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 003 — Related work v0.1 (missing relation & capability expansion)

Positioning doc · cite in Introduction · complements paper003_lit_positioning_v0.1.md (execution-horizon entry framing — a different, narrower comparison) Core wedge: failure-triggered relational expansion, validated by capability threshold crossing — not "learn a relational world model" alone, and not "mean error went down" alone.


Novelty (defensible)

Avoid: "we add a graph/relational module to a world model" (not novel by itself — see §1) · "relational reasoning improves generalization" (not novel by itself — see §2) · "capability emerges from scale" (different mechanism, different literature — see §3).

Use:

Failure-triggered relational-adequacy testing, validated by achievable-task-space growth: determining when parameter repair and mode-level expansion are both insufficient because the model class is missing a relation between entities, and testing whether the resulting expansion converts previously unachievable task variants into achievable ones — not only whether it lowers prediction error.

Scope defense:

Paper 003 does not perform open-ended relational/causal discovery. It reuses Paper 002's restricted setting: a prepared relation-module operator, invoked only after a prespecified gate (residual survives parameter repair and mode expansion, and correlates with a second entity's state). The novel measurement is capability threshold crossing, not the expansion mechanism itself.


1. Relational / graph-based world models (how to represent relations — not our contribution)

These lines establish how to build relational structure into a dynamics model. Paper 003 does not compete on this axis — it assumes a prepared relation-module operator exists (same posture as Paper 002 toward its dynamics expert).

Work Focus Relation to Paper 003
Neural Relational Inference (NRI) — Kipf et al. Unsupervised discovery of an interaction graph + dynamics, jointly, from observational data Learns relations from scratch across many examples; Paper 003 tests whether one failure-triggered switch to a prepared relation module beats parameter/mode repair — not open discovery
Interaction Networks / Neural Physics Engines / Graph Networks as Learnable Physics Engines (survey coverage) Graph-based reasoning over objects and pairwise relations for physical prediction Foundational relational-dynamics architecture; assumes the graph is relevant from step 1, not "detected as missing after failure"
Factored World Models for Zero-Shot Generalization in Robotic Manipulation Decomposes scenes into objects, models pairwise interactions, zero-shot generalizes to novel object combinations Closest prior on "relations → generalization to new task configurations"; does not test a repair-vs-expansion gate, and measures generalization accuracy broadly, not a preregistered 0%→achievable threshold
TriRelVLA Triadic (object-hand-task) relational structure for compositional generalization across unseen scenes/objects/tasks Relational structure is designed in from the start for a VLA policy; no failure-triggered adequacy test
FOCUS-style object-centric world models (see WM robot-learning survey) Explicit object representations for efficient exploration of object dynamics Object-centricity as a design choice, not a response to detected model-class inadequacy

Gap these leave open: none of these ask when a relational representation becomes necessary — they either assume it from the start or learn it from broad multi-scenario data. Paper 003's question is narrower and prior-conditioned: given a system that already believes entities are independent, and has already tried parameter repair and mode expansion, is the next failure signature specifically diagnostic of a missing relation?


2. Compositional / relational generalization (measuring success — partial overlap)

Work What it measures vs Paper 003
Compositional generalization surveys (e.g. survey) Whether a system generalizes to novel combinations of known parts Closest framing neighbor to "achievable task space," but typically reported as aggregate accuracy on a held-out combination set, not as threshold crossing on individually preregistered variants
Relational Structural Causal Models Causal structure over relational (multi-entity) data Causal-discovery framing, broader than Paper 003's single prepared operator

Differentiation: Paper 003 borrows the intuition that relational structure is what unlocks compositional/OOD task success, but narrows the claim to a controlled repair-vs-expansion comparison with a prespecified capability-crossing threshold, matching Paper 002's confirmatory discipline rather than a broad generalization benchmark.


3. Capability-boundary framing (methodological precedent, different domain)

This is where the "beyond error reduction" measurement idea has precedent — in LLM-agent RL, not embodied world models. Cite as methodological precedent, not as competing work.

