# Mathematical Framework for Custom Reward Functions in Job Application Evaluation using Reinforcement Learning

arXiv: https://arxiv.org/abs/2511.16073 · HTML: https://mlsanity.com/paper/2511.16073/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Placed in Reinforcement learning for reasoning, one of 17 directions in LLM reasoning.
- Builds on 7 of its 11 references that are in the map.
- First posted on arXiv 20 Nov 2025.

## Placed in

- LLM reasoning / [Reinforcement learning for reasoning](https://mlsanity.com/lr/reinforcement-learning-for-reasoning/)

## Builds on (most cited first, 7 of 7)

- [Training language models to follow instructions with human feedback](https://mlsanity.com/paper/2203.02155/) (2022)
- [DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models](https://mlsanity.com/paper/2402.03300/) (2024)
- [Learning to summarize from human feedback](https://mlsanity.com/paper/2009.01325/) (2020)
- [Scaling Laws for Reward Model Overoptimization](https://mlsanity.com/paper/2210.10760/) (2022)
- [Open Problems and Fundamental Limitations of Reinforcement Learning from Human Feedback](https://mlsanity.com/paper/2307.15217/) (2023)
- [Small Language Models are the Future of Agentic AI](https://mlsanity.com/paper/2506.02153/) (2025)
- [Reward Hacking Mitigation using Verifiable Composite Rewards](https://mlsanity.com/paper/2509.15557/) (2025)

## Abstract (arXiv)

Most of the traditional Applicant Tracking Systems (ATS) depend on strict matching using keywords, where candidates that are highly qualified are many times disqualified because of minor semantic differences. In this article, the two-stage process of developing a more comprehensive resume assessment system based on a small language model that is trained with fewer than 600M parameters is introduced and fine-tuned by using GRPO with a uniquely designed reward function. The initial stage is Supervised Fine-Tuning (SFT), which is used to create a strong base model with the ability to perceive resumes beyond superficial overlap of keywords. This SFT model is further optimized in the second step with Reinforcement Learning (RL) via GRPO with the help of multi-component-based rewarding, which will not be considered as a commission of tokens matching. In the initial RL experiments, we found a severe difficulty in the shape of reward hacking: overly aggressive penalty terms resulted in unstable training dynamics and prohibitively negative model behavior. This was solved by trial-and-error refinement of the reward and careful training hyperparameter tuning, which led to a stable and controlled process of gentle polishing. The GRPO-refined model shows high real-life performance, as it shows an accuracy of 91% on unseen data used for testing. It has a high recall of 0.85 on the SELECTED class with a perfect precision of 1.0, which highlights its high reliability for identifying qualified applicants. These findings demonstrate that an appropriately structured two-step fine-tuning pipeline can effectively be used to transfer a small language model into human-like candidate evaluation, surpassing the shortcomings of both traditional ATS systems and unrefined uses of reinforcement learning.