Pith. sign in

REVIEW 4 major objections 4 minor 35 references

A deterministic symbolic geometric checker can replace stochastic vision-language critics for spatial layout in LLM-generated educational animations, cutting per-iteration cost by 6×–18×.

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 · deepseek-v4-flash

2026-08-01 15:56 UTC pith:J4N5REAF

load-bearing objection A well-engineered symbolic verification layer for Manim pipelines, but its quantitative claims rest on an unvalidated metric that the method directly optimizes, so treat the MVQS gains as internal until human validation appears. the 4 major comments →

arxiv 2607.18116 v1 pith:J4N5REAF submitted 2026-07-20 cs.AI cs.CVcs.GRcs.MAcs.MM

SGA: Plug&Play Geometric Verification for Educational Video Synthesis

Lopez Jhon , Hinojosa Carlos , Ghanem Bernard This is my paper
classification cs.AI cs.CVcs.GRcs.MAcs.MM
keywords Symbolic Geometric Agentscene graph extractiongeometric verificationeducational video generationManimLLM code generationvisual quality metricaxis-aligned bounding box
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 proposes the Symbolic Geometric Agent (SGA), a plug-and-play verification layer for code-centric educational animation pipelines. Instead of rendering videos and asking a vision-language model to critique frames, SGA partially executes the generated animation code to extract a symbolic scene graph, detects geometric overlaps with exact bounding-box analysis, and compiles targeted coordinate-level fixes for the LLM. The paper also introduces the Manim Visual Quality Score (MVQS), a deterministic rendering-free metric that scores spatial integrity. Experiments on the MMMC-Code benchmark across four LLMs and two pipelines report that SGA improves MVQS in 7 of 8 configurations and reaches a peak of 73.11 — a 16.1% relative gain — while reducing refinement overhead by 6×–18× per iteration. The paper argues that spatial correctness is precisely what pixel-level perceptual evaluation misses.

Core claim

The central claim is that spatial layout errors in LLM-generated educational animations — overlapping equations, labels pushed off-canvas, colliding primitives — can be detected and corrected deterministically in the symbolic code domain, without rendering and without stochastic VLM judgments. SGA builds a scene graph by fault-tolerantly executing the animation script, represents every object as a semantic label plus an axis-aligned bounding box, and flags a conflict whenever two boxes intersect beyond an epsilon threshold. For each conflict it computes the minimum separation vector and packages it, with a ±5-line source context window, into a patch prompt that forces the LLM to make localiz

What carries the argument

The carrying mechanism is the combination of: (1) partial execution, where the Manim runtime is repurposed as a symbolic engine that wraps every statement in try-except, so a scene graph can be extracted even from buggy LLM code; (2) exact AABB conflict detection with a two-stage semantic filter that suppresses intentional overlaps (decorative enclosures, annotation pointers, containers) and spatio-temporal sampling that ignores transient mid-animation contacts; (3) a feedback compiler that converts violations into minimum separation vectors and localized patch prompts anchored to source line numbers; and (4) MVQS, a weighted composite (0.5·spatial + 0.2·layout + 0.3·educational) computed pu

Load-bearing premise

The entire argument rests on the assumption that the Manim Visual Quality Score (MVQS), a hand-weighted combination of geometric descriptors, reflects the visual and pedagogical quality a human viewer would perceive — and the paper explicitly states that MVQS has not yet been validated against human expert annotations.

What would settle it

A concrete falsifier would be a human-subject study where experts rate the visual legibility of the same animations from RAW, VLM-critic, and SGA conditions, and the ordering of MVQS (SGA highest) does not reproduce the human ordering, or where MVQS ranks a blatantly misleading but overlap-free animation above a legible one. Alternatively, if a rendering-based metric or human scores showed that SGA's corrections often degrade compositional readability by dispersing objects to avoid overlap at the cost of destroying intended grouping, the claim that spatial integrity equals visual quality would

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

If this is right

  • SGA improves MVQS in 7 of 8 backbone × pipeline configurations, reaching a peak 16.1% relative improvement over the raw baseline.
  • Deterministic symbolic verification outperforms VLM-based critique in spatial layout quality, while costing 6×–18× less per iteration (symbolic checks under 5 seconds vs 30–90 seconds for a full rendering pass).
  • MVQS separates configurations over a much wider range than VLM-Judge aesthetics (58.7–73.1 vs 85.9–96.4) and the two axes are largely orthogonal, showing that perceptual critics miss geometric violations.
  • SGA gains concentrate in the spatial-correctness and educational-utility sub-scores, while VLM feedback sometimes introduces regressions such as out-of-canvas displacement.
  • SGA is model-agnostic and integrates into existing pipelines as a spatial guardrail, achieving a completion ratio of 1.000 across all tested scenes without requiring architectural changes.

