Position: It is Time to Virtualize Foundation Models with a Self-evolving Operating System Layer
Suparna Bhattacharya, Tarun Kumar, Cong Xu, Satish Kumar Mopur, Jiahao Li, Ashish Mishra, Aalap Tripathy, Annmary Justine Koomthanam, et al.
cs.AI · cs.LG · cs.MA · cs.OS
Quick verdict
Position paper proposing an FMOS layer to virtualize foundation models: compelling fragmentation diagnosis, but no prototype, evaluation, or evidence.
Summary
This position paper argues that compound agentic AI systems have outgrown their software stacks and need a Foundation Model Operating System (FMOS): a system layer that virtualizes foundation models the way virtual machines abstract hardware. The authors, mostly from HPE, diagnose current fragmentation: frameworks like LangChain, AutoGen, and Claude Code each embed implicit runtime semantics for state, memory, budgets, and guardrails, while connectivity protocols (MCP, A2A) standardize integration but not execution or governance. The proposed abstraction, Virtual Foundation Models (VFMs), exposes a stable interface giving applications the illusion of dedicated, trustworthy models with unbounded capabilities, while the FMOS internally orchestrates knowledge across memory tiers, model routing and composition, and verification/policy enforcement. Drawing on Popek and Goldberg's virtualization theory, the paper reformulates the fidelity requirement as progressive quality gain through self-evolution, with learned traps switching between a fast inference path and a deliberative slow path (System 1/System 2). Appendix A sketches formal support: a virtualizability condition (sensitive operations contained in privileged operations), a sample-complexity argument for shared policy learning, and Lagrangian co-optimization with shadow prices. The main limitation is that the work is purely conceptual: no prototype, API specification, benchmark, or empirical evidence is presented, and the cited self-evolution benchmarks (Evo-Memory, Evo-Test) show current models fail at exactly the capability the FMOS assumes its control layer can supply.
Review complete
Assessment at a glance
5
risk flags identified
● high 2
● medium 2
● low 1
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Agent frameworks each embed framework-specific implicit runtime semantics for state, memory, budgets, and guardrails, recreating a pre-OS fragmentation that blocks portable behavior and governance (Section 2.1).
- MCP and A2A standardize connectivity but deliberately omit execution semantics and governance, leaving the system layer unbuilt (Sections 2.1 and 7).
- A virtualization boundary enforcing sensitive-operations-in-privileged-operations (Eq. 1) yields efficiency, resource control, and interface-level equivalence for FM workloads (Section 4.1, Appendix A.1).
- FM virtualization should replace hardware-fidelity with progressive quality gain through self-evolution, with the FMOS control plane rather than the model making evolution decisions (Sections 3 and 4.6).
- Shared FMOS-level learning of privileged-mechanism policies converges faster than per-application learning because trapped-operation samples pool across applications (Appendix A.2, Eq. 3).
Strengths
Concrete and timely fragmentation diagnosis
Section 2.1 enumerates exactly which portable contracts are missing (state/memory semantics, observability, resource governance, trust enforcement, model mediation) and correctly notes MCP/A2A stop at connectivity; Section 2.2's workflows-vs-loops substrate split matches real enterprise pain.
Intellectually honest adaptation of classical virtualization theory
Section 4.1 maps Popek & Goldberg (1974) efficiency/resource-control/fidelity onto FM resources and explicitly replaces fidelity with progressive quality gain instead of pretending equivalence; Appendix A.1 discusses graceful degradation when Eq. 1 is only partially satisfied.
Formal scaffolding beyond hand-waving
Serious engagement with alternative views
Cites counter-evidence against its own premise
Weaknesses
Zero empirical content for a systems-feasibility claim
Load-bearing mechanism underspecified
Formalization is textbook-thin and not FM-specific
Circular reliance on self-evolution
Evidence base skews gray literature
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This is a position paper (ICML 2026, PMLR 306; HPE + University of Chicago/Argonne) arguing that compound agentic AI systems need a new system layer — a Foundation Model Operating System (FMOS) — that virtualizes foundation models behind a stable Virtual Foundation Model (VFM) interface, analogous to how virtual machines abstract hardware. Central claims: (1) agent frameworks (LangChain, AutoGen, Claude Code) embed framework-specific runtime semantics for state, memory, budgets, and guardrails, reproducing a pre-OS fragmentation (Section 2.1); (2) connectivity protocols (MCP, A2A) standardize integration but not execution semantics or governance (Sections 2.1, 7); (3) a virtualization boundary enforcing A_sen ⊆ A_priv (Appendix A.1, Eq. 1) yields efficiency, resource control, and interface-level equivalence; (4) FM virtualization should replace hardware fidelity with progressive quality gain via self-evolution (Section 4.1); (5) shared FMOS-level policy learning beats per-application learning through sample pooling (Appendix A.2, Eq. 3). Section 8 closes with a community agenda: define abstractions, build open reference implementations, establish system-level benchmarks, align governance with standards bodies.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is FMOS (Foundation Model Operating System)?
FMOS is a proposed system layer introduced in this ICML 2026 position paper by HPE and University of Chicago/Argonne researchers. Like an operating system virtualizes hardware, FMOS would virtualize foundation models: applications see a stable 'Virtual Foundation Model' interface that appears to be a dedicated, trustworthy model with effectively unbounded capabilities, while the FMOS internally handles context/memory management, model selection and routing, verification, and policy enforcement.
Is FMOS a real system I can use today?
No. The paper is a position paper — it presents an argument, an architecture sketch, and formal motivation, but no prototype, no code, and no experimental results. The authors call on the community to build reference implementations, benchmarks, and standards. Earlier systems it cites as partial steps include AIOS, MemGPT/Letta, and Llumnix, each addressing a narrower slice.
How is FMOS different from MCP (Model Context Protocol) and A2A?
According to the paper, MCP and A2A standardize connectivity — how agents discover and call tools and each other — but do not define execution semantics: state and memory persistence, observability, resource budgets, trust enforcement, or model routing. FMOS is proposed as the missing layer beneath those protocols providing these shared, enforceable contracts uniformly across frameworks.
What evidence supports the FMOS idea?
None empirically. Support is a conceptual analogy to classical virtualization theory (Popek & Goldberg, 1974), a formal sketch in the appendix (a virtualizability condition, a sample-complexity argument for shared policy learning, and constrained-optimization framing), and cited evidence that today's agent stacks are fragmented. Notably, the benchmarks the paper itself cites (Evo-Memory, Evo-Test) show current models fail at self-evolution — the capability FMOS's control layer is supposed to supply.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Text extraction is truncated: parts of Sections 4.2-4.3, 5.3-5.4, 6.1, and Appendices B and D are missing; the trap-mechanism section (5.4) is repeatedly referenced but not fully present, limiting assessment of the paper's core mechanism.
- Figures 1-6 are available only as embedded caption/label text; the architecture diagrams could not be assessed visually.
- As a position paper there are no quantitative results to verify; confidence applies to the structure and consistency of the argument, not to empirical validity.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.