The GUI Is Not the State: Diagnosing State Aliasing in GUI World Models

1 INTRODUCTION

• Observation-only GUI world models condition on the current interface and action, although the GUI can omit information needed to predict later transitions.For example, a moved file may no longer be visible even though its file-system state affects a later prediction.Source: §1 · p. 1
• State aliasing occurs when identical observable conditions correspond to different transition-relevant states and valid futures (Jin et al., 2026).The visible interface may not reveal what state is currently true.Source: §1 · pp. 1–2
• StateAliasBench uses strict pairs to test whether the same current observation and next action can require different validated futures.It covers four state families, 25 causal templates, and 800 verified strict pairs.Source: §1 · p. 2
• The recovery method infers structured predictive state from interaction history and uses it to condition a frozen world model.Family-specific specialists are consolidated through multi-teacher distillation into a unified estimator.Source: §1 · p. 2
• Predictive-state augmentation substantially improves state-sensitive future prediction, preserves next-observation fidelity, and improves two WM-assisted AndroidWorld agents.The approach augments frozen GUI world models without retraining them.Source: §1 · p. 2

2.1 STATE ALIASING AND STRICT-PAIR DIAGNOSIS

• An observation-only world model approximates the controlled next-observation law using the current GUI and action rather than the full interaction history.The law is defined for a reachable history and an intervention on the next action.Source: §2.1 · p. 3
• Observation-conditioned prediction is sufficient only when the current observation preserves information needed to distinguish relevant future transition laws.The formulation describes a controlled process and defines interaction history as observations and prior actions.Source: §2.1 · pp. 2–3
• State aliasing is the case where histories are observation-equivalent but not predictively equivalent, so a shared action can yield different controlled future laws.The distinction is between matching current observations and matching future behavior over the action scope.Source: §2.1 · p. 3
• When two histories have different true future laws, an observation-only model must make the same conditional prediction for both and incurs log-loss from that mismatch.The shared prediction is constrained by insufficient conditioning information, not by a particular prediction architecture.Source: §2.1 · p. 3
• Strict pairs serve as controlled empirical witnesses of state aliasing, using structured transition-relevant states and corresponding semantic outcomes.Ground-truth state values come from read-only environment evidence and are used for construction, validation, and supervision, not as observation-only WM inputs.Source: §2.1 · p. 3
• The strict-pair diagnosis tests observation–state mismatch over a specified action scope rather than requiring estimation of complete transition laws.The available process is described through histories, observations, actions, and controlled next-observation laws.Source: §2.1 · p. 3

A strict pair Pf

• A strict pair matches the model-visible condition across two real environment branches while varying hidden state and semantic outcome.The matched observation and action remove branch information from the observation-only prediction input; semantic validation checks the realized outcome contrast.Source: A strict pair Pf · p. 3

2.2 STATEALIASBENCH

• StateAliasBench covers Persistent Artifact, Clipboard, Ordered Collection, and Temporal Deadline state families.The benchmark summarizes each family's operational structured state and semantic outcomes.Source: §2.2 · p. 3, Table 1 · p. 4
• The benchmark constructs pairs with distinct hidden states, matched complete WM-visible observations and next actions, and independently validated real-environment outcomes.It contains 25 causal templates and 800 verified strict pairs; equality is checked after each evaluated WM's exact input transformation.Source: §2.2 · p. 4
• The evaluation also tests cross-application transfer, long-history stress, and semantic interference, and audits separation between training and evaluation examples.The audit found no duplicate examples or complete trajectories.Source: §2.2 · p. 4

3.1 PREDICTIVE INFORMATION STATE

• The predictive state captures transition-relevant information in history that is not determined by the current observation.The distribution-level gap measures predictive information omitted by the GUI.Source: §3.1 · pp. 4–5
• The canonical predictive representation preserves controlled next-observation laws and is minimal among deterministic sufficient representations, up to a deterministic mapping.Predictive sufficiency is scoped to one-step controlled transitions over the evaluated action set; long-history evaluation tests recovery of more distant evidence, not multi-step sufficiency.Source: §3.1 · p. 5

