ZGCM-1 is a 7B dense foundation model trained from scratch with extreme data, system, and algorithmic efficiency. It uses a core premise that compact models can overcome capacity limits by combining deliberate internal thinking with active external tool use, supported by a 256K context and an end‑to‑end high‑efficiency training recipe that includes interleaved gated sliding‑window and full attention, a stable FP8 Muon optimizer, progressive curriculum scaling, and reformulation of interaction traces into Markov Decision Processes. The model is competitive with much larger frontier models on challenging mathematical reasoning and agentic search tasks, offers a ~4.2× efficiency improvement in pre‑training time‑to‑loss, and its weights, checkpoints, training code, data recipes, and logs are fully open‑source to support community research.
By Jiyan He, Guang Liang, Hao Liu, Haoxiang Guan, Jinbo Sun, Junyi Guo, Wenjun Feng, Yantai Xie, Yifei Shen, Bin Shao, Chuyang Wei, Kai Chen, Kexin Zhou, Minghang Zhu, Shuxin Zheng, Tie-Yan Liu, Taine Zhao, Wenhui Zhu, Xueyin Xu, Xiaoqing Zhang, Yatao Li, Yuxuan Ren
arXiv:2608. 10357v1 Announce Type: cross Abstract: Long-horizon tool-using agents must reason over user goals, domain policies, tool calls, simulator state, and delayed verifiable rewards.
By Zelei Cheng, Amritansh Mishra, Sambit Sahu, William Campbell
arXiv:2607. 28582v1 Announce Type: new Abstract: On-policy self-distillation (OPSD) is a promising approach to improve reasoning language models, but it remains brittle in practice: making it work reliably often requires substantial engineering effort.
By Jiawei Xu, Minghui Liu, Juzheng Zhang, Tom Goldstein, Furong Huang
arXiv:2510. 05592v2 Announce Type: replace Abstract: Outcome-driven reinforcement learning has advanced reasoning in large language models (LLMs), but prevailing tool-augmented approaches train a single, monolithic policy that interleaves thoughts and tool calls under full context; this scales poorly with long horizons and diverse tools and generalizes weakly to new scenarios.
By Zhuofeng Li, Haoxiang Zhang, Seungju Han, Sheng Liu, Jianwen Xie, Yu Zhang, Yejin Choi, James Zou, Pan Lu
arXiv:2606. 03892v1 Announce Type: cross Abstract: Training LLMs to orchestrate multi-step tool calls is held back by three coupled obstacles: realistic stateful execution environments are costly to build, synthetic training queries are often detached from the server's actual state (so the generated tool calls fail to execute), and recall-based RL rewards incentivize verbose tool-calling patterns.
By Ibrahim Abdelaziz, Asim Munawar, Kinjal Basu, Maxwell Crouse, Chulaka Gunasekara, Suneet Katrekar, Pavan Kapanipathi
Instella‑MoE is a fully open Mixture‑of‑Experts language model with 16 billion total parameters and 2.8 billion active parameters per token, trained from scratch on AMD Instinct GPUs. It incorporates a sparsely activated MoE design with Gated Multi‑head Latent Attention and FarSkip‑Collective connectivity, and follows a multi‑stage pipeline that includes pre‑training, long‑context extension, supervised fine‑tuning, direct preference optimization, and reinforcement learning with Multi‑Teacher On‑Policy Distillation. The model achieves an average score of 76.7 on pre‑training benchmarks and 73.2 on instruction‑following, reasoning, math, coding, and chat benchmarks, outperforming comparable fully open and open‑weight models, and its full training pipeline, weights, and code are released for reproducibility.
By Jiang Liu, Sudhanshu Ranjan, Prakamya Mishra, Yonatan Dukler, Gowtham Ramesh, Jialian Wu, Ximeng Sun, Wen Xie, Chaojun Hou, Vikram Appia, Zhenyu Gu, Zicheng Liu, Emad Barsoum
The paper introduces T1, a 122‑billion‑parameter Mixture‑of‑Experts model trained with reinforcement learning to perform long‑horizon terminal tasks such as coding and scientific discovery. T1 operates a real shell in a cloud sandbox, making over 300 tool‑call turns per task and receiving rewards from task‑specific verifiers. The authors detail a training recipe that includes aggressive warm‑starting, TITO construction with drift repair, and rollout‑routing replay, achieving significant performance gains on Terminal‑Bench 2.1 and surpassing GPT‑5.4 and GLM‑5.1 on the Long‑Horizon Terminal Bench.
By Junyao Yang, Yucheng Shi, Zhongzhi Li, Ruhan Wang, Zongxia Li, Haitao Mi, Leowei Liang
arXiv:2606. 18284v1 Announce Type: cross Abstract: The limiting resource for training agents via reinforcement learning (RL) is increasingly frontier task supply: valid, solvable tasks just difficult enough to train the current model.
By Lorenz Wolf, Connor Watts, Roger Creus Castanyer, Geoffrey Bradway, Maxwill Lin, Augustine N. Mavor-Parker, Matthew Daborn-Sargent
arXiv:2608. 16798v1 Announce Type: cross Abstract: Agent harnesses have substantially improved performance on long-horizon tasks by coordinating agent interactions with the environment.
By Huatong Song, Fei Bai, Ming Yang, Renyuan Li, Jia Deng, Jujie He, Zhange Zhang, Daixuan Cheng, Yan Xing, Qi Yun, Xuxing Chen, Danyang Li, Feng Chang, Chuan Hao, Ran Tao, Jian Yang, Bryan Dai, Wayne Xin Zhao, Mingjie Tang, Ji-Rong Wen
arXiv:2607. 20468v1 Announce Type: new Abstract: AI agents are increasingly used to automate research and development tasks, yet existing benchmarks typically evaluate them on prescribed workflows or narrow action spaces.
By Jehyeok Yeon, Ben Rank, Maksym Andriushchenko
The paper introduces SMITH, a reinforcement learning framework that jointly trains large language models to create and use tools within a single policy. By alternating between build and use tasks and employing separate reward signals for schema, code, and outcome failures, SMITH enables a 4B Qwen3 model to achieve state‑of‑the‑art accuracy on procedural reasoning benchmarks, outperforming larger untrained models and improving performance on downstream tasks when its tools are applied.
By Zhi Rui Tam, Chieh-Yen Lin, Yun-Nung Chen, Shao-Hua Sun, Hung-yi Lee
The paper introduces EvoResearcher, a training‑free, inference‑time protocol that enables a frozen large language model to perform cost‑bounded self‑reflection and early stopping. By iterating through generate → self‑critique → revise steps until a maximum depth or a CONFIRMED sentinel is reached, the model can self‑verify its answers within a strict compute budget. The protocol incorporates four self‑reflective meta‑reward components—correctness, efficiency, reflection depth, and tool‑call diversity—implemented as prompt‑level mechanisms, and is validated on Big‑Bench Hard, GSM8K, and MATH benchmarks, achieving comparable accuracy while terminating 82‑88% of items early with only about 2.1 generations per question.
By Wei Yu, Suxing Liu, Minjie Yu, Jiahao Wang, Zhijian Zheng, Haocheng Deng, Bing Li