Tsinghua & Tencent Propose GPS for Reasoning Training
jiqizhixin · x · 2026-07-19
Researchers from Tsinghua University and Tencent propose GPS (Generalizable Predictive Prompt Selection). It uses a lightweight predictive model based on shared training history to first estimate prompt difficulty, then prioritizes moderately difficult and diverse samples to guide the RL post-training of large models.
The paper claims this approach reduces expensive trial-and-error rollouts, outperforming strong baselines on multiple reasoning benchmarks while improving training efficiency, final performance, and test-time speed. The diagram illustrates the complete training pipeline: difficulty prediction, batch selection, generation and reward calculation, RLVR updates, and updates to history and PPM parameters.
More from Infra
- AI Infra Meme: While Countries Need TW-Scale Power for Datacenters, People Argue About Lightbulbs — mgostIH · 2026-10-27
- Legora runs 2,000 parallel agent evals on Modal Sandboxes for legal AI — blelbach · 2026-10-27
- Cloud release velocity scores: AWS 95, Azure 90, Google Cloud 82 — DavidLinthicum · 2026-10-27
- Selling a 5090 to scale local inference: one business weighs 5 hardware paths — Anarchaotic · 2026-10-27
- AWS tutorial: fleet-wide edge CV model lifecycle with SageMaker SDK v3, ~$0.07 per run — AWS ML Blog · 2026-10-27
- Qwen3.8-27B hits 40 tok/s on Jetson Thor with INT4 + MTP, accuracy beats FP16 — No-Dragonfly6246 · 2026-10-27