3.2 LEARNING A STRUCTURED PREDICTIVE STATE

• Family-specific specialist teachers are distilled into a unified estimator that predicts a canonical JSON structured-state representation from interaction history and a family/schema instruction.The unified estimator supports single-model deployment, while specialists remain diagnostic baselines.Source: §3.2 · p. 5
• Training combines the exact structured-state target with soft supervision from the corresponding family specialist.Each example receives soft supervision only from its matching family specialist.Source: §3.2 · p. 6

3.3 STATE-AUGMENTED GUI WORLD MODELING

• A deterministic, model-compatible interface expresses recovered state through family- and WM-compatible surface constraints and serialization.Its branch-dependent content is derived from recovered state and prediction-time inputs; it adds no environment information.Source: §3.3 · p. 6
• The risk decomposition separates information still missing after conditioning from the frozen world model's residual error given that conditioning.Predictive-state augmentation improves over observation-only prediction when it removes enough missing predictive information and the frozen WM can exploit it with sufficiently small residual error.Source: §3.3 · pp. 6–7

4.1 EXPERIMENTAL SETUP

• Experiments evaluate two frozen GUI world models, compare observation, raw-history, oracle-state, and recovered-state conditioning, and measure strict-pair prediction, fidelity, recovery, robustness, and downstream task success.Oracle State uses unavailable ground-truth state and serves as a diagnostic upper bound.Source: §4.1 · p. 7

4.2.1 STATEALIASBENCH

• 79.38% macro PairCorrect on Code2World-8B and 80.63% on gWorld-8B, versus 0% under Observation-only conditioning; predictive state restores state-sensitive prediction across four families.The unified estimator also achieves 88.63% exact branch-level structured-state recovery.Source: §4.2.1 · p. 8, Table 2 · p. 8
• Predictive State improves both DINOv2 and SigLIP similarity for Code2World-8B and gWorld-8B, preserving next-observation fidelity alongside state-sensitive correctness.Table 3 reports next-observation fidelity.Source: §4.2.1 · p. 8, Table 3 · p. 8

4.2.2 DOWNSTREAM GUI AGENTS

• Task success rises from 58.62% to 60.34% (+1.72 pp) for GPT-5.4 and from 44.83% to 47.41% (+2.58 pp) for Qwen3-VL-8B with predictive-state conditioning.These correspond to Relative Gains of 66.4% and 74.8%, respectively.Source: §4.2.2 · p. 8, Table 4 · p. 9

4.3 STATE RECOVERY AND ROBUSTNESS

• Multi-teacher distillation raises worst-family PairCorrect from 69.5% to 82.0% and macro average from 83.63% to 84.88% over family-balanced Joint CE.Independently deployed specialists remain stronger on average at 87.88%, reflecting a trade-off with a single unified inference-time model.Source: §4.3 · p. 9
• Under matched Code2World-8B evaluation, MTD raises final-WM macro PairCorrect from 76.75% to 79.38% and worst-family PairCorrect from 69.50% to 71.50% over Joint CE.The improvement varies by family: Joint CE remains stronger on Ordered Collection.Source: §4.3 · p. 9
• 79.75% average Exact under cross-application transfer is below 85.25% under long-history stress and 84.88% under semantic interference, making application-context change the most challenging evaluated shift.The distribution-shift suites use Exact on trajectories and are distinct from Strict PairCorrect results.Source: §4.3 · p. 9, Table 5 · p. 9

5 CONCLUSION

• Predictive-state recovery improves frozen GUI-WMs across state-sensitive and downstream evaluations, supporting attention to hidden transition state as well as visible interfaces.Source: §5 · p. 9

AI USE STATEMENT

• The authors used generative AI for language polishing and phrasing assistance, reviewed and revised the edits, and retain responsibility for the manuscript and scientific content.Source: AI USE STATEMENT · p. 10

ETHICS STATEMENT

• The study uses controlled Android interaction environments, model-generated predictions, and benchmark-derived state annotations, without human-subject studies or end-user deployment.Source: ETHICS STATEMENT · p. 10

