Taste did not make the AI researcher better — it only changed the route

An autonomous AI researcher does not necessarily become more effective when the user's "taste" is added. In the experiment the best result in training a quadruped robot's navigation policy went precisely to the version without a preference oracle — the dedicated module that stores research preferences.

But that does not make the oracle useless. Both versions ran the same cycle across eleven research directions, and both refuted roughly three quarters of their own hypotheses. The difference lay not in the honesty of the testing or in the final score but in where the system decided to look at all.

The authors added three elements to the cycle. Every experiment gets an immutable card: the prediction and the result are recorded in it in advance, so a failed hypothesis cannot be rewritten after the fact as a success. Subagents get narrow mechanical roles, and subjective decisions are passed only to kkanbu — the preference oracle, cast as a typed knowledge graph of the user's preferences.

It is that separation that looks most useful. The cards and constraints stop the research from sliding imperceptibly into endless optimisation of one metric. And kkanbu sets the taste: the version with it was the only one to explore test-time adaptation, produced the best designs in the directions where it led the search itself, and carried lessons between streams instead of rediscovering the same ideas.

The upshot is an important trade-off: "taste" is not an accelerator of quality by itself. It is more of a navigator. You may arrive at more interesting or promising hypotheses, but that does not guarantee a better measurable result — in this experiment the system without the oracle won.

The limitations here are substantial. The work tests the approach in a simulation of quadruped robot navigation, not in a real laboratory. The available description gives no detail on the statistical robustness of the results or on how far the conclusions would carry to other fields. So this is for now an argument for the architecture of a research cycle, not proof that AI with a "memory of taste" conducts research better overall.

If you had to pick just one thing for your own AI agent — protection against rewriting failures, or passing on your research taste — which would you put first? Source: cs.AI updates on arXiv.org