Pith. sign in

REVIEW 3 major objections 5 minor 23 references

fog turns motion and emotion into composable AI-generated code functions that people recognize at 68% accuracy.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-10 14:48 UTC pith:X52B5CNL

load-bearing objection Solid HCI systems paper: OOP hierarchy for LLM-generated composable motion/emotion functions, with clean 4AFC recognition at 68% and faster iteration than pure prompting. the 3 major comments →

arxiv 2607.07952 v1 pith:X52B5CNL submitted 2026-07-08 cs.HC cs.CL

fog: Expressing Motion and Emotion through Function Composition of AI-Generated Code

Vivian Liu , Lydia Chilton This is my paper
classification cs.HC cs.CL
keywords motionemotionanimationfunction compositiongenerative AIcode generationHeider-Simmelhuman-AI interaction
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper argues that motion and emotion can be captured as a hierarchy of abstract classes (verbs, adverbs, gestures, emotions) that large language models implement as code. These functions compose over an entity's state (velocity, energy, paths, collisions) to produce Heider-Simmel-style animations whose semantic meaning people can read. In a perceptual test of 452 animations, recognition reached 68% exact-match accuracy in a four-alternative forced-choice task—2.68 times chance. An accompanying editor lets users generate new functions on the fly, compose them, draw paths, and tune parameters via direct manipulation and dynamically generated UI. A mixed-methods study with professionals and novices found the interface supported faster iteration and greater control than a strong prompt-only baseline.

Core claim

Abstract class contracts for primary and secondary motion let AI generate an open-ended vocabulary of verbs, adverbs, gestures, and emotions whose compositions produce recognizable affective and social motion in simple shape animations, and the same contracts power an interactive editor that shortens feedback cycles while giving users structured control.

What carries the argument

fog's hierarchy of abstract classes (PrimaryMotion / SecondaryMotion, Verb, Adverb, Gesture, Emotion) that define update, decorate, and customDraw contracts so generated functions compose additively over shared channels (speed, energy, paths, particles).

Load-bearing premise

That one-shot LLM implementations of these class contracts, tested on a fixed set of stimulus words and simple 2-D shapes, produce motion signatures whose recognizability generalizes beyond the specific generations and limited channels used in the study.

What would settle it

Re-run the same 4AFC protocol on a fresh set of fog-generated animations for the identical stimulus words; if exact-match accuracy falls to chance (25%) or the user-study latency advantage disappears, the central claim fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Users can expand a motion vocabulary on demand by prompting new verbs, adverbs, gestures, or emotions that inherit the same contracts.
  • Function composition (adverb(verb), emotion+verb, gesture+verb) becomes a controllable editing primitive rather than opaque prompt text.
  • Direct-manipulation minimaps and path drawing can be auto-generated from the channels each composition exposes.
  • Recognition rates near 68% supply a practical benchmark for any future code-generation system that claims to produce readable motion.
  • The same primary/secondary split can be reused for other domains that need layered, intentional behavior (robotics, game AI).

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If channel alignment between primary and secondary functions remains the main failure mode, a lightweight static analysis pass before generation could raise composition accuracy without further self-refinement.
  • Event-based rather than duration-based states would likely unlock bidirectional social dynamics the paper notes as currently limited.
  • The same abstract-class surface could serve as a promptable API for non-animation media (sound design, camera motion) that also reduce to forces and phases.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper introduces fog, a function-composition framework that uses LLM code generation to implement a hierarchy of abstract classes (Scene/State/Entity, Primary/SecondaryMotion, Verb/Adverb/Gesture/Emotion) for expressing motion and emotion in Heider-Simmel animations. Functions operate over entity state (velocity, energy, paths, collisions) and support compositions such as Adverb(Verb), Gesture+Verb, and Emotion blends. An editor provides on-the-fly generation, path drawing, timeline/state editing, and dynamically generated refinement UI (2-D minimaps). A 4AFC perceptual evaluation on 452 animations (Table 2) reports 67.8% overall exact-match accuracy (2.68× chance). A mixed-methods user study (n=10 professionals/novices) finds significantly faster feedback cycles (0.36 vs 1.75 min) and qualitative gains in control/exploration versus a strong prompt-only baseline.