REPRODUCIBILITY STATEMENT

• Reproducibility materials cover the theory, benchmark construction, state recovery and distillation, conditioning interface, and evaluation protocols.Appendices B–G provide details intended to support reproduction of these components and evaluations.Source: REPRODUCIBILITY STATEMENT · p. 10

A RELATED WORK

• GUI world-model research spans semantic transition modeling, visual or sketch-based prediction, renderable-code generation, transition-grounded code updates, and retrieved knowledge for long-horizon rollouts.Examples include MobileWorldBench (Li et al., 2025), ViMo, UISim, and MobileDreamer (Luo et al., 2026; Xiang et al., 2025; Cao et al., 2026), gWorld, Code2World, and AppDeltaWorld (Koh et al., 2026; Zheng et al., 2026; Xu et al., 2026a), and R-WoM (Mei et al., 2026).Source: A RELATED WORK · p. 13
• Strict pairs control the model-visible observation and action while varying transition-relevant state, testing whether state recovery can repair frozen GUI world models.The paper distinguishes this input-sufficiency question from improving generation, retrieval, or representation alone.Source: A RELATED WORK · p. 13
• The formulation builds on partial-observability and predictive-state ideas, specializing them to residual predictive information missing from the current GUI for one-step prediction over a specified action scope.Source: A RELATED WORK · p. 13
• Agent-memory methods retain history to help select actions, whereas this work conditions the world model after an action is specified to predict its consequences.StateAliasBench tests predictive sufficiency of world-model inputs, not long-horizon agent memory.Source: A RELATED WORK · pp. 13–14
• Predictive state retains only history-dependent environment information relevant to controlled next-state transitions over the evaluated action scope; it is not a replacement for agent memory or history retrieval.The two mechanisms address different interfaces and can be complementary: agent memory supports decision making, while predictive-state recovery addresses information missing from world-model simulation.Source: A RELATED WORK · p. 14

B PROOFS FOR PREDICTIVE STATE RECOVERY

• The proof appendix assumes the observation is determined by interaction history and that relevant conditional entropies and KL divergences are finite under the fixed evaluation distribution.Source: B PROOFS FOR PREDICTIVE STATE RECOVERY · p. 14

B.1 PROOF OF PREDICTIVE SUFFICIENCY AND MINIMALITY

• The canonical predictive equivalence class is the coarsest observation-preserving sufficient representation, and the predictive information gap equals the log-loss advantage of full history over observation-only conditioning.Any other deterministic sufficient representation must determine the same predictive equivalence class once the current observation is known.Source: §B.1 · p. 15, §B.2 · p. 15

B.3 PROOF OF THE PREDICTIVE-RISK DECOMPOSITION

• World-model log-loss decomposes into representation insufficiency and model mismatch, which vanish respectively when conditioning is sufficient and the frozen model matches the conditional transition law.The representation term measures information missing relative to full history; the model term is the conditional KL divergence from the true transition law.Source: §B.3 · p. 16

C.1 BENCHMARK COMPOSITION AND CAUSAL TEMPLATES

• StateAliasBench contains 800 strict pairs across 25 templates in four state families, and an independent audit found every template present, all validators passing, and all pairs release-valid.The families are Persistent Artifact (9 templates), Clipboard (5), Ordered Collection (5), and Temporal Deadline (6); the released parameter support is grounded in traces and recorded evidence rather than template names alone.Source: §C.1 · p. 17, §C.1.1 · p. 17

C.1.2 STRICT-PAIR CONSTRUCTION AND ACCEPTANCE CRITERIA

• Strict pairs match the complete model-visible observation and shared action while differing in structured state and validated semantic outcome; failed or unstable executions are rejected.Observation equality is checked after the model-specific input transformation, and incidental screenshot variation is not accepted as evidence of aliasing.Source: §C.1.2 · p. 17

C.2 TRAIN–EVALUATION SPLIT INTEGRITY

• All five train–evaluation leakage criteria had zero intersection, while shared low-level GUI states and action primitives were intentionally permitted across splits.The audit tested record, trajectory, target, content-fingerprint, and normalized history-trajectory identities.Source: §C.2 · p. 18

