Meituan's ACL Outstanding Paper: GeoRA Solves RLVR Geometric Misalignment
美团技术团队 · wechat · 2026-08-27
The Meituan technical team won the ACL 2026 Outstanding Paper Award for "GeoRA: Geometry-Aware Low-Rank Adaptation for RLVR," proposing a new solution to the geometric misalignment between RLVR (Reinforcement Learning from Verifiable Rewards) and existing Parameter-Efficient Fine-Tuning (PEFT) methods.
Background & Problem
RLVR has become a key paradigm for enhancing LLM reasoning capabilities, but its training cost is high. PEFT methods like LoRA, designed primarily for SFT, lead to suboptimal results, capability forgetting, or even training collapse when applied to RLVR. This is because RLVR updates tend to be sparse and avoid the principal directions of pre-trained weights, a geometric difference LoRA ignores.
GeoRA Method
- Geometric Subspace Construction: Uses spectral prior (stability) and Euclidean prior (plasticity) to select stable and plastic parameter regions from pre-trained weights.
- Low-Rank Approximation: Performs SVD on the selected subspace, initializing the low-rank adapter with the top-r singular components.
- Residual Anchor: Freezes remaining weights to ensure function invariance at initialization, smoothing cold starts and protecting pre-trained structure.
Experimental Results
- On RLVR tasks for math, medicine, and coding, GeoRA consistently outperforms baselines like LoRA and PiSSA across models from 1.5B to 32B, even surpassing full-parameter fine-tuning (e.g., Qwen3-8B reached 23.75% on AIME24).
- Out-of-Distribution (OOD) Performance: Significantly reduces capability forgetting, with improvements observed on HumanEval.
- Efficiency: Reduces trainable parameters by 99.5%, memory by 28.5%, and step time by 19.9% compared to full-parameter tuning, avoiding the computational overhead of sparse fine-tuning.
Business Application
The method has been deployed in Meituan's AI recruitment Agent (AgenticRL) scenario, achieving full-parameter training results at LoRA costs, with a 12% improvement over LoRA and significantly reduced memory usage.
More from Companies & People
- Researcher Jason Wei joins Google DeepMind to focus on RL and post-training — _jasonwei · 2026-08-27
- Monetizing without chasing views: sharing AI knowledge on X leads to paid opportunities — AryHHAry · 2026-08-27
- NVIDIA Covers Full AI Stack via Licenses and Investments — himanshustwts · 2026-08-27
- Report: OpenAI Building an Interface Platform Inside ChatGPT — ryanmerket · 2026-08-27
- Liam Fedus rejoins Google DeepMind to focus on RL — LiamFedus · 2026-08-27
- Concerns Arise Over HuggingFace's Hardware Neutrality After NVIDIA Acquisition — QuixiAI · 2026-08-27