TBNAI
Speedrunning Relational Foundation Model Pretraining
Publication: cs.AI updates on arXiv.org
Author: Gabriel Wedson Mendon\c{c}a de Souza Veloso, Luan Borges Teodoro Reis Sena, Denis Oliveira Correa, Francisco Galuppo Azevedo
Source published: 22/10/2026, 04:00:00 UTC
Briefing created: 22/10/2026, 04:09:00 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 22/10/2026, 04:03:24 UTC.
Briefings
A new speedrunning benchmark for relational transformer pretraining, modded-nanoRT, was introduced to remove barriers to entry by ranking entries based on wall-clock time rather than training resources. The benchmark fixed the database selection issue by calibrating six candidates using a 45-minute learning curve to separate improvement from noise. It achieved a record of 305.63 seconds on a single NVIDIA H200 GPU, representing an 8.11x speedup over the reference implementation and winning 13 of 15 pairs in the development database.
Key points
- The benchmark modded-nanoRT uses a calibrated 45-minute learning curve to distinguish meaningful improvements from seed noise, selecting six candidate databases for ranking.
- The record was achieved in 305.63 seconds on a single NVIDIA H200 GPU, an 8.11x speedup over the reference implementation, winning 13 of 15 pairs in the development database.
Context
Speedrunning Relational Foundation Model Pretraining allows decentralized competition by ranking entries based on wall-clock time rather than training resources, removing barriers to entry. The benchmark fixes the database selection issue by calibrating candidates using a 45-minute learning curve to separate improvement from noise.
Limitations
The benchmark relies on a single NVIDIA H200 GPU for the record, and the 45-minute calibration period may not capture all relevant performance variations. The 13/15 win rate is specific to the development database, and the benchmark is still in preprint status with no frozen artifacts released yet.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 22/10/2026, 04:09:00 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeifyo6ojuh7rhac7xxgda2b6jbwqlrxyu52ns5w6aohzpbasepizaa