After the Spill

How far should an embodied system change after unexpected experience? Research OS, Mismatch Lab, and Builder Lab by Seonhye Gu.

View the Project on GitHub higuseonhye/research-os

Paper 002 — Related work v0.2 (WM expansion)

Positioning doc · cite in Introduction · supersedes v0.1 (mock→physics · archived)
Core wedge: failure-conditioned model-adequacy testing · not “add expert after failure” alone


Novelty (defensible)

Avoid: failure-driven world-model reconstruction · add expert after failure · latent representation expansion.

Use:

Failure-conditioned model-adequacy testing for embodied world models: determining when parameter repair is insufficient and restricted structural expansion is necessary.

Long form (manuscript):

We study whether persistent, structured task failures can provide evidence that an embodied agent’s current dynamics model is inadequate, and whether selecting a prepared structural expansion operator improves prediction and control beyond parameter-only adaptation without degrading nominal behavior.

Scope defense:

Paper 002 does not autonomously invent an unconstrained causal ontology. It tests a restricted first step: whether failure evidence can justify switching from parameter repair to a prepared structural expansion operator.

Closest prior work already adds modules, routes mixtures, or integrates new world models at test time. Our contribution is whether and when expansion is warranted — validated on downstream prediction · planning · recoverability · nominal regression.


1. Latent predictive world models

World models compress high-dimensional observations into latent state, learn action-conditioned dynamics, and support planning or policy learning via imagined rollouts; prediction error often drives refinement. Recent surveys treat WM roles as policy learning, planning, simulation, evaluation, and synthetic data generation (World Model for Robot Learning survey).

Fixed model class (typical):

fixed representation / architecture
+ new observations
→ better latent state or parameters

Paper 002:

persistent task failure
→ is the fixed model class still adequate?
→ repair parameters or change model structure?

Latent representation learning is background, not the novelty claim.


2. Continual and adaptive world models

Representative lines:

Work Focus Trigger
DRAGO Retain past dynamics across tasks New task · same state space
EvoAgent Self-reflection · curriculum on Minecraft Long-horizon experience
CAMA / embodied continual Confidence-based adaptation Behavior / environment change

These ask how to absorb new experience, preserve old knowledge, or adapt to change — not whether the current model class is fundamentally wrong.

Paper 002 gap: repair within class vs structural expansion — a decision before continual update strategy.


3. Closest structural adaptation work (required citations)

This subsection is mandatory. “Add expert after anomaly” alone is not novel.

TMoW — Test-Time Mixtures of World Models

arXiv:2601.22647 · routes among world models at test time · few-shot integration of new models for dynamic environments.

TMoW Paper 002
How to expand · route · refine prototypes Whether / when expansion is warranted
Unseen domain · test-time mixture Task failure after failed parameter repair
Embodied task outcomes + L1 vs L3 arm comparison · nominal regression guardrail

MuSix — Multi-scale Mixture of World Models

arXiv:2607.00457 · selects WM scale by novelty / experiential distance · different update rates per scale.

MuSix Paper 002
Environment change · experience distance Structured residual after K parameter updates
Scale selection Model-class adequacy test

Worldscape-MoE

arXiv:2607.03964 · diffusion-transformer WM · progressive control-specific experts for new action modalities.

Worldscape-MoE Paper 002
Action modality coverage · shared physics Failure diagnosis · repair vs expansion
Progressive MoE tuning Deployment task failure as evidence

L3 arm (add drift expert) is structurally similar — differentiation is trigger + adequacy test + behavioral validation, not expert addition per se.

LMC — Local Module Composition

OpenReview · input-distribution outlier → new module · local structural relevance.

LMC Paper 002
Observation OOD Action-conditioned prediction failure + task consequence
Module creation Parameter repair insufficiency as gate

4. Structural and causal world-model adaptation (program horizon)

Longer-term neighbors — Paper 002 is a restricted confirmatory cell, not full causal discovery:

Work Link
Variational Causal Dynamics OpenReview
Dynamic predicate invention OpenReview
Compositional world models OpenReview

These target invariant mechanisms, symbolic repair, or compositional abstractions. Paper 002 tests one prepared operator (drift expert + gating) under hidden-mode control.


