Generative Fluid Intelligence (GFI)

• Generative Fluid Intelligence (GFI) is a paradigm that combines generation, abstraction, and novelty handling to model non-linear dynamics and perform zero-shot rule induction.
• GFI methodologies enable rapid fluid simulation and geometric generation by leveraging latent-space interpolation, high compression rates, and efficient reconstruction techniques.
• Benchmarking reveals that while GFI models excel in familiar tasks, they face challenges in abstract reasoning and require dual-channel architectures and neuro-symbolic interventions for improved generalization.

Generative Fluid Intelligence (GFI) denotes a family of concepts at the intersection of generation, abstraction, and novelty handling. In one usage, originating in neural fluid simulation, it refers to a generative model that captures non-linear fluid dynamics in a parameterized latent form, enabling interpolation, latent-space time integration, and fast reconstruction without explicit PDE solves (Kim et al., 2018). In a second usage, developed in zero-shot geometric generation, multimodal image generation, and large-language-model evaluation, it denotes the capacity to induce latent rules from immediate context, execute ad-hoc constraints, and generate correct outputs for unseen cases rather than relying on memorized patterns or training-distribution familiarity (Chidambaram et al., 2018). Across these usages, GFI is consistently associated with systematic generalization under novel constraints, although the operational definitions, benchmarks, and failure analyses differ substantially by domain.

1. Conceptual scope and competing definitions

The literature uses GFI in multiple, partially overlapping senses. In psychometric and benchmark-oriented work, the core contrast is between fluid and crystallized intelligence. Fluid intelligence is defined as the capacity to solve novel, abstract problems without relying on prior domain knowledge, whereas crystallized intelligence is performance on inputs drawn from the training distribution and therefore more amenable to retrieval or memorization (Wu et al., 11 Feb 2025). In visual-generation work, GFI is formalized as an overview of three primitives—Inductive Inference, Abstract Dynamic Reasoning, and Adaptive Inhibition—written as $GFI=(I,II,III)$ (An et al., 11 Feb 2026). In dual-channel reasoning architectures, the “GFI channel” is the open-ended, probabilistic generator, while a distinct crystallized channel encodes explicit chain-of-thought procedures as a white-box graph $G=(V,E)$ (Zhou et al., 12 Apr 2025).

Context	Operational meaning of GFI	Representative source
Fluid simulation	Parameterized latent generation of non-linear fluid behaviors	(Kim et al., 2018)
Zero-shot geometry	Rendering unseen concepts from textual descriptions by recombining primitives	(Chidambaram et al., 2018)
ARC-style LLM reasoning	Inducing an unknown transformation rule from few examples and generating outputs for unseen inputs	(Wu et al., 11 Feb 2025)
Multimodal visual generation	Inducing patterns, executing ad-hoc constraints, adapting to contextual knowledge	(An et al., 11 Feb 2026)
Dual-channel AI systems	Stochastic hypothesis generation coupled to explicit procedural verification	(Zhou et al., 12 Apr 2025)

A common formulation in rule-induction settings is: given example pairs $S_{\text{train}}=\{(x_i,y_i)\}_{i=1}^k$ generated by an unknown rule $R$, the model must generalize $R$ to novel inputs $S_{\text{test}}$ (Yang et al., 3 Jun 2025). This makes GFI an operational property of behavior on unseen transformations rather than a statement about model size, parametric capacity, or conventional generation quality.

The distinction from crystallized intelligence is sharpened in chess-based analyses. Pleiss et al. define crystallized intelligence as model performance when $x\sim p_{\text{train}}(x)$ and fluid intelligence as performance when $x\sim p_{\text{test}}(x)$ with $p_{\text{test}}\neq p_{\text{train}}$, measuring the separation by the generalization gap $$\Delta_{\text{gen}}=E_{x\sim p_{\text{test}}}[L(f(x))]-E_{x\sim p_{\text{train}}}[L(f(x))]$$ (Pleiss et al., 23 Jan 2026). This formulation makes explicit that GFI is not merely competence on difficult tasks, but competence under distributional displacement.