Where Pith is reading between the lines

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

  • Editorial inference: Because SGA works on a symbolic scene graph with source-line provenance, the same interception pattern could extend beyond spatial checks to verify semantic references (e.g., catching animations that reference objects never created) — a natural next step the paper does not claim.
  • Editorial inference: MVQS's determinism means it could serve as an inexpensive pre-render gate in curriculum-generation workflows, letting educators evaluate dozens of draft scripts for layout quality before committing to full rendering; this is a practical corollary the paper leaves implicit.
  • Editorial inference: The partial-execution + AABB + minimum-separation-vector recipe is domain-generic and could testably transfer to other structured visual domains, such as slides, infographics, or UI mockups, where overlapping boxes are the dominant failure mode.

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

4 major / 4 minor

Summary. This paper addresses spatial layout errors in LLM-generated Manim educational videos. It proposes SGA, a plug-and-play module that intercepts LLM code, performs partial execution to construct a symbolic scene graph, detects AABB-based spatial conflicts, computes Minimum Separation Vectors, and prompts the LLM with localized patch instructions. It also introduces MVQS, a deterministic rendering-free weighted composite of spatial, layout, and educational sub-scores. Experiments on MMMC-Code with four LLMs and two pipelines (TheoremExplainAgent, Code2Video) report that SGA improves MVQS in 7 of 8 pipeline/backbone configurations, including a 16.1% relative gain for Code2Video+GPT-5.1, and reduces refinement latency by 6–18× per iteration. The appendix provides weight-sensitivity, completion-ratio, and scene-level distribution analyses.

Significance. If the reported gains reflect true visual/pedagogical quality, the contribution is practically valuable: SGA is deterministic, cheap, model-agnostic, and offers a concrete alternative to expensive VLM critics for geometric refinement. The paper has clear strengths: a well-specified algorithm (Algorithm 1), reproducible symbolic evaluations, a weight-sensitivity analysis (Table 2), a completion-ratio check (Table 5), and full scene-level distributions (Table 6). However, the central evaluation rests on MVQS, a hand-weighted proxy that has not been validated against human judgment. Because SGA directly minimizes the geometric quantities that MVQS measures, the headline gains are partly by construction; the weak MVQS–VLM correlation (ρ=+0.24, p=0.27) does not supply external validity. If the authors add a human evaluation or sharply reframe the claims, the work could be a solid engineering contribution.

