Human-LLM Collaborative Feature Engineering for Tabular Learning

Conference: ICLR 2026 arXiv: 2601.21060 Code: None Area: AutoML / Tabular Learning Keywords: Feature Engineering, Human-AI Collaboration, Bayesian Optimization, LLM, Tabular Data

TL;DR

This paper proposes a human-LLM collaborative feature engineering framework that decouples the proposal and selection of feature operations. A Bayesian neural network models operation utility and uncertainty to guide selection, with selective human preference feedback incorporated when appropriate. The framework achieves 8.96%–11.23% average error rate reduction across 18 tabular datasets.

Background & Motivation

Background: LLMs are widely adopted in tabular learning for automated feature engineering, leveraging semantic understanding to generate meaningful feature transformation operations (e.g., CAAFE, OCTree).

Limitations of Prior Work: Existing methods use LLMs simultaneously as both proposers and selectors of feature operations, relying entirely on LLM-internal heuristics without calibrated estimates of operation utility and uncertainty. This leads to repeated exploration of low-gain operations and poor performance under limited iteration budgets.

Key Challenge: LLMs excel at generating diverse candidate feature transformations but are ill-suited for making optimal selections among them — a fundamental tension between strong proposal ability and weak selection ability.

Goal: To decouple LLM-based operation proposal from selection, and to effectively integrate human expert knowledge into the selection process to improve feature engineering efficiency.

Key Insight: Drawing on Bayesian optimization, an explicit surrogate model replaces the LLM's implicit selection, and a selective human feedback mechanism is designed to control the cost of expert involvement.

Core Idea: The LLM is responsible solely for proposing candidate feature operations; selection is guided by a UCB strategy over a Bayesian neural network, with human preference feedback selectively queried under high uncertainty.

Method

Overall Architecture

In each feature engineering round: (1) the LLM generates \(N\) candidate feature transformation operations based on task descriptions, feature semantics, and historical performance; (2) a Bayesian neural network (BNN) surrogate estimates the expected utility \(\mu_t(e)\) and uncertainty \(\sigma_t^2(e)\) for each operation; (3) a UCB strategy selects operations and selectively queries human preference feedback when conditions are met; (4) the actual utility of the selected operation is evaluated and the surrogate model is updated.

Key Designs

1. Feature Operation Encoding and BNN Surrogate:

   • Function: Maps LLM-generated natural language feature operations to vector representations and estimates utility via a Bayesian neural network.
   • Mechanism: Operation embeddings are formed by concatenating semantic embeddings \(\phi_{\text{embedding}}(e)\) (text-embedding-3-small) and column usage encodings \(\phi_{\text{column}}(e) \in \{0,1\}^d\). The BNN learns a variational posterior \(q_t(\boldsymbol{\theta}) = \mathcal{N}(\boldsymbol{\theta}; \boldsymbol{M}_t, \boldsymbol{\Sigma}_t)\) via variational inference, yielding predictive mean \(\mu_t(e)\) and variance \(\sigma_t^2(e)\).
   • Design Motivation: Gaussian processes scale poorly in the high-dimensional language-derived feature space and are less suited for modeling non-stationarity; BNNs address this limitation. Column usage encodings resolve ambiguity when multiple columns have similar semantic descriptions.

2. Selective Human Preference Feedback Mechanism:

   • Function: After UCB selects the best candidate \(e_t^a\), determines whether to query a human expert for preference feedback.
   • Mechanism: Querying is triggered only when two conditions are simultaneously satisfied — (C1) confidence interval overlap: \(\text{UCB}_t(e_t^b) > \text{LCB}_t(e_t^a)\), ensuring non-trivial uncertainty; (C2) sufficiently large uncertainty: \(\sqrt{\beta_t}(\sigma_t(e_t^a) + \sigma_t(e_t^b)) \geq \gamma_\kappa\), ensuring potential gain exceeds query cost.
   • Design Motivation: Indiscriminate querying imposes unnecessary cognitive burden; human intervention is solicited only when feedback can yield a significant utility gain.

3. Posterior Update via Preference Feedback:

   • Function: Incorporates human preference feedback \(Z_t\) into the surrogate model's posterior distribution.
   • Mechanism: Preference feedback is modeled via a probit likelihood \(\mathcal{P}(Z_t | \boldsymbol{\theta}, e_t^a, e_t^b) = \Phi(\eta Z_t [\hat{g}(\phi(e_t^a); \boldsymbol{\theta}) - \hat{g}(\phi(e_t^b); \boldsymbol{\theta})])\); the variational posterior \(q_t'(\boldsymbol{\theta})\) is updated and the updated UCB values are used for final selection.
   • Design Motivation: Probabilistic treatment of human feedback is more robust than direct adoption, smoothing the effect of noisy responses.

