ANNIE BURKE

DOCTORAL RESEARCH

Research question
FLORIDA INTERNATIONAL UNIVERSITY

Inferential Discovery and Design Adaptation in Generative Workflows

2024 — 2026
Research question

How does a designer’s relationship to their visual work in progress shift when inferential discovery can recur across a generative workflow?

The object of study is the designer’s response to inferential discovery. The direction and significance of any change remain open to investigation. Agency is examined through the designer’s responses and their consequences.

Inference names both the process of drawing a conclusion from evidence and the conclusion drawn. In his account of human reasoning with beliefs, Boghossian argues that one belief causing another is insufficient for inference: the person must take the former to support the latter.

In machine learning, inference names a computational process: applying a trained model to supplied input to produce a prediction or response. I do not take the use of ‘inference’ in computing as evidence that a machine reasons as a person does. Material produced through computational inference can provide grounds for the designer’s interpretation of the work. My interest is in how this material becomes significant for subjective continuation of design work.

Monochrome abstract image of translucent planes, intersecting folds, light and shadow.

In generative workflows, designers encounter an abundance of information in visual material produced or transformed through model inference.

Abundance includes the information available in generated material, the possibility of further discovery, and the range of responses available across blended practice. A blended practice combines generative computation with other ways of making and editing, within one application or across several. A designer can pursue something found in a generated result through another kind of operation. The response may develop the existing material, change the conditions of further production, or change the intention guiding the work. Further discovery can remain possible whichever response the designer chooses.

What the designer wants may develop through the encounter rather than remain a fixed standard against which outputs are assessed. A single encounter can matter.

More available material does not establish what matters to the designer. An undesirable result can occasion a consequential discovery; a discovery need not be adopted.

The designer makes an inferential discovery when determining the significance of an inference for continuation of visual work in progress.

Types and levels of automation describe which functions are automated and to what extent (Parasuraman, Sheridan and Wickens). Automating production does not, by itself, explain what the designer discovers or what follows from their response.

How the designer continues the work depends partly on their reasons and agency—their capacity to act and how they exercise it. The amount of automated production alone does not establish that capacity.

Earlier decisions shape what the designer encounters, and acting on an interpretation can require changes to work that depends on those decisions. What pursuing that finding requires depends partly on the possibilities for further work that remain after an operation. Those requirements can affect whether and how the designer continues.

A designer’s response can change what they encounter in later decisions. Further discovery can occur after work has come to depend on an earlier decision. Reconsidering that decision can then affect the work that depends on it.

Further relationships can become apparent through rearranging material already available. Further work may change what the designer understands and wants from it (Schön). A discovery can change the designer’s understanding without being incorporated into a finished artifact.

Crop A: luminous diagonal fold at the left of the image.
A
Crop B: dark overlapping planes at the centre of the image.
B
Crop C: luminous vertical fold at the right of the image.
C

Continuation concerns what becomes of the work as the designer responds to it, including the possibility of deciding not to proceed. The work can be sufficient for its purpose while further discovery remains possible. A decision to stop can therefore rest on what the work accomplishes. Looking only at a finished output would leave much of the designer’s response and its consequences unexplained.

Proposed contribution

The proposed vocabulary would help designers describe and compare their responses to inferential discovery. Documented examples would connect designers’ accounts of their decisions to the consequences in their work. Comparisons would retain the circumstances that explain those decisions and consequences.

Research proposal

Inference Resolution Map

The map compares what a designer specifies and controls with what conventional computation and trained models determine in the result of a software operation. Each assessment identifies the implementation, inputs and settings, traces how the result is produced, and explains which model predictions remain in it. These differences describe the conditions under which a designer encounters material that may matter for the continuation of their work.

Placement criteria

X = 15[1 − R((S+U+J)/30 + (p+c)/4)]

Scale 0 to 15.

Scope (S): what the model determines, from settings or relationships governing supplied material to reconstruction, restricted generation or generation throughout the result.

Latitude (U): what the learned task requires, from recognizing observed content to recovering missing information, choosing an interpretation, reconstructing plausible detail or constructing new content.

Coordination (J): the relationships the model resolves together, from independent or local predictions to relationships across supplied structures, synthesized scenes, time or views.

Each rating runs from 0 to 5.

R

1 when a learned decision remains in the returned result; 0 otherwise, giving X = 15 without counts

Later arithmetic that uses learned values retains their dependency.

P

Distinct active parameters, biases included, in the learned components whose output reaches the returned result for the stated configuration. Count shared parameters once. Exclude components that do not run or whose output does not reach the result. Include every learned component that contributes (for example text encoders, decoders and control networks). For a fitted representation, count the fitted parameters of the documented example scene or fit.

Counting a parameter is a capacity proxy; it does not measure how much that parameter changes the output.

C

Neural multiply-accumulates (matrix multiplications and convolutions, attention score and value products included; bias additions, activations, normalization, softmax, interpolation and memory traffic excluded) executed by those components for one invocation at the stated output resolution and step count, divided by the number of output RGB pixels.

p = log2(1 + P/P0) / [1 + log2(1 + P/P0)]

c = log2(1 + C/C0) / [1 + log2(1 + C/C0)]

P0 = 16,697,987 parameters. C0 = 1,120,428 convolution multiply-accumulates (MACs) per output RGB pixel. These are the original ESRGAN reference generator's quantities, used as a fixed, inspectable calibration reference.

Their selection is a proposed scale convention. At either reference value the corresponding transformed quantity is .5; doubling a quantity does not double its score.

This pixel-based calibration is not established for text, audio or geometric-model outputs; applying it across all catalogue media would need compatible declared benchmarks and validation.

The index does not measure a percentage of AI or the probability of inferential discovery.

0–5 rating definitions

Assess all learned processing that contributes to the result returned by this operation. Use the highest applicable scope rating. Unknown evidence remains unresolved.

S — scope

RatingDefinition
0No in-scope learned determination remains in the returned result.
1A learned category, scalar or bounded set of global settings; no dense field, structured relations or content construction is learned.
2A learned selection, constraint/dimension set, coordinate or transformation field governing supplied content or geometry, without learned reconstruction of its appearance or new scene content.
3A missing attribute channel or a fixed identified component/property is learned, while the surrounding observed instance defines the task: e.g. chroma, a face-recovery component or a driven mouth region. Evidence must establish this functional restriction.
4Reconstruct an observed instance or its fitted representation, or synthesize a completion under an effective restriction retaining source information outside the generative prediction. Reconstruction includes registered images and scene representations fitted to supplied observations. It does not assert exact pixel copying or unique recovery.
5The operation can generate the entities/parts and arrangement throughout the output domain, with no documented protection/reinstatement restricting retained generative predictions to a component or region. A prompt or reference alone does not establish protection.