2. GFI in computational fluid dynamics

The earliest explicit use of the term in the supplied literature is the fluid-mechanics sense instantiated by “Deep Fluids” (Kim et al., 2018). There, a convolutional generator $G=(V,E)$0 maps a low-dimensional parameter vector $G=(V,E)$1—such as buoyancy, inflow speed, obstacle position, or time—into a velocity representation. For incompressible settings, the network outputs a stream function or stream vector $G=(V,E)$2 and reconstructs velocity as

$G=(V,E)$3

which enforces $G=(V,E)$4 by construction. The architecture combines fully connected layers, reshaping into a low-resolution feature volume, and cascades of “Big Blocks” composed of residual “Small Blocks” with upsampling (Kim et al., 2018).

Training aligns both velocities and spatial derivatives through

$G=(V,E)$5

with typical normalized weights $G=(V,E)$6 (Kim et al., 2018). For extended scenes, an encoder maps a frame $G=(V,E)$7 to a latent code $G=(V,E)$8, where $G=(V,E)$9 is unsupervised and $S_{\text{train}}=\{(x_i,y_i)\}_{i=1}^k$0 contains supervised control parameters. A separate time-integration network predicts latent residuals, $S_{\text{train}}=\{(x_i,y_i)\}_{i=1}^k$1, and advances the state by $S_{\text{train}}=\{(x_i,y_i)\}_{i=1}^k$2 (Kim et al., 2018). This supports latent-space resimulation, continuous parameter interpolation, and time re-sampling.

The reported performance characteristics are central to this formulation of GFI. Reconstructed velocity fields are generated up to $S_{\text{train}}=\{(x_i,y_i)\}_{i=1}^k$3 faster than re-simulating with the underlying CPU solver, with a conservatively normalized figure of approximately $S_{\text{train}}=\{(x_i,y_i)\}_{i=1}^k$4 after accounting for memory-bandwidth differences; batch evaluation can generate 16 frames in about 2 ms on a GTX 1080; and the trained network plus latent codes occupies at most about 30 MB, achieving up to $S_{\text{train}}=\{(x_i,y_i)\}_{i=1}^k$5 compression relative to stored solver output (Kim et al., 2018). The declared applications include rapid prototyping of parameter sweeps, real-time or interactive fluid effects in games and VR, time-resampling, latent-space resimulation under arbitrary user controls, and compression of fluid simulation data (Kim et al., 2018).

A later, statistically oriented formulation appears in “Generative AI for fast and accurate statistical computation of fluids,” where GFI is framed as replacing costly ensemble-based CFD by a learned generative model that samples from the statistical solution of turbulent flows (Molinaro et al., 2024). GenCFD uses an end-to-end conditional score-based diffusion model, training a denoiser $S_{\text{train}}=\{(x_i,y_i)\}_{i=1}^k$6 under a weighted denoising loss and sampling via a reverse diffusion process with predictor–corrector steps (Molinaro et al., 2024). Rather than predicting a single trajectory, it targets statistical quantities of interest: mean flow, variance, one-point PDFs, two-point correlations and Reynolds stresses, and the 3D energy spectrum. In the reported experiments, GenCFD accurately recovers these using Monte Carlo averages over approximately 100–200 samples, resolves spectra up to Nyquist, outperforms deterministic baselines by factors of 2–10 on multiple error metrics, and yields per-sample speed-ups of about 10–1,000 over DNS/LES, with sampling around 0.45 s on an NVIDIA RTX 4090 (Molinaro et al., 2024).

Taken together, these two fluid-mechanics lines define GFI as a generative surrogate for a structured solution manifold: in one case a low-dimensional parameterized manifold for incompressible flows, in the other a conditional distribution over turbulent states.

