Affine launches live GLM distillation competition with public corpus and scoring
const_reborn · x · 2026-08-24
Affine has launched a live research competition for post-training and distillation. The competition uses GLM-4.5-Air-FP8 as the frozen teacher model, with the corpus and scoring code made public. Participants must train models to generate reasoning that aligns more closely with the teacher's preferred actions than the reigning champion. The 1,300-turn evaluation slice is determined after submission to prevent overfitting. Methods like distillation, RL, synthetic trajectories, and rejection sampling are allowed.
More from Research
- DataSpace Benchmark: Strong LLMs Fail as Reliable Data Agents — rohanpaul_ai · 2026-08-24
- Engineering gas vesicles into CAR-T cells lets ultrasound track them inside living organs — NikoMcCarty · 2026-08-24
- Alibaba PAI open-sources unified ControlNet for MiniMax-H3: one 7GB checkpoint, five control modes — linoy_tsaban · 2026-08-24
- SparsePR: Training-Free Sparse Attention for Video Generation — TexasAMUniversity · 2026-08-24
- Pure reinforcement learning enables zero-shot robot transfer — chris_j_paxton · 2026-08-24
- Mathematicians have low costs for using AI to solve problems, not relying on big labs — littmath · 2026-08-24