U — inferential latitude

RatingDefinition
0Recognize or locate observed content, or emulate a specified fixed transformation; no restoration, preference assignment or new-content task is included.
1Restore a corrupted observed signal on its existing sampling support under a fidelity target.
2Estimate missing samples of an observed instance under a documented fidelity target, including increased image sampling. Do not infer a fidelity target merely because training examples exist.
3

Infer an unobserved attribute, coordinate value, preference or design-intent relation among feasible choices not uniquely specified by the supplied instance and rules; no new semantic scene construction is the task.

Latitude 3 applies when the model chooses which of several observed candidates the result is about, with no location from the person and no class named by the operation; selection of a named class and selection the person marks stay at latitude 0.

4Reconstruct plausible detail or appearance under a documented generative/perceptual/distributional role beyond fidelity-only recovery, or synthesize new content following an explicit supplied structural guide. Such a guide prescribes locations, shapes, poses, trajectories or event ordering for identifiable output parts, in words, coordinates or media; it is more than a high-level subject, style or action request. Evidence must establish that the assessed mechanism uses it. Guidance need not be hard protection. Predicting unobserved samples alone does not establish this anchor.
5Construct a new content/configuration instance without an observed recovery target or an explicit supplied structural guide. Semantic instructions and surrounding context can constrain the task without specifying its internal arrangement.

J — learned coordination

RatingDefinition
0Separable learned responses; no learned coordination between locations, components or states is established as part of the task.
1Learned local neighboring values/features are resolved together.
2

Learned relationships across extended regions or scales of registered supplied content are resolved together.

Coordination 2 applies to fitted predictors whose features are the output's own spatial coordinates, linear fits included, when the fitting happens inside the operation whose returned result is assessed.

3Integrate context across the supplied task instance or structure in learned predictions, including a face, a whole subject, a sketch or an observed scene. Establish this functional integration with product-linked task, mechanism or experimental evidence; an undisclosed architecture is not automatically zero. Record the reference domain.
4The learned task jointly constructs semantic and spatial/structural relationships in synthesized content. Mere consistent appearance in a screenshot does not establish the mechanism.
5A learned joint mechanism synthesizes evolving content relationships across time or views. Autoregressive internal tokens, conventional playback or repeated independent calls alone do not establish this.
Current X review table (CSV) ↓ Current X review records (JSON) ↓

Reviewed operations

ComfyUI masked inpainting

Proposed X 3.7

The designer supplies an image, prompt and mask. In the assessed configuration, the effective mask has a protected region where it is zero and an active region where it is positive. The sampler restores the source’s encoded representation in the protected region after each model prediction; generation remains in the active region. Learned encoding and decoding also contribute to the returned image, so this does not establish exact preservation of the original pixels. Protected regional generation and image reconstruction give S=4; new-content construction gives U=5. SD3.5 processes text and image representations together through attention, supporting joint semantic and spatial generation, J=4.

X = 15 − 13 × 13/15 ≈ 3.7.

ComfyUI sampler, SD3 architecture paper.

ComfyUI text-to-image

Proposed X 2.9

The designer specifies desired content through prompts and selects settings. SD3.5 processes text and image representations together through attention to generate content and arrangement throughout the image. Whole-image generation gives S=5; constructing a new image gives U=5; joint semantic and spatial generation gives J=4. This coordination rating distinguishes the still-image task from a learned mechanism that jointly generates relationships across time or views.

X = 15 − 13 × 14/15 ≈ 2.9.

SD3.5 Large model card, SD3 architecture paper.

Fusion AutoConstrain

Proposed X 8.1

The designer supplies sketch geometry and existing constraints, then chooses among candidate results. The model predicts missing constraints and dimensions; the conventional solver applies the selected predictions to the sketch. This assessed configuration has Modify Geometry off. The enabled mode requires a separate dependency account. Determining relationships on supplied geometry gives S=2; choosing a design-intent interpretation gives U=3; integrating relationships among the supplied primitives gives J=3. The predicted relationships remain in the updated sketch.

X = 15 − 13 × 8/15 ≈ 8.1.

Autodesk’s AutoConstrain research, AutoConstrain controls.

Map

Browse operations by X value

Each group lists operations at one proposed X value, with unresolved and awaiting-reassessment entries in separate groups. An unresolved entry identifies the information needed before a coordinate can be assigned.

Missing P or C leaves X unresolved.

Operations

Previous versions
4 October 2026 · X

Placement criteria

The map combines the 4 October review of operations previously at X = 2 with the earlier reference-based placements for other operations. Open an operation for its method, evidence and any unresolved requirement.

Decision-category review · 4 October 2026

The review examines the 145 operations previously placed at X = 2. It distinguishes properties carried from required inputs into generated output from properties determined by the trained model. Copied or untouched surroundings do not count as preservation within generated output.

PROPOSED 2–5 RULE · DOCUMENTATION REVIEW
X = 2 + 3 × kept categories ÷ (applicable categories − 1)

The six categories are content, layout and viewpoint, shape, surface appearance, fine detail, and motion and timing. Under this rule, more kept categories give a higher X when the number of applicable categories stays the same.

Kept requires evidence that generated output carries the property from a required input. Model decides requires evidence that the model determines it. Unknown records missing or conflicting evidence. Depends on setting records an assignment that varies with a documented setting or instruction. Not applicable requires an explanation of why the category does not apply to the assessed output.

Where applicability is established, the range shows the lowest and highest values permitted by unknown assignments and documented variation across settings. Each operation identifies which categories are unknown and which depend on settings. An unresolved output definition, disputed applicability, a zero denominator, or an endpoint outside the equation’s conditions prevents a complete coordinate. Where a lower bound is established, it is recorded without inventing an upper bound. These operations remain under “No coordinate,” with the requirement for resolving them.

The review yields 57 working points, 50 ranges and 38 unresolved records. These are results under a proposed rule; equal category weights and the 2–5 interval remain research choices. They do not quantify computation, effort, image area or a percentage of inference. Equal coordinates can still describe different decision profiles. The other operations retain their earlier placements.

This review checked documentation, not outputs from running every tool. Before disagreements were resolved, the v5 coders agreed on 88% of category assignments (κ = 0.835) and gave the same value or range for 99 of 145 operations. These figures describe the two coding passes before disagreements were resolved; they do not validate the coordinate formula.

The earlier reference-based method compares the model’s task with the reference operations below. Lower X records greater assessed responsibility for determining the result. The assessment covers a specified operation: Photoshop Select Subject, for example, is assessed before manual selection refinement.

REFERENCE BASED X EQUATION · EARLIER PLACEMENTS
X = XL + t(XH − XL)

This equation applies when trained-model predictions within the assessed software operation still affect its final output. XL and XH are the lower and higher comparison references. t, from 0 to 1, is the recorded expert judgment of position between them.

X = 15 when no trained-model prediction within the assessed software operation affects its final output. An unknown model contribution or unsupported comparison remains unresolved. An assessment can identify a range rather than a single coordinate.

