DynaWeave: Learning to Adapt through Dynamic Scenario Synthesis

Xuan Yang; Chunxiao Liu; Hongsheng Xin; Hao Xu; Ning Miao

cs.CLOct 21, 2026

DynaWeave: Learning to Adapt through Dynamic Scenario Synthesis

Authors: Xuan Yang, Chunxiao Liu, Hongsheng Xin, Hao Xu, Ning Miao

Organizations: Department of Data Science, City University of Hong Kong. · Hong Kong Institute of AI for Science, City University of Hong Kong. · Independent Researcher. · Li Auto Inc.

Abstract

Large language model (LLM) agents increasingly use tools to complete tasks in stateful environments. However, many training environments remain largely static apart from changes caused by the agent's own actions, whereas external events introduced by other users, services, or sudden contingencies in real environments can invalidate ongoing plans. To address this gap, we introduce DynaWeave, an automated synthesis framework that transforms existing static tasks into verifiable dynamic scenarios. The framework generates exogenous events relevant to each task, covering changes in task requirements, environment state, and external information. It uses directed acyclic graphs (DAGs) to represent dependencies between events and an execution engine to activate them during agent interaction. A dynamic verifier revises only the evaluation checks affected by events that actually occur and retains all other checks. Using DynaWeave, we collect 2,095 verified interaction trajectories across 36 environments and use them for supervised fine-tuning of Qwen3.5 models. Experiments on Gaia2, CostBench, and ToolMaze show that DynaWeave achieves consistent improvements in overall performance across all three evaluated model sizes, increasing the task completion rate on CostBench by 67.34 percentage points and even exceeding the performance of Gemini-3.1-Pro-Preview on ToolMaze. These results support dynamic scenario synthesis as a practical approach to training agents to adapt to changing environments. Our code is available at https://github.com/Miaow-Lab/DynaWeave.

Figures & tables

Figure 1: Overview of DynaWeave. (a) The framework generates task-grounded exogenous events from an existing static task, validates their specifications, and organizes compatible events into a dependency graph. (b) An execution engine activates events during agent interaction through action completion, virtual time, or random activation. (c) The dynamic verifier adapts checks affected by realized events while retaining unaffected requirements. (d) Agents interact through the original tools, receiving tool results and permitted event notifications; verified teacher trajectories provide demonstrations for supervised fine-tuning.

Figure 2: Statistics of the dynamic scenario corpus generated by DynaWeave.

Advanced models
	Gaia2						CostBench			ToolMaze
Model	Execution	Search	Ambiguity	Adaptability	Time	Overall	Cost Gap $\downarrow$	TCR	ITUR $\downarrow$	Avg
GPT-5.5	79.4	96.2	46.9	55.0	4.4	56.4	0.000	95.01	0.00	35.94
Gemini-3.1-Pro-Preview	78.6	92.8	40.6	45.9	2.1	52.0	0.010	95.28	1.42	32.19
GLM-5.1	71.2	83.8	39.4	56.9	1.2	50.5	0.030	93.18	0.29	32.28
GPT-4o	8.3	17.5	4.4	6.2	5.8	7.4	1.179	90.29	19.96	31.29

Table 1: Main results on Gaia2, CostBench, and ToolMaze. For CostBench, Cost Gap measures excess tool execution cost relative to the optimal reference; TCR (Task Completion Ratio) is the percentage of correctly completed tasks; and ITUR (Invalid Tool-Use Ratio) is the percentage of invalid tool calls. ToolMaze Avg denotes its official composite score. Detailed ToolMaze results and the definition of Avg are provided in Appendix A .

Figure 3: Effect of training scenario quantity on Qwen3.5-4B performance on CostBench (left) and ToolMaze (right). Zero scenarios denotes the baseline Qwen3.5-4B.

		CostBench			ToolMaze
Training subset	Data	Gap $\downarrow$	TCR $\uparrow$	ITUR $\downarrow$	NP $\uparrow$	TSR $\uparrow$	RC $\downarrow$	Avg $\uparrow$
Mixed	1,048	0.286	87.08	1.27	61.75	24.13	73.29	28.91
w/o Task Requirement	1,048	0.278	50.39	1.41	61.25	22.31	75.83	26.86
w/o Environment State	1,048	0.287	82.68	1.24	61.00	19.06	77.87	24.85
w/o External Information	1,048	0.285	80.58	0.31	64.50	21.19	76.54	26.45
Mixed	524	0.423	80.89	8.97	59.75	22.81	74.25	27.64

Table 2: Effect of event-source composition on Qwen3.5-4B performance. Data denotes the number of training trajectories; Mixed includes all three event sources. For ToolMaze, NP (No Perturbation) reports task success without injected failures; TSR (Task Success Rate) reports average task success under tool failures or corrupted outputs; and RC (Recovery Cost) reports average normalized recovery cost under the same conditions. Bold marks the best value in each metric column.

