Sparse autoencoders interpret differences in early-career LinkedIn job ads
Original title: Beyond Job Titles: Interpreting Early-Career Job Advertisements with Sparse Autoencoders and Semantic Clustering
The study covers 36,081 US postings from December 2025 to October 2026 and checks hypotheses on unseen companies.
Researchers compare PERC embedding and clustering with sparse-autoencoder features on early-career LinkedIn advertisements. GPT-6 Luna labels passages as team context, role task, or qualifications, while a pretrained SAE extracts features from layer 19 of Llama-3.1-8B-Instruct. GPT-6 Luna proposes hypotheses and GPT-6.1 Terra tests them against held-out descriptions. Reported patterns include greater emphasis on model development for data scientists, reporting for data analysts, and deployment for machine-learning engineers.
Unconfirmed: The authors say dataset scale and uneven coverage limit interpretation of monthly differences.
- Maximum Mean Discrepancy ranks differences between embedding distributions.
- SAE features come from Llama-3.1-8B-Instruct layer 19.
- The collection window runs from December 2025 through October 2026.
- Analyst-role differences were also observed across selected months.
- Uneven coverage limits interpretation of those monthly comparisons.
- Organizations
- Countries
- United States
From the original article
1 min read at arxiv.orgJob titles and job descriptions alone provide limited insight into how technical roles differ in a changing job market. To assess how such roles differ in their work and industry applications, we design and compare two methods based on deep learning and transformer representations using 36,081 early-career job postings collected from LinkedIn in the United States between December 2025 and October 2026.
By Yisong Cao · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.