X comparison cases and assigned coordinates
XComparison case
15No trained-model prediction affects the assessed software output
12HDRNet: fixed Local Laplacian operator emulation
10SAM v1: image segmentation
8DnCNN-S: image denoising at a specified noise level
5Original GAN-trained ESRGAN: reconstruction of plausible image detail
2Text-conditioned LDM: learned image synthesis

The reference order and numerical spacing are declared research conventions. For example, this ordering treats reconstruction of image appearance at X = 8 as greater model responsibility than selecting an image region at X = 10. That ordering is a comparison judgment; it is not determined by the model architecture.

Zero remains an unassigned boundary. The generation reference at 2 is a specified comparison, not a universal minimum or a score assigned merely because a prompt is present. Reference matches do not establish that a product uses the reference architecture.

Method and limitations

The assessment identifies what the model predicts, what it was trained to do, and how its prediction affects the result. It then explains which references provide the closest comparison and why that comparison fits better than a plausible alternative. The equation records that judgment as a coordinate or range.

Input amount, prompt detail and earlier asset generation do not add or subtract X points. X expresses the recorded comparison, rather than a percentage of AI use, a count of computations or the designer’s agency. Operations with the same X share a placement under these criteria; their controls, outputs and consequences for the designer can differ.

The earlier reference-based placements use published documentation. That review did not test each operation in software or establish agreement between independent reviewers. Some entries describe related operations or the same operation in different software hosts. Where a source could only be read through a search result or another page, the operation’s source note identifies that limitation.

A range preserves more than one supported comparison position; it is not a statistical confidence interval. Its recorded bounds remain visible without inserting a point to stand for the range. Conventional processing does not automatically reduce the assessed role of model predictions that still affect the result.

Comparison rules used in the 3 October review

Generative reconstruction of detail in supplied media uses the plausible-detail reference at X = 5. GAN and diffusion architectures do not receive different coordinates merely because their architectures differ.

Reconstruction intended to preserve the source while producing missing samples follows the Lightroom comparison: t = 0.75 between X = 5 and X = 8, displayed as 7.3, with sensitivity 6.5–8.

For language-model construction, the review uses 6–10 when the model selects or arranges existing elements, 2–6 when it creates new content such as text or layout geometry, and 2–10 for new logic or forms or an undocumented role. The division at 6 is a declared comparison convention.

For representations fitted to one captured scene without a pretrained network, the treatment of scene-specific training remains unresolved in these placement criteria. Those records retain unresolved X values.

Displayed X table (CSV) ↓ Coordinate review records (JSON) ↓

X coordinates for Photoshop and Lightroom operations

PHOTOSHOP · GENERATIVE FILL / FIREFLY

Image synthesis

X 2.0–3.5

Photoshop generates image content inside the selection. The v5 coding records content, layout and viewpoint, and surface appearance as model-decided; shape and fine detail remain unknown. These assignments give X = 2.0–3.5 under the proposed rule.

Required: establish whether the generated shape and fine detail are carried from the required input or determined by the model. This requires evidence for this operation and version, or a controlled test.

Previous coordinate explanation

This software operation matches the image-synthesis comparison at X = 2.

2 + 0 × (5 − 2) = 2

LIGHTROOM · SUPER RESOLUTION / RAW

Image super-resolution

X 7.3

Lightroom reconstructs a higher-resolution photograph. The recorded comparison gives its source-preserving task a position closer to denoising (X = 8) than plausible-detail reconstruction (X = 5): t = 0.75. This is the comparison judgment used to calculate the coordinate.

5 + 0.75 × (8 − 5) = 7.25

Sensitivity: t = 0.5–1 gives X = 6.5–8.

PHOTOSHOP · SELECT SUBJECT / CLOUD

Automatic subject selection

X 10.0

Photoshop predicts which pixels belong to the image’s subjects. The automatic selection matches the SAM image-segmentation comparison at X = 10. Manual refinement is assessed separately; this comparison does not identify Adobe’s model as SAM.

8 + 1 × (10 − 8) = 10

Map

X review · 4 October 2026

The reviewed operations use the proposed decision-category rule. Unresolved operations appear under “No coordinate”; their earlier placement at 2 is retained in each record’s history. How these coordinates are calculated

Browse operations by X value

Each group lists operations at one X value or across an X range. Expand a group to browse its operations.

Image-synthesis reference at 2 · No model prediction carried into the result at 15. “No coordinate” means the recorded evidence or rule does not yet support a placement.

Operations

September 29 data ↓

Historical X method · September 21–27

The superseded five-aspect equation calculated a provisional X coordinate from the five ratings of opportunity for inference.

HISTORICAL EQUATION / X
X = 15 − 3∑Oᵢ

Oᵢ is the assessed opportunity for inference in each of five aspects: subject, composition, form, appearance, and detail. Each is rated from 0 to 1, in provisional quarter steps. Higher opportunity places the operation farther left; lower opportunity places it farther right.

The sum ranges from 0 to 5. Multiplying by 3 maps it onto 0–15; subtracting from 15 preserves the map’s orientation. Each aspect contributes equally, with a maximum contribution of 3 coordinate units.

Worked calculation

ComfyUI text-to-image example: a scoped assessment, not a rating of every ComfyUI operation.

SubjectCompositionFormAppearanceDetail
0.751.000.750.501.00

∑Oᵢ = 4.00
X = 15 − (3 × 4.00) = 3.00

A change of 0.25 in one aspect changes X by 0.75. For example, lowering detail opportunity from 1.00 to 0.75 produces X = 3.75. This sensitivity calculation shows the consequence of a disputed rating; it does not change the plotted assessment.

Different profiles can have the same total. The five ratings and their rationales therefore remain available beside the coordinate.

Earlier equations

Initial placements

The initial survey used shared axis descriptions and reference examples to guide coordinate judgments. A calculation for every original decimal coordinate has not been recovered. Those placements are part of the instrument’s history.

Four-component proposals

Xautomation = (10 / 8)(T + G + C + R)

Task delegation, generative resolution, decision closure, and run/continuation autonomy were each rated 0–2.

Xspecification = (10 / 8)(D + R + C + U)

A later proposal used direct input, representation, constraint control, and retention of user decisions, again rated 0–2. Both convert a maximum sum of 8 to a 0–10 scale. With integer component ratings, possible totals are separated by 1.25. These are historical proposals, not the equation then used for revised points.

Five-aspect specification score · September 20

Xspecification = ∑Sᵢ   ·   Sᵢ ∈ {0, 1, 2, 3}

Subject, composition, form, appearance, and detail were scored by how they were supplied: unspecified, described in words, constrained by a guide or source, or explicitly constructed. Their sum gave a 0–15 specification score; the accompanying notes proposed 15 − X as opportunity for inference.

Five-aspect opportunity assessments · September 21 onward

The September 21 examples changed the assessment from supplied inputs to what generative AI resolves. They document the operation, evidence, rating rationale and sensitivity.

Judgment, consistency, and calibration

Here, the coordinate describes an operation’s production conditions; it is not a quality or taste score.