Appendix figures & tables1 asset

Supplementary material from the paper’s appendix.

Appendix

	TSR $\uparrow$						PRR $\uparrow$					RC $\downarrow$					Overall
Model	NP	P1	P2	P3	P4	Avg	P1	P2	P3	P4	Avg	P1	P2	P3	P4	Avg	Avg $\uparrow$
GPT-5.5	70.00	67.25	26.75	16.75	14.00	31.19	86.09	29.49	18.89	14.75	37.31	35.61	73.02	81.33	83.83	68.45	35.94
Deepseek-V4-Pro	69.75	68.00	34.75	12.00	9.75	31.12	84.70	37.74	12.41	12.00	36.71	29.98	63.55	85.19	87.55	66.57	36.33
GLM-5.1	75.75	60.00	32.50	8.00	8.50	27.25	67.19	35.33	9.04	10.15	30.43	37.86	66.47	88.65	89.11	70.52	32.28
Gemini-3.1-Pro-Preview	71.25	59.75	30.00	6.75	9.00	26.38	73.49	33.52	8.09	10.84	31.48	37.53	67.64	88.64	87.24	70.26	32.19
GPT-4o	59.25	47.00	33.50	9.25	11.50	25.31	70.49	27.72	12.72	12.71	30.91	35.97	78.34	81.77	80.54	69.16	31.29

Appendix

Table 3: Detailed results on ToolMaze.

LLM Agent Training

Large language model agent training teaches models to plan, use tools, and learn from multi-step interaction with environments.

See all 586 papers

AgentMercury: Your Agent Can Synthesize Verifiable Environments for Business Scenarios at scale
Aug 21, 2026Minbyul Jeong, Chanwoong YoonLLM Agent Training RL for LLM Agents
Meridian Intelligence Global Inc. · University of Massachusetts Amherst
EnvCraft: Synthesizing Executable Environments in Agentic RL for Claw-like Agent
Sep 4, 2026Yirong Zeng, Shen You, Jinhang Feng +8LLM Agent Training Agentic RL
Harbin Institute of Technology, SCIR Lab · Huawei Technologies Co., Ltd · Peking University +1
EnvFactory: Scaling Tool-Use Agents via Executable Environments Synthesis and Robust RL
May 18, 2026Minrui Xu, Zilin Wang, Mengyi DENG +12Tool-Augmented Language Model Agents Agentic RL
LARK, HKUST (GZ) · University of Cambridge · UCL +2

Aug 21, 2026J/K move · Enter open · S save

AgentMercury: Your Agent Can Synthesize Verifiable Environments for Business Scenarios at scale

Minbyul Jeong, Chanwoong Yoon

Meridian Intelligence Global Inc. · University of Massachusetts Amherst

LLM Agent Training RL for LLM Agents

Open paper AskPDF arXiv

Agents learn to act through interaction with environments, yet the environments used for training are often manually constructed or synthesized around predefined tasks and benchmarks. This task-centric paradigm makes it difficult to scale environments that reflect realistic and evolving workflows where diverse tasks can naturally emerge from the underlying world. We introduce AgentMercury, a scalable framework for synthesizing executable environments from high-level business scenarios. Rather than constructing an environment for a specific task, AgentMercury first instantiates a persistent world with entities, services, tools, state, and executable cross-service invariants, from which diverse tasks and interaction trajectories can subsequently emerge. We construct 4,783 executable environments spanning 14 industries and 50 countries, and use them as training substrates for reinforcement learning. Despite being generated without targeting the evaluation benchmarks, policies trained on these business-oriented environments improve substantially on both enterprise workflows and out-of-domain benchmarks spanning reasoning, coding, scientific computing, and tool use. In our experiments, Qwen3.5-4B improves from 12.3 to 15.7 on EnterpriseOps-GYM and from 45.9 to 56.0 on AIME26 after training on AgentMercury environments. We further show that the construction process itself can be learned: fine-tuning Qwen3.5-35B-A3B on construction traces increases executable-world authoring success from 3.3% to 83.3% on held-out business scenarios. These results show that scenario-grounded environments can provide useful and generalizable learning signals beyond benchmark-specific training, while their construction can itself become a learnable capability.

Abstract

LLM Agent Training

Explore similar work

AgentMercury: Your Agent Can Synthesize Verifiable Environments for Business Scenarios at scale

EnvCraft: Synthesizing Executable Environments in Agentic RL for Claw-like Agent

EnvFactory: Scaling Tool-Use Agents via Executable Environments Synthesis and Robust RL