3. Benchmarking GFI as zero-shot and dynamic reasoning

In zero-shot image generation, “Geometric Generalization Based Zero-Shot Learning Dataset Infinite World” treats GFI as a model’s ability to imagine and render novel concepts from unseen textual descriptions by recombining learned geometric primitives such as lines, points, and connectivity (Chidambaram et al., 2018). The dataset pairs simple 2D figures with template-based textual descriptions and scales in the number of sides or lines, image size, and sample count. Its “3–9 World” subset consists of 64×64 images, approximately 41,000 training samples, 567 train-text variants, and 91 unseen test descriptions (Chidambaram et al., 2018). Evaluation is based on the Zero-Shot Intelligence Metric $S_{\text{train}}=\{(x_i,y_i)\}_{i=1}^k$7, where $S_{\text{train}}=\{(x_i,y_i)\}_{i=1}^k$8 is awarded by a rule-based evaluator using contour detection, Canny edge extraction after Gaussian blur, and RGB k-means color analysis (Chidambaram et al., 2018).

For LLMs, ARC-based work operationalizes fluid intelligence as inducing a hidden transformation $S_{\text{train}}=\{(x_i,y_i)\}_{i=1}^k$9 from a few 2D grid examples and producing the output for a held-out test input (Wu et al., 11 Feb 2025). Performance is measured by

$R$0

and by the shape-mismatch rate

$R$1

On a controlled subset of 100 ARC tasks, reported accuracies are 19% for GPT-4o, 6% for GPT-3.5, 9% for Llama-3-70B, and 2% for Mistral-7B, compared with a human average of 75 (Wu et al., 11 Feb 2025). The same study introduces ARAOC, an atomic-operation benchmark with Move, Change Color, Copy, Mirror, Fill Internal, and Scale, each instantiated in 100 micro-tasks (Wu et al., 11 Feb 2025).

DRE-Bench extends this benchmarking program by organizing 36 abstract reasoning tasks into four cognitive levels—Attribute, Spatial, Sequential, and Conceptual—each with multiple dynamic variants generated from the same latent rule (Yang et al., 3 Jun 2025). Representative average accuracies illustrate a monotonic decline with abstraction level: Claude-3.7 reports 58.8%, 58.4%, 44.1%, and 7.96% across Levels 1–4; GPT-4o reports 51.2%, 9.9%, 7.61%, and 0%; o1 reports 62.5%, 58.9%, 28.9%, and 2.65%; DeepSeek-R1 reports 57.9%, 62.8%, 35.6%, and 0.53%; humans report 75.6%, 71.1%, 62.2%, and 47.4% (Yang et al., 3 Jun 2025). This benchmark is designed to separate low-level competence from higher-level abstraction and working-memory demands.

For multimodal visual generation, GENIUS defines GFI as “the capacity to induce patterns, reason through ad-hoc constraints, and adapt to novel scenarios on the fly” (An et al., 11 Feb 2026). It contains 510 expert-curated samples over 20 sub-tasks, grouped into Implicit Pattern Generation, Symbolic Constraint Generation and Visual Constraint Generation, and Prior-Conflicting Generation and Multi-Semantic Generation (An et al., 11 Feb 2026). Evaluation uses three 0/1/2 metrics—Rule Compliance, Visual Consistency, and Aesthetic Quality—with weighted overall score

$R$2

The strongest proprietary model reported, Nano Banana Pro, scores 57.19 out of 100, while the leading open-source model, Bagel, scores 26.74 (An et al., 11 Feb 2026).

4. Empirical deficits, failure modes, and limits

Across these benchmarks, the dominant empirical result is that contemporary models often perform substantially better on familiar or low-level tasks than on tasks requiring compositional novelty. ARC analysis identifies three major limitations in LLMs: limited skill composition, unfamiliarity with abstract input formats, and the intrinsic deficiency of left-to-right decoding (Wu et al., 11 Feb 2025). Fine-tuning on atomic operations does not solve the composition problem: on the simple composition Move $R$3 Copy, GPT-4o achieves 2% accuracy and a fine-tuned Llama-3 variant reaches 5%; on full ARC tasks, a model fine-tuned on all atomic operations still remains at 2% ARC accuracy, rising only to 6% with additional ARC fine-tuning (Wu et al., 11 Feb 2025). Input representation matters sharply: for GPT-4o, converting grid coordinates to natural-language triples raises Move accuracy from 13% to 59% and Copy accuracy from 15% to 40% (Wu et al., 11 Feb 2025).

