Zero-shot TabPFN beats fine-tuned LLM NCPL at predicting training curves, results show
ZimingLiu11 · x · 2026-10-31
Ziming Liu's team released a comparison table for training-curve prediction across four corpora (Marin, StepLaw, Marin-MoE, DataDecide). A pre-trained TabPFN used zero-shot (frozen weights, labeled training runs as context) predicts final loss and trajectories, beating NCPL, XGBoost, and parametric scaling laws on most ID/OOD metrics.
- On Marin, TabPFN reaches final-loss (ID) MAE 0.0080 vs NCPL's paper-reported 0.0109; it also leads on OOD curve prediction
- NCPL reproduction matches paper numbers, but MoE needs a one-seed fine-tune and degrades notably
- Takeaway: a tabular foundation model already works as a meta model for curve prediction with OOD generalization, no fine-tuning needed
Related event: TabPFN's Pre-registered Training Curve Predictions Validated(2 posts)→
More from Models
- 5 cheap open models vote vs Opus: 89.6% accuracy at 1/10th the cost — PomegranateAble4339 · 2026-10-31
- Gemini 4 Carbon reportedly scores 53 FCI on FrontierCS 2, big jump from Argon in weeks — clmt · 2026-10-31
- Gemini 3.5 Pro reportedly delayed into oblivion while rivals took share, says dev — tarantulae · 2026-10-31
- Dev suspects longer training degraded Qwen backbone multilingual abilities, plans more multilingual data — antoine_chaffin · 2026-10-31
- LiquidAI's LFM3-VL d1-2B edge vision-language model trends on Hugging Face — LiquidAI · 2026-10-31
- OpenAI's new API credits make ChatGPT Pro tiers a no-brainer for API users — Quiet-Recording-9269 · 2026-10-31