with proximal policy optimization, the method optimizes circuit architectures based on leave‑one‑group‑out cross‑validation error, achieving the lowest mean absolute error across 11 targets compared to six classical baselines. The results show significant reductions in error and variability for key device metrics (Ioff, VTH, SS) on both HEMT and NWFET datasets, demonstrating the viability of RL‑selected quantum circuits as compact, physically consistent surrogates without explicit physical constraints.
By Rushat Rai, Yun-Yuan Wang, Autsada Kakaen, Pei-Jie Chang, Doan Viet Nguyen, Yuan-Chieh Chiu, Doldet Tantraviwat, Niall Tumilty, Simon See, Wen-Jay Lee, Tai-Yue Li, Nan-Yow Chen, Tian-Li Wu
The paper reviews 66 studies on large language models (LLMs) applied to HVAC operations in building energy systems, categorizing them by application and method families and evaluating their evidence realism and deployment readiness. It finds that most work focuses on building energy modelling, with only four studies reaching pilot-level evidence and none reporting sustained operational deployment. LLMs are currently best suited as semantic and workflow layers—such as point‑name normalisation and document‑grounded operator support—rather than autonomous HVAC controllers, and future research should target field‑validated benchmarks and safe, low‑latency LLM‑MPC/RL integrations.
By Alexander Neubauer, Tianzhen Hong, Han Li, Mengbo Yu, Amin Darbandi, Yannick F\"urst, Martin Kriegel