The arithmetic is repeatable once the five ratings are fixed. The ratings depend on the specified operation and inputs, evidence for each aspect, and the treatment of uncertainty. Independent agreement and calibration have not yet been established for this working rubric.

Equal weights and equally spaced quarter steps are provisional assumptions. They allow an explicit calculation, but do not establish that the aspects are equally consequential or that perceived differences are equally spaced. A decimal coordinate is not a percentage of AI contribution or a direct count of inferences.

The earlier website snapshot contained 81 revised X examples. This archived map retains its original Y assessments.

29 AUGUST 2026

Photoshop operation comparison

EARLIER CATALOGUE PLACEMENTS

Before X revisions

AUGUST 2026 / PROTOTYPE

Earlier map

1 October 2026 Y · Revision capacity

October 1 working Inference Resolution MapX runs from 0 to 15; Y runs from 0 to 100. Blue points show fixed coordinates. Vertical lines show Y ranges; horizontal lines show X ranges. Outlined areas show ranges on both axes. Select a mark to inspect its assessment.

X · Model contribution to software output
Greater model contribution ← → Less model contribution

Catalogue

Software revision capacity: Y

HISTORICAL Y EQUATION · VERSION 3 · 1 OCTOBER 2026
Y = 100 × (R + T + L) / 3

Each component runs from 0 to 1 and carries equal weight. The resulting coordinate runs from 0 to 100.

R / RECOVER SOFTWARE INPUTS, SETTINGS, AND EDITS

Recover content from before the software operation, reuse stored tool settings, and revisit its layer, node, or edit record. Score each before closing and after reopening. Higher scores require reuse or revision; the highest also keeps the current version.

T / EDITABLE SOFTWARE OUTPUT

The operation’s own controls can edit a whole output, selected area, separate component such as a layer or node, or one component property. The score records the finest supported control.

L / KEEP SUBSEQUENT SOFTWARE EDITS

Change an earlier input, setting, or edit without manually rebuilding subsequent software edits that depend on it.

Component calculations and uncertainty

R = sum of applicable recovery codes / (3 × number of applicable checks). The six checks cover original software inputs, stored tool settings, and editable operation records, each before closing and after reopening. Codes: 0 = documented loss; 1 = restored; 2 = available for reuse or revision; 3 = reuse or revision while retaining the current version.

T = target code / 3. Codes: 0 = whole software output; 1 = selected area or group; 2 = separate component; 3 = one component property.

L = yes answers / eligible checks for retaining subsequent software edits. Recovery checks reaching 2 or 3 are eligible. Unknown evidence contributes to lower and upper bounds; it is not recorded as a confirmed absence. Non-applicable recovery checks are excluded.

Vertical lines show Y bounds, horizontal lines show X comparison ranges, and outlined areas show both.

Y development

The documentation pilot examined recovery and reuse across different software operations. Recovery alone reached its maximum for 57 of the 80 assessed operations. The editable output and retention of subsequent software edits distinguish what can change and which edits remain applied afterward.

This version used an equal-weight composite. The October 1 model replaced the trial Rasch approach. Recovery remains valuable without treating historical distance or the number of undo steps as an automatic measure of value. A deliberate decision to nest or flatten work does not itself establish a lack of capacity.

October 1 coordinate data (CSV) ↓Full assessment data (JSON) ↓

Software operations and designers’ responses

The operation records describe how software produces a result and what the designer can subsequently edit or recover. A documented case can connect those conditions to what the designer encountered, their response to inferential discovery, and any consequences observed in the developing work.

Case documentation remains in development.

Next steps

2024 — 2025

Earlier research

2025

Modalities & spatial structures

Multimodal Spatial Intelligence
PRESENTATION / DECEMBER 2025Full size ↗
Multimodal Spatial Intelligence presentation, slide 1 of 21
1 / 21

Earlier work examined how different input modalities supply partial spatial information, how generative systems resolve that information, and how people interpret the resulting structures.

Read 2025 paper ↗
2024–25

Teaching, tools, and individual learning

While teaching architecture in Miami, I explored how people could learn to use tools through experiences tailored to them. My early research considered visualization and gamification as ways of making learning interactive, with climate education as an initial context. The emphasis subsequently moved toward personalized education and how learning could respond to individual needs.

As generative practice became a stronger focus, my attention shifted toward what designers encounter while working with these tools. The current dissertation focuses on the designer’s response to inferential discovery.

Research timeline
OCTOBER–NOVEMBER 2024

Visualization and interactive learning

Literature review and research-paper drafts examined 3D visualization, gamification, and AI-supported personalization in climate education.

DECEMBER 2024

Architectural education and AI

A Digital Futures paper considered prospective changes to architectural education, curricula, and relationships between educators and students.

MARCH 2025

Personalized learning

Personalized Learning in Higher Education: AI-Driven Analytics and Competency-Based Models as Catalysts for Reform placed adaptive learning and individual needs at the center of the inquiry.

Personal reading & sources

Listening is an important part of how I learn and develop my work. My reading and listening span technology, creativity, perception, culture, and lived experience.

A selection from my reading and listening, including Algorithms to Live By, Visual Thinking, Unmasking AI, music biographies, and cultural history
Browse titles