major comments (4)
  1. [§3.3–§3.4, Eq. (3), Limitations] The load-bearing evaluation is partly circular: SGA’s conflict detection and MVQS’s overlap/legibility/canvas components use the same AABB geometry, thresholds, and separation logic. SGA is explicitly optimizing the quantities that MVQS rewards, so its MVQS gains are expected by construction. The paper’s own Limitations state that MVQS has not been validated against human expert annotations, and §4.2’s only external bridge is the Sed–Educational Logic correlation (ρ=+0.58), while the overall MVQS–VLM correlation is negligible (ρ=+0.24, p=0.27). Consequently, the abstract’s claim that SGA improves visual legibility is currently a claim about an internal index. I recommend either adding a human preference/rating study on a sample of scenes or reframing the conclusions as improvements of a geometric proxy.
  2. [Table 1 and Table 6] Table 1 reports mean scores without error bars or significance tests. The scene-level distributions in Table 6 show large standard deviations; for example, Code2Video GPT-5.1 has RAW MVQS 62.9±9.8 and SGA 73.0±16.3, and GPT-5 mini has 63.6±9.9 → 66.4±9.0. Several reported gains are within one standard deviation of the baseline. To support the “7 of 8 configurations” claim, the authors should report paired significance tests across the 2,252 scene evaluations (e.g., Wilcoxon signed-rank) and effect sizes. Without this, the reader cannot distinguish systematic improvement from noise, and the claimed consistency is not established.
  3. [§4.1] The text states: “SGA improves Sed over the VLM Critic by 15.0 points in TheoremExplainAgent and 10.5 points in Code2Video, and increases overall MVQS by 1.50 and 4.43 points, respectively.” Averaging over the four backbones in Table 1 gives Sed gains of 10.8 (TEA) and 5.9 (C2V), and MVQS gains of 0.8 (TEA) and 3.7 (C2V), not 15.0/10.5 and 1.50/4.43. These figures are a central quantitative summary of SGA’s advantage over the VLM critic, yet they are not reproducible from the table. Please correct the numbers or specify the subset/calculation used.
  4. [§4.2, Fig. 3] The near-zero MVQS–VLM correlation is interpreted as evidence that VLM critics are insensitive to geometry. Given that MVQS is unvalidated, the data are equally consistent with MVQS being orthogonal to perceptually salient quality. The significant Sed–EL correlation is a partial anchor, but only for one sub-score; Ssp and Sly—the dimensions most affected by SGA—show no external correlation. Please report a human validation for these sub-scores or temper the interpretation. The current one-sided reading overstates what the evidence supports.
minor comments (4)
  1. [Abstract, §4.1] “7 of 8 backbone × pipeline configurations” should clarify that this compares SGA against the raw baseline per pipeline/backbone, not all 24 rows in Table 1.
  2. [Appendix §8] The appendix refers to “Eq. 1 of the main paper” for the MVQS weighted sum, but in the main text Eq. (1) is the object record; the MVQS equation is Eq. (3) in §3.4. Please renumber/reference consistently.
  3. [Table 6] SGA scene counts are exactly 2× the RAW counts because each topic yields a pre-patch and post-patch evaluation. Please state explicitly whether the RAW and VLM Critic rows in Table 6 come from the same single-iteration protocol, since the comparison of variances depends on this.
  4. [§4.1] The sentence “SGA and the VLM Critic differ by fewer than 1.5 points on every VLM sub-score” is only true for the overall means across backbones/pipelines; individual configurations exceed 1.5 (e.g., TheoremExplainAgent GPT-5.1 Educational Logic: 90.41 vs 86.40). Please specify the aggregation.

Circularity Check

0 steps flagged

No significant circularity; the MVQS/SGA overlap is a construct-validity concern, not a derivation-circle.

full rationale

The paper does not contain a circular derivation. SGA is a deterministic geometric refiner: it extracts AABB overlaps from a symbolic scene graph and asks the LLM to apply MSV-based fixes (Sec. 3.3, 3.5). MVQS is a separately defined weighted sum of spatial, layout, and educational sub-scores (Eq. 3, Sec. 3.4). The overlap-penalty component of S_sp is computed from the same AABB intersections SGA repairs, so some MVQS gain is unsurprising; but this is an evaluation-alignment / construct-validity issue, not an equation-level reduction. MVQS also includes components SGA does not directly optimize (legibility threshold, separation, density, distribution, labeling completeness, animation quality), and SGA's feedback uses AABB/MSV, not the MVQS formula. The empirical outcomes are not forced: SGA decreases MVQS in one configuration (TEA+Claude: 63.84 -> 62.83) and produces small gains in others (e.g., GPT-5 mini Code2Video +2.79), showing the headline 16.1% improvement is not identical to the method's objective by construction. The paper's own Limitations explicitly state 'MVQS has not yet been validated against human expert annotations'; this is a validity limitation, not circularity. There are no self-citations used as load-bearing evidence, no imported uniqueness theorems, and no fitted parameter renamed as a prediction. The evaluation is run against an external benchmark (MMMC-Code) with fixed budgets and baselines, so the central claim is independently checkable even though the metric itself is not yet human-validated.

