Research preprint · October 2026

JevSpawn

Adaptive Agentic Inference through Compositional Action Spaces

Haoyang Su^1,3Weiran Huang^2,3

¹ Fudan University² Shanghai Jiao Tong University³ Shanghai Innovation Institute

Read the paper ↗ GitHub ↗ Hugging Face ↗ BibTeX ↓

An agent can explore many actions without writing each action from scratch. JevSpawn brings finite probabilistic prediction to multi-turn interaction through action spaces that are inferred from task context and adapted through feedback.

Task-balanced quality versus end-to-end latency, and quality accumulated over elapsed time, for JevSpawn and seven agent baselines. — Task quality and end-to-end latency across eight benchmark tasks.

8benchmark tasks

1,761evaluation instances

0additional training

From task meaning to finite actions

Natural-language rules are expressed as composable action fields. Finite probabilities drive parallel spawning, and returned observations guide continuation, revision, and recovery. Known alternatives share model computation instead of requiring a separate text generation for every branch.

Key results

Qwen3.8-27B · Four H100 GPUs

Maze success

0.96

40.91 s E2E latency

Grid success

0.95

40.58 s E2E latency

Highest task score

5 / 8

Against seven agent baselines

All task scores

Task score ↑. PPNL, Maze, and Grid report success rates. Other tasks use benchmark rewards or scores. Bold denotes the best result; underlining denotes the second best.
Method	PPNL	Maze	Grid	LightsOut	RushHour	Sokoban	2048	Nullify
LATS	0.842	0.000	0.000	0.010	0.000	0.000	0.000	0.010
LLMCompiler	0.929	0.200	0.480	0.080	0.147	0.350	5.000	0.020
AgentPrune	0.956	0.280	0.280	0.340	0.453	0.790	1.240	0.030
HiAgent	0.930	0.400	0.320	0.033	0.172	0.140	0.000	0.090
FoldAgent	0.464	0.520	0.730	0.020	0.240	0.105	0.000	0.090
DyFlow	0.828	0.040	0.020	0.040	0.060	0.200	0.120	0.060
LatentMAS	0.902	0.320	0.120	0.030	0.113	0.095	0.000	0.050
TypeSafe Jev	0.955	0.960	0.930	0.660	0.290	0.250	296.000	0.170
JevSpawn	0.950	0.960	0.950	0.610	0.390	0.150	305.120	0.210

TypeSafe Jev uses the same JevSpawn architecture with API-based finite scoring.

End-to-end latency

E2E latency in seconds ↓. Bold denotes the best result; underlining denotes the second best.
Method	PPNL	Maze	Grid	LightsOut	RushHour	Sokoban	2048	Nullify
LATS	131.34	300.00	300.00	299.01	298.57	300.00	300.00	298.18
LLMCompiler	13.61	235.32	117.16	99.26	223.42	123.99	33.51	172.85
AgentPrune	24.23	47.91	79.57	233.19	115.48	142.53	202.57	42.90
HiAgent	16.25	209.06	206.05	254.54	20.32	150.76	300.00	98.18
FoldAgent	43.55	191.69	61.44	84.17	22.74	78.32	300.07	48.10
DyFlow	125.47	285.43	269.57	294.18	289.03	280.57	287.35	287.37
LatentMAS	48.70	269.86	281.61	292.64	150.25	269.05	289.03	185.05
TypeSafe Jev	43.96	63.98	78.39	153.95	187.45	214.76	249.05	189.41
JevSpawn	21.88	40.91	40.58	111.05	89.20	122.49	169.94	104.21

E2E latency includes failed tasks and timeouts. API communication is included for TypeSafe Jev.

Watch the interaction unfold

Follow eight paper case studies through action declarations, spawned branches, observations, and answers.

The video replays recorded traces. Run the demo locally to explore each turn.

bash demo/run.sh

Run the demo ↗

Citation

Download .bib

@misc{su2026jevspawnadaptiveagenticinference,
      title={JevSpawn: Adaptive Agentic Inference through Compositional Action Spaces},
      author={Haoyang Su and Weiran Huang},
      year={2026},
      eprint={2610.00437},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2610.00437},
}