78 titles

  • Feeling BeautyG. Gabrielle Starr
  • RangeDavid Epstein
  • The Nvidia WayTae Kim
  • PeakAnders Ericsson and Robert Pool
  • A Paradise Built in HellRebecca Solnit
  • Algorithms to Live ByBrian Christian and Tom Griffiths
  • The Anxious GenerationJonathan Haidt
  • Building a StoryBrandDonald Miller
  • The Thinking MachineStephen Witt
  • The Infinity MachineSebastian Mallaby
  • Worldwide Evil and Misery – The Legacy of the 13 Satanic BloodlinesRobin de Ruiter and Fritz Springmeier
  • Life on EarthDavid Attenborough
  • Breaking the Habit of Being YourselfJoe Dispenza
  • Blue MindWallace J. Nichols
  • OceanDavid Attenborough and Colin Butfield
  • Africa Is Not a CountryDipo Faloyin
  • Exactly What to SayPhil M. Jones
  • GoogooshGoogoosh and Tara Dehlavi
  • QuantumManjit Kumar
  • Visual ThinkingTemple Grandin
  • The Let Them TheoryMel Robbins
  • Only God Can Judge MeJeff Pearlman
  • It Was All a DreamJustin Tinsley
  • Tupac ShakurStaci Robinson
  • Unmasking AIJoy Buolamwini
  • Energy Healing for AnimalsJoan Ranquet
  • How the Word Is PassedClint Smith
  • CasteIsabel Wilkerson
  • Re-RegulatedAnna Runkle
  • Miracle and Wonder (Special Edition)Malcolm Gladwell and Bruce Headlam
  • David and GoliathMalcolm Gladwell
  • Claim Your PowerMastin Kipp
  • Your Debt PlanRachel Rodgers
  • AbundanceEzra Klein and Derek Thompson
  • WordslutAmanda Montell
  • Reclaim Your Nervous SystemMastin Kipp
  • The Mountain Is YouBrianna Wiest
  • The Creative ActRick Rubin
  • BlinkMalcolm Gladwell
  • I Hate the Ivy LeagueMalcolm Gladwell
  • Data FeminismCatherine D’Ignazio and Lauren F. Klein
  • OutliersMalcolm Gladwell
  • SapiensYuval Noah Harari
  • ShowboatRoland Lazenby
  • Revenge of the Tipping PointMalcolm Gladwell
  • NexusYuval Noah Harari
  • The Tipping PointMalcolm Gladwell
  • What the Dog SawMalcolm Gladwell
  • Finding MeViola Davis
  • SupercommunicatorsCharles Duhigg
  • DuneFrank Herbert
  • The Deep Learning RevolutionTerrence J. Sejnowski
  • Weapons of Math DestructionCathy O’Neil
  • Invisible WomenCaroline Criado Perez
  • Stop Living on AutopilotAntonio Neves
  • It’s All in Your HeadRuss
  • The Broken LadderKeith Payne
  • Evolve Your BrainJoe Dispenza
  • The Four Horsemen
  • We Should All Be MillionairesRachel Rodgers
  • Steve JobsWalter Isaacson
  • Elon MuskAshlee Vance
  • Healing Your Attachment WoundsDiane Poole Heller
  • The AlchemistPaulo Coelho
  • The Body Keeps the ScoreBessel van der Kolk
  • The Autobiography of Malcolm XMalcolm X and Alex Haley
  • Shoe DogPhil Knight
  • WillWill Smith and Mark Manson
  • Born a CrimeTrevor Noah
  • JAY-ZMichael Eric Dyson
  • Dark PsychologyValerie Glossner
  • MetahumanDeepak Chopra
  • Atomic HabitsJames Clear
  • AttachedAmir Levine and Rachel Heller
  • Trillion Dollar Coach
  • The Spark and the GrindErik Wahl
  • Summary: 12 Rules for LifeEpicread
  • Pitch AnythingOren Klaff
Sources

