Omni-Feature Generator (OFG)

• OFG is a paradigm that constructs novel feature spaces from raw inputs using techniques like generator-critic duets, adaptive multi-agent reasoning, and deformable attention fusion.
• It spans diverse applications in tabular learning, robot perception, and atomistic modeling by leveraging both unsupervised pseudo-supervision and label-driven optimization.
• OFG methods iteratively refine feature transformations through symbolic operations and statistical cues, achieving measurable performance gains in downstream tasks.

Omni-Feature Generator (OFG) is a designation used in recent machine-learning literature for systems that construct new feature spaces from raw observations, either by explicit feature engineering, by multi-view feature fusion, or by physically grounded featurization. In tabular machine learning, OFG denotes LLM-centered pipelines that apply unary and binary operations to raw variables, sometimes under generator-critic duet-play without labels and sometimes through multi-agent Tree-of-Thoughts, Monte Carlo Tree Search, or two-stage rationale-and-code generation (Gong et al., 30 Apr 2025, Zhang et al., 2024, Ko et al., 30 Mar 2025). In robot manipulation, the term names a deformable attention module that lifts $M$ image feature maps into a unified bird’s-eye-view (BEV) 3D feature volume (Mao et al., 16 Aug 2025). In atomistic machine learning, it has also been used as a unifying interpretation of the Gaussian-multipole (GMP) featurization scheme, which maps local electron-density structure to fixed-length, rotationally invariant descriptors (Lei et al., 2021).

1. Scope of the term and core mathematical formulations

The literature does not use OFG as a single canonical algorithm. Instead, the name is attached to several mechanisms whose common function is feature-space construction. In the tabular setting, one formulation starts from a labeled dataset $D=\langle F,y\rangle$ with original feature set $F=\{f_1,\dots,f_n\}$ and seeks a reconstructed feature set $F^*$ maximizing downstream performance,

$F^*=\arg\max_{\hat F}\,\theta_{R(\hat F,y)},$

where $R$ denotes the downstream task and $\theta$ its evaluation criterion (Zhang et al., 2024). A second formulation is explicitly unsupervised: with $X\in\mathbb{R}^{n\times d}$ and operator set $O=\{+, -, *, /, \log, \exp, \dots\}$, a critic implements $k:\mathbb{R}^{n\times d}\to\Theta$, where $D=\langle F,y\rangle$0 is textual advice, and a generator implements $D=\langle F,y\rangle$1, where $D=\langle F,y\rangle$2 is a token sequence representing transformed features (Gong et al., 30 Apr 2025).

The same naming pattern appears outside tabular ML. In OmniD, OFG is a 3D-feature-fusion core that maps multi-view image features $D=\langle F,y\rangle$3 to a fused BEV tensor $D=\langle F,y\rangle$4 (Mao et al., 16 Aug 2025). In GMP featurization, the “OFG” interpretation refers to a fixed-dimensional per-atom representation derived from multipole projections of approximate valence-electron density (Lei et al., 2021).

Context	Input $D=\langle F,y\rangle$5 output	Core mechanism
Unsupervised tabular transformation	$D=\langle F,y\rangle$6	generator-critic duet-play, in-context generation
Dynamic adaptive tabular generation	$D=\langle F,y\rangle$7	multi-agent ToT, feedback loop, MCTS
Reason-generation tabular engineering	raw variables $D=\langle F,y\rangle$8 rationale $D=\langle F,y\rangle$9 pandas code	two-stage conversational dialogue, SFT, DPO
Multi-view robot perception	$F=\{f_1,\dots,f_n\}$0	deformable attention-based BEV fusion
Atomistic featurization	local electron density $F=\{f_1,\dots,f_n\}$1 fixed-length per-atom vector	Gaussian multipole projections, MCSH, radial Gaussians

2. Generator-critic OFG for unsupervised feature transformation

In "Unsupervised Feature Transformation via In-context Generation, Generator-critic LLM Agents, and Duet-play Teaming" (Gong et al., 30 Apr 2025), the OFG approach is also referred to as LPFG. Its architecture consists of a critic agent, a generator agent, and an iterative refinement loop. The critic diagnoses the raw dataset along two axes—semantic, defined by feature names versus task description, and distributional, defined by feature-space statistics—and emits “advice” or “textual gradients” about how the feature space could be improved. The generator tokenizes original features and operators into a symbolic grammar and produces new feature-transformation expressions such as $F=\{f_1,\dots,f_n\}$2, $F=\{f_1,\dots,f_n\}$3, and $F=\{f_1,\dots,f_n\}$4.

