6 papers
Trust-Region Behavior Blending for On-Policy Distillation
Daniil Plyusov, Alexey Gorbatovski, Alexey Malakhov +4
On-policy distillation (OPD) trains a student on prefixes sampled from its own policy while matching a stronger teacher. This addresses the prefix mismatch of offline distillation,…
F-GRPO: Don't Let Your Policy Learn the Obvious and Forget the Rare
Daniil Plyusov, Alexey Gorbatovski, Boris Shaposhnikov +4
Reinforcement Learning with Verifiable Rewards (RLVR) is commonly based on group sampling to estimate advantages and stabilize policy updates. In practice, computational limits oft…
Small Vectors, Big Effects: A Mechanistic Study of RL-Induced Reasoning via Steering Vectors
Viacheslav Sinii, Nikita Balagansky, Gleb Gerasimov +6
The mechanisms by which reasoning training reshapes LLMs' internal computations remain unclear. We study lightweight steering vectors inserted into the base model's residual stream…
ESSA: Evolutionary Strategies for Scalable Alignment
Daria Korotyshova, Boris Shaposhnikov, Alexey Malakhov +7
Online alignment of large language models (LLMs) is dominated by reinforcement learning from human feedback (RLHF) with gradient-based optimizers such as PPO or GRPO. While effecti…
Steering LLM Reasoning Through Bias-Only Adaptation
Viacheslav Sinii, Alexey Gorbatovski, Artem Cherepanov +3
We show that training a single -dimensional steering vector per layer with reinforcement learning, while freezing all base weights, matches the accuracy of fully RL-tuned reason…
The Differences Between Direct Alignment Algorithms are a Blur
Alexey Gorbatovski, Boris Shaposhnikov, Viacheslav Sinii +2
Direct Alignment Algorithms (DAAs) simplify LLM alignment by directly optimizing policies, bypassing reward modeling and RL. While DAAs differ in their use of SFT (one-stage vs. tw…