CoFEE: Reasoning Control for LLM-Based Feature Discovery

2026-04-23 · Source: Artificial Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, quick

Summary

The CoFEE (Cognitive Feature Engineering Engine) framework introduces reasoning control for Large Language Models (LLMs) to enhance feature discovery from complex unstructured data. This method addresses the challenge of identifying predictive abstractions while avoiding data leakage, proxies, and post-outcome signals. CoFEE enforces specific cognitive behaviors, such as backward chaining, subgoal decomposition, verification against observability and leakage criteria, and explicit backtracking, acting as structured inductive biases. In controlled comparisons, CoFEE-generated features demonstrated a 15.2% higher average Success Rate Score compared to unconstrained vanilla LLM prompts. Furthermore, CoFEE reduced the number of generated features by 29% and cut costs by 53.3%, indicating improvements in both quality and efficiency of LLM-based feature discovery.

Key takeaway

For AI Engineers and Research Scientists developing LLM-based feature engineering pipelines, integrating reasoning control frameworks like CoFEE can significantly enhance feature quality and reduce operational costs. You should consider implementing structured cognitive behaviors to guide LLM reasoning, leading to more predictive features and a more efficient discovery process.

Key insights

Enforcing cognitive behaviors in LLMs significantly improves feature discovery quality and efficiency.

Principles

• Feature discovery is a reasoning problem.
• Cognitive behaviors act as inductive biases.

Method

CoFEE enforces cognitive behaviors like backward chaining, subgoal decomposition, verification, and backtracking during LLM-based feature discovery to improve feature quality and efficiency.

In practice

• Implement backward chaining from outcomes.
• Decompose subgoals for complex features.
• Verify features against observability criteria.

Topics

• CoFEE
• Feature Discovery
• Large Language Models
• Reasoning Control
• Cognitive Behaviors

Best for: AI Engineer, Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• Benchmarking LLMs at the Game Of Science (Eleusis) — LLM scientific reasoning involves distinct inductive ability and metacognition, measurable through iterative hypothesis testing.
• CogRAG+: Cognitive-Level Guided Diagnosis and Remediation of Memory and Reasoning Deficiencies in Professional Exam QA — Aligning RAG with human cognitive hierarchies significantly boosts LLM performance on professional exams without fine-tuning.
• Feature Engineering with LLMs: Techniques & Python Examples — LLMs automate semantic feature engineering, transforming raw data into context-rich representations for improved ML performance.
• Behavior Cue Reasoning: Monitorable Reasoning Improves Efficiency and Safety through Oversight — Behavior Cues improve LLM reasoning monitorability and controllability without performance cost, enabling efficient and safer AI systems.
• LLM-as-an-Investigator: Evidence-First Reasoning for Robust Interactive Problem Diagnosis — LLM-as-an-Investigator uses evidence-first reasoning to counter user-driven sycophancy in technical problem diagnosis.
• CogManip: Benchmarking Manipulative Behavior in Multi-Turn Interactions with Large Language Model — CogManip benchmarks LLM psychological manipulation across 15 strategies in 1,000 multi-turn scenarios, revealing varied risks and prompt sensitivity.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.