The chess study strengthens this diagnosis under a controlled distributional taxonomy (Pleiss et al., 23 Jan 2026). Positions are partitioned into within-distribution, near-distribution, and out-of-distribution subsets, and performance is measured by centipawn loss, illegality rate, and normalized ACPL. Raw ACPL degrades sharply with novelty: near-distribution ACPL is approximately $R$4 within-distribution, and out-of-distribution ACPL is approximately $R$5 within-distribution (Pleiss et al., 23 Jan 2026). Except for GPT-5 in within-distribution settings, no LLM outperforms a random legal-move baseline outside the familiar regime, and even GPT-5 produces illegal OOD moves more than 30% of the time (Pleiss et al., 23 Jan 2026). Reasoning-augmented inference improves average ACPL and reduces illegal moves, but the marginal improvement per token falls from $R$6 in WD to $R$7 in ND and $R$8 in OOD (Pleiss et al., 23 Jan 2026).

Zero-shot image generation exhibits analogous weaknesses. In Infinite World, a Reed et al. text-to-image GAN reaches near-100% train-set accuracy but only approximately 20–30% $R$9 on irregular polygons, 35–45% on regular polygons, and 50–60% on parallel lines; AttnGAN achieves high Inception scores on natural data yet low and unstable $R$0 on these geometric tasks, sometimes below the vanilla GAN (Chidambaram et al., 2018). The authors attribute this to convolutional filters latching onto regular, highly symmetric patterns while failing on random-edge polygons and to conditional generators memorizing train classes rather than learning connectivity as a concept (Chidambaram et al., 2018).

The fluid-simulation literature reports a different class of limits. Deep Fluids requires a sufficiently dense training set of parameterized runs; overly coarse sampling causes ghosting and loss of small eddies; small-scale splashes and high-frequency features may be smoothed out; and extrapolation beyond $R$1 of the training range produces noticeable artifacts (Kim et al., 2018). There is also no hard enforcement of boundary penetration for unseen parameters, although the paper reports empirically less than 1% velocity-penetration error (Kim et al., 2018).

A recurrent pattern across these studies is that high within-distribution accuracy, high Inception-like scores, or good conventional generation quality do not by themselves establish GFI. The literature suggests that GFI must be probed under held-out rules, immediate-context constraints, or explicit distribution shifts.

5. Architectural responses and intervention strategies

The surveyed work proposes several architectural responses to these deficiencies. In fluid simulation, Deep Fluids couples a divergence-free generator with an encoder for extended parameterizations and a latent-space time integrator, thereby embedding both interpolation and temporal evolution into the learned manifold (Kim et al., 2018). Its future extensions include learning surface boundaries or multi-phase SDFs jointly with velocities, using physics-informed adversarial or Sobolev losses for sharper vortical detail, employing GAN architectures to hallucinate small-scale turbulence, and replacing the latent integrator with recurrent or attention-based variants for improved long-term stability (Kim et al., 2018).

In geometric zero-shot generation, the proposed “Proactive Optimizer” maintains two parameter sets, reactive $R$2 and proactive $R$3, and updates the latter using auxiliary cues, meta-information, and an external memory of past non-convex solutions:

$R$4

The stated goal is to fuse fast, symbolic-like adjustments with slower gradient-based learning (Chidambaram et al., 2018).

For LLM-based GFI, proposed structural modifications include hierarchical reasoning modules, multi-axis attention with 2D positional encodings, bidirectional or non-autoregressive decoding, and abstract-representation pretraining on synthetic ARC/ARAOC-style data (Wu et al., 11 Feb 2025). Proposed algorithmic strategies include explicit decomposition prompts, chain-of-thought with structured intermediate representations, and retrieval of symbolic subroutines for operations such as Move, Copy, and Mirror (Wu et al., 11 Feb 2025).