5. Generative and diffusion world models

Industry and academia move toward foundation-scale video / action world models (Cosmos, World Action Models). Worldscape-MoE combines diffusion transformers with experts.

Boundary: diffusion expresses multimodal futures within current support. Paper 002 asks when the generative / dynamics family itself is inadequate.

confirmatory v0.1:  modular dynamics expert (L3)
extension:          deterministic vs diffusion vs discrete experts

Diffusion remains an expansion operator candidate, not the v0.1 confirmatory arm (taxonomy).


6. Failure detection, uncertainty, and model inadequacy

Three distinctions (must be explicit in method):

Type Meaning Paper 002?
Distribution shift Input distribution changed Related · not sufficient alone
Epistemic uncertainty Model lacks knowledge Related · not sufficient alone
Structural inadequacy No stable explanation within current model class Primary novelty

Adequacy gate (pre-specified · not τ alone):

structured residual persists
after K parameter-update attempts
AND is conditionally predictable under alternative model class
(provisional mode cluster / prepared expert explains held-out trajectories)

Evidence must show: repeats · mode-associated · not absorbed by L1 repair · L3 explains out-of-sample.


Positioning matrix

Line What changes Trigger Structure Behavior link vs Paper 002
Latent WM latent / weights training loss low planning / policy adequacy not tested
Continual WM knowledge / weights new task / data limited yes no repair vs expansion test
LMC modules input outlier yes task perf OOD not task failure
TMoW mixture / router unseen domain yes embodied no failure-after-repair gate
MuSix multi-scale WM experiential novelty yes adaptation no L1 vs L3 comparison
Worldscape-MoE control experts new modality yes generation / control modality not deployment failure
Causal WM mechanisms intervention yes prediction unconstrained discovery
Paper 002 param or expert unexplained task failure core compare pred · control · recoverability nominal regression

Manuscript paragraph (draft)

World models learn predictive representations of action-conditioned environmental dynamics and support planning, policy learning, simulation, evaluation, and synthetic experience generation. Recent robot-learning research has expanded this paradigm from compact latent dynamics models toward foundation-scale video and action world models. However, most approaches assume a fixed representational or architectural family and adapt its latent state or parameters as new observations arrive.

Continual world-model methods focus primarily on incorporating new experience while preserving previously acquired dynamics or task knowledge. Modular approaches move beyond monolithic adaptation by routing among multiple models or adding experts. Test-time Mixtures of World Models refine routing and construct new world models for unseen domains, while multi-scale mixtures update world knowledge at different rates across abstraction levels. Recent diffusion-transformer world models also support progressive extension through shared and control-specific experts. These methods establish that modularity can improve adaptation and scalability, but they generally begin from detected domain novelty, new task distributions, or additional control modalities rather than testing whether observed task failure specifically invalidates the current model class after parameter repair has failed.

A related body of work studies compositional and causal world models. Such methods factorize transition mechanisms, identify sparse environmental changes, or invent symbolic predicates to repair an insufficient causal representation. These approaches motivate structural adaptation but often assume predefined intervention settings or address unconstrained causal and symbolic discovery, which lies beyond the scope of the present study.

We instead study a restricted model-adequacy decision in an embodied control setting. Given an unexplained failure, we ask whether parameter-level repair within the current dynamics model is sufficient or whether a prepared model-class expansion operator is required. Our central comparison is therefore not between alternative latent representations alone, but between no adaptation, parameter adaptation, and modular structural expansion. Latent reorganization provides mechanistic evidence, while the primary outcomes are prediction, action selection, recoverability, and retention on nominal conditions.


References (anchor · verify at submission)

ID Work URL
TMoW Test-Time Mixture of World Models arXiv:2601.22647
MuSix Multi-scale Mixture of World Models arXiv:2607.00457
Worldscape-MoE Unified MoE World Model arXiv:2607.03964
LMC Local Module Composition OpenReview
WM survey World Model for Robot Learning arXiv:2605.00080
WAM World Action Models arXiv:2605.12090
Paper 001 Same-state recoverability @ S ../paper1/status.md

Doc Path
Description paper002_description_wm_expansion_v0.1.md
Industry context (brief) paper002_industry_context_v0.1.md
Pre-reg paper002_prereg_wm_expansion_v0.1.md
Mock→physics RW (archived) paper002_related_work_v0.1.md