# Peft Rlvr Evaluation

> 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.

- Skill: `adu2021/peft-rlvr-evaluation` (Agent Skill)
- Install (CLI): `npx skillmds@latest add adu2021/peft-rlvr-evaluation`
- Raw SKILL.md: https://api.skillmd.com/api/skills/adu2021/peft-rlvr-evaluation/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Coding & Dev Tools
- License: MIT
- Author: adu2021 (https://skillmd.com/u/adu2021)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/adu2021/peft-rlvr-evaluation

---


## Overview

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

## Core Technique

**Structural Variant Comparison:**

```python
# 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