Work Framing Precedent value
ProRL: Prolonged RL Expands Reasoning Boundaries in LLMs Extended RL training solves problems the base model could not solve at all, not just faster/more accurately Direct precedent for "does the intervention cross a solve-at-all boundary," the same shape as capability threshold crossing
Does RL Expand the Capability Boundary of LLM Agents? A Pass@(k,T) Analysis Explicit metric distinguishing capability-boundary expansion from within-boundary performance gain Closest metric-design precedent for capability threshold crossing — worth citing directly when defining the metric in paper003_description_v0.1.md

Boundary: these papers study LLM reasoning/agents under RL scaling, not embodied structural world-model expansion. Paper 003 imports the measurement logic (solved-at-all vs solved-better), not the mechanism.


Positioning matrix

Line What changes Trigger Failure-gated? Success metric vs Paper 003
NRI / Interaction Networks learn relation graph training data no prediction accuracy assumes relation relevant from start
Factored World Models object decomposition design choice no zero-shot generalization (aggregate) no repair-vs-expansion gate
TriRelVLA relational policy structure design choice no task success (aggregate) no failure-triggered test
Compositional generalization (general) combinatorial reuse design / data diversity no held-out combination accuracy not threshold-crossing framed
ProRL / Pass@(k,T) RL training regime scale / compute no capability-boundary crossing (LLM) same metric shape, different domain/mechanism
Paper 002 mode expert task failure after L1 repair yes prediction, recoverability missing mode, not relation
Paper 003 relation module task failure after L1 and L3-mode repair yes prediction, recoverability, + capability threshold crossing this paper

Manuscript paragraph (draft)

Relational and graph-based world models represent entities and their pairwise interactions explicitly, supporting generalization across novel object counts and configurations. Neural relational inference methods learn the interaction graph and dynamics jointly from observational data, and object-centric or factored world models decompose a scene to model relative interactions, showing that relational structure supports zero-shot generalization to new task configurations. Task-specific relational representations further extend this to compositional manipulation policies. These approaches establish that relational structure is useful, but they generally assume relational structure is warranted from the outset, or infer it from broad multi-scenario training data, rather than asking whether a specific pattern of task failure — one that already survives parameter repair and mode-level expansion — provides evidence that a relation is specifically missing.

A separate line of work, in LLM-agent reinforcement learning, distinguishes capability-boundary expansion (solving problems previously unsolvable at all) from within-boundary performance improvement, and designs explicit metrics for this distinction. We import this measurement logic into embodied structural world-model expansion: rather than reporting only mean prediction-error reduction, we test whether a prepared relation-module expansion converts task variants with near-zero baseline success into achievable ones, under a preregistered threshold.

We therefore study a restricted, failure-gated relational-adequacy decision: given a task failure that survives both parameter repair and mode-level structural expansion, does invoking a prepared relation-module operator expand the achievable-task space beyond what either prior arm reaches, without regressing performance on relation-independent tasks.


References (anchor · verify at submission)

ID Work URL
NRI Neural Relational Inference for Interacting Systems arXiv:1802.04687
Dynamics review Learning-based dynamics models for robotic manipulation Science Robotics
Factored WM Factored World Models for Zero-Shot Generalization in Robotic Manipulation arXiv:2202.05333
TriRelVLA Triadic Relational Structure for Generalizable Embodied Manipulation arXiv:2605.05714
RSCM Relational Structural Causal Models arXiv:2606.14892
WM survey World Model for Robot Learning: A Comprehensive Survey arXiv:2605.00080
Compositional survey A Survey on Compositional Generalization in Applications arXiv:2302.01067
ProRL Prolonged RL Expands Reasoning Boundaries in LLMs arXiv:2505.24864
Pass@(k,T) Does RL Expand the Capability Boundary of LLM Agents? arXiv:2604.14877
Paper 002 Missing dynamic mode — detect · expand · validate paper002_related_work_v0.2.md

Open item


Doc Path
RQ paper003_rq_v0.1.md
Method draft paper003_description_v0.1.md
Lit positioning (entry framing, narrower) paper003_lit_positioning_v0.1.md