Source: https://dhruvipaprunia.com/series/

Series

# 30 days, 30 papers.

One paper a day, explained simply.

[![Day 1: Stop agents from gaming the test](https://dhruvipaprunia.com/series/31-day1.jpg)](https://x.com/Dhruvi1505/status/2103153744296714699)

Day 1

## Stop agents from gaming the test

RRSI: Regularized Recursive Self-Improvement of Agent Harnesses

[X ↗](https://x.com/Dhruvi1505/status/2103153744296714699) [LinkedIn ↗](https://www.linkedin.com/feed/update/urn:li:activity:7508920525260492801/) [Paper ↗](https://arxiv.org/abs/2609.24972)

[![Day 2: Teach models to win clean](https://dhruvipaprunia.com/series/2-32-day2.png)](https://x.com/Dhruvi1505/status/2103535740101578836)

Day 2

## Teach models to win clean

MiMo-V2.6: Scaling Up Reinforcement Learning for Self-Improvement

[X ↗](https://x.com/Dhruvi1505/status/2103535740101578836) [LinkedIn ↗](https://www.linkedin.com/feed/update/urn:li:activity:7509268690672967682/) [Paper ↗](https://mimo.mi.com/docs/en-US/news/latest/v2-6)

[![Day 3: A week old, already tested](https://dhruvipaprunia.com/series/3-33-day3.png)](https://x.com/Dhruvi1505/status/2103868590546182455)

Day 3

## A week old, already tested

JEV-as-a-Judge: Accept When Confident, Escalate When Unsure

[X ↗](https://x.com/Dhruvi1505/status/2103868590546182455) [LinkedIn ↗](https://www.linkedin.com/feed/update/urn:li:activity:7509630224989609984/) [Paper ↗](https://arxiv.org/abs/2609.26550)

[![Day 4: The rich get richer](https://dhruvipaprunia.com/series/1-34-day4.png)](https://x.com/Dhruvi1505/status/2104227997309907355)

Day 4

## The rich get richer

Never Give Up: Learning to Solve Hard Problems in RL for LLMs

[X ↗](https://x.com/Dhruvi1505/status/2104227997309907355) [LinkedIn ↗](https://www.linkedin.com/feed/update/urn:li:activity:7509992205630226432/) [Paper ↗](https://arxiv.org/abs/2609.13443)

[![Day 5: Grade the process, not the output](https://dhruvipaprunia.com/series/day5-osworld-pro.png)](https://x.com/Dhruvi1505/status/2104582521467892015)

Day 5

## Grade the process, not the output

OSWorld-Pro: evaluating where computer-use agents fail

[X ↗](https://x.com/Dhruvi1505/status/2104582521467892015) [LinkedIn ↗](https://www.linkedin.com/feed/update/urn:li:activity:7510350857729708032/) [Paper ↗](https://arxiv.org/abs/2609.24890)

[![Day 6: Train the search, keep the model frozen](https://dhruvipaprunia.com/series/day6-beyond-repeated-sampling.png)](https://x.com/Dhruvi1505/status/2104948948817301780)

Day 6

## Train the search, keep the model frozen

Beyond Repeated Sampling: concepts guide the search

[X ↗](https://x.com/Dhruvi1505/status/2104948948817301780) [LinkedIn ↗](https://lnkd.in/p/dwRnynHN) [Paper ↗](https://arxiv.org/abs/2609.26704)

[![Day 7: Reward the progress](https://dhruvipaprunia.com/series/day7-published.png)](https://x.com/Dhruvi1505/status/2105330181795955036)

Day 7

## Reward the progress

ProCredit: rerun task checks after each turn, so useful progress earns credit.

[X ↗](https://x.com/Dhruvi1505/status/2105330181795955036) [LinkedIn ↗](https://www.linkedin.com/feed/update/urn:li:activity:7511099442071109632/) [Paper ↗](https://arxiv.org/abs/2609.27532)

[![Day 8: How much KV cache is enough?](https://dhruvipaprunia.com/series/day8-published.png)](https://x.com/Dhruvi1505/status/2105678171475726398)

Day 8

## How much KV cache is enough?

KVSet: one pass over requests predicts the cache size you need.

[X ↗](https://x.com/Dhruvi1505/status/2105678171475726398) [LinkedIn ↗](https://www.linkedin.com/feed/update/urn:li:activity:7511446667502505985/) [Paper ↗](https://arxiv.org/abs/2609.27746)

[![Day 9: Partial credit, done right](https://dhruvipaprunia.com/series/day9-published.png)](https://x.com/Dhruvi1505/status/2106036541784998235)

Day 9

## Partial credit, done right

Progressive Point Matching: credit milestones from one correct solution.

[X ↗](https://x.com/Dhruvi1505/status/2106036541784998235) [LinkedIn ↗](https://www.linkedin.com/feed/update/urn:li:activity:7511851099398627328/) [Paper ↗](https://arxiv.org/abs/2609.07303)
