UniToolCall introduces a unified framework for tool-use in large language model agents, standardizing toolset construction, dataset generation, and evaluation. The framework aggregates over 22,000 tools and creates a hybrid training corpus of more than 390,000 instances by combining ten public datasets with synthetically generated, structurally controlled trajectories. It models diverse interaction patterns—single‑hop vs. multi‑hop, single‑turn vs. multi‑turn, serial vs. parallel execution—and adds an Anchor Linkage mechanism to enforce cross‑turn dependencies, while converting seven public benchmarks into a common Query–Action–Observation–Answer format for fine‑grained evaluation.
By Yijuan Liang, Xinghao Chen, Yifan Ge, Ziyi Wu, Hao Wu, Changyu Zeng, Wei Xing, Xiaoyu Shen
arXiv:2609.05837v1 Announce Type: new
Abstract: LLM-based agents are increasingly deployed in real-world applications through tool-use APIs, yet training them for specific environments remains fundam...
By Zhiyi Lyu, Yewen Li, Longtao Zheng, Shengtian Yang, Lang Feng, Lei Feng, Peng Jiang, Kun Gai, Qingpeng Cai, Bo An
SWE-Prime introduces a two-stage supervised fine-tuning data selection process for large language models tackling software issues. The first stage filters entire trajectories by quality and representativeness, while the second stage selects meaningful semantic segments based on contribution, learnability, and risk. Experiments on SWE-Bench Pro and Verified demonstrate that training on just 10% of trajectories chosen by SWE-Prime surpasses full-dataset training, achieving up to 12.2% and 24.2% performance gains.
By Dewu Zheng, Ruizhe Ye, Yanlin Wang, Yang Ye, Hongyu Zhang, Ensheng Shi, Xilin Liu, Yuchi Ma, Jianxing Yu, Zibin Zheng
SWE-Prime introduces a two‑stage, multi‑granularity supervised fine‑tuning (SFT) data selection process for large language models tackling real‑world software problems. The first stage filters entire trajectories based on process quality, result quality, and representativeness, while the second stage evaluates semantic segments for contribution, learnability, and risk, keeping all segments in context but only penalizing selected ones during training. Experiments on SWE‑Bench Pro and Verified demonstrate that training on just 10% of trajectories chosen by SWE‑Prime outperforms full‑dataset training, achieving up to 12.2% and 24.2% relative gains.
PG-SFT: Balancing Capability Acquisition and Retention in Offline Agent Fine-Tuning explores how to maintain a model’s existing abilities while teaching it new ones through supervised fine‑tuning on offline agent trajectories. The authors compare standard SFT, KL‑penalty, and update‑magnitude constraints, finding that these methods still degrade non‑target capabilities. They introduce Privilege‑Guided SFT (PG‑SFT), which uses turn‑level information gain to modulate supervision strength, achieving a better trade‑off between acquiring new skills and preserving existing ones, though with a slight drop in target‑task performance.
By Ronghua Li, Zi Liang, Zhishan Li, Shinan Liu
The paper investigates how multi‑turn reinforcement learning (RL) improves large‑language‑model agents, focusing on the sensitivity of performance gains to interaction constraints such as sampling breadth ($K$) and interaction depth ($T$). Using the metric $Pass@(K,T)$, the authors find that trajectory‑level RL benefits require sufficient turn budgets and that RL mainly enhances error recovery and long‑horizon resilience rather than single‑step decision quality. To address the budget dependence, they discuss process supervision signals and propose BudgetMix RL, which varies interaction budgets during training and shows robust performance across different execution settings.
By Kejian Zhu, Zhuoran Jin, Shangqing Tu, Tianyi Men, Yupu Hao, Jiachun Li, Kang Liu, Jun Zhao
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
arXiv:2609.22951v1 Announce Type: cross
Abstract: Enterprise agentic systems that route every trajectory step to a frontier model waste 60-80% of their inference budget on subtasks that smaller model...
By Rudrendu Kumar Paul, Sourav Nandy
MetaCurate is a training‑free framework that automatically evolves curation filters for agentic coding data by using pairwise large language model judgments to refine its rules, rubric, and aggregation policy. Starting from no initial filter, it discovers a strategy that scores trajectories independently, requiring pairwise comparisons only during the search phase. Experiments on an uncurated trajectory pool and the Nemotron v3 SWE dataset show that MetaCurate achieves the highest mean SWE‑Verified score compared to random selection, hand‑authored filtering, and a training‑loop baseline, while incurring lower search costs.
By Donghoon Kim, Seunghyuk Oh, Kevin Galim, Minseo Kim, Minjae Lee, Younghyun Kim, Wooseok Seo, Hyung Il Koo, Wonjun Kang
arXiv:2608. 11552v1 Announce Type: cross Abstract: Uncertainty quantification (UQ) methods for language models are typically evaluated on single-turn outputs, where uncertainty is attached to one generated answer.
By Dylan Bouchard, Mohit Singh Chauhan
Selective Regenerative Decoding (SRD) is a new inference-time decoding method that improves large language model reasoning by allowing segment-level intervention on candidate trajectories. Instead of discarding or keeping entire trajectories, SRD selectively refines only the degraded suffix while preserving useful prefixes, leading to higher expected trajectory quality and better sample efficiency. Experiments on MATH500, GPQA Diamond, HotpotQA, and AlpacaEval show that SRD matches Best-of-N accuracy with fewer generated tokens and outperforms speculative rejection in low‑compute settings.
By Sophia Xiao Pu, Yumo Xu, Sailik Sengupta, Millennium Bismay, Ruixue Lian, James Gung, Yi-an Lai, Arshit Gupta
SWE-VERIFIER is a lightweight generative model that selectively triggers candidate sampling and verification for coding agents, scoring proposed actions to decide when additional sampling is needed and ranking the resulting actions before execution. Trained on Qwen3.5-4B with preference data, it improves resolution on MiniMax-M2.7 from 75.47% to 77.20% at 1.09× cost and on DeepSeek-V4-Flash from 72.60% to 74.27% at 1.13× cost, while verifying fewer than 4% of steps. The approach demonstrates that selective sampling and lightweight verification can enhance coding‑agent performance and inference efficiency with minimal overhead.
By Kaijian Zou, Vaibhav Chennareddy, Yunxiang Zhang, Naihao Deng, Longju Bai, Jiaxin Pei, Rada Mihalcea, Lu Wang