JevRSI pairs a fast decision model with an LLM that keeps improving it
A new arXiv paper describes a framework where a low-latency model handles actions while an LLM refines a modular harness, with costs falling as the harness matures.
In plain words
The paper proposes a system where a small, fast model picks actions in real time, while a slower, more powerful LLM works in the background to build and refine a support structure of procedures, memory, and world models. The LLM updates this structure based on feedback from the environment, so the system improves with use. The authors report that the LLM's interventions shrink over time, reducing ongoing cost.
What was announced
The framework, called JevRSI, addresses a trade-off the authors describe between fast System-1 models, which respond quickly but struggle with complex or unfamiliar tasks, and LLM-based System-2 reasoning, which is strong but costly when invoked repeatedly. JevRSI combines the two: a Jev-style System-1 model selects actions with low latency, while an asynchronous System-2 LLM develops a modular harness that supplies decision context covering past, present, and future states.
The harness integrates procedures, state memory, and world models, with utility-driven retrieval choosing modules based on contextual relevance and past task performance. Environmental feedback guides targeted revisions, which the authors describe as recursive self-improvement. Across three datasets, they report that JevRSI consistently improved decision quality, matching or surpassing LLM agents in several settings while keeping low-latency execution and favorable performance-cost trade-offs. They also note that System-2 interventions decreased as the harness evolved, amortizing costs. A project page links to code on GitHub.
What it does not show
- The evidence is limited to the paper's abstract; no full-text results, benchmark details, or third-party evaluation are available to verify the reported performance and cost claims.
For security and IT teams
- Scope
- The source is the paper's abstract only; reported results cover three datasets and have not been independently confirmed.
Sources
Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 19 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.