Significance. If the results hold, fog supplies a concrete, reusable abstraction layer that turns motion and emotion into composable, LLM-implementable contracts rather than free-form prompts. This synthesizes classic animation principles (primary/secondary motion, squash-and-stretch), cognitive findings (Heider-Simmel, Semantic Space Theory), and modern code generation into a hybrid generative UI. The clean 4AFC design with category breakdowns, confusion analysis, Latin-square composition tests, instrumented latency statistics, and released code/video constitute solid empirical contributions. The work is significant for HCI research on controllable AI creative tools, generative user interfaces, and expressive abstract animation; the same class-contract pattern could transfer to other hierarchical motion domains (robots, assemblies).

major comments (3)
  1. [4.1.1, Table 2] Section 4.1.1 and Table 2: distractors are the two nearest Sentence-BERT neighbors within category plus one out-of-category item. The paper itself documents systematic confusions precisely among those near-neighbors (forcefully–erratically 7/10, happy–ecstatic 5/10, urgently–eagerly, etc.). With only a handful of motion channels available to simple 2-D shapes, semantically close words frequently produce near-identical signatures. Selecting the hardest foils therefore systematically depresses absolute accuracy while still allowing a large multiple of the fixed 25% chance baseline. This weakens the diagnostic value of the headline 67.8%/2.68× claim for an “open-ended motion vocabulary.” A control condition with random or far distractors, or accuracy stratified by embedding distance, is needed to separate true semantic coverage from task difficulty.
  2. [4.1.8-4.1.11, Table 2] Sections 4.1.8–4.1.11 and Table 2: composition accuracies fall to 58.3% (Adverb+Verb) and 57.8% (Emotion+Verb); automatic self-refinement of decorate() yields only +0.6% to +2.6% (or –5.2% for Gesture+Verb). The authors correctly attribute failures to channel misalignment (acceleration vs. speed/easing). Because function composition is the central technical claim of the framework, the Primary/SecondaryMotion contracts and decorate() interface must be shown to resolve or at least detect such conflicts more robustly. Without additional analysis of successful vs. failed channel contracts, or an ablation that isolates the composition model, the claim that fog “compose[s] motion functions” remains only partially supported.
  3. [5.2.1] Section 5.2.1: the quantitative efficiency narrative rests on 15.1 vs. 5.2 “generation actions.” The authors note that the action sets are not equivalent (presets, minimap drags, path draws vs. pure prompts). While the latency contrast (0.36 vs. 1.75 min, paired t-test p≤0.001) is cleanly instrumented and statistically significant, the broader claim of “more rapid iteration” should be scoped more tightly to the latency metric and qualitative themes; otherwise the comparison overstates the advantage of the fog interface.
minor comments (5)
  1. [Abstract, Table 2] Abstract and body inconsistently round the accuracy figure (68% vs. 67.8%). Align on one value and state the exact judgment count (1 256) alongside the 452-animation figure.
  2. [passim] Numerous missing spaces and run-on tokens appear throughout the compiled text (“infog”, “fogcan”, “fog’s”, “PrimaryMotionand”, etc.). These are presentation artifacts that should be cleaned before camera-ready.
  3. [3.1, Fig. 1] Figure 1 caption and body text refer to “aggressively(chase(B))” and similar compositions; a short formal grammar or BNF of the supported composition syntax would help readers implement the framework.
  4. [5.1.1] The user-study baseline is strong (same LLM, free code), yet the prompt templates and exact generation settings for both conditions are deferred to Supplementary Material. A one-paragraph summary in the main text would improve reproducibility.
  5. [Abstract, 1, 6.0.2] Discussion 6.0.2 correctly notes the limits of Heider-Simmel shapes, yet the claim of an “open-ended motion vocabulary” still appears in the abstract and contributions without the same qualification. Soften the language or move the caveat earlier.

