Research map / Large language models / Language, safety and interpretability
Enhancing Job Recommendation through LLM-Based Generative Adversarial Networks
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes LGIR, which uses ChatGLM-6B to complete job seekers' resumes using interaction-derived context and a GAN-based module that aligns low-quality resume representations with high-quality ones for job recommendation.
Problem
Job recommendation on online recruitment platforms suffers from low-quality user resumes and few-shot users with scarce interaction records; directly using LLMs to complete resumes causes fabricated content.
Approach
Complete resumes with an LLM conditioned on both self-description and the job requirements a user interacted with (interactive resume completion), then use a classifier-gated GAN to map low-quality (few-shot) resume representations toward high-quality ones before a pairwise recommender.
Key results
- LGIR achieves map@5 0.2887 on Designs vs. 0.2664 for the best baseline LightGCN (about 8.4% relative gain), with significance at p<0.01.
- On Sales, LGIR reaches map@5 0.1751 vs. 0.1629 for LightGCN; on Tech, 0.5086 vs. 0.4719 for LightGCN+SRC.
- Ablation on Designs: BASE 0.2627, SRC 0.2601 (no gain), IRC 0.2859, full LGIR 0.2887 map@5, showing interactive completion drives most of the gain.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use LGIR instead of the strongest baseline | Job recommendation · Designs | map@5: 0.2664 -> 0.2887* | Helped |
| Use LGIR instead of the strongest baseline | Job recommendation · Sales | map@5: 0.1629 -> 0.1751* | Helped |
| Use LGIR instead of the strongest baseline | Job recommendation · Tech | map@5: 0.4719 -> 0.5086* | Helped |
| Replace original self-description with simple resume completion (ablation) | Two-tower text matching model · Designs | map@5: 0.2627 -> 0.2601 | Hurt |
| Replace original self-description with simple resume completion (ablation) | Two-tower text matching model · Sales | map@5: 0.1617 -> 0.1652 | Helped |
| Replace original self-description with simple resume completion (ablation) | Two-tower text matching model · Tech | map@5: 0.4994 -> 0.5048 | Helped |
| Add interactive resume completion using user interaction behavior (ablation) | Two-tower text matching model · Designs | map@5: 0.2627 -> 0.2859 | Helped |
| Add interactive resume completion using user interaction behavior (ablation) | Two-tower text matching model · Sales | map@5: 0.1617 -> 0.1671 | Helped |
| Add interactive resume completion using user interaction behavior (ablation) | Two-tower text matching model · Tech | map@5: 0.4994 -> 0.5056 | Helped |
| Add GAN-based alignment of low-quality resumes (ablation) | IRC · Designs | map@5: 0.2859 -> 0.2887 | Helped |
| Add GAN-based alignment of low-quality resumes (ablation) | IRC · Sales | map@5: 0.1671 -> 0.1751 | Helped |
| Add GAN-based alignment of low-quality resumes (ablation) | IRC · Tech | map@5: 0.5056 -> 0.5086 | Helped |
Surprising findings
- Simple LLM resume completion (SRC) gives little or no improvement over the base text and sometimes hurts, e.g. Designs map@5 drops from 0.2627 to 0.2601.
- Text-only LLM-based methods (SGPT variants) perform poorly relative to interaction-based baselines, suggesting that resume text alone is insufficient.
Limitations
- Few-shot GAN alignment gains are modest and results are reported without variance beyond averages over five runs.
- Relies on unpaired resume alignment with heuristic thresholds (kappa1=30, kappa2=5) selected by hyper-parameter search on the test-style analysis.
- Evaluation limited to three datasets from a single platform, and the datasets are not released.
Experimental setup
Datasets: Designs Sales Tech
Benchmarks: Designs Sales Tech
Metrics: map@5 ndcg@5 mrr
Baselines: SGPT-BE SGPT-ST SGPT-ST+SRC BPJFNN MF NCF PJFFF SHPJF LightGCN LightGCN+SRC
Builds on: GAN LightGCN SGPT NCF BERT
Closest papers by meaning
- Exploring Large Language Model for Graph Data Understanding in Online Job Recommendations (AAAI 2024 · similarity 0.84)
- Sparse Enhanced Network: An Adversarial Generation Method for Robust Augmentation in Sequential Recommendation (AAAI 2024 · similarity 0.82)
- Improving LLMs for Recommendation with Out-Of-Vocabulary Tokens (ICML 2025 · similarity 0.82)
- Actions Speak Louder than Words: Trillion-Parameter Sequential Transducers for Generative Recommendations (ICML 2024 · similarity 0.81)
- Language Ranker: A Lightweight Ranking framework for LLM Decoding (NeurIPS 2025 · similarity 0.81)
- Can LLMs Outshine Conventional Recommenders? A Comparative Evaluation (NeurIPS 2025 · similarity 0.81)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.