OLMo

Timeline

1. Oct 15, 2026 · Research paper · 1 source
   Training-Free On-Policy Distillation
   We show that the consolidation does not need training: with forward passes only, the capabilities of separately tuned specialists can be aggregated into one model that is, on average, more accurate than either specialist alone.
2. Oct 14, 2026 · Research paper · 1 source
   Function-Space Temporal Self-Distillation for Policy Optimization
   We introduce Function-Space Policy Optimization (FSPO), a single-policy framework that combines current privileged adaptation with function-space temporal consolidation.
3. Oct 13, 2026 · Research paper · 1 source
   Geometric Calibration of Supervised Fine-Tuning
   Within this view, supervised fine-tuning (SFT) is shown to induce dense updates that perturb the spectral geometry of pre-SFT weights along their principal directions, potentially disrupting the model's prior knowledge.
4. Oct 8, 2026 · Research paper · 1 source
   Which Skill to Distill? SGUID: Selecting a Compact Skill Bank for Model-Skill Co-Evolution
   Skills, reusable procedural guidance added at inference, can substantially improve LLM downstream performance (Li et al., 2026).
5. Sep 9, 2026 · Open-source release · 1 source
   vllm-project/vllm v0.29.0
   MRV2 also gained CUDA graph memory profiling for KV cache auto-sizing (#53306), batch-sharded sampling that cuts per-step logits memory by 1/TP (#50465), prompt embeds (#42963), extracthiddenstates speculation (#49811), padded FULL cudagraph dispatch for uniform decode under spec decode (#53407), and DP-sync skipping before EAGLE/MTP draft prefill (#53694).
6. Jul 27, 2026 · Open-source release · 1 source
   vllm-project/vllm v0.26.0
   New Inkling model family with a full support stack: base modeling (#48799), piecewise CUDA graph support (#48822), Hopper FA4 relative attention (#48858), MTP=1 speculative decoding (#48869), LoRA (#48884), and standard ModelOpt NVFP4 quantization (#48990).

Often appears with