AI Agents Designed to Enhance Job Referral Requests Show 14% Predicted Success Boost for Weaker Submissions

Global: AI Agents Designed to Enhance Job Referral Requests Show 14% Predicted Success Boost for Weaker Submissions

Researchers have introduced a pair of artificial‑intelligence agents aimed at improving the quality of job referral requests posted in professional online communities. The work, authored by Ross Chu and Yuting Huang, was submitted to arXiv on December 28, 2025 (arXiv:2601.10726) and falls under the categories Artificial Intelligence (cs.AI) and Computation and Language (cs.CL).

System Architecture

The proposed framework consists of an “improver” agent that rewrites a user’s original referral request and an “evaluator” agent that scores the revised text. The evaluator relies on a predictive model trained to estimate the probability that a request will elicit a referral from another community member.

Use of Large Language Models and RAG

Both agents are built on a large language model (LLM). To mitigate the risk of harmful edits, the authors augment the LLM with Retrieval‑Augmented Generation (RAG), which supplies relevant external examples during the rewriting process.

Experimental Findings

Testing on a dataset of simulated requests revealed that LLM‑generated revisions increased the predicted success rate for weaker submissions by 14%, while the same revisions slightly lowered the predicted success for stronger submissions. Incorporating RAG prevented the degradation of stronger requests and amplified the gains for weaker ones, resulting in an overall improvement without harming high‑quality inputs.

Limitations

The authors caution that the model’s predicted success does not guarantee actual referral outcomes in real‑world deployments. The study relies on simulated data and a proxy success metric, which may not capture all factors influencing human decision‑making.

Future Directions

Further research is suggested to validate the approach with live user experiments, explore ethical considerations of automated persuasion, and refine the evaluation model to better reflect real referral behavior.

This report is based on information from arXiv, licensed under Academic Preprint / Open Access. Based on the abstract of the research paper. Full text available via ArXiv.