Yes, partly. But there’s an annoying catch, because reality insists on being complicated.

The quote you pasted is basically right for models trained only on polished artifacts. If all you feed the system is papers, accepted code, and “winner” trajectories, then it mostly learns the visible positive manifold. It does not automatically learn the holes around it, which is where a lot of human taste lives. A 2026 case study of autonomous research attempts found repeated failures from “bias toward training data defaults,” “insufficient domain intelligence,” and “weak scientific taste,” which is almost exactly your negative-space complaint in lab coat form. (arXiv)

But agent swarms can get around that if they generate their own negative space instead of waiting for the internet to hand it to them.

Recent systems are starting to do exactly that. Google’s AI co-scientist uses a multi-agent “generate, debate, and evolve” setup with tournament-style self-improvement, and AI Scientist-v2 uses progressive agentic tree search rather than a single straight-line shot. OR-Agent goes even more explicitly in this direction: it organizes research as branching hypotheses with systematic backtracking, iterative experimentation, and reflections that accumulate lessons from failures. In other words, the swarm can create the missing corpus of “bad but informative” attempts by running them, criticizing them, and storing the traces. (arXiv)

So the sharp answer is:

Can swarms develop taste?

Yes, but usually not from static pretraining alone. They need an interactive loop where they:

• try many paths,
• observe which ones fail and why,
• keep structured memory of those failures,
• and use judges or reward models that score intermediate decisions, not just final outcomes.

That last part matters a lot. OpenAI’s process-supervision work found that supervising intermediate reasoning steps significantly beat outcome-only supervision on hard math, and improved data efficiency too. That is basically “teach the model what a wrong turn looks like before the cliff, not after the splat.”

As for your distillation idea: yes, distillation can absolutely transfer a kind of taste. There is long-standing evidence from policy distillation that a smaller student can inherit an expert policy, and newer work on algorithm distillation shows exploration behavior can be transferred from oracle trajectories into LLMs. There is even very recent work showing that reasoning capability can be distilled from black-box model outputs by reconstructing synthetic traces, which again suggests that useful policy-level structure can be transferred to a later model. (arXiv)

But the important caveat is this:

You can only distill taste that the teacher system has already earned.

If the teacher only read papers, then distillation mostly compresses polished-survivor bias.

If the teacher is a swarm that explored, failed, backtracked, and logged why, then distillation can compress something much closer to real taste.

That means there are really two different things people call “taste”:

1. Retrospective taste