How do expert priors constrain human researchers from exploring novel concepts?
This explores why deep expertise can narrow the space of ideas a researcher is willing to explore — and what the corpus says about systems (human or machine) that get trapped by what they already know.
This explores why deep expertise can narrow the space of ideas a researcher is willing to explore. The clearest evidence in the corpus is a head-to-head study: when 100+ NLP researchers were pitted against language models, the model's ideas were rated *more* novel, though slightly less feasible Do language models generate more novel research ideas than experts?. The interpretation isn't that the model is smarter — it's that expert knowledge acts as a filter. Experts have internalized what 'should' work, which combinations are dead ends, and what the field considers serious. That same knowledge that makes their ideas feasible is what keeps them inside familiar conceptual territory.
The mechanism becomes clearer when you look at a parallel finding about machine agents. Systems trained only on expert demonstrations are 'locked into the imagination of the training data' — they can't generalize past what their curators already conceived, because their competence is capped by demonstrated scenarios rather than by their own capacity to explore Can agents learn beyond what their training data shows?. Swap 'curator' for 'a researcher's training and field' and you have a tidy analogy: a strong prior is a high-quality demonstration set, and it bounds the search to what's already been imagined.
What's striking is that the very tendency we usually call a flaw can become an advantage once you remove the prior. Fine-tuned LLMs out-predict neuroscience experts at guessing which experimental results actually occurred — and the corpus argues this is the *same* loose pattern-integration that produces hallucination in backward-looking tasks. Unconstrained recombination looks like error when you're retrieving known facts, but like genuine foresight when you're reaching for something new Can LLMs predict novel scientific results better than experts?. Expert priors suppress exactly this kind of promiscuous combination.
There's a deeper reason priors stick, though, and it's social rather than cognitive. The force of an argument in a field rides on the authority of the person making it — reputation, track record, standing — not just the idea's content Can language models distinguish expert arguments from common assumptions?. A novel concept that violates expert consensus carries reputational risk, so the prior isn't only 'what I believe is true' but 'what my peers will accept.' That's a constraint experts feel and models simply don't.
The useful twist for the curious reader: priors aren't pure liability. Reasoning that actually *generalizes* leans on broad, transferable procedural knowledge rather than narrow memorized facts Does procedural knowledge drive reasoning more than factual retrieval? — and one experiment shows new capabilities emerging when independent 'expert' models are recombined through collaborative search, solving problems none could solve alone Can language models discover new expertise through collaborative weight search?. The lesson isn't to discard expertise but to recombine it loosely. Novelty seems to live not in having fewer priors, but in being willing to mix them in ways the field hasn't sanctioned yet.
Sources 6 notes
A statistically significant study of 100+ NLP researchers found LLM-generated ideas rated as more novel than human expert ideas (p<0.05), though slightly lower on feasibility. Expert knowledge constrains novelty, while LLMs explore wider conceptual combinations.
Agents trained on static expert datasets cannot learn from their own failures or generalize beyond demonstrated scenarios because they never interact with environments during training. Competence is capped by what curators imagined, not by agent capacity.
BrainBench benchmarks show fine-tuned LLMs outperform neuroscience experts at predicting which experimental results actually occurred. The same pattern-integration tendency that causes hallucination in retrieval tasks enables genuine prediction in forward-looking scenarios.
LLMs lose the social context that gives expert claims their force—reputation, track record, and standing—because they process only text, not the social world where expertise is built and evaluated.
Analysis of 5 million pretraining documents shows reasoning relies on broad, transferable procedural knowledge from diverse sources, unlike factual recall which depends on narrow, document-specific memorization of target facts.
PSO-inspired swarms of LLM particles moving through weight space discover composed experts with new capabilities—including answering questions all initial experts failed on—using only 200 validation examples and no gradient-based training.