AI
Predictive Credit: Measuring What Scientific Explanations Add to Experimental Forecasts
Publication: Hugging Face Daily Papers
Author: Not available in collected metadata
Source published: 28/09/2026, 20:00:00 UTC
Briefing created: 02/10/2026, 12:10:22 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 02/10/2026, 02:18:24 UTC.
Briefings
A study evaluated predictive credit in scientific forecasting using 336 prospective states across 12 Tox21 endpoints and 24 OpenML tasks. Researchers compared five experimental conditions: description, matched explanation, donor context, and two AI models (DeepSeek V4 Pro and Flash). Key findings include unmet ROC AUC interval-score harm tests for Tox21 and OpenML, unconfirmed matched point-accuracy gains, and zero intervals for seed-donor matches. AI models reduced secondary Tox21 drift by 64.5% and 59.1% respectively, while Flash replay increased MAE and widened OpenML intervals. Natural-explanation credit remained unconfirmed at tested donor resolutions.
Key points
- Tox21's preregistered ROC AUC interval-score harm test was unmet with a D-M of -.0026 (95 percent interval [-.0174, .0104]) across 12 endpoints.
- Matched point-accuracy gains over description remained unconfirmed, and seed-donor intervals spanned zero for both Tox21 and OpenML tasks.
Context and limitations
Snapshot data from a controlled learning environment with 336 prospective states; natural-explanation credit unconfirmed at tested donor resolutions; preprint status with uncertainty.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 02/10/2026, 12:10:22 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeidk53hjg6z2csflmoic7mlcnl4p2isz2yzlkjg3c43rx6l4wmthdm