Axiom & Free-Parameter Ledger

10 free parameters · 7 axioms · 3 invented entities

The pipeline rests on ten hand-chosen weights/thresholds, seven domain assumptions about Manim execution and evaluation validity, and one unvalidated invented metric (MVQS). No code is released. The paper's own Limitations section concedes the metric lacks human validation, which is the most significant uncharged debt in the ledger.

free parameters (10)
  • MVQS composite weights w_sp, w_ly, w_ed = {0.5, 0.2, 0.3}
    Hand-set in Eq. 3; controls the headline MVQS ranking; sensitivity analysis shows rank stability for w_sp in [0.40,0.60].
  • Ssp sub-weights = [0.50, 0.25, 0.15, 0.10]
    Hand-set weights for Overlap, Canvas Boundary, Legibility, Separation (§3.4).
  • Sly sub-weights = [0.40, 0.60]
    Hand-set weights for Density and Distribution.
  • Sed sub-weights = [0.40, 0.60]
    Hand-set weights for Labeling Completeness and Animation Quality.
  • Overlap threshold epsilon = 0
    Default 0 in Eq. 2; any positive intersection counts as conflict; controls severity classification.
  • Severity thresholds = HIGH>0.30, MEDIUM>0.10
    Discretization of coverage ratios; hand-chosen.
  • Minimum legibility scale delta = 0.15 units
    Objects below this scale penalize Legibility.
  • Spatio-temporal sample count K = 11 frames
    Sampling across animation time; conflicts at t∈{0,1} are persistent.
  • Desired density range = [0.10, 0.35]
    Density sub-metric favors occupancy inside this interval.
  • Refinement budget n = 1 iteration
    All experiments use one generate-refine-evaluate loop; results might change with more iterations.
axioms (7)
  • domain assumption Partial execution with per-statement try-except faithfully represents the rendered Manim scene.
    §3.2; if statements fail silently, the extracted scene graph may miss objects and bias MVQS; object-level completeness is not reported.
  • domain assumption AABB intersection geometry is a sufficient approximation of visual occlusion in these 2D scenes.
    §3.3; ignores rotation, complex paths, and 3D; the paper scopes itself to 2D Manim primitives.
  • ad hoc to paper The P1–P5 structural patterns and placement/motion keyword vocabulary cover the intentional-overlap cases in the benchmark.
    §3.3; hand-crafted filter without ablation; misclassification would suppress real errors or patch intentional layouts.
  • ad hoc to paper Persistent conflicts at t∈{0,1} and transient mid-trajectory overlaps can be separated with K=11 frame sampling.
    §3.3; the resting-frame rule is a modeling choice with no external justification.
  • ad hoc to paper MVQS's weighted linear aggregation reflects educational visual quality.
    §3.4/Limitations; explicitly unvalidated against human expert annotations.
  • domain assumption The LLM backbones can apply ±5-line patch prompts without introducing new semantic errors.
    §3.5; patch rollback on parse failure mitigates but assumes edits preserve meaning and animation logic.
  • domain assumption Gemini 3.0 Flash VLM-Judge scores are a valid perceptual reference for aesthetics.
    §4; the VLM judge itself is not calibrated against human ratings in this paper.
invented entities (3)
  • MVQS composite metric no independent evidence
    purpose: Render-free evaluation of spatial integrity; the main outcome measure for all headline claims.
    Unvalidated against human expert annotations (Limitations); the only partial indirect evidence is correlation of Sed with VLM Educational Logic (ρ=0.58).
  • Minimum Separation Vector (MSV) no independent evidence
    purpose: Axis-aligned displacement that resolves an overlap; used to ground LLM patch prompts.
    A deterministic computational artifact derived from AABB geometry; not externally verified but mathematically simple.
  • Semantic taxonomy Phi={TEXT, GEOM, MARK, DECO} no independent evidence
    purpose: Labels extracted mobjects for filtering and labeling-completeness scoring.
    Hand-defined classes with no validation of taxonomy adequacy or completeness.

pith-pipeline@v1.3.0-alltime-deepseek · 15005 in / 12623 out tokens · 129555 ms · 2026-08-01T15:56:44.371641+00:00 · methodology