The iterative loop is specified as

$F=\{f_1,\dots,f_n\}$5

with initialization $F=\{f_1,\dots,f_n\}$6, repetition for $F=\{f_1,\dots,f_n\}$7 iterations, and termination when convergence or no meaningful advice is observed. The tokenization grammar is expressed informally as

$F=\{f_1,\dots,f_n\}$8

with $F=\{f_1,\dots,f_n\}$9 and $F^*$0.

A central claim of this formulation is that pseudo-supervision can be derived from unlabeled data. Semantic pseudo-labels come from feature names and task description; distributional pseudo-labels come from summary statistics such as skewness and correlation structure; and the textual advice $F^*$1 serves as the only supervision. The paper states that the duo of critic and generator therefore constitutes a “pseudo model,” “pseudo objective,” and “pseudo optimization.” Although the critic output is textual, the method is described as implicitly guided by distributional cues such as variance improvement $F^*$2, correlation reduction, and cluster separation estimated by silhouette score on $F^*$3. The same framework can be generalized to human-agent collaborative generation by replacing the critic agent with human experts.

3. Dynamic, adaptive, and reason-generative OFG in tabular machine learning

In "Dynamic and Adaptive Feature Generation with LLM" (Zhang et al., 2024), OFG is an end-to-end pipeline around a large-language-model core, augmented with prompted agent-creation, Tree-of-Thoughts (ToT) reasoning, and Monte Carlo Tree Search (MCTS). The user supplies raw feature set $F^*$4, label $F^*$5, and operation set $F^*$6. The LLM is prompted using a ToT template to spawn $F^*$7 “expert” agents $F^*$8. Each agent inspects an assigned subset $F^*$9, applies operations $F^*=\arg\max_{\hat F}\,\theta_{R(\hat F,y)},$0 in sequence, emits an extended subset $F^*=\arg\max_{\hat F}\,\theta_{R(\hat F,y)},$1, and records full step-by-step reasoning. The subsets are unioned into $F^*=\arg\max_{\hat F}\,\theta_{R(\hat F,y)},$2, which is then evaluated by downstream model $F^*=\arg\max_{\hat F}\,\theta_{R(\hat F,y)},$3 via $F^*=\arg\max_{\hat F}\,\theta_{R(\hat F,y)},$4, after which the agents exchange feedback and revise subsequent operations. In parallel or at the end, an MCTS runs over the generation tree to explore promising branches.

The adaptive control layer uses both self-evaluation and peer-learning. Agent $F^*=\arg\max_{\hat F}\,\theta_{R(\hat F,y)},$5 tracks $F^*=\arg\max_{\hat F}\,\theta_{R(\hat F,y)},$6 versus $F^*=\arg\max_{\hat F}\,\theta_{R(\hat F,y)},$7; if $F^*=\arg\max_{\hat F}\,\theta_{R(\hat F,y)},$8, its strategy is reinforced. MCTS uses standard UCB selection,

$F^*=\arg\max_{\hat F}\,\theta_{R(\hat F,y)},$9

where $R$0 denotes visits to node $R$1, $R$2 visits to its parent, and

$R$3

The operation set can be extended beyond numeric transforms to handle categorical, textual, timestamp, and image inputs, and the same pipeline is reported for classification, regression, and ranking. Interpretability is explicit: each agent produces a fully documented ToT, the prompt enforces “Markdown and Organization,” and the final feature set is accompanied by a human-readable log of transforms and $R$4 improvements.

In "FeRG-LLM : Feature Engineering by Reason Generation LLMs" (Ko et al., 30 Mar 2025), OFG appears as a two-stage conversational dialogue. Stage 1 takes domain, task type, raw-feature descriptions, and a machine-learning objective, then generates a Chain-of-Thought rationale consisting of numbered definitions for candidate feature ideas. Stage 2 asks the model to create new features in Python code using those ideas; the output is executable pandas code that realizes each idea as a new column. The underlying model stack is Llama 3.1 8B, fine-tuned with LoRA on approximately $R$5 two-turn dialogues and then aligned by Direct Preference Optimization (DPO) to prefer rationales that improve downstream AUC or MSE. The DPO objective is given as

$R$6

with

$R$7

This variant emphasizes local deployability, lower inference time, and the avoidance of cloud-hosted APIs.

4. OFG as a representation-construction module beyond tabular data

In "OmniD: Generalizable Robot Manipulation Policy via Image-Based BEV Representation" (Mao et al., 16 Aug 2025), OFG is not a symbolic feature-engineering engine but the 3D-feature-fusion core of a three-stage visuomotor policy. Stage 1 performs 3D feature learning: $R$8 monocular RGB images are passed through a shared ResNet-18 backbone to yield 2D feature maps $R$9, and OFG lifts and fuses these maps into a unified BEV 3D feature volume. Stage 2 performs dimensional compression by channel pooling, and Stage 3 performs conditional denoising by concatenating the compressed BEV feature with proprioceptive robot states and feeding the result into a diffusion-based action head.