Loss & Training

The BNN is trained by minimizing the ELBO: \(\text{KL}(q_t(\boldsymbol{\theta}) \| \mathcal{P}(\boldsymbol{\theta})) - \mathbb{E}_{q_t(\boldsymbol{\theta})}[\log \mathcal{P}(H_t | \boldsymbol{\theta})]\). The UCB exploration coefficient is \(\beta_t = 2\log(|\mathcal{S}_t|\pi^2 t^2 / 3\delta)\) with \(\delta=0.1\). The human query cost threshold is \(\gamma_\kappa=4\).

Key Experimental Results

Main Results

Dataset	Metric	Ours (w/o human)	Ours (w/ human)	Best Baseline	Gain (w/ human)
13 classification datasets (MLP)	AUROC (%)	85.3	85.5	84.7 (OCTree)	Error rate ↓ 8.96%
13 classification datasets (XGBoost)	AUROC (%)	87.4	87.4	86.7 (OCTree)	Error rate ↓ 11.23%
flight (MLP)	AUROC (%)	96.9	97.3	94.8 (OCTree)	Error rate ↓ +48.1%
conversion (XGBoost)	AUROC (%)	93.5	93.9	92.4 (OCTree)	Error rate ↓ +11.5%

Ablation Study

Configuration	Metric	Notes
Different LLM backbone (GPT-5)	MLP avg. 85.9→86.5	Ours (w/ human) achieves best performance under GPT-5 backbone
Different LLM backbone (GPT-3.5)	MLP avg. 84.6→85.1	Advantage is maintained even with a weaker backbone
User study (ALG vs. Control)	Performance: p=0.011	ALG framework significantly improves user performance
User study (ALG vs. Control)	Completion time: p<0.001	ALG framework significantly reduces completion time

Key Findings

• LLM-based methods consistently outperform traditional AutoML approaches (OpenFE, AutoGluon), validating the value of semantic understanding for feature engineering.
• Explicitly modeling utility and uncertainty outperforms relying purely on LLM heuristics by 7.24% and 9.02% error rate reduction, respectively.
• Human preference feedback consistently provides additional gains, while the computational overhead of BNN+UCB accounts for only 2.2% of total runtime.

Highlights & Insights

• Introducing Bayesian optimization principles into LLM-driven feature engineering and decoupling proposal from selection is an elegant system design. The theoretical guarantees of UCB for exploration–exploitation trade-offs render the selection process transparent and principled.
• The two conditions of the selective querying mechanism (confidence interval overlap + uncertainty gating) are grounded in solid theoretical analysis (Lemma 3.1–3.2), achieving an optimal trade-off between human cognitive cost and information gain.

Limitations & Future Work

• Human feedback in experiments is simulated by GPT-4o; the actual user study is conducted on only a single dataset, limiting generalizability.
• BNN surrogate calibration quality may be poor in early rounds with sparse data; the cold-start problem is not sufficiently discussed.
• The framework models only single-operation utility and does not capture interaction effects among combinations of multiple operations.

Related Work & Insights

• vs. CAAFE: CAAFE uses the LLM for both proposing and selecting feature operations, making it prone to local optima; the proposed decoupling enables continuous discovery of high-value operations.
• vs. OCTree: OCTree uses decision tree feedback to guide the LLM but still relies on LLM-internal heuristics for selection; this work employs a BNN to provide better-calibrated utility estimates.
• vs. Traditional Bayesian Optimization: Conventional BO uses Gaussian processes as surrogate models, which are effective in low-dimensional spaces; this work applies BNNs to handle high-dimensional language embedding spaces.

Rating

• Novelty: ⭐⭐⭐⭐ The framework design of decoupling proposal and selection with integrated human feedback is novel, supported by complete theoretical analysis.
• Experimental Thoroughness: ⭐⭐⭐⭐ Validated across 18 datasets with user studies and computational scalability analysis from multiple perspectives.
• Writing Quality: ⭐⭐⭐⭐ Problem motivation is clear, theoretical derivations are rigorous, and experimental presentation is comprehensive.
• Value: ⭐⭐⭐ The practical application scenario is well-defined, though LLM API costs are required; the method demonstrates good generality.

Related Papers

• [NeurIPS 2025] CLIMB: Class-Imbalanced Learning Benchmark on Tabular Data
• [ICLR 2026] TabStruct: Measuring Structural Fidelity of Tabular Data
• [ICLR 2026] LLM Unlearning with LLM Beliefs
• [ICLR 2026] Unpacking Human Preference for LLMs: Demographically Aware Evaluation with the HUMAINE Framework
• [AAAI 2026] MAPS: Multi-Agent Personality Shaping for Collaborative Reasoning