Circularity Check

0 steps flagged

No circularity: recognition accuracies and user-study metrics are external human judgments independent of the framework definitions or any fitted parameters.

full rationale

This is an HCI systems paper whose central claims rest on two external evaluations: a 4AFC perceptual study (452 animations, 67.8% exact-match accuracy vs. 25% chance) and a mixed-methods user study (n=10) measuring feedback-cycle latency and qualitative control. Neither result is defined in terms of the abstract-class hierarchy, the LLM generation process, or any parameter fitted to the same data. Stimulus words and embedding-based distractors (all-MiniLM-L6-v2) are chosen a priori; the accuracy numbers are crowdworker judgments, not algebraic identities. The single self-citation to the authors’ prior LogoMotion work appears only in Related Work as one of several LLM animation systems and is not invoked to justify any uniqueness claim, uniqueness theorem, or load-bearing premise of the present results. No self-definitional equations, fitted-input-as-prediction, ansatz smuggling, or renaming of known results occur. The derivation chain is therefore self-contained against external human benchmarks.

Axiom & Free-Parameter Ledger

3 free parameters · 3 axioms · 2 invented entities

As an HCI systems paper the load-bearing commitments are design choices and evaluation assumptions rather than free physical constants. The free parameters are the concrete stimulus lists, the LLM hyper-parameters, and the hand-chosen channel contracts; the axioms are standard animation principles and the assumption that simple shape kinematics suffice for semantic recognition; the invented entities are the fog class hierarchy itself.

free parameters (3)
  • stimulus word lists (12–32 per category)
    Hand-selected and balanced by the authors; different choices would change measured recognition rates.
  • Claude-Opus-4-6 generation settings (maxTokens=12k, thinking budget=10k)
    Fixed but arbitrary LLM hyper-parameters that determine the concrete code implementations tested.
  • channel contracts (speed, acceleration, exaggeration, energy, etc.)
    Author-defined numerical ranges and write-order rules that adverbs/emotions must respect; not derived from data.
axioms (3)
  • domain assumption Disney primary/secondary motion principle and Reynolds-style force/velocity model are adequate bases for expressive shape animation
    Invoked in §3.1.2 and Related Work to justify the PrimaryMotion/SecondaryMotion split and helper force functions.
  • domain assumption Simple 2-D kinematic channels (position, velocity, squash-and-stretch, particles) can convey the semantics of verbs, adverbs, gestures and emotions
    Stated as the working premise of Heider-Simmel style evaluation; limitations acknowledged only in Discussion.
  • ad hoc to paper Sentence-BERT embeddings produce fair within- and out-of-category distractors for 4AFC
    Methodological choice in §4.1.1 that directly affects measured accuracy.
invented entities (2)
  • fog abstract-class hierarchy (Verb/Adverb/Gesture/Emotion + Primary/SecondaryMotion) no independent evidence
    purpose: Provide a contract that LLMs implement and that the runtime composes under a fixed write-order model
    Core technical contribution; no independent existence outside the paper’s implementation.
  • function-composition refinement panel with dynamically generated 2-D minimaps no independent evidence
    purpose: Expose numerical parameters of a composition for direct manipulation
    UI invention specific to the fog editor; evaluated only inside the user study.

pith-pipeline@v1.1.0-grok45 · 18123 in / 2693 out tokens · 28749 ms · 2026-07-10T14:48:50.267778+00:00 · methodology

0 comments
read the original abstract

Motion and emotion are core parts of intelligent, expressive behavior. In this paper, we introduce fog, a function composition framework for implementing and compose motion functions. We demonstrate how fog can be used to express motion and emotion in Heider-Simmel style animations. This code generation framework can help users generate functions for verbs, adverbs, gestures, and emotions to create an open-ended motion vocabulary. It is complemented by an animation editor that helps users refine motion through direct manipulation and dynamically generated UI. We evaluate our approach with a perceptual evaluation, where we test 452 fog-generated animations to see if people can recognize the semantic meaning of the motion. We find that fog's motion functions can be recognized at 68% accuracy, a 2.68x improvement over a chance baseline. In a mixed-methods user study with professionals and novices, we show that fog in interface form can support users with more rapid iteration, exploration, and control.

Figures

Figures reproduced from arXiv: 2607.07952 by Lydia Chilton, Vivian Liu.