A stronger separation of roles appears in “Continuum-Interaction-Driven Intelligence,” which proposes a dual-channel architecture combining a probabilistic “Fluid Channel” with a white-box “COT Channel” (Zhou et al., 12 Apr 2025). The fluid channel samples sequences according to

$R$5

while the crystallized channel encodes procedural reasoning in an explicit graph. An API mediates the interaction: the fluid side proposes candidates, and the COT side accepts, rejects, or queries for clarification (Zhou et al., 12 Apr 2025). The paper states that multi-turn interaction is a necessary condition for intelligence emergence and gives both a theorem asserting $R$6 for alignment score as a function of interaction depth and empirical results in legal QA and medical diagnosis: at a five-turn average, hallucination rate falls from 24% to 6%, alignment score rises from 62 to 89, and audit-trail completeness rises from 0% to 100% (Zhou et al., 12 Apr 2025).

GENIUS contributes a training-free attention intervention rather than a new base architecture (An et al., 11 Feb 2026). Its three stages are keyword distillation, relevance mapping, and bias injection into pre-softmax attention logits,

$R$7

On Bagel, the intervention raises the overall score from 26.74 to 32.92, improving Rule Compliance and context-sensitive tasks without parameter fine-tuning (An et al., 11 Feb 2026).

6. Theoretical interpretations and open questions

A central theoretical question is whether apparent fluid intelligence reflects abstract reasoning or a more minimal generative substrate. “The minimal computational substrate of fluid intelligence” reports that LaMa, a 51M-parameter self-supervised in-painting network trained only on masked natural scenes, attains a raw RAPM score of 8 on Raven’s Advanced Progressive Matrices Set I, with a psychometric threshold lying within the 95% credibility intervals of healthy humans and right-frontal lesion patients (Nelson et al., 2023). Removing Fast Fourier Convolution collapses performance to approximately chance and reproduces characteristic right-frontal-like error patterns (Nelson et al., 2023). The paper therefore argues that RAPM may admit computationally simple solutions based on spatial pattern completion rather than formal abstract reasoning. This directly complicates any simple equation of benchmark success with GFI.

A different theoretical account appears in GenCFD, where the success of generative models on turbulent flow is tied to denoising rather than point prediction (Molinaro et al., 2024). The paper shows that the optimal denoiser under additive Gaussian noise is the posterior mean, that the zero-noise limit approaches projection onto the data manifold, and that diffusion objectives remain near-optimal even when deterministic approximators fail because the underlying solution operator is unstable or highly oscillatory (Molinaro et al., 2024). In toy models, deterministic mean-squared-error fitting collapses to the mean, whereas score-based denoisers recover the correct statistical spread and spectral content (Molinaro et al., 2024). In this line of work, GFI is less a symbolic faculty than a generative capacity to represent distributions over unstable multiscale phenomena.

The open problems recorded across the literature are correspondingly diverse. In geometric reasoning they include bridging the human–machine gap on irregular shapes, automated extraction of symbolic constraints, and meta-learners that autonomously induce geometric axioms (Chidambaram et al., 2018). In LLM evaluation they include neuro-symbolic integration, memory-augmented reasoning, curriculum and dynamic training, and simulation-in-the-loop support for grounded conceptual tasks such as gravity and reflection (Yang et al., 3 Jun 2025). In chess-based testing, the central implication is that scaling and reasoning-augmented inference alone appear insufficient to close the crystallized–fluid gap, indicating a need for new representational or algorithmic primitives beyond scale (Pleiss et al., 23 Jan 2026).

Taken together, the literature does not support a single canonical definition of GFI. It does, however, converge on a common technical problem: building generative systems that can infer latent structure from immediate evidence, respect novel constraints, and remain competent when memorized priors are unavailable or actively misleading.