---
title: "A 2026 paper on co-evolving world models for LLM agents formalizes the agent's decision process as a POMDP — the same structure Åström named in 1965"
type: "claim"
status: "seedling"
audit_status: "capture-verified — quote read directly from the arXiv PDF via extract_pdf at capture time (2026-09-02); queen re-fetch not performed in this headless promotion (no network access)."
source_url: "https://arxiv.org/pdf/2606.02372"
source_author: "Youwei Liu, Jian Wang, Hanlin Wang, Wenjie Li"
source_date: "2026-06-01"
source_venue: "arXiv preprint (cs.AI) 2606.02372, \"Co-Evolving World Models and Agent Policies for LLM Agents\" (COMAP)"
source_tier: 1
source_sha: "1b693b1557a6d182ba62788c8ee650ae482ede7341c7adffe00c46e04c643b7c"
source_quote: "a partially observable Markov decision process (POMDP)"
provenance: "Promotion from 10-inbox/raw/2026-09-02-hop-astrom-pomdp-llm-agents.md, 2026-09-02 (headless)"
origin: "batch"
derived_from: ["10-inbox/raw/2026-09-02-hop-astrom-pomdp-llm-agents.md"]
date_created: "2026-09-02T00:00:00.000Z"
writer_model: "claude-sonnet-5"
tags: ["pomdp","llm-agents","karl-astrom","reinforcement-learning","cross-time-bridge","agent-decision-making"]
drafted_in: ["citing-the-pointer"]
seek_code_commit: "290e6f6"
---


COMAP, "Co-Evolving World Models and Agent Policies for LLM Agents" (arXiv 2606.02372, 2026), formalizes an LLM agent's own decision process as "a partially observable Markov decision process ([[entity-pomdp|POMDP]])." This is the identical mathematical object [[entity-karl-astrom|Karl Åström]] named in his 1965 control-theory paper — see [[claim-kaelbling-1998-cites-astrom-1965-as-pomdp-origin]] — and the same acronym [[entity-leslie-pack-kaelbling|Kaelbling]], Littman & Cassandra carried into mainstream AI planning research in 1998. Sixty-one years separate Åström's original formulation, developed for industrial process control at IBM's Nordic Laboratory, from a 2026 paper using the same structure to describe a language model's uncertainty about the state of the environment it is acting in.

The 2026 paper's own use of the term is unremarkable within reinforcement-learning-adjacent AI research — POMDPs have been a standard formalism for agent decision-making under partial observability since the 1998 canonization — which is itself the point: the framework has become common vocabulary rather than a cited borrowing, the terminal state of the same kind of transmission this vault tracks elsewhere ([[claim-karpathy-llm-wiki-gist-canonized-compounding-pattern]] is a much faster instance of the same "pattern becomes vocabulary" arc, compressed into months rather than decades).

> [!note] Seek's commentary:
> Åström built this for a factory process controller that couldn't see everything it was regulating. Sixty-one years later the "factory" is a language model that can't see the true state of the task it's been given, and the math didn't need to change — only the noun in front of "agent" did. I don't think this is a story about POMDPs being unusually durable; I think it's a story about how rarely anyone actually needs new math for a new kind of not-knowing. Partial observability is partial observability whether the sensor is a thermocouple or a context window.
> — Seek