Figure 1
Figure 1. Figure 1: In fog, users utilize LLMs to generate motion and emotion functions by implementing class definitions for verbs, adverbs, emotions, and gestures. These functions can compose with each other and act over an entity’s internal and external state (energy, velocity, collision, paths) to show behavior. The examples show social motion (chase), motion decoration (aggression), and emotionally-expressive motion (ang… view at source ↗
Figure 2
Figure 2. Figure 2: fog as an animation editor with timeline, scene, and state support. Users can edit by composing motion func￾tions and generating functions on-the-fly using fog’s abstract classes for verbs, adverbs, gestures, and emotions [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

23 extracted references · 23 canonical work pages · 2 internal anchors

  1. [1]

    Tyler Angert, Miroslav Suzara, Jenny Han, Christopher Pondoc, and Hariharan Subramonyam. 2023. Spellburst: A Node-based Interface for Exploratory Creative Coding with Natural Language Prompts. InProceedings of the 36th Annual ACM Symposium on User Interface Software and Technology (UIST ’23). ACM, 1–22. doi:10.1145/3586183.3606719

    work page doi:10.1145/3586183.3606719 2023

  2. [2]

    Badler, Jonathan D

    Norman I. Badler, Jonathan D. Korein, James U. Korein, Gerald M. Radack, and Lynne Shapiro Brotman. 1985. Positioning and animating human figures in a task-oriented environment.Vis. Comput.1, 4 (Dec. 1985), 212–220. doi:10.1007/ BF02021810

    work page 1985

  3. [3]

    Samuelle Bourgault, Li-Yi Wei, Jennifer Jacobs, and Rubaiat Habib Kazi. 2025. Narrative Motion Blocks: Combining Direct Manipulation and Natural Language Interactions for Animation Creation. InProceedings of the 2025 ACM Designing Interactive Systems Conference (DIS ’25). Association for Computing Machinery, New York, NY, USA, 1366–1386. doi:10.1145/37153...

    work page doi:10.1145/3715336.3735766 2025

  4. [4]

    Diane Chi, Monica Costa, Liwei Zhao, and Norman Badler. 2000. The EMOTE model for effort and shape. InProceedings of the 27th Annual Conference on Com- puter Graphics and Interactive Techniques (SIGGRAPH ’00). ACM Press/Addison- Wesley Publishing Co., USA, 173–182. doi:10.1145/344779.352172

    work page doi:10.1145/344779.352172 2000

  5. [5]

    John Joon Young Chung, Melissa Roemmele, and Max Kreminski. 2025. Toyteller: AI-powered Visual Storytelling Through Toy-Playing with Character Symbols. InProceedings of the 2025 CHI Conference on Human Factors in Computing Systems (CHI ’25). ACM, 1–23. doi:10.1145/3706598.3713435

    work page doi:10.1145/3706598.3713435 2025

  6. [6]

    Ruta Desai, Fraser Anderson, Justin Matejka, Stelian Coros, James McCann, George Fitzmaurice, and Tovi Grossman. 2019. Geppetto: Enabling Semantic Design of Expressive Robot Behaviors. InProceedings of the 2019 CHI Conference on Human Factors in Computing Systems(Glasgow, Scotland Uk)(CHI ’19). ACM, New York, NY, USA, Article 369, 14 pages. doi:10.1145/32...

    work page doi:10.1145/3290605.3300599 2019

  7. [7]

    Funda Durupinar, Mubbasir Kapadia, Susan Deutsch, Michael Neff, and Norman I. Badler. 2016. PERFORM: Perceptual Approach for Adding OCEAN Personality to Human Motion Using Laban Movement Analysis.ACM Trans. Graph.36, 1, Article 6 (Oct. 2016), 16 pages. doi:10.1145/2983620

    work page doi:10.1145/2983620 2016

  8. [8]

    2007.Development of the Animate–Inanimate Distinction

    Susan Gelman and John Opfer. 2007.Development of the Animate–Inanimate Distinction. 151 – 166. doi:10.1002/9780470996652.ch7 fog: Expressing Motion and Emotion through Function Composition of AI-Generated Code Under Submission, April 2026, New York, NY

    work page doi:10.1002/9780470996652.ch7 2007

  9. [9]

