PoTRE: Test-Time Reasoning inspired by Cognitive Heterogeneity

Large Language Models (LLMs) often struggle with complex, multi-step reasoning, frequently failing when tasks require long-horizon planning or novel abstractions. Standard prompting techniques and homogeneous ensembles often suffer from "topological mode collapse," where multiple agents share the same blind spots and reinforce incorrect reasoning. This paper investigates whether orchestrating diverse, heterogeneous reasoning strategies can improve performance without requiring the massive parameter scaling typical of frontier models.

To address these limitations, the authors introduce PoTRE (Poly-Topological Reasoning Ensembles). Instead of relying on a single reasoning style, PoTRE decouples inference into four specialized agents that operate in parallel:

• Adversarial Refinement Agent: Uses a multi-turn debate between a proposer and a verifier to iteratively improve solution quality.
• Hierarchical Strategic Planning Agent: Employs a team structure (Planner, Executor, Verifier, Overseer) to decompose complex tasks and perform active meta-reasoning to pivot when stuck.
• Spectrum Search Agent: Generates multiple candidate solutions using stochastic sampling and filters them through constraint-guided or semantic judging.
• Direct Chain Agent: Provides a standard step-by-step reasoning baseline.

A final Task-Adaptive Aggregation Layer reconciles these diverse outputs using methods like candidate selection, semantic synthesis, or neuro-symbolic verification to produce a robust global solution.

PoTRE demonstrates that architectural heterogeneity can effectively substitute for parameter scale. By applying this framework to the Gemini-3-Flash-Preview model, the authors achieved state-of-the-art results on the Humanity's Last Exam (HLE) benchmark (49.92% accuracy), outperforming larger, homogeneous baselines. Furthermore, the framework is modular; the authors show that pruning specific sub-agents based on the domain can reduce inference token consumption by up to 85% while maintaining or improving accuracy, highlighting the efficiency of structured test-time compute allocation.