KACE: Knowledge-Adaptive Context Engineering for Mathematical Reasoning

· Source: Artificial Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

Knowledge-Adaptive Context Engineering (KACE) addresses the context bloat limitation in large language models performing mathematical reasoning by separating knowledge storage from usage. Unlike existing methods that conflate these aspects, KACE organizes learned guidance through difficulty- and domain-based stratification. Offline, a self-reflective learning loop constructs an "epistemic tree," a knowledge base of typed cards categorized by problem difficulty and epistemic domain, based on failure origins. During evaluation, a tiered self-consistency mechanism dynamically classifies problems as easy, medium, or hard using per-tier agreement gates. Easy problems are solved directly, while harder problems retrieve relevant branches from the epistemic tree. This approach achieves 62.2% accuracy on AIME 2025, representing a 10.4-point absolute gain over fixed Best-of-5 self-consistency and a 5.6-point gain over Tiered + GEPA, both at comparable solver-call budgets. KACE also shows consistent gains on MATH-HARD and OlymMATH, classifying problem difficulty with 78% pairwise concordance.

Key takeaway

For Machine Learning Engineers developing LLMs for mathematical reasoning, if you are struggling with context bloat and performance ceilings, KACE provides a robust solution. By separating knowledge storage from usage and dynamically retrieving difficulty- and domain-specific guidance, your models can achieve substantial accuracy gains, such as the 10.4-point increase observed on AIME 2025. You should explore implementing a tiered self-consistency and epistemic tree approach to enhance your model's problem-solving capabilities.

Key insights

KACE improves mathematical reasoning in LLMs by adaptively retrieving knowledge from a difficulty- and domain-stratified epistemic tree.

Principles

Method

KACE uses an offline self-reflective loop to build an epistemic tree of typed cards. At evaluation, tiered self-consistency classifies problem difficulty to retrieve matching knowledge branches.

In practice

Topics

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

Related on AIssential

Open in AIssential →

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