On-policy distillation is quietly eating GRPO's lunch — and DPO's tax bill just got itemized

The latest research and engineering on post-training optimization has been published between October 30, 2026 and November 6, 2026. Notably, a study on task-selective capability loss under length-normalized DPO with SFT retention highlights that the "alignment tax" is not uniformly graded. While extraction remains intact, routing tasks experience a 30-point drop in a specific reward-margin band. This research provides practical guidance on identifying collateral damage from DPO+SFT-retention.

Another significant development is the introduction of a deterministic benchmark to measure the marginal value of verifier-based RL beyond supervised fine-tuning. With 1,200 gold demonstrations, SFT already solves most issues, and 60 steps of GRPO offer no detectable in-distribution gain, only a slight out-of-distribution bump. This evidence can inform decisions on whether an RLVR pipeline is worth the cost when good demonstration data is already available.

In the realm of model distillation, on-policy distillation (OPD) is quietly eroding the competitive edge of gradient-reversed policy optimization (GRPO). OPD matches GRPO's accuracy with a fraction of the computational cost, particularly when the teacher is the model's own GRPO checkpoint. Multi-teacher OPD further consolidates several domain experts into a single model, requiring 42% fewer total FLOPs than joint GRPO training.

Additionally, a new technique called EpiLoRA has been developed to fuse many LoRA adapters into low-precision GEMMs for serving and RL. This innovation cuts up to 24% of the per-token decode latency on Hopper GPUs, outperforming vLLM's best configuration and beating SGLang/TensorRT-LLM at high tenant counts. Furthermore, EpiLoRA includes an RL trainer that reuses serving engine FP8 weights, enabling a 27B model to train with RL on a single 48GB GPU.

Stepwise-Querying offers a novel approach to distilling reasoning from black-box teachers that only provide final answers and short rationales. By issuing independent, step-specific queries to the teacher, this method reconstructs, validates, and corrects intermediate derivations before assembling trajectories for standard SFT. This technique has proven effective in 14 out of 16 comparisons against three baselines on math and code benchmarks.

Lastly, a new study on dynamic rank in LoRA challenges existing methods and proposes GEAR, a technique that combines tangent-space optimization with a matrix-free spectrum estimate in the normal space. GEAR effectively distinguishes under-optimization within the current subspace from the need for new directions, raising the average commonsense score from 86.61 to 88.45 while reducing trainable parameters from 42.0M to 14.4M on LLaMA3-8B.