112 sources

  • Ameisen, E., Lindsey, J., Pearce, A., Gurnee, W., Turner, N. L., Chen, B., Citro, C., Abrahams, D., Carter, S., Hosmer, B., Marcus, J., Sklar, M., Templeton, A., Bricken, T., McDougall, C., Cunningham, H., Henighan, T., Jermyn, A., Jones, A., … Batson, J. (2025). Circuit tracing: Revealing computational graphs in language models. Transformer Circuits Thread. https://transformer-circuits.pub/2025/attribution-graphs/methods.html
  • Bahmani, S., Park, J. J., Paschalidou, D., Yan, X., Wetzstein, G., Guibas, L. J., & Tagliasacchi, A. (2023). CC3D: Layout-conditioned generation of compositional 3D scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision (pp. 7137–7147). https://doi.org/10.1109/ICCV51070.2023.00659
  • Bar-Tal, O., et al. (2024). Lumiere: A space-time diffusion model for video generation. ArXiv.org.https://arxiv.org/abs/2401.12945
  • Barron, J. T., et al. (2021). Mip-NeRF: A multiscale representation for anti-aliasing neural radiance fields. ArXiv.org. https://arxiv.org/abs/2103.13415
  • Batra, H., Tu, H., Chen, H., Lin, Y., Xie, C., & Clark, R. (2025). SpatialThinker: Reinforcing 3D reasoning in multimodal LLMs via spatial rewards [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2511.07403
  • Behrens, T. E. J., Muller, T. H., Whittington, J. C. R., Mark, S., Baram, A. B., Stachenfeld, K. L., & Kurth-Nelson, Z. (2018). What is a cognitive map? Organizing knowledge for flexible behavior. Neuron, 100(2), 490–509. https://doi.org/10.1016/j.neuron.2018.10.002
  • Bowker, G. C., & Star, S. L. (1999). Sorting things out: Classification and its consequences. MIT Press.
  • Bruce, J., et al. (2024). Genie: Generative interactive environments. ArXiv.org. https://arxiv.org/abs/2402.15391
  • Chattopadhyay, A., Zhang, X., Wipf, D. P., Arora, H., & Vidal, R. (2023). Learning graph variational autoencoders with constraints and structured priors for conditional indoor 3D scene generation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (pp. 785–794). https://doi.org/10.1109/WACV56688.2023.00085
  • Chen, B., Xu, Z., Kirmani, S., Ichter, B., Sadigh, D., Guibas, L., & Xia, F. (2024). SpatialVLM: Endowing vision-language models with spatial reasoning capabilities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 14455–14465). https://doi.org/10.1109/CVPR52733.2024.01370
  • Chen, S., Zhu, T., Zhou, R., Zhang, J., Gao, S., Niebles, J. C., Geva, M., He, J., Wu, J., & Li, M. (2025). Why is spatial reasoning hard for VLMs? An attention mechanism perspective on focus areas. In A. Singh, M. Fazel, D. Hsu, S. Lacoste-Julien, F. Berkenkamp, T. Maharaj, K. Wagstaff, & J. Zhu (Eds.), Proceedings of the 42nd International Conference on Machine Learning (Vol. 267, pp. 9910–9932). PMLR. https://proceedings.mlr.press/v267/chen25cr.html
  • Chen, Y., Nguyen, H. T., Voleti, V., Jampani, V., & Jiang, H. (2025). HouseCrafter: Lifting floorplans to 3D scenes with 2D diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (pp. 28440–28450).
  • Cheng, A.-C., Yin, H., Fu, Y., Guo, Q., Yang, R., Kautz, J., Wang, X., & Liu, S. (2024). SpatialRGPT: Grounded spatial reasoning in vision-language models. In A. Globerson, L. Mackey, D. Belgrave, A. Fan, U. Paquet, J. Tomczak, & C. Zhang (Eds.), Advances in Neural Information Processing Systems (Vol. 37, pp. 135062–135093). Curran Associates, Inc. https://doi.org/10.52202/079017-4293
  • Chiriatti, M., Ganapini, M., Panai, E., Ubiali, M., & Riva, G. (2024). The case for human–AI interaction as system 0 thinking. Nature Human Behaviour, 8(11), 1995–1997. https://doi.org/10.1038/s41562-024-01995-5
  • Christian, B., & Griffiths, T. (2016). Algorithms to live by: The computer science of human decisions. Henry Holt and Company.
  • Clark, A. (1997). Being there: Putting brain, body, and world together again. MIT Press.
  • Daston, L., & Galison, P. (2007). Objectivity. Zone Books.
  • Deng, A., Cao, T., Chen, Z., & Hooi, B. (2025). Words or vision: Do vision-language models have blind faith in text? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. https://arxiv.org/abs/2503.02199
  • Doshi-Velez, F., & Kim, B. (2017). Towards a rigorous science of interpretable machine learning [Preprint]. arXiv. https://arxiv.org/abs/1702.08608
  • Dourish, P. (2001). Where the action is: The foundations of embodied interaction. MIT Press.
  • Durante, Z., et al. (2025). Agent AI: Surveying the horizons of multimodal interaction. ArXiv.org. https://arxiv.org/abs/2504.01512
  • Erdem, U., et al. (2019). Applications of spatial and temporal reasoning in cognitive robotics. Cognitive Robotics Lab.
  • Eslami, S. M. A., et al. (2018). Neural scene representation and rendering. Science, 360(6394), 1204–1210. https://doi.org/10.1126/science.aar6170
  • Feng, M., Hou, H., Zhang, L., Wu, Z., Guo, Y., & Mian, A. (2023). 3D spatial multimodal knowledge accumulation for scene graph prediction in point cloud. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 9182–9191). https://doi.org/10.1109/CVPR52729.2023.00886
  • Gao, J., et al. (2021). Dynamic view synthesis from dynamic monocular video. ArXiv.org. https://arxiv.org/abs/2105.06468
  • Garcin, S., Walker, T., McDonagh, S., Pearce, T., Bilen, H., He, T., Wang, K., & Bian, J. (2026). Beyond pixel histories: World models with persistent 3D state [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2603.03482
  • Gibson, J. J. (1979). The ecological approach to visual perception. Houghton Mifflin.
  • Gibson, J. J. (1979). The theory of affordances. In The ecological approach to visual perception (pp. 127–143).
  • Goodwin, C. (1994). Professional vision. American Anthropologist, 96(3), 606–633.
  • Green, T. R. G., & Petre, M. (1996). Usability analysis of visual programming environments: A ‘cognitive dimensions’ framework. Journal of Visual Languages & Computing, 7(2), 131–174. https://doi.org/10.1006/jvlc.1996.0009
  • Ha, D., & Schmidhuber, J. (2018). World models. ArXiv.org. https://arxiv.org/abs/1803.10122
  • Hafner, D., et al. (2023). Mastering diverse domains through world models. ArXiv.org. https://arxiv.org/abs/2311.01460
  • Hassabis, D., Kumaran, D., Summerfield, C., & Botvinick, M. (2017). Neuroscience-inspired artificial intelligence. Neuron, 95(2), 245–258. https://doi.org/10.1016/j.neuron.2017.06.011
  • Herweijer, C. (2026, July 15). World models are AI's next frontier. TIME.
  • Hu, S., et al. (2024). Simulating the real world: A unified survey of multimodal generative models. ArXiv.org. https://arxiv.org/abs/2405.14034
  • Hu, Y., et al. (2025). EWMBENCH: Evaluating scene, motion, and semantic quality in embodied world models. ArXiv.org. https://arxiv.org/abs/2505.09694
  • Huang, Z., et al. (2024). EnerVerse: Envisioning embodied future space for robotic manipulation. ArXiv.org. https://arxiv.org/abs/2407.08768
  • Hutchins, E. (1995). Cognition in the wild. MIT Press.
  • Jewitt, C. (Ed.). (2009). The Routledge handbook of multimodal analysis. Routledge.
  • Jiang, L., et al. (2024). EnerVerse-AC: Envisioning embodied environments with action-conditioned world models. ArXiv.org. https://arxiv.org/abs/2407.08769
  • Kerbl, B., et al. (2023). 3D Gaussian splatting for real-time radiance field rendering. ArXiv.org. https://arxiv.org/abs/2303.13495
  • Khan, H., & Asif, S. (2026). GenCtrl: A formal controllability toolkit for generative models [Preprint]. arXiv. https://arxiv.org/abs/2601.05637
  • Kirsh, D. (1995). The intelligent use of space. Artificial Intelligence, 73(1–2), 31–68.
