Entity · technique

DRIFT

techniqueactivedrift-1852b6e0·1 events·first seen Jun 1, 2026

Aliases: DRIFT

Co-occurring entities

KL-regularized RL Reinforcement Learning importance-weighted supervised fine-tuning

More like this (12)

self-anchored drift Drifting Models Stein Drift deng2026drifting Drifting Preference Optimization Conservative Drifting Method World-Action Drift Attacks Anti-Drift Rectification ModelRift DFly EdgeFlow DRFLOW

Recent events (1)

6arXiv · cs.CL·Jun 1, 2026·source ↗

DRIFT: Decoupled Rollouts and Importance-Weighted Fine-Tuning for Efficient Multi-Turn Optimization

DRIFT is a training framework that bridges online RL and offline SFT for multi-turn LLM optimization by exploiting the theoretical equivalence between KL-regularized RL and importance-weighted supervised learning. It decouples rollout generation from policy optimization: trajectories are sampled from a fixed reference policy offline, weighted by return-based importance scores, and used for weighted SFT. Empirically, DRIFT matches or exceeds multi-turn RL baselines while retaining the efficiency and simplicity of standard supervised fine-tuning. Code is publicly released.

Inference Economics Agent and Tool Ecosystem KL-regularized RL Reinforcement Learning DRIFT +2 more