C.3 GROUND-TRUTH STRUCTURED STATE AND PROVENANCE

• Ground-truth states come from independent environment-side evidence, while the estimator receives only the family/schema instruction and model-visible history.Temporal Deadline state records lifecycle and effective remaining time, distinguishing visually similar timers that evolve differently under the same action.Source: §C.3 · pp. 18–19

C.4 FAMILY-SPECIFIC SEMANTIC VALIDATION

• Family-specific validators measure the semantic consequence of a shared probe using independent post-action evidence, rather than arbitrary screenshot inequality.Generated predictions are read from the rendered initial viewport using deterministic scoring without OCR or an LLM-based judge.Source: §C.4 · pp. 19–20
• Semantic validation requires branch-specific visible clipboard evidence, an exact ordered collection match, or internally consistent visible timer evidence, and a strict pair is correct only when both futures are recovered.A timer prediction must be grounded on a recognizable notification-shade surface; contradictory timer-state evidence is invalid.Source: §C.4 · pp. 20–21

C.5 COMPLETE MODEL-VISIBLE PAIR MATCHING

• Across all 800 strict pairs, Code2World and gWorld received identical canonical observation-only request hashes for branches A and B in 800/800 pairs each.The hash covers the complete client-side request payload; the branches nevertheless require different validated semantic futures.Source: §C.5 · pp. 21–22

C.6 PREDICTIVE-STATE INFORMATION ACCESS

• The estimator uses ordered GUI observations and actions through the current neutral observation, but excludes the prediction-boundary action and benchmark-side state evidence.Family-specific schemas specify the recovered state, while the same unified estimator parameters serve all families.Source: §C.6 · p. 22

C.7 TEMPORAL EXECUTION AND TIME INFORMATION

• For Temporal Deadline, the estimator may use user-visible timing evidence but not simulator timing state; audited logical-time advancement executes both branches and validates their distinct outcomes.Accepted strict pairs reach the same model-visible boundary and shared action while retaining different canonical timer states.Source: §C.7 · p. 22

C.8 REPEATED-EXECUTION STABILITY

• All 50 replayed branches had identical family-specific semantic outcomes across three executions, and all 150 executions reproduced expected outcomes under faithful trajectory reconstruction.The audit covered all 25 causal templates and observed no failed action, reset failure, capture failure, or within-branch semantic disagreement.Source: §C.8 · p. 23

C.9 REPRESENTATIVE STRICT-PAIR TRAJECTORIES

• Across four state families, different histories converge on the same prediction-boundary GUI, yet the same action produces different validated outcomes.The displayed histories include model-visible observations and actions; ground-truth structured states are shown separately as audit-only evidence, not as world-model input.Source: §C.9 · p. 23
• A file-move pair returns to the same empty search interface, but searching reveals whether the tracked file was moved.The branches differ in file-system state, despite their indistinguishable visible interface.Source: §C.9 · p. 23, Figure 3 · p. 24
• Two histories copy different clipboard payloads before reaching the same blank editor; the shared paste action inserts the branch-specific text.The clipboard contents are hidden in the common observation.Source: §C.9 · p. 23, Figure 4 · p. 24
• Different interaction histories construct different queue orders, which the same queue-opening action reveals from a shared Now Playing interface.Source: §C.9 · p. 23, Figure 5 · p. 24
• Different effective timer deadlines yield different timer-fire outcomes under the same controlled wait-and-probe action, although both branches reach the same Date & Time interface.The histories contain user-visible temporal evidence, while simulator-side logical-time state appears only in audit annotations.Source: §C.9 · p. 23, Figure 6 · p. 26

D.1 TRAINING DATA AND SPLIT CONSTRUCTION

• Training uses 2,538 examples across the four state families, with evaluation examples disjoint from training records and complete trajectories.Low-level GUI states, actions, and update primitives may recur compositionally across splits.Source: §D.1 · pp. 27–28, Table 9 · p. 28
• Training examples pair interaction histories with canonical structured-state targets, and evaluation tests composing update semantics across more complex histories.Training emphasizes atomic updates, while strict evaluation includes composed operations, interference, hidden suffixes, resets, and histories that converge on a common prediction boundary.Source: §D.1 · p. 27

