Spork post-training recipe lifts chess puzzle solve rate 100% over best GRPO setup
aidenybai · x · 2026-10-14
Applied Compute introduced Spork, an efficient post-training recipe that improves chess puzzle solve rates by 100% compared to its best GRPO setup.
Key points:
- Starts with a GRPO warmup
- Splits task-specific datasets into partitions and trains specialist models on each
- Consolidates the specialists via multi-teacher on-policy distillation
Named after the spork utensil, the method shows gains can come from simply consolidating specialists rather than a new paradigm. Aiden Bai said he wants to try it on Go puzzles next.
More from Research
- elder_plinius launches V3R1T4S: agents tackle open problems, proof or it didn't happen — Kyrannio · 2026-10-14
- RigidFormer accepted at NeurIPS: mesh-free Transformer beats GNN simulators for rigid dynamics — camenduru · 2026-10-14
- Anthropic's 2021 research roadmap surfaces: 1T models, alignment scaling laws already on the list — srush_nlp · 2026-10-14
- On-policy supervised fine-tuning: RL-level generalization at SFT-level compute — burkov · 2026-10-14
- Two teams independently find AI user proxies overshare, user modeling emerges as research area — niloofar_mire · 2026-10-14
- Study: model collapse doesn't happen when synthetic data accumulates, even without new real data — AndrewLampinen · 2026-10-14