The internal OFG mechanism is deformable attention over BEV queries. A fixed set of $\theta$0 spatial query points is embedded into $\theta$1. Each 3D query $\theta$2 is projected into each camera view by intrinsics and extrinsics, producing a 2D reference $\theta$3. A lightweight MLP predicts offsets $\theta$4 and attention scores $\theta$5, which are normalized as

$\theta$6

The final fused feature is

$\theta$7

where $\theta$8. OFG is trained end-to-end, and the sole loss is the diffusion policy’s MSE denoising loss

$\theta$9

In "A Universal Framework for Featurization of Atomistic Systems" (Lei et al., 2021), OFG appears as a unified reading of the Gaussian-multipole featurization scheme. Each atom is associated with an approximate valence-electron density $X\in\mathbb{R}^{n\times d}$0 represented as a fixed linear combination of atom-centered Gaussians. Radial probes $X\in\mathbb{R}^{n\times d}$1 and angular probes given by Maxwell-Cartesian spherical harmonics $X\in\mathbb{R}^{n\times d}$2 are used to compute multipole moments

$X\in\mathbb{R}^{n\times d}$3

The per-atom feature vector groups components of the same multipole order $X\in\mathbb{R}^{n\times d}$4 and enforces rotation invariance through a weighted Euclidean norm,

$X\in\mathbb{R}^{n\times d}$5

The total feature dimension remains fixed regardless of the number of chemical elements, and the features interpolate smoothly across element types through the Gaussian decomposition of isolated valence densities.

5. Experimental protocols and quantitative performance

The empirical literature evaluates OFG under different downstream criteria: classification accuracy and AUC in tabular learning, MSE in regression, success rate and out-of-distribution robustness in robot manipulation, and MAE versus runtime or feature count in atomistic prediction (Gong et al., 30 Apr 2025, Zhang et al., 2024, Ko et al., 30 Mar 2025, Mao et al., 16 Aug 2025, Lei et al., 2021).

Work	Benchmark setting	Reported result
Unsupervised generator-critic OFG	12 public tabular binary-classification sets; baselines TTG, AutoFeat, GRFG, OpenFE, CAAFE	unsupervised yet outperforms all supervised baselines on 8/12 datasets, matches second-best on the rest; average accuracy lift over original features: $X\in\mathbb{R}^{n\times d}$6–$X\in\mathbb{R}^{n\times d}$7; consistently $X\in\mathbb{R}^{n\times d}$8s per dataset
Dynamic adaptive OFG	UCI/Kaggle: Ionosphere, Amazon Reviews, Abalone, Diabetes Indicators; RF accuracy	Raw $X\in\mathbb{R}^{n\times d}$9; RL $O=\{+, -, *, /, \log, \exp, \dots\}$0; OFG-3 iter $O=\{+, -, *, /, \log, \exp, \dots\}$1; OFG $O=\{+, -, *, /, \log, \exp, \dots\}$2 iters $O=\{+, -, *, /, \log, \exp, \dots\}$3
FeRG-LLM OFG	binary classification AUC and regression MSE	dataset-averaged AUC: XGB $O=\{+, -, *, /, \log, \exp, \dots\}$4, CAAFE $O=\{+, -, *, /, \log, \exp, \dots\}$5, FeatLLM $O=\{+, -, *, /, \log, \exp, \dots\}$6, Llama 3.1 70B $O=\{+, -, *, /, \log, \exp, \dots\}$7, FeRG-LLM $O=\{+, -, *, /, \log, \exp, \dots\}$8; MSE reductions include Concrete $O=\{+, -, *, /, \log, \exp, \dots\}$9 and Realestate $k:\mathbb{R}^{n\times d}\to\Theta$0
OmniD OFG	in-distribution, OOD, few-shot robot manipulation	average improvement over best baseline: $k:\mathbb{R}^{n\times d}\to\Theta$1, $k:\mathbb{R}^{n\times d}\to\Theta$2, and $k:\mathbb{R}^{n\times d}\to\Theta$3; removing OFG drops multi-view BCDE performance from $k:\mathbb{R}^{n\times d}\to\Theta$4 to $k:\mathbb{R}^{n\times d}\to\Theta$5
GMP as OFG	MD17, QM9, OC20	MD17 Aspirin: BP+HDNN $k:\mathbb{R}^{n\times d}\to\Theta$6–$k:\mathbb{R}^{n\times d}\to\Theta$7 meV at $k:\mathbb{R}^{n\times d}\to\Theta$8–$k:\mathbb{R}^{n\times d}\to\Theta$9 ms/image, GMP+SNN $D=\langle F,y\rangle$00 meV at $D=\langle F,y\rangle$01 ms/image and $D=\langle F,y\rangle$02 meV at $D=\langle F,y\rangle$03 ms/image; QM9 reaches $D=\langle F,y\rangle$04 meV at $D=\langle F,y\rangle$05k, below chemical accuracy $D=\langle F,y\rangle$06 meV