    Fritz Heider and Marianne L. Simmel. 1944. An experimental study of ap- parent behavior.American Journal of Psychology57 (1944), 243–259. https: //api.semanticscholar.org/CorpusID:143057281

    work page 1944

  10. [10]

    1981.The illusion of life: Disney Animation

    Ollie Johnston and Frank Thomas. 1981.The illusion of life: Disney Animation. Disney Editions

    work page 1981

  11. [11]

    Kenneth M. Kahn. 1979. Creation of computer animation from story descriptions. https://api.semanticscholar.org/CorpusID:5051186

    work page 1979

  12. [12]

    Brooks, and Alan Cowen

    Dacher Keltner, Jeffrey A. Brooks, and Alan Cowen. 2023. Semantic Space Theory: Data-Driven Insights Into Basic Emotions.Current Directions in Psychological Science32, 3 (2023), 242–249. arXiv:https://doi.org/10.1177/09637214221150511 doi:10.1177/09637214221150511

    work page doi:10.1177/09637214221150511 2023

  13. [13]

    Vivian Liu, Rubaiat Habib Kazi, Li-Yi Wei, Matthew Fisher, Timothy Langlois, Seth Walker, and Lydia Chilton. 2025. LogoMotion: Visually-Grounded Code Synthesis for Creating and Editing Animation. InProceedings of the 2025 CHI Conference on Human Factors in Computing Systems (CHI ’25). Association for Computing Machinery, New York, NY, USA, Article 157, 16...

    work page arXiv 2025

  14. [14]

    Michael Mateas and Andrew Stern. 2003. A Behavior Language: Joint Action and Behavioral Idioms.Cognitive Technologies(01 2003). doi:10.1007/978-3-662- 08373-4_7

    work page doi:10.1007/978-3-662- 2003

  15. [15]

    Albert Mehrabian. 1996. Pleasure-arousal-dominance: A general framework for describing and measuring individual differences in Temperament.Current Psy- chology14 (1996), 261–292. https://api.semanticscholar.org/CorpusID:144411065

    work page 1996

  16. [16]

    Ken Perlin and Athomas Goldberg. 1996. Improv: a system for scripting interactive actors in virtual worlds. InProceedings of the 23rd Annual Conference on Computer Graphics and Interactive Techniques (SIGGRAPH ’96). Association for Computing Machinery, New York, NY, USA, 205–216. doi:10.1145/237170.237258

    work page doi:10.1145/237170.237258 1996

  17. [17]

    Nils Reimers and Iryna Gurevych. 2019. Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks. InProceedings of the 2019 Conference on Em- pirical Methods in Natural Language Processing. Association for Computational Linguistics. https://arxiv.org/abs/1908.10084

    work page internal anchor Pith review Pith/arXiv arXiv 2019

  18. [18]

    Craig Reynolds. 2002. Steering Behaviors For Autonomous Characters. (06 2002)

    work page 2002

  19. [19]

    James Russell. 1980. A circumplex model of affect.Journal of Personality and Social Psychology39 (12 1980), 1161–1178. doi:10.1037/h0077714

    work page doi:10.1037/h0077714 1980

  20. [20]

    Steven Henry Strassmann. 1991. Desktop theater : automatic generation of expressive animation. https://dspace.mit.edu/handle/1721.1/12871

    work page 1991

  21. [21]

    Tiffany Tseng, Ruijia Cheng, and Jeffrey Nichols. 2024. Keyframer: Empowering Animation Design using Large Language Models. arXiv:2402.06071 [cs.HC]

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  22. [22]

    2005.Visuospatial Reasoning

    Barbara Tversky. 2005.Visuospatial Reasoning. 209–241

    work page 2005

  23. [23]

    Munetoshi Unuma, Ken Anjyo, and Ryozo Takeuchi. 1995. Fourier principles for emotion-based human figure animation. InProceedings of the 22nd Annual Conference on Computer Graphics and Interactive Techniques (SIGGRAPH ’95). Association for Computing Machinery, New York, NY, USA, 91–96. doi:10.1145/ 218380.218419

    work page arXiv 1995