  • Kress, G., & van Leeuwen, T. (2001). Multimodal discourse: The modes and media of contemporary communication. Arnold.
  • Kumar, A., et al. (2019). Consistent generative query networks. ArXiv.org. https://arxiv.org/abs/1807.02033
  • Lee, H.-P., Sarkar, A., Tankelevitch, L., Drosos, I., Rintel, S., Banks, R., & Wilson, N. (2025). The impact of generative AI on critical thinking: Self-reported reductions in cognitive effort and confidence effects from a survey of knowledge workers. In Proceedings of the CHI Conference on Human Factors in Computing Systems. ACM. https://doi.org/10.1145/3706598.3713778
  • Li, C., Peh, E., & Fernando, B. (2025). HMR3D: Hierarchical multimodal representation for 3D scene understanding with large vision-language model [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2511.22961
  • Li, C., Wu, W., Zhang, H., Xia, Y., Mao, S., Dong, L., Vulić, I., & Wei, F. (2025). Imagine while reasoning in space: Multimodal visualization-of-thought. In A. Singh, M. Fazel, D. Hsu, S. Lacoste-Julien, F. Berkenkamp, T. Maharaj, K. Wagstaff, & J. Zhu (Eds.), Proceedings of the 42nd International Conference on Machine Learning (Vol. 267, pp. 36340–36364). PMLR. https://proceedings.mlr.press/v267/li25cz.html
  • Li, C., Zhang, C., Zhou, H., Collier, N., Korhonen, A., & Vulić, I. (2024). TopViewRS: Vision-language models as top-view spatial reasoners. In Y. Al-Onaizan, M. Bansal, & Y.-N. Chen (Eds.), Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing (pp. 1786–1807). Association for Computational Linguistics. https://doi.org/10.18653/v1/2024.emnlp-main.106
  • Li, M., Xie, C., Wu, Y., Zhang, L., & Wang, M. (2025). FiVE-Bench: A fine-grained video editing benchmark for evaluating emerging diffusion and rectified flow models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (pp. 16672–16681).
  • Liapis, A., Yannakakis, G. N., Alexopoulos, C., & Lopes, P. (2016). Can computers foster human users' creativity? Theory and praxis of mixed-initiative co-creativity. Digital Culture & Education, 8(2), 136–153.
  • Lin, J., Zhu, C., Xu, R., Mao, X., Liu, X., Wang, T., & Pang, J. (2025). OST-Bench: Evaluating the capabilities of MLLMs in online spatio-temporal scene understanding [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2507.07984
  • Lin, Z., Ehsan, U., Agarwal, R., Dani, S., Vashishth, V., & Riedl, M. (2023). Beyond prompts: Exploring the design space of mixed-initiative co-creativity systems [Preprint]. arXiv. https://arxiv.org/abs/2305.07465
  • Ling, L., Ge, Y., Sheng, Y., & Bera, A. (2025). I-Scene: 3D instance models are implicit generalizable spatial learners [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2512.13683
  • Linghu, X., Huang, J., Niu, X., Ma, X., Jia, B., & Huang, S. (2024). Multi-modal situated reasoning in 3D scenes. In A. Globerson, L. Mackey, D. Belgrave, A. Fan, U. Paquet, J. Tomczak, & C. Zhang (Eds.), Advances in Neural Information Processing Systems (Vol. 37, pp. 140903–140936). Curran Associates, Inc. https://doi.org/10.52202/079017-4473
  • Linghu, X., Huang, J., Zhu, Z., Jia, B., & Huang, S. (2026). SceneCOT: Eliciting grounded chain-of-thought reasoning in 3D scenes. In Proceedings of the International Conference on Learning Representations. https://openreview.net/forum?id=U9meoc0Sau
  • Lipton, Z. C. (2018). The mythos of model interpretability. Queue, 16(3), 31–57. https://doi.org/10.1145/3236386.3241340
  • Liu, R., Wu, R., Van Hoorick, B., Tokmakov, P., Zakharov, S., & Vondrick, C. (2023). Zero-1-to-3: Zero-shot one image to 3D object. In Proceedings of the IEEE/CVF International Conference on Computer Vision (pp. 9264–9275). https://doi.org/10.1109/ICCV51070.2023.00853
  • Liu, Y., et al. (2024). Aligning cyberspace with the physical world: A comprehensive survey on embodied AI. IEEE TPAMI.
  • Liu, Y., Li, X., Zhang, Y., Qi, L., Li, X., Wang, W., Li, C., Li, X., & Yang, M.-H. (2025). Controllable 3D outdoor scene generation via scene graphs. In Proceedings of the IEEE/CVF International Conference on Computer Vision (pp. 28052–28062).
  • Long, X., Guo, Y.-C., Lin, C., Liu, Y., Dou, Z., Liu, L., Ma, Y., Zhang, S.-H., Habermann, M., Theobalt, C., & Wang, W. (2024). Wonder3D: Single image to 3D using cross-domain diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 9970–9980). https://doi.org/10.1109/CVPR52733.2024.00951
  • Lu, Y., Luo, W., Tu, P., Li, H., Zhu, H., Yu, Z., Wang, X., Chen, X., Peng, X., Li, X., & Chen, Z. (2025). 4DWorldBench: A comprehensive evaluation framework for 3D/4D world generation models [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2511.19836
  • Lyu, R., Lin, J., Wang, T., Yang, S., Mao, X., Chen, Y., Xu, R., Huang, H., Zhu, C., Lin, D., & Pang, J. (2024). MMScan: A multi-modal 3D scene dataset with hierarchical grounded language annotations. In A. Globerson, L. Mackey, D. Belgrave, A. Fan, U. Paquet, J. Tomczak, & C. Zhang (Eds.), Advances in Neural Information Processing Systems (Vol. 37, pp. 50898–50924). Curran Associates, Inc. https://doi.org/10.52202/079017-1611
  • Ma, W., Chen, H., Zhang, G., Chou, Y.-C., Chen, J., de Melo, C., & Yuille, A. (2025). 3DSRBench: A comprehensive 3D spatial reasoning benchmark. In Proceedings of the IEEE/CVF International Conference on Computer Vision (pp. 6924–6934).
  • Martin-Brualla, R., et al. (2021). NeRF in the Wild: Neural radiance fields for unconstrained photo collections. ArXiv.org. https://arxiv.org/abs/2008.02268
  • Mildenhall, B., et al. (2020). NeRF: Representing scenes as neural radiance fields for view synthesis. ArXiv.org. https://arxiv.org/abs/2003.08934
  • Morris, B. (2026, July 2). AI is 'not smart' so what's next in artificial intelligence? BBC News. https://www.bbc.com/news/articles/cj6gr0xkyr3o
  • Naik, S., Shukla, P., Obi, I., Backus, J., Rasche, N., & Parsons, P. (2025). Tracing the invisible: Understanding students' judgment in AI-supported design work. In Creativity and Cognition (C&C '25). ACM. https://doi.org/10.1145/3698061.3734399
  • Nguyen, T., Michaels, J., Fiterau, M., & Jensen, D. (2025). Challenges in understanding modality conflict in vision-language models [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2509.02805
  • Nooralahzadeh, F., Rohanian, O., Zhang, Y., Fürst, J., & Stockinger, K. (2026). Arbitration failure, not perceptual blindness: How vision-language models resolve visual-linguistic conflicts [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2604.09364
  • O’Keefe, J., et al. (1978). The hippocampus as a cognitive map. Oxford University Press.
  • Parasuraman, R., Sheridan, T. B., & Wickens, C. D. (2000). A model for types and levels of human interaction with automation. IEEE Transactions on Systems, Man, and Cybernetics—Part A: Systems and Humans, 30(3), 286–297. https://doi.org/10.1109/3468.844354
  • Park, J., Jang, K. J., Alasaly, B., Mopidevi, S., Zolensky, A., Eaton, E., Lee, I., & Johnson, K. (2025). Assessing modality bias in video question answering benchmarks with multimodal large language models. Proceedings of the AAAI Conference on Artificial Intelligence, 39(19), 19821–19829. https://doi.org/10.1609/aaai.v39i19.34183
  • Park, K., et al. (2021). HyperNeRF: A higher-dimensional representation for topologically varying neural radiance fields. ArXiv.org. https://arxiv.org/abs/2103.16376
  • Poole, B., et al. (2022). DreamFusion: Text-to-3D using 2D diffusion. ArXiv.org. https://arxiv.org/abs/2209.14988
  • Qin, C., et al. (2024). WorldSimBench: Towards video generation models as world simulators. ArXiv.org. https://arxiv.org/abs/2403.12031
