Beyond Reweighting: Can RLVR Discover?

rl

Does RLVR only reweight and sharpen pre-existing solutios, or can it discover new solutions?

Published

March 5, 2026

See discovery_slides.pdf.

Population-based approach: population_rlvr_slides.pdf.
GLM5-inspired population-based approach:
Multi-objective RLVR (vector of rewards, multi-task, in parallel):

Motivation

Why is this worth pursuing? What gap does it fill?

RLVR can only help climb the hill more efficiently, not help you climb a new hill.

Core Idea

Describe the approach. Include equations, pseudocode, or diagrams as needed.

Open Questions

Potential Experiments

---
title: "Beyond Reweighting: Can RLVR Discover?"
description: "Does RLVR only reweight and sharpen pre-existing solutios, or can it discover new solutions?"
date: "2026-03-05"
categories: [rl]
status: "exploring"
---

See [discovery_slides.pdf](./discovery_slides.pdf). 

- Population-based approach: [population_rlvr_slides.pdf](./population_rlvr_slides.pdf). 
- GLM5-inspired population-based approach: 
- Multi-objective RLVR (vector of rewards, multi-task, in parallel): 
- 

## Motivation

Why is this worth pursuing? What gap does it fill?

> RLVR can only help climb the hill more efficiently, not help you climb a new hill. 

## Core Idea

Describe the approach. Include equations, pseudocode, or diagrams as needed.

## Open Questions

-

## Potential Experiments

-

## Related Work

-