Within the unsupervised duet-play setting, ablations isolate the value of textual guidance. Replacing the critic with downstream accuracy feedback (LPFG-a), with feature importance from Random Forest (LPFG-i), or removing the critic entirely so that a single LLM diagnoses and generates (LPFG-o) all degrade performance; LPFG-o falls behind by $D=\langle F,y\rangle$07–$D=\langle F,y\rangle$08 on average, and LPFG-a and LPFG-i underperform even LPFG-o on many datasets (Gong et al., 30 Apr 2025). In the dynamic adaptive setting, OFG-3 already captures most of the gain in three rounds, while the full system adds further $D=\langle F,y\rangle$09–$D=\langle F,y\rangle$10 and yields similar gains of $D=\langle F,y\rangle$11–$D=\langle F,y\rangle$12 absolute in Precision, Recall, and F1 (Zhang et al., 2024). In FeRG-LLM, inference speed is reported as $D=\langle F,y\rangle$13 s per dataset on $D=\langle F,y\rangle$14A6000 in cloud settings and approximately $D=\langle F,y\rangle$15 s locally, compared with $D=\langle F,y\rangle$16 s for Llama 3.1 70B on $D=\langle F,y\rangle$17A6000 and $D=\langle F,y\rangle$18 s on $D=\langle F,y\rangle$19A100 (Ko et al., 30 Mar 2025). In OmniD, OFG yields up to $D=\langle F,y\rangle$20 success under background OOD on four novel textures and $D=\langle F,y\rangle$21 zero-shot success under position OOD, whereas competitors are reported at $D=\langle F,y\rangle$22 in both settings (Mao et al., 16 Aug 2025).

6. Interpretive issues, limitations, and future directions

A common misconception is that OFG denotes one standardized architecture. The literature instead spans at least four distinct paradigms: unsupervised generator-critic feature transformation, supervised multi-agent adaptive feature generation, reason-generation followed by executable code synthesis, and deformable attention-based feature fusion; GMP adds a physically motivated featurization interpretation (Gong et al., 30 Apr 2025, Zhang et al., 2024, Ko et al., 30 Mar 2025, Mao et al., 16 Aug 2025, Lei et al., 2021). This suggests that OFG is best understood as a family resemblance among methods that reorganize representation spaces, rather than as a single algorithmic specification.

Another misconception is that OFG is inherently supervised or label-driven. The LPFG formulation explicitly derives pseudo-supervision from feature names, task description, and summary statistics without numeric $D=\langle F,y\rangle$23, whereas the dynamic adaptive and FeRG-LLM variants optimize downstream performance with labels, and OmniD trains OFG only through the diffusion policy’s denoising loss without separate feature-reconstruction or attention-regularization losses (Gong et al., 30 Apr 2025, Zhang et al., 2024, Ko et al., 30 Mar 2025, Mao et al., 16 Aug 2025). The tabular papers also position OFG against brute-force or fixed-policy exploration: prior methods are described as enumerating operation combinations in BFS/DFS, using RL over expression DAGs, or relying on single-policy heuristics, whereas OFG variants separate diagnosis from generation, run simultaneous reasoning branches, or revisit decisions through tree search (Gong et al., 30 Apr 2025, Zhang et al., 2024).

The limitations are domain-specific. In dynamic adaptive LLM-based generation, repeated LLM queries and MCTS tree expansions can be expensive on very large datasets; noisy or poorly curated data may misguide the LLM; and reliance on GPT-3.5 Turbo risks embedding undesirable biases (Zhang et al., 2024). In GMP featurization, manual selection of Gaussian widths and maximum multipole order remains a limitation (Lei et al., 2021). Reported future directions include scaling through smaller distilled LLMs or adapter-based fine-tuning, integrating symbolic reasoning engines for guaranteed algebraic correctness, extending to fully mixed data such as audio and video via multi-modal LLMs, conducting deeper causal analysis of generated feature semantics, using self-consistent electron densities, adding spin-density probes for magnetic materials, embedding electrostatic or long-range physics directly into featurization, and automating hyperparameter optimization (Zhang et al., 2024, Lei et al., 2021). In the unsupervised duet-play setting, an additional extension is explicit human-agent collaborative generation by swapping the critic agent for human experts (Gong et al., 30 Apr 2025).