peft-rlvr-evaluation

name	peft-rlvr-evaluation
title	Evaluating Parameter Efficient Methods for RLVR
version	0.0.2
engine	skillxiv-v0.0.2-claude-opus-4.6
license	MIT
url	https://arxiv.org/abs/2512.23165
keywords	["parameter-efficient","reinforcement-learning","adapter","peft"]
description	Comprehensively evaluate 12+ parameter-efficient fine-tuning methods for RL with Verifiable Rewards (RLVR). Show DoRA/AdaLoRA outperform LoRA, SVD-based methods fail on RL, extreme reduction creates bottlenecks—providing empirical evidence that geometric-aware adapters align better with RL's off-principal update dynamics.

Overview

Reveals LoRA's limitations in RL contexts through systematic PEFT evaluation.

Core Technique

Structural Variant Comparison:

# Standard LoRA sometimes suboptimal
methods = [LoRA, DoRA, AdaLoRA, PiSSA, VeRA, ...]

# Evaluate on RLVR (sparse binary rewards)
results = benchmark_all_methods(methods, rlvr_training)

# Finding: DoRA/AdaLoRA outperform

When to Use

Use when: RL fine-tuning, parameter efficiency critical, method selection.

References

• Comprehensive PEFT method evaluation
• RL-specific optimization dynamics
• Geometric-aware adapter superiority