0 comments
read the original abstract

Recent work leverages Large Language Models (LLMs) to generate executable code for pedagogical animations using libraries such as Manim. However, ensuring spatial correctness and visual legibility remains challenging, as existing frameworks emphasize pedagogical content while overlooking geometric occlusions. We propose the Symbolic Geometric Agent (SGA), a plug-and-play module for code-centric animation pipelines that intercepts LLM-generated code, performs partial execution to extract symbolic scene graphs, and applies targeted refinement when spatial conflicts are detected. We further introduce the Manim Visual Quality Score (MVQS), a deterministic rendering-free proxy for spatial integrity. Experiments on the MMMC-Code benchmark across four LLM backbones and two agentic pipelines show that SGA achieves a peak MVQS of 73.11 (Code2Video + GPT-5.1), corresponding to a 16.1% relative improvement over the raw baseline, and improves MVQS in 7 of 8 backbone x pipeline configurations.

Figures

Figures reproduced from arXiv: 2607.18116 by Ghanem Bernard, Hinojosa Carlos, Lopez Jhon.

Figure 1
Figure 1. Figure 1: Current SOTA agents produce syntactically valid but visually occluded animations. (Left) [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The SGA acts as a neuro-symbolic interceptor for spatial-linguistic grounding. (Left) The pipeline [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Perceptual vs. symbolic evaluation across [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Comparison on Code2Video (Chen et al., 2025) scenes. Raw frames exhibit label-equation collisions and occluded annotations; VLM-Critic introduces regressions, including out-of-canvas displacement and missing seman￾tic elements; SGA resolves spatial conflicts while preserving mathematical notation and geometric relationships. material (Sec. 1). 4.4 Computational Efficiency Analysis SGA provides a substantia… view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative comparison on TheoremExplainAgent ( [PITH_FULL_IMAGE:figures/full_fig_p012_5.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

35 extracted references · 1 canonical work pages

  1. [1]

    arXiv preprint arXiv:2510.01174 , year=

    Code2Video: A Code-centric Paradigm for Educational Video Generation , author=. arXiv preprint arXiv:2510.01174 , year=

    arXiv

  2. [2]

    Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Theoremexplainagent: Towards video-based multimodal explanations for llm theorem understanding , author=. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  3. [3]

    Technology, Knowledge and Learning , volume=

    Video analytics in digital learning environments: Exploring student behaviour across different learning contexts , author=. Technology, Knowledge and Learning , volume=. 2024 , publisher=

    2024

  4. [4]

    International Journal of Artificial Intelligence in Education , pages=

    A closer look into recent video-based learning research: A comprehensive review of video characteristics, tools, technologies, and learning effectiveness , author=. International Journal of Artificial Intelligence in Education , pages=. 2025 , publisher=

    2025

  5. [5]

    International conference on artificial intelligence in education , pages=

    Generative AI for learning: Investigating the potential of learning videos with synthetic virtual instructors , author=. International conference on artificial intelligence in education , pages=. 2023 , organization=

    2023

  6. [6]

    Education Sciences , volume=

    Bridging disciplines: Exploring interdisciplinary curriculum development in STEM teacher education , author=. Education Sciences , volume=. 2025 , publisher=

    2025

  7. [7]

    arXiv preprint arXiv:2505.13851 , year=

    A Challenge to Build Neuro-Symbolic Video Agents , author=. arXiv preprint arXiv:2505.13851 , year=

    Pith/arXiv arXiv

  8. [8]

    arXiv preprint arXiv:2210.02303 , year=

    Imagen video: High definition video generation with diffusion models , author=. arXiv preprint arXiv:2210.02303 , year=

    Pith/arXiv arXiv

  9. [9]

    European Conference on Computer Vision , pages=

    Photorealistic video generation with diffusion models , author=. European Conference on Computer Vision , pages=. 2024 , organization=

    2024

  10. [10]

    ACM Computing Surveys , volume=

    A survey on video diffusion models , author=. ACM Computing Surveys , volume=. 2024 , publisher=

    2024

  11. [11]

    arXiv preprint arXiv:2412.14169 , year=

    Autoregressive video generation without vector quantization , author=. arXiv preprint arXiv:2412.14169 , year=

    Pith/arXiv arXiv

  12. [12]

    arXiv , author=

    STIV: Scalable text and image conditioned video generation. arXiv , author=

  13. [13]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Text2video-zero: Text-to-image diffusion models are zero-shot video generators , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  14. [14]

    doi:10.5281/zenodo.3689999 , url =

    The Manim Community Developers , title =. doi:10.5281/zenodo.3689999 , url =

    work page doi:10.5281/zenodo.3689999

  15. [15]

