Optimal Reasoning Research | Optimal Reasoning

This is a working research agenda, not a claim that AGI is imminent or that quantum computing currently improves general-purpose AI reasoning. The aim is to turn an ambitious hypothesis into questions that can be measured and disproved.

Working thesis

The core idea

Generality may depend on the quality of the reasoning loop, not only the size of the model.

A capable system should recognise when a problem is difficult, choose how much computation to spend, search beyond its first answer, test intermediate states and update its plan when evidence changes.

ARC’s original framing is useful because it defines intelligence through skill-acquisition efficiency, generalisation difficulty, priors and experience—not task performance alone.[François Chollet / arXiv, 5 November 2019] ARC-AGI-2 extends that challenge by pairing capability with resource efficiency.[ARC Prize Foundation, 24 March 2025]

My hypothesis is that optimising this full loop could contribute to more general intelligence. It would not, by itself, establish AGI. A benchmark score is evidence about a capability, not a universal definition of intelligence.

Research model

A loop that earns each next step.

The system should decide when to continue, when to call a tool, when to revise and when to stop.

Frame

Represent the objective, constraints, available evidence and acceptable failure conditions.

Allocate

Estimate difficulty and spend additional reasoning compute only where it is likely to improve the result.

Explore

Generate competing plans or representations instead of committing to the first plausible path.

Evaluate

Use tests, tools, environment feedback or independent verifiers—not confidence alone.

Correct

Identify the failed assumption, revise the plan and retain a bounded memory of what changed.

Transfer

Test whether the correction improves performance on unfamiliar tasks rather than only the original case.

Evidence and limits

What current research supports—and what it does not.

The evidence is strongest for adaptive computation and feedback-supported correction. Claims about general self-improvement or quantum advantage remain much narrower.

Adaptive compute

More thinking helps selectively.

Research on test-time scaling shows that the best inference strategy depends on both the problem and the base model.[ICLR / OpenReview, ICLR 2025] The design question is therefore how to allocate compute, not how to maximise it everywhere.

External evaluation

Reflection needs evidence.

Intrinsic self-correction can fail or even degrade an answer when the model has no external feedback.[ICLR / OpenReview, 16 January 2024] Agent work using task feedback and memory suggests a more useful direction.[NeurIPS, NeurIPS 2023]

Tool mediation

MCP exposes capability; it does not create judgement.

MCP provides lifecycle management, capability negotiation and structured access to tools and resources.[Model Context Protocol, 26 March 2025 revision] The agent still needs routing logic, permissions, evaluation and a reason to call the tool.

General intelligence

A benchmark is a probe, not a finish line.

ARC is valuable for testing unfamiliar abstraction tasks and efficiency. It does not measure every capability required for safe, autonomous or broadly useful intelligence.

Conditional research path

Quantum-assisted planning—only when the structure earns it.

The research question is deliberately narrow: could a quantum service exposed through MCP improve a specific planning subproblem after a classical-first routing decision?

AgentDefines the planning problem

RouterTests relevance and constraints

Classical firstSolver or search baseline

Optional MCP toolQuantum or hybrid optimiser

VerifierChecks the candidate plan

QAOA was introduced for approximate combinatorial optimisation.[Farhi, Goldstone and Gutmann / arXiv, 14 November 2014] That makes constrained scheduling, routing or allocation a more plausible research target than open-ended language reasoning.

The broader literature stresses careful benchmarking against appropriate classical methods.[Nature Reviews Physics, 28 October 2024] Even promising work on constraint satisfaction documents substantial hardware, depth and comparative-performance challenges.[PRX Quantum, 10 September 2024]

MCP would be the controlled interface to such a service. It would not be the source of the optimisation advantage and would not give an agent permission to execute the resulting plan.

The routing gate

A quantum call should occur only when all five conditions are met.

The planning subproblem can be expressed as a constrained optimisation or search problem.
A quantum or hybrid method is benchmarked against strong classical solvers on the same objective.
Expected quality, latency or resource gains justify the integration and operational overhead.
The returned candidate can be independently checked before it affects a real workflow.
Data handling, permissions and auditability remain within the agent’s operating boundary.

Current programme

What I am investigating now.

The work is currently at the research-design stage: literature synthesis, architecture, evaluation criteria and experiment definition. I am not presenting unpublished benchmark results.

01
Can a controller allocate test-time compute from calibrated difficulty and uncertainty signals?
02
Which external evaluators reduce false corrections without creating brittle dependence on one verifier?
03
Can episodic error memory improve transfer while avoiding the accumulation of incorrect lessons?
04
Which planning problems retain enough formal structure to justify a specialised optimisation tool?
05
How should progress be measured across accuracy, action efficiency, compute, latency and correction quality?