D.2 INPUT REPRESENTATION AND STATE SCHEMAS

• The estimator receives temporally ordered GUI observations and historical actions plus a family-specific schema instruction, then generates the structured state autoregressively.The future shared action is excluded from the estimator input; the canonical state is the training target.Source: §D.1 · p. 27, §D.2 · p. 28

D.3 FAMILY-SPECIFIC SPECIALIST TEACHERS

• Four family-specific specialists use parameter-efficient adapters on a fixed 2B Qwen3.5 base model and also serve as teachers for unified estimation.The base model is frozen and only low-rank attention adapters are optimized.Source: §D.3 · pp. 29–30

D.4 UNIFIED MULTI-TEACHER DISTILLATION

• Multi-teacher distillation consolidates family-specific supervision into one estimator sharing a parameter set across all schemas.Each example receives its own family specialist’s soft supervision alongside the hard canonical-state target; teacher models are not needed at inference.Source: §D.4 · pp. 30–31

D.5 COMPARISON WITH DIRECT JOINT TRAINING

• 84.88% macro and 82.0% worst-family PairCorrect under MTD exceed Joint CE’s 83.63% macro and 69.5% worst-family PairCorrect, while specialists average 87.88%.Table 12 reports family-level BranchExact and PairCorrect; its average and worst values are computed from PairCorrect.Source: §D.5 · p. 31, Table 12 · p. 32
• 79.38% macro PairCorrect and 71.50% worst-family PairCorrect under MTD exceed Joint CE’s 76.75% and 69.50% in matched Code2World-8B evaluation.The improvement is not uniform: MTD improves Clipboard and modestly improves Persistent Artifact and Temporal Deadline, while Joint CE is stronger on Ordered Collection.Source: Table 13 · p. 32, §D.5 · p. 32
• On pairs where only one estimator recovers state correctly, final PairCorrect strongly favors that estimator; both methods score 0% when both recovered states are incorrect.On the 48 Joint CE-only pairs, scores are 93.75% for Joint CE and 22.92% for MTD; on the 58 MTD-only pairs, scores are 0% and 94.83%, respectively.Source: §D.5 · p. 32

D.6 COMPUTATIONAL COST

• Recovered-state conditioning adds an average 183.93 prompt tokens, while estimator generation and decoding takes a mean 0.768 seconds per transition.The prompt-token measure excludes the separate multimodal input to the predictive-state estimator; latency excludes input preparation and model initialization.Source: §D.6 · p. 33

E.1 DETERMINISTIC STATE CONDITIONING

• The deterministic interface normalizes the action, computes its semantic consequence from recovered state, selects a GUI surface, projects relevant context, and serializes the result for the frozen WM.The evaluated WMs were not trained to consume the structured state schemas directly.Source: §E.1 · p. 34
• Family rules preserve recovered payloads and collection order, resolve artifact presence for search, and compute timer outcomes from active status and remaining time.Source: §E.1 · pp. 34–35

E.2 MODEL-SPECIFIC SERIALIZATION

• Code2World-8B and gWorld receive the same recovered state and deterministic consequence but use different native request conventions.The gWorld temporal realization fixes a compact notification-shade layout; Code2World uses its fixed-canvas generation protocol.Source: §E.2 · p. 35, Table 15 · p. 36

E.3 PREDICTION-TIME INFORMATION PROVENANCE

• At prediction time, branch-dependent conditioning derives from recovered state, candidate action, context, and fixed rules—not future observations, labels, or ground-truth state.Source: E STATE-CONDITIONED WORLD-MODEL… · p. 33, §E.3 · p. 36