    The eleventh international conference on learning representations , year=

    React: Synergizing reasoning and acting in language models , author=. The eleventh international conference on learning representations , year=

  16. [16]

    URL https://arxiv

    Gorilla: Large language model connected with massive apis, 2023 , author=. URL https://arxiv. org/abs/2305.15334 , year=

    Pith/arXiv arXiv 2023

  17. [17]

    2023 , publisher=

    E-learning and the science of instruction: Proven guidelines for consumers and designers of multimedia learning , author=. 2023 , publisher=

    2023

  18. [18]

    arXiv preprint arXiv:2410.03439 , year=

    Toolgen: Unified tool retrieval and calling via generation , author=. arXiv preprint arXiv:2410.03439 , year=

    Pith/arXiv arXiv

  19. [19]

    International conference on machine learning , pages=

    Learning transferable visual models from natural language supervision , author=. International conference on machine learning , pages=. 2021 , organization=

    2021

  20. [20]

    arXiv preprint arXiv:2306.01815 , year=

    Prototyping the use of Large Language Models (LLMs) for adult learning content creation at scale , author=. arXiv preprint arXiv:2306.01815 , year=

    Pith/arXiv arXiv

  21. [21]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Genmac: compositional text-to-video generation with multi-agent collaboration , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  22. [22]

    Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=

    Creativity in llm-based multi-agent systems: A survey , author=. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=

    2025

  23. [23]

    Advances in neural information processing systems , volume=

    Reflexion: Language agents with verbal reinforcement learning , author=. Advances in neural information processing systems , volume=

  24. [24]

    Educational Psychology Review , volume=

    The past, present, and future of the cognitive theory of multimedia learning , author=. Educational Psychology Review , volume=. 2024 , publisher=

    2024

  25. [25]

    Advances in neural information processing systems , volume=

    Self-refine: Iterative refinement with self-feedback , author=. Advances in neural information processing systems , volume=

  26. [26]

    International Conference on Learning Representations , volume=

    MetaGPT: Meta programming for a multi-agent collaborative framework , author=. International Conference on Learning Representations , volume=

  27. [27]

    arXiv preprint arXiv:2305.13655 , year=

    Llm-grounded diffusion: Enhancing prompt understanding of text-to-image diffusion models with large language models , author=. arXiv preprint arXiv:2305.13655 , year=

    Pith/arXiv arXiv

  28. [28]

    Findings of the Association for Computational Linguistics: ACL 2024 , pages=

    Prometheus-vision: Vision-language model as a judge for fine-grained evaluation , author=. Findings of the Association for Computational Linguistics: ACL 2024 , pages=

    2024

  29. [29]

    arXiv preprint arXiv:2501.05435 , year=

    Neuro-symbolic AI in 2024: A systematic review , author=. arXiv preprint arXiv:2501.05435 , year=

    Pith/arXiv arXiv 2024

  30. [30]

    Forty-first International Conference on Machine Learning , year=

    Mllm-as-a-judge: Assessing multimodal llm-as-a-judge with vision-language benchmark , author=. Forty-first International Conference on Machine Learning , year=

  31. [31]

    arXiv preprint arXiv:2503.05977 , year=

    Is your video language model a reliable judge? , author=. arXiv preprint arXiv:2503.05977 , year=

    Pith/arXiv arXiv

  32. [32]

    2004 , publisher=

    Real-time collision detection , author=. 2004 , publisher=

    2004

  33. [33]

    arXiv preprint arXiv:2401.03003 , year=

    Ast-t5: Structure-aware pretraining for code generation and understanding , author=. arXiv preprint arXiv:2401.03003 , year=

    Pith/arXiv arXiv

  34. [34]

    Psychology of learning and motivation , volume=

    Multimedia learning , author=. Psychology of learning and motivation , volume=. 2002 , publisher=

    2002

  35. [35]

    Learning and instruction , volume=

    Cognitive load theory, learning difficulty, and instructional design , author=. Learning and instruction , volume=. 1994 , publisher=

    1994