Aleksei Romanov (@gfactor)

Multi-Reward RL, Part 3: GDPO + CISPO + REPO-R at 27B, and the Advantage Floor That Stopped Learning

Part 1 explained how PPO, GRPO, DAPO and GDPO turn several rewards into one learning signal. Part 2 benchmarked the trainers on a 14B model for 50 steps and found a trap: CISPO won the training curves

Oct 9 · 15 min readgfactor.hashnode.dev

Pushing the Limits: Extreme Inference Speedup of Qwen 3.8 27B on NVIDIA B300 (100 to 10k+ tok/s)

Deploying a 27-billion parameter reasoning model like Qwen 3.8 27B on modern hardware presents a stark paradox. If you boot a default configuration on an NVIDIA B300 SXM6 GPU and send a solitary strea

Oct 7 · 22 min readgfactor.hashnode.dev

Multi-Reward RL, Part 2: Benchmarking GRPO, DAPO, and CISPO on Unseen Tasks

Follow-up: Part 3 scales the CISPO + REPO-R recipe to Qwen3.8-27B and 600 steps, with a one-change-per-run holdout ladder and an advantage-floor failure we found in a harsher environment. Part 1 analy

Oct 6 · 14 min readgfactor.hashnode.dev

The Verifier Design Playbook: How to Build RLVR Gyms That Models Can't Game

In traditional machine learning, your loss function is a clean mathematical equation: mean squared error, cross-entropy, or cosine distance. The math is simple, deterministic, and impossible for the m

Oct 5 · 14 min readgfactor.hashnode.dev

The Specialized Agent Shift: How Open Models Are Beating Frontier APIs in Production

For the past two years, the standard enterprise AI strategy looked suspiciously like a parlor trick: take an 8,000-token system prompt packed with JSON schemas, API documentation, and polite formattin

Oct 4 · 18 min readgfactor.hashnode.dev

Aleksei Romanov

About

Available for

Aleksei Romanov's blogs

Multi-Reward RL, Part 3: GDPO + CISPO + REPO-R at 27B, and the Advantage Floor That Stopped Learning

Pushing the Limits: Extreme Inference Speedup of Qwen 3.8 27B on NVIDIA B300 (100 to 10k+ tok/s)

Multi-Reward RL, Part 2: Benchmarking GRPO, DAPO, and CISPO on Unseen Tasks

The Verifier Design Playbook: How to Build RLVR Gyms That Models Can't Game

The Specialized Agent Shift: How Open Models Are Beating Frontier APIs in Production

Aleksei Romanov

About

Available for

Aleksei Romanov's blogs

Multi-Reward RL, Part 3: GDPO + CISPO + REPO-R at 27B, and the Advantage Floor That Stopped Learning

Pushing the Limits: Extreme Inference Speedup of Qwen 3.8 27B on NVIDIA B300 (100 to 10k+ tok/s)

Multi-Reward RL, Part 2: Benchmarking GRPO, DAPO, and CISPO on Unseen Tasks

The Verifier Design Playbook: How to Build RLVR Gyms That Models Can't Game

The Specialized Agent Shift: How Open Models Are Beating Frontier APIs in Production

Aleksei Romanov

About

Available for

Aleksei Romanov's blogs

Multi-Reward RL, Part 3: GDPO + CISPO + REPO-R at 27B, and the Advantage Floor That Stopped Learning

Pushing the Limits: Extreme Inference Speedup of Qwen 3.8 27B on NVIDIA B300 (100 to 10k+ tok/s)

Multi-Reward RL, Part 2: Benchmarking GRPO, DAPO, and CISPO on Unseen Tasks

The Verifier Design Playbook: How to Build RLVR Gyms That Models Can't Game

The Specialized Agent Shift: How Open Models Are Beating Frontier APIs in Production

Search Hashnode

Aleksei Romanov

About

Available for

Aleksei Romanov's blogs

Multi-Reward RL, Part 3: GDPO + CISPO + REPO-R at 27B, and the Advantage Floor That Stopped Learning

Pushing the Limits: Extreme Inference Speedup of Qwen 3.8 27B on NVIDIA B300 (100 to 10k+ tok/s)

Multi-Reward RL, Part 2: Benchmarking GRPO, DAPO, and CISPO on Unseen Tasks

The Verifier Design Playbook: How to Build RLVR Gyms That Models Can't Game

The Specialized Agent Shift: How Open Models Are Beating Frontier APIs in Production