E.4 PROGRESSIVE REALIZATION OF THE STATE-CONDITIONING INTERFACE

• The Canonical-to-+Consequence and +Consequence-to-+Surface controls are byte-for-byte request-matched, passing audits for 400/400 branches in each evaluated model–family combination.All conditions use the same family-specific specialist predictions, holding the recovered-state source fixed.Source: §E.4 · pp. 36–37
• PairCorrect rises most from +Surface to Full: Persistent Artifact improves from 26.0% to 78.5% on Code2World-8B and 40.5% to 85.5% on gWorld-8B; Temporal Deadline improves from 0.5% to 84.0% and 22.5% to 88.5%.Full uses the production WM-native compilation path, so this comparison does not isolate a single component’s effect.Source: §E.4 · p. 37, Table 17 · p. 37

E.4.1 SEMANTIC AND SURFACE DECOMPOSITION OF THE INTERFACE GAIN

• For Full-interface gains, Temporal Deadline semantic corrections account for 311/319 Code2World and 186/193 gWorld gains; gWorld Persistent Artifact has 97/117, while Code2World Persistent Artifact is mixed: 135/200 semantic, 65/200 surface-only.The decomposition follows the existing scorer exactly and is not a counterfactual rescoring experiment.Source: §E.4.1 · p. 38

E.5 FULL-INTERFACE CONSTRUCTION AND REPRODUCIBILITY

• The Full interface deterministically combines a fixed family transition contract with a model-native realization of recovered state, resolved consequence, and target surface.For Persistent Artifact, the request includes the current image, an authoritative state/result block, and WM-native transition text; its templates instantiate fields deterministically rather than as per-example free-form prompts.Source: §E.5 · p. 39, §E.5.1 · pp. 39–40

E.5.1 EXACT MODEL-NATIVE REQUEST STRUCTURE

• Temporal requests insert a deterministically resolved post-action result into the model-native prompt, while the dedicated gWorld temporal request uses an authoritative result specification.The template selects the none/running/fired outcome without consulting a target observation or evaluation label; the frozen world model then renders the resolved consequence as a next UI.Source: §E.5.1 · pp. 42–43

E.5.2 REQUEST RECONSTRUCTION AUDIT

• 16/16 reconstructed request payloads matched archived SHA-256 hashes, and removing each added consequence or surface component recovered its preceding request in 400/400 branches per model–family combination.The archive stores hashes and prediction metadata rather than duplicate raw request bodies; the audit verifies the reconstructed specification at the original model-facing request boundary.Source: §E.5.2 · p. 43

E.6 RENDERING, SCORING, AND EVALUATION COVERAGE

• All 6,400 staged predictions had score records, with zero inference, rendering, scoring, or self-containedness failures and no silently removed or replaced branches.Strict semantic evidence is taken from the initial rendered viewport, excluding hidden, materially clipped, substantially occluded, and below-the-fold content.Source: §E.6 · p. 44

E.7 WORLD-MODEL CONFIGURATION

• The staged interface experiment uses frozen released world-model checkpoints and deterministic decoding, with configuration details reported in Table 21.The experiment separates predictive-state estimation from the deterministic interface that converts recovered state into a model-compatible post-action specification.Source: §E.7 · p. 44, Table 21 · p. 44, §E.8 · p. 44

E.8 INTERPRETATION AND SCOPE OF THE INTERFACE STUDY

• The largest gain appears only when recovered state is compiled through the complete WM-native interface, reflecting both state-dependent semantic correctness and faithful GUI-surface realization.The Full–Surface difference reflects the complete compilation path—including action-conditioned consequence resolution, surface integration, context projection, and native constraint placement—not any single internal component; the staged study holds recovered state fixed.Source: §E.8 · pp. 44–45

F ADDITIONAL ROBUSTNESS AND FIDELITY EVALUATION

• The additional evaluation protocols cover cross-application transfer, long-history stress, semantic interference, and next-observation fidelity while preserving family-specific state semantics.The appendix also specifies the frozen evaluation subset, rendering protocol, and feature metrics for fidelity measurement.Source: F ADDITIONAL ROBUSTNESS AND FIDELITY… · p. 45

F.1 EVALUATION PROTOCOL AND METRICS

• 79.75% average Exact under cross-application transfer is below 85.25% under long-history stress and 84.88% under semantic interference for the unified estimator.Strict uses PairCorrect and is reported separately, so its percentages are not a numerical baseline for these shift-suite comparisons.Source: §F.1 · p. 45, Table 22 · p. 46

