Materials language models often repeat facts without true reasoning
CARAT: Do Materials LLMs Reason or Recite?
Artificial Intelligence
Summary
When asked about crystal structures, materials language models often copy answers they've seen instead of reasoning from the structure itself. The authors created tests to compare how these models use different inputs and found the models mostly repeat known relations. Even tricks designed to make the models rely on evidence showed models often ignore that evidence and just recite learned answers. This means current materials models may not genuinely understand or reason about crystal structures as hoped.
What this means in practice
- •For materials informatics engineers: Assess if materials LLM outputs reflect true reasoning versus memorization to improve trust in automated materials analysis tools.
- •For software developers for scientific ai: Design better input representations and training workflows to promote reasoning rather than rote copying in domain-specific language models.
Authors
Jiajun Wu, Jian Yang, Zixiang Ni, Zhenzhu Li, Bin Chong
Abstract
When a materials LLM answers a question about crystal structure, does it reason from the structure or copy an answer already printed in its input? Accuracy cannot tell: a structural description often prints the very field it is scored against. CARAT holds question and gold answer fixed across eight matched views, names each structural relation separately in GraphSpace, and adds matched fine-tuning, answer masking, evidence injection, paired inference, and a rule that can withhold claims. First, on the benchmark's hardest families the grounded view is worth 17.3 points over formula inputs. Second, we turn that scrutiny on ourselves. GraphSpace beats a plain periodic graph by 19.3 points, but that margin is two effects at once: where the plain rendering carries everything the question needs it is 1.96 points, and where it omits those fields entirely, 46.7 points. The headline mostly measures what the baseline lacked, not how evidence is presented. Third, we attack our own benchmark. A rule that skips the link and reads the list directly answers four of seven hardened families, so we rebuilt it until eleven such shortcuts sat near chance. The frozen model quotes that link yet answers the same when we redirect it, on 95.6% of paired cases: it repeats the relation without using it. After matched supervision it reaches 99.8%, and deleting the link drops it to 23.4%, below the 27.0% the best shortcut reaches: both steps are learnable.