Personal
AI Behavioral Runtime
Steering a model's reasoning along principles defined in advance. It lifts the benchmark, and every step pays for it.
Interim resultSteering works. Forcing a model to reason along defined principles lifted the benchmark without touching native reasoning, which is the result the experiment was after. It is not free: the steering is paid on every step, and anything the principles never described is exactly as unsolved as it was before.
- Stack
- Reasoning runtime / Gemini 3.1 Pro
- Aspects compared
- 4
- Open limits
- 4
01 / The question
Can a defined set of principles aim a model's reasoning in a better direction, without touching the reasoning itself?
A Flash-class model, or any model that is not frontier, can hold plenty of knowledge and still fail because it aimed the reasoning the wrong way. Knowing the material and choosing what to do with it are not the same capability.
This is not only a small-model problem. Gemini 3.1 Pro is one observed example: the capability of a large model, and still a run that ends at the wrong answer because it started down the wrong path.
02 / How it ran
Step through the run.
Four moves, in the order they happened.
02 / How it ran / step 01 of 04
Watch where the run goes wrong.
Not which facts were missing — which turn the reasoning took. The failures that mattered were direction failures: a plausible first move that nothing later could recover from.
03 / The approaches
Two builds, one question.
Switch between them. Same slots, same order, so the difference is the content.
The model on its own
The model gets the problem and picks its own direction. Cheap, short, and entirely dependent on the first move happening to be the right one.
- No steering overhead — you pay for the reasoning and nothing else.
- Direction is whatever the model chose this time.
- A wrong first move is rarely recovered later in the run.
- Nothing about the failure tells you how to prevent the next one.
04 / What came out
Aspect by aspect.
Choose a row to read why it went the way it did.
Principle-guided run — clear
Direction of reasoning
Defined principles keep the run pointed at the problem it was given. This is the whole reason the runtime exists, and it is the part that worked.
Read against the same problems, on the same model. The benchmark moved; the figures behind it are not cleared for publication yet.
05 / Trade-offs
Nothing here was free.
What each gain cost, on the same row as the gain.
What it bought
Principles that are actually written down help the model work in a more correct direction, run after run.
What it cost
They have to be written down first, and a principle that is wrong is now wrong consistently rather than occasionally.
What it bought
The benchmark for that model moves up without changing the model.
What it cost
Reasoning gets more expensive: the additional steering is carried through every step of the run, so the lift is bought with tokens.
What it bought
Native reasoning is left intact, so the runtime rides on whatever the model already does well.
What it cost
It also inherits whatever the model does badly. The runtime aims the reasoning; it does not improve it.
06 / Still open
What is not solved.
4 limits the experiment has not answered.
A case that falls outside the defined principles is still unsolved. The runtime has no fallback for it and does not detect that it has left its own coverage.
The lift was read on one model. Whether the same principles help across model families has not been tested.
The cost of steering was not separated per principle, so there is no evidence yet about which principles are earning their keep.
No benchmark figures are cleared for publication.
Interim conclusion
Steering works. Forcing a model to reason along defined principles lifted the benchmark without touching native reasoning, which is the result the experiment was after. It is not free: the steering is paid on every step, and anything the principles never described is exactly as unsolved as it was before.
Process notes
- The runtime directs the process only. Native reasoning is left exactly as it was.
- The benchmark moved up, and the cost of reasoning moved up with it.
- A case the principles never described is still unsolved, and the runtime does not notice it has left its coverage.
