Fine-tuning with just 3 data points cuts thermal conductivity error from 47% to 2%
bravo_abad · x · 2026-09-06
Póta et al. argue that scientific foundation models should be judged not by zero-shot accuracy but by how much expensive high-quality data they amortize.
- A pretrained atomistic foundation model serves as a reusable prior; only material-specific parts are fine-tuned, with the learned potential supplying interatomic forces while the Wigner transport equation keeps heat-transfer physics explicit
- Thermal conductivity is a demanding test since it hinges on subtle anharmonic atomic interactions
- For LiBr, adaptation needed just 3 new DFT configurations, dropping the thermal-conductivity error from 47% to 2%
Takeaway: the real test of a scientific foundation model may be how little new data it needs to become trustworthy on a new problem.
More from Research
- JEPA may keep unpredictable features: new paper gives set-level characterization — hisspikeness · 2026-10-23
- GAFU Uses Generation-Side Representations to Boost Fine-Grained Visual Understanding in Unified Multimodal Models — Tengju Ru · 2026-10-23
- NTU Study: Learned Unmasking Policies Improve Speed-Accuracy Trade-offs in Diffusion Language Models — nanyang-technological-university-singapore · 2026-10-23
- NBER releases 'The Economics of Transformative AI' edited by Agrawal and Brynjolfsson — JMateosGarcia · 2026-10-23
- Swapping only the eval harness moves results 30%: designing agent evals like a scientist — AI Engineer · 2026-10-23
- Jane Street and NVIDIA speed up Mamba-2 with custom cuTile kernels — blelbach · 2026-10-23