  • Rezwana, J., & Maher, M. L. (2023). Designing creative AI partners with COFI: A framework for modeling interaction in human-AI co-creative systems. ACM Transactions on Computer-Human Interaction, 30(5), 1–28. https://doi.org/10.1145/3519026
  • Rombach, R., et al. (2022). High-resolution image synthesis with latent diffusion models. ArXiv.org. https://arxiv.org/abs/2112.10752
  • Sajjadi, M. S. M., et al. (2022). OSRT: Object scene representation transformer. NeurIPS. https://osrt-paper.github.io
  • Sajjadi, M. S. M., et al. (2022). Scene representation transformer: Geometry-free novel view synthesis through set-latent scene representations. CVPR. https://srt-paper.github.io
  • Sargent, K., Li, Z., Shah, T., Herrmann, C., Yu, H.-X., Zhang, Y., Chan, E. R., Lagun, D., Li, F.-F., Sun, D., & Wu, J. (2024). ZeroNVS: Zero-shot 360-degree view synthesis from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 9420–9429). https://doi.org/10.1109/CVPR52733.2024.00900
  • Sarkar, A. (2024). Intension is the new attention: A cognitive perspective on natural language interaction with generative AI. In Proceedings of the CHI Conference on Human Factors in Computing Systems. ACM.
  • Sheng, Z., Han, X., Zhang, Z., Xiong, Z., Ding, Y., Ping, A., Li, X., Guo, T., & Mao, Y. (2026). InEdit-Bench: Benchmarking intermediate logical pathways for intelligent image editing models [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2603.03657
  • Shi, Y., Wang, P., Ye, J., Mai, L., Li, K., & Yang, X. (2024). MVDream: Multi-view diffusion for 3D generation. In The Twelfth International Conference on Learning Representations. https://openreview.net/forum?id=FUgrjq2pbB
  • Sim, M. Y., Zhang, W. E., Dai, X., & Fang, B. (2025). Can VLMs actually see and read? A survey on modality collapse in vision-language models. In Findings of the Association for Computational Linguistics: ACL 2025 (pp. 24452–24470). Association for Computational Linguistics. https://aclanthology.org/2025.findings-acl.1256/
  • Spelke, E. S., et al. (2007). Core knowledge. Developmental Science. https://doi.org/10.1111/j.1467-7687.2007.00569.x
  • Stogiannidis, I., McDonagh, S., & Tsaftaris, S. A. (2025). Mind the gap: Benchmarking spatial reasoning in vision-language models [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2503.19707
  • Suchman, L. A. (1987). Plans and situated actions: The problem of human-machine communication. Cambridge University Press.
  • Sun, Y., Jang, E., Ma, F., & Wang, T. (2024). Generative AI in the wild: Prospects, challenges, and strategies. In Proceedings of the CHI Conference on Human Factors in Computing Systems (pp. 1–16). ACM. https://arxiv.org/abs/2404.04101
  • Suresh, H., & Guttag, J. (2021). A framework for understanding sources of harm throughout the machine learning life cycle. In Equity and Access in Algorithms, Mechanisms, and Optimization (EAAMO '21) (pp. 1–9). ACM. https://doi.org/10.1145/3465416.3483305
  • Tankelevitch, L., Kewenig, V., Simkute, A., Scott, A. E., Sarkar, A., Sellen, A., & Rintel, S. (2024). The metacognitive demands and opportunities of generative AI. In Proceedings of the CHI Conference on Human Factors in Computing Systems. ACM. https://doi.org/10.1145/3613904.3642902
  • Tewari, A., et al. (2020). State of the art on neural rendering. ArXiv.org. https://arxiv.org/abs/2004.04776
  • Tversky, B. (2011). Visualizing thought. Topics in Cognitive Science, 3(3), 499–535. https://doi.org/10.1111/j.1756-8765.2010.01113.x
  • Ullman, S. (1984). Visual routines. Cognition.
  • Van de Maele, T., Dhoedt, B., Verbelen, T., & Pezzulo, G. (2024). A hierarchical active inference model of spatial alternation tasks and the hippocampal-prefrontal circuit. Nature Communications, 15, Article 9892. https://doi.org/10.1038/s41467-024-54257-3
  • Wang, C., Zhou, Y., Wang, Q., Wang, Z., & Zhang, K. (2025). ComplexBench-Edit: Benchmarking complex instruction-driven image editing via compositional dependencies. In Proceedings of the 33rd ACM International Conference on Multimedia (pp. 13391–13397).
  • Wang, Q., Zhang, Y., Holynski, A., Efros, A. A., & Kanazawa, A. (2025). Continuous 3D perception model with persistent state. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. https://arxiv.org/abs/2501.12387
  • Wang, Z., et al. (2025). SITE: Towards spatial intelligence through evaluation. ArXiv.org. https://arxiv.org/abs/2503.09733
  • Whittington, J. C. R., Muller, T. H., Mark, S., Chen, G., Barry, C., Burgess, N., & Behrens, T. E. J. (2020). The Tolman-Eichenbaum machine: Unifying space and relational memory through generalization in the hippocampal formation. Cell, 183(5), 1249–1263.e23. https://doi.org/10.1016/j.cell.2020.10.024
  • Wu, G., et al. (2024). 4D Gaussian splatting for real-time dynamic scene rendering. CVPR. https://guanjunwu.github.io/4dgs
  • Wu, M., Ji, J., Huang, O., Li, J., Wu, Y., Sun, X., & Ji, R. (2024). Evaluating and analyzing relationship hallucinations in large vision-language models. In R. Salakhutdinov, Z. Kolter, K. Heller, A. Weller, N. Oliver, J. Scarlett, & F. Berkenkamp (Eds.), Proceedings of the 41st International Conference on Machine Learning (Vol. 235, pp. 53553–53570). PMLR. https://proceedings.mlr.press/v235/wu24l.html
  • Wu, X., Liang, D., Feng, T., Xia, K., Zhang, Y., Li, X., Tan, X., & Bai, X. (2026). Generation models know space: Unleashing implicit 3D priors for scene understanding [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2603.19235
  • Yang, J., Yang, S., Gupta, A. W., Han, R., Li, F.-F., & Xie, S. (2025). Thinking in space: How multimodal large language models see, remember, and recall spaces. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 10632–10643). https://doi.org/10.1109/CVPR52734.2025.00994
  • Yang, M., et al. (2024). Cambrian-S: Towards spatial supersensing in video. ArXiv.org. https://arxiv.org/abs/2403.12843
  • Yannakakis, G. N., Liapis, A., & Alexopoulos, C. (2014). Mixed-initiative co-creativity. In Proceedings of the 9th International Conference on the Foundations of Digital Games.
  • Yu, A., et al. (2021). PlenOctrees for real-time rendering of neural radiance fields. ArXiv.org. https://arxiv.org/abs/2103.14024
  • Yuan, D., Graham, C., Yakura, H., & Lim, A. (2026). Using the blackbox in embodied AI art practice: Uncertainty at the interface. In Designing Interactive Systems Conference (DIS ’26) (pp. 4181–4201). ACM. https://doi.org/10.1145/3800645.3813020
  • Zhang, W., Huang, Y., Xu, Y., Huang, J., Zhi, H., Ren, S., Xu, W., & Zhang, J. (2025). Why do MLLMs struggle with spatial understanding? A systematic analysis from data to architecture [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2509.02359
  • Zhang, Z., Zhou, W., Zhao, J., & Li, H. (2025). Robust multimodal large language models against modality conflict. In Proceedings of the 42nd International Conference on Machine Learning. https://arxiv.org/abs/2507.07151
  • Zhao, R., Zhang, Z., Xu, J., Chang, J., Chen, D., Li, L., Sun, W., & Wei, Z. (2025). SpaceMind: Camera-guided modality fusion for spatial reasoning in vision-language models [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2511.23075
  • Zhao, X., et al. (2024). Pseudo-generalized dynamic view synthesis from a video. ICLR. https://arxiv.org/abs/2310.08587
  • Zhen, X., et al. (2024). 3D-VLA: A 3D vision-language-action generative world model. ArXiv.org. https://arxiv.org/abs/2403.15952

August 2026 · Archived prototype

 August 29 presentation figure: inputs supplied versus recurrence of inference, with highlighted Photoshop operations

29 August 2026 · Presentation figure