Evolutionary Task Discovery: Advancing Reasoning Frontiers via Skill Composition and Complexity Scaling

2026-05-12

AI Summary

Current large language models exhibit reasoning capabilities constrained by the diversity and complexity of their training data, and conventional data synthesis approaches often yield homogeneous examples that fail to systematically expand reasoning boundaries. This work formulates task generation as a directed search over a manifold defined by algorithmic skill and complexity axes. It introduces structured crossover and parameterized mutation operators, coupled with a dynamic proximal zone-of-development filtering mechanism, to enable innovative skill composition and controllable complexity growth. Built upon an evolutionary algorithm framework, the proposed system integrates skill composition, complexity scaling, and learnability-based filtering into a unified task generation and curriculum learning pipeline. Extensive experiments demonstrate consistent and significant improvements in reasoning performance across diverse model architectures, pretraining strategies, and scales, with strong generalization to unseen tasks.

Abstract

The reasoning frontier of Large Language Models (LLMs) has advanced significantly through modern post-training paradigms (e.g., Reinforcement Learning from Verifiable Rewards (RLVR)). However, the efficacy of these methods remains fundamentally constrained by the diversity and complexity of the training data. One practical solution is data synthesis; yet, prevalent methods relying on unstructured mutation or exploration suffer from homogeneity collapse, failing to systematically expand the reasoning frontier. To overcome this, we propose Evoutionary Task Discovery (EvoTD), a framework that treats data synthesis as a directed search over a dual-axis manifold of Algorithmic Skills and Complexity Attributes. We introduce structured evolutionary operators to navigate this space: a Crossover operator that synthesizes novel skill compositions to enhance diversity, and a Parametric Mutation operator that scales structural constraints (e.g., input size, tree depth) to drive robust generalization. Crucially, we integrate a dynamic Zone of Proximal Development filter, ensuring tasks lie within the learnable region of the model. Empirically, EvoTD delivers substantial reasoning gains that generalize consistently across model architectures, pretraining regimes, and scales, demonstrating that structured evolutionary curricula can effectively support reasoning improvement. We release our code on https://github.com/liqinye/EvoTD.

Problem

• reasoning frontier
• data synthesis
• homogeneity collapse
• task diversity
• complexity scaling

Innovation

• Evolutionary Task Discovery
• Skill Composition
• Complexity Scaling
• Structured Evolutionary Operators
• Zone of Proximal Development