JevSpawn: Adaptive Agentic Inference through Compositional Action Spaces
Haoyang Su, Weiran Huang
Fudan University · Shanghai Jiao Tong University · Shanghai Innovation Institute
Published September 30, 2026 · on the feed October 2, 2026 · ▲ 6 · arXiv 2610.00437
The abstract
LLM agents generate intermediate reasoning and actions token by token, making extended interactions slow and computationally expensive. Jev-style models offer fast probabilistic predictions over finite fields, but require those fields to be specified in advance. This requirement limits autonomous task solving, where the available actions must be derived from natural language instructions and adapted through interaction. We introduce JevSpawn, a compositional policy that connects natural language task specifications to finite probabilistic exploration. Parallel action spawning is coupled with feedback driven branch selection, representation revision, and recovery from retained alternatives. Shared action structure and model prefixes reduce repeated generation and context computation without additional training. Evaluations on eight benchmark tasks against seven agent baselines and a TypeSafe Jev variant establish JevSpawn as a promising approach to structured agentic inference, with improved task performance and faster navigation.
The Brief
1:54A short podcast summary of the paper. Written and voiced by AI.
Transcript
JevSpawn. Adaptive Agentic Inference through Compositional Action Spaces. By Haoyang Su and one other, from Fudan University, Shanghai Jiao Tong University, and one other institution. Published on the thirtieth of September. Its code repository is worked out: about ninety source files, with a readme.
Language model agents produce their intermediate reasoning and their actions one token at a time, and over an extended interaction that becomes slow and computationally expensive. Jev-style models offer a faster route: quick probabilistic predictions over finite fields. Their catch is that those fields have to be specified in advance. In autonomous task solving they cannot be, because the available actions have to be derived from natural language instructions and then adapted as the interaction goes on.
JevSpawn is a compositional policy built to bridge that gap. It connects task specifications written in natural language to finite probabilistic exploration, so the fast predictive approach can work over actions that are derived from the instructions rather than fixed beforehand.
Actions are spawned in parallel. That spawning is coupled with branch selection driven by feedback, with revision of representations, and with recovery from alternatives that were kept in reserve. Efficiency comes from sharing. Shared action structure and shared model prefixes cut down on repeated generation and on context computation, and none of this requires additional training.
Evaluation covers eight benchmark tasks, against seven agent baselines and a TypeSafe Jev variant. The authors report improved task performance together with faster navigation, and present JevSpawn as a promising approach to structured agentic inference.
The pitch rests on one move. Jev-style prediction was fast but needed its fields fixed ahead of time. JevSpawn lets those fields be built from language and revised through interaction, while shared structure keeps the cost of building them down.
Paper Quest · 7 chapters · about 41 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
An agent that writes out every move is an agent that waits. Here is the bottleneck.
Your language model agent writes every thought and action token by token. Over a long task that is slow and expensive, even though many candidate actions differ in only a few arguments.
The road ahead
- 1 · The problem ~7 min Language model agents write every action as text, which is slow, while faster models that score fixed fields need those fields defined in advance.
- 2 · What came before ~3 min
- 3 · The key idea ~9 min JevSpawn lets the model derive finite action fields from the instructions, try several field-value actions in parallel, and use feedback to choose, recover or revise.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Each round derives fields, spawns several actions in parallel, selects a branch from execution feedback, and then revises the fields or recovers a retained branch.
- 6 · The experiment ~6 min You’ll guess: Against seven agent baselines on eight benchmark tasks, how did JevSpawn trade off speed against task performance?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.