Recursive Models Are Not Your Architecture

Date: 2026-04-25 Author: Limina Engineering Team

English

We audited the latest Recursive Language Model paper. The proposal: slice prompts, recurse, manage state explicitly. Fine for benchmarks. Useless for systems that must not break.

The trap is engineering for problems you do not have. RLM turns inference into a program of sub-tasks. We do not need recursive thinking. We need deterministic flow.

Our stack relies on two primitives:

Context purification. Strip ambiguity locally before anything reaches the LLM. Do not delegate cleanup to a model that hallucinates.
Intent routing. Decide: direct answer, memory lookup, or structural split. One layer. No recursion.

RLM adds branch depth. Each branch is a point of failure where token drift cascades. Latency balloons not from compute limits but from recovery logic trying to reconcile inconsistent recursive states.

Robust agents are built on:

Input hygiene. Garbage in, garbage out. Recursion does not sanitize garbage; it multiplies it.
Minimal moving parts. Every recursive split is state you must version, checkpoint, and reconcile.
ROI discipline. Tuning retrieval and tightening system prompts yields more reliability than rebuilding inference as a state machine.

We keep the agent flow thin. Context pure. Execution predictable.

Architecture is not measured by concepts integrated. It is measured by layers removed until nothing can break silently.

中文

我们审阅了最新的递归语言模型论文。方案：切分提示，递归，显式状态管理。跑分好看，系统却更容易崩。

陷阱在于为不存在的问题写代码。RLM 把推理变成子任务程序。我们不需要递归思考，只需要确定性流。

我们的栈只依赖两个原语：

上下文净化。在到达 LLM 之前就地消除歧义。不要把清理工作交给会幻觉的模型。
意图路由。判定：直答、记忆检索、结构拆分。一层逻辑，不要递归。

RLM 增加分支深度。每个分支都是故障点，token 漂移会在那里级联。延迟上升不是算力触顶，而是恢复逻辑试图对齐不一致的递归状态。

健壮的系统建立在：

输入卫生。垃圾进，垃圾出。递归不会净化垃圾，只会把垃圾复制。
最小活动部件。每个递归切分都是你必须版本化、落点、和解的状态。
ROI 纪律。优化检索和收紧系统提示，比把推理重构成状态机更可靠。

保持流细，上下文净，执行可预测。

架构的价值不在于堆叠了多少概念，而在于削掉多少层直到没有东西能静默崩坏。