DRIFT: A New Framework from Tsinghua and Beike for Continuous LLM Self-Evolution
青稞AI · wechat · 2026-07-26
A joint research team from Tsinghua University, Beike, and ENS Paris-Saclay introduced DRIFT, an online self-evolution post-training framework for LLMs. It aims to solve the challenge of how models can continuously improve after RL without collapsing.
The framework features several core mechanisms:
- Dynamic Difficulty Routing & Rhythm-Gated Exploration: These help the model identify critical reasoning nodes, determining when to apply RL versus self-distillation.
- Success Buffer: Combined with curriculum learning, it ensures more stable policy optimization.
Experiments show that DRIFT enables self-correction and continuous evolution without external expert supervision, achieving new SOTA on multiple complex reasoning and Tool Use benchmarks. The authors will host an upcoming webinar to dive deep into the algorithm's design and theoretical motivations.
More from Research
- Meta FAIR's Self-Direction turns agent failures into reusable LoRA skills, +13.8pp on Terminal-Bench — karen_ullrich · 2026-10-27
- DONIC hits 86.79% on bimanual dexterous tracking, crushing prior SOTA's 9.43% — liuziwei7 · 2026-10-27
- Why don't more LLMs ship 10M-token context? A researcher asks about attention tradeoffs — agarwl_ · 2026-10-27
- PSNR below 15 dB in papers likely means the metric was misused, researcher warns — kwangmoo_yi · 2026-10-27
- Ortet's first health eval: 34 open-weight models, 50 tasks, generic rankings fall short — thegautamkamath · 2026-10-27
- Scaling-fit discrepancy: output-head FLOPs are ~65% of compute at the smallest rung, ~3% at 535B — dlwh · 2026-10-27