PulseAugur
EN
LIVE 11:48:02

New orthonormal initialization method boosts RLVR training stability

By PulseAugur Editorial · [2 sources] ·

Researchers have developed a new method for initializing low-rank adaptation (LoRA) matrices in Reinforcement Learning with Verifiable Rewards (RLVR). This approach, called geometry-preserving orthonormal initialization, aims to improve training stability and performance compared to standard LoRA and other variants like PiSSA and MiLoRA, which can underperform or become unstable in RLVR settings. The proposed method, leading to new variants RLPO and RLMO, is supported by theoretical analysis and demonstrated through experiments on mathematical reasoning benchmarks, showing enhanced stability and superior results. AI

IMPACT This research could lead to more stable and effective fine-tuning of large language models for reinforcement learning tasks.

RANK_REASON The cluster contains an academic paper detailing a new method for fine-tuning large language models.

Read on arXiv cs.AI →

AI-generated summary · Google Gemini · from 2 sources. How we write summaries →

New orthonormal initialization method boosts RLVR training stability

COVERAGE [2]

  1. arXiv cs.AI TIER_1 English(EN) · Ruijia Zhang, Jiacheng Zhu, Hanqing Zhu, Laixi Shi ·

    Geometry-Preserving Orthonormal Initialization for Low-Rank Adaptation in RLVR

    arXiv:2606.31813v1 Announce Type: cross Abstract: Low-rank adaptation (LoRA) and its variants enable parameter-efficient fine-tuning of large language models under the supervised fine-tuning (SFT) paradigm. However, their efficacy and behavior under Reinforcement learning with ve…

  2. arXiv cs.AI TIER_1 English(EN) · Laixi Shi ·

    Geometry-Preserving Orthonormal Initialization for Low-Rank Adaptation in RLVR

    Low-rank adaptation (LoRA) and its variants enable parameter-efficient fine-tuning of large language models under the supervised fine-tuning (SFT) paradigm. However, their efficacy and behavior under Reinforcement learning with verifiable rewards (RLVR) are less well understood. …

RELATED ENTITIES

RELATED TOPICS