> *"Standard single-stream prompting appears brittle in the face of novel abstraction or strict domain constraints. We introduce PoTRE — a heterogeneous framework that decouples reasoning into four agents: an Adversarial Refinement Agent, a Hierarchical Strategic Planning Agent, a Spectrum Search Agent, and a Direct Chain Agent. A final task-adaptive aggregation layer dynamically orchestrates these perspectives."*
The Lone Genius vs. the Diverse Team
Imagine facing an extremely difficult puzzle alone. You may be brilliant, but your thinking style is singular — you reason left-to-right, front-to-back. When a puzzle requires leaps of thought, multi-angle verification, or simultaneous reasoning across abstraction levels, you get stuck.
Now imagine four friends with very different styles: a devil's advocate who hunts for flaws, a strategist who decomposes big problems, a diverger who brainstorms wild possibilities, and an intuitionist who answers directly on gut feeling. Each solves the puzzle independently, then they compare notes: the devil's advocate exposes holes in the strategist's plan; the diverger inspires the intuitionist; the strategist organizes everything into an executable plan. The group's answer beats any individual's.
That is the core idea of PoTRE (Poly-Topological Reasoning Ensembles): cognitive heterogeneity is the key to complex reasoning.
The Four-Agent Reasoning Circus
🎭 Adversarial Refinement Agent
Like a harsh reviewer of your paper draft, this agent ignores your overall brilliance and only hunts for weaknesses: "Does this assumption hold? Is there a counterexample? Is this a skipped step?" It takes preliminary reasoning and systematically attacks its weak points, forcing rigor. In mathematical proofs this role is crucial — many errors come from "obviously" true assumptions, and the adversary's job is to challenge every "obviously."
🗺️ Hierarchical Strategic Planning Agent
Like a mountain expedition leader, it does not just say "climb up." It analyzes: where to set base camp, where the dangerous ice walls are, when to summit, how much oxygen is needed. It decomposes the big problem into subgoals and concrete steps, evaluating feasibility and risk. For long-horizon planning problems (like abstract reasoning in ARC-AGI-2), this role is vital — without a map, reasoning easily falls into local optima, wandering in a small valley while the exit is over the next ridge.
🌈 Spectrum Search Agent
Like a radio sweeping the whole band, this agent does not analyze deeply — it quickly browses the solution space and flags interesting directions. It guarantees no important possibility is missed, not that each direction is correct. On problems requiring broad exploration (cryptanalysis, combinatorial optimization), it provides necessary breadth: the voice reminding everyone not to overlook that strange option.
⚡ Direct Chain Agent
Like a veteran doctor who reads a patient's complexion at the door and is already mostly sure of the diagnosis, this agent does "fast thinking": no lengthy intermediate steps, direct answers from pattern matching. For simple or familiar problems, intuitive reasoning is more efficient than systematic analysis. Its value lies not in solving the hardest problems, but in quickly handling the "obvious" cases, leaving cognitive resources for problems that genuinely need deep thought.
Task-Adaptive Aggregation: Dynamic Fusion, Not Simple Voting
Four agents give four answers — what now? Majority voting won't do, because the best solution method differs by problem. PoTRE's Task-Adaptive Aggregation Layer acts like a wise meeting chair, deciding dynamically how to combine the four perspectives:
- Final candidate selection: for problems needing precise answers (e.g., math proofs), pick the most reliable of the four candidates.
- Semantic synthesis: for problems needing creative answers (open-ended questions), fuse the essence of the four perspectives into a more comprehensive response.
- Neural-symbolic verification: for problems requiring strict guarantees, verify the consistency of candidates with symbolic logic to ensure no logical gaps.
- Humanity's Last Exam (HLE): an extremely difficult benchmark spanning math, physics, chemistry, biology, and computer science — domain experts are said to not always get everything right. The previous best official score was about 45%. PoTRE achieved 49.92% — a new state of the art.
- ARC-AGI-2: Francois Chollet's abstract reasoning corpus testing generalization to brand-new abstraction patterns — precisely the weakest area of current LLMs. PoTRE also performed strongly here.
- PRBench Finance: a professional finance reasoning benchmark combining domain knowledge and logical reasoning. PoTRE again showed strong capability.
- Adversarial refinement provides logical constraints
- Hierarchical planning provides structural constraints
- Spectrum search provides possibility constraints
- Direct chaining provides experiential constraints
- Kankariya, A., & Arık, S. Ö. (2026). PoTRE: Test-Time Reasoning inspired by Cognitive Heterogeneity. arXiv:2607.20268. Accepted at TMLR 2026.
- Chollet, F. (2019). On the Measure of Intelligence. arXiv:1911.01547.
- Surowiecki, J. (2004). The Wisdom of Crowds. Doubleday.
- Kahneman, D. (2011). Thinking, Fast and Slow. Farrar, Straus and Giroux.
This layer is not a fixed rule — it is dynamic and context-sensitive, like a good team lead who knows when to decide, when to merge, and when to send work back for revision.
Historic Results on the Hardest Exams
PoTRE was evaluated on three frontier benchmarks:
Most strikingly: PoTRE uses a comparable or smaller number of reasoning tokens than heavily scaled homogeneous baselines. The heterogeneous architecture improves not only performance but efficiency — four specialized "small experts" working together beat one giant "generalist."
Why It Works: Echoes of Cognitive Science
PoTRE's design philosophy aligns with Multiple Constraints Theory in cognitive science: humans solve problems by simultaneously activating multiple knowledge sources (semantic, syntactic, contextual, world knowledge) that mutually constrain and reinforce each other, converging on a solution. PoTRE's four agents can be seen as an engineering realization:
When these constraints act simultaneously, the solution space shrinks dramatically, and the converged solution is more likely to be globally — not locally — optimal.
It can also be understood via the Wisdom of Crowds (Surowiecki, 2004): independently formed, diverse judgments combined through an aggregation mechanism often outperform any individual. PoTRE satisfies all three conditions: the agents reason independently, use diverse strategies, and converge through the adaptive layer.
A Bridge from Single-Modal to Multi-Topological
PoTRE signifies a paradigm shift: from "how to build a better single reasoning engine" to "how to orchestrate multiple complementary reasoning engines." Like early industrial factories that moved from one all-purpose machine to specialized machines on an assembly line, PoTRE trains specialized reasoners plus a smart scheduler.
This resembles Mixture-of-Experts (MoE) superficially, but differs fundamentally: MoE "experts" are implicit, router-selected, and share the same architecture. PoTRE's agents are explicit, functionally defined, and architecturally distinct — true heterogeneity, not just a quantitative division of labor.
Looking Ahead: AI's "Multiple Personalities"
PoTRE opens a door: future AI systems may not be "one model" but "a team" — skeptics, planners, explorers, intuitionists, like human teams. This has deep implications:
1. Interpretability: ask "why this answer?" and observe the four agents' debate — more convincing than a black box's "I just know." 2. Safety: if the adversarial agent reliably flags flaws while the direct-chain agent rushes answers, safety mechanisms can reject direct-chain output when the adversary raises serious objections. 3. Continual learning: new agents can join the team — future "creative association agents," "causal reasoning agents," "empathy agents" — continually expanding cognitive capability. 4. Human-AI collaboration: humans can participate in aggregation. In medical diagnosis, a doctor's intuition could act as a fifth "agent" alongside the AI's four perspectives.