F.2 CROSS-APPLICATION GENERALIZATION

• Cross-application evaluation changes application context while keeping the target state schema fixed; lower aggregate accuracy indicates context remains a meaningful shift, particularly for Persistent Artifact and Temporal Deadline.Persistent Artifact operations move from Android Files to Markor, while Ordered Collection and Temporal Deadline predictions end in unrelated applications.Source: §F.2 · pp. 46–47

F.3 LONG-HISTORY STRESS

• Long-history stress tests recovery of state evidence farther back in the interaction history, not multi-step predictive sufficiency of the canonical state.The controlled history depth varies through family-specific trajectory constructions.Source: §F.3 · p. 47

F.4 SEMANTIC INTERFERENCE

• Semantic interference fixes history depth at d = 8 while making distractor events increasingly similar to target-state updates.The final state schema and evaluation size remain fixed across interference levels.Source: §F.4 · pp. 47–48

F.5 NEXT-OBSERVATION FIDELITY

• Next-observation fidelity is measured alongside state-dependent semantic correctness to check whether state-sensitive gains sacrifice image-level fidelity.The comparison uses the same frozen 100 branches and recorded next observations for both conditioning methods.Source: §F.5 · p. 48, Table 23 · p. 49

G DOWNSTREAM EVALUATION IN ANDROIDWORLD

• AndroidWorld evaluation compares AGENT, AGENT+WM, and AGENT+WM+STATE on 116 tasks, with predictive state added only to the Code2World request.The recovered state's downstream effect is mediated through the predicted future GUI; Table 24 reports task-success results and the additional gain from state conditioning.Source: G DOWNSTREAM EVALUATION IN ANDROIDWORLD · p. 49, §G.1 · p. 50, Table 24 · p. 50

G.1 EXPERIMENTAL PROTOCOL

• AGENT+WM+STATE retains the AGENT+WM pipeline and adds recovered transition-relevant state before prediction; the state is not given directly to the candidate-generating agent or action selector.Code2World-8B supplies future-GUI predictions in both WM-assisted conditions.Source: G DOWNSTREAM EVALUATION IN ANDROIDWORLD · p. 49, §G.1 · p. 50

G.2 DOWNSTREAM RESULTS

• Predictive-state conditioning raises full-benchmark task success by +1.72 percentage points for GPT-5.4 and +2.58 percentage points for Qwen3-VL-8B over AGENT+WM.The final success rates are 60.34% and 47.41%, respectively, over 116 tasks.Source: §G.2 · p. 50
• The additional gains equal 66.4% of the WM-only improvement for GPT-5.4 and 74.8% for Qwen3-VL-8B.Source: §G.2 · p. 50
• For Qwen3-VL-8B, three tasks were rescued by state conditioning and none regressed, so the aggregate gain reflects a small number of additional solved tasks.Source: §G.2 · p. 50

G.3 REPRESENTATIVE RESCUE TRAJECTORY

• In FilesDeleteFile, standard prediction selects SWIPE UP after a near-name match, while predictive-state conditioning leads to SEARCH and successful deletion of the target.Both conditions share the same interaction prefix and current observation at the divergence point; recovered state is supplied only to the world model.Source: §G.3 · p. 50, Figure 7 · p. 51

• StateAliasBench covers Persistent Artifact, Clipboard, Ordered Collection, and Temporal Deadline, but these represent only a subset of latent state variables in interactive systems.Broader coverage of application-specific and cross-application state mechanisms remains necessary for a more comprehensive characterization.Source: H LIMITATIONS · p. 51
• The implemented family-specific state is an approximation to canonical predictive state and may not recover all transition-relevant information from interaction history.Residual representation error may remain even after state recovery.Source: H LIMITATIONS · p. 51
• Supplying transition-relevant state does not eliminate a frozen world model's approximation error, so state conditioning does not imply complete recovery of the transition law.Source: H LIMITATIONS · p. 51