PoTRE introduces heterogeneous multi-agent framework for test-time reasoning

The authors introduce PoTRE (Poly-Topological Reasoning Ensembles), a framework designed to address the limitations of standard single-stream prompting in complex reasoning tasks. PoTRE decouples inference into four distinct agents: an Adversarial Refinement Agent, a Hierarchical strategic Planning Agent, a Spectrum Search Agent, and a Direct Chain Agent.

The system employs a Task-Adaptive Aggregation Layer that dynamically reconciles the outputs of these agents via candidate selection, semantic synthesis, or neuro-symbolic verification.
Evaluations on ARC-AGI-2, Humanity's Last Exam (HLE), and PRBench Finance show PoTRE achieves state-of-the-art accuracy of 49.92% on HLE.
The framework demonstrates improved reasoning performance using similar or fewer inference tokens compared to heavily scaled homogeneous baselines.

This architectural heterogeneity allows for robust global solutions in novel abstractions and rigorous domain constraints without requiring the massive scale of traditional models.

Benchmarks

Benchmark	Model	Score
Humanity's Last Exam	PoTRE	49.92%

Benchmark

Model

Score

Humanity's Last Exam

PoTRE

49.92%

Benchmarks

Deep Interaction enables precise human correction of LLM reasoning errors

Ring-Zero scales zero RL to 1T parameters for emergent reasoning

Qwen3-30B-A3B achieves 8.3% on FukuyamaBench for mechanistic reasoning

Theoretical framework explains Transformer inductive reasoning via invariant manifold

TreeThink: modular async tree search library for LLM theorem proving