Pith. sign in

REVIEW 3 major objections 4 minor 67 references

STEP, a sequence model that conditions on job durations and education level, predicts the next occupation in a career trajectory, and the paper reports it beats all baselines on four public benchmarks, with top-10 recall of 46.6% to 70.3%.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-02 06:47 UTC pith:NFWPEOTF

load-bearing objection Deserves a careful referee: solid dataset and honest empirics, but the SOTA claim is only as strong as the unquantified label noise in the ground truth. the 3 major comments →

arxiv 2607.11722 v2 pith:NFWPEOTF submitted 2026-07-13 cs.CL

STEP: Career-Path Recommendation via Temporal and Educational Trajectory Modeling

Iman Johary , Guillaume Bied , Alexandru C. Mara , Tijl De Bie This is my paper
classification cs.CL
keywords career path predictionnext job recommendationtemporal sequence modelingeducational attainmentESCO taxonomyoccupational embeddingscontrastive learningpublic benchmark
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper's central claim is that next-job prediction improves when a model uses how long each job lasted and what degree the person holds, not just the sequence of job titles. To show this, it introduces STEP, a lightweight sequence model whose recurrent cell decays its memory according to elapsed time and whose hidden states are modulated by education level, trained on top of ROUTE, a two-stage procedure that adapts a multilingual sentence encoder to career text. The paper reports that STEP outperforms existing baselines on all four public career-trajectory datasets—Decorte, Karrierewege, JobHop, and the new JobHop v2—with top-10 recall between 46.6% and 70.3%. It also releases JobHop v2, a 355k-trajectory benchmark with quarter-level timing and five-level education labels, giving the field a large public resource for reproducible career-path research. If the claim holds, job recommenders and employment services can make more precise suggestions by exploiting temporal and educational signals available in resumes.

Core claim

The paper claims that career trajectories are better modeled as time- and education-conditioned sequences than as plain token sequences or static embeddings. STEP encodes each job with frozen ROUTE embeddings, passes them through a GRU whose hidden state is exponentially decayed by inter-job gaps, applies FiLM conditioning on the highest degree attained, pools the hidden states by attention, and scores all 3,007 ESCO occupations through a learnable temperature-scaled cosine similarity. Trained end-to-end with cross-entropy and label smoothing, STEP reports the best MRR, Recall@5, and Recall@10 on all four benchmarks—for instance 46.6% R@10 on Decorte, 70.3% on Karrierewege, 40.4% on JobHop,

What carries the argument

The central mechanism is a time-decay GRU cell: before each update the hidden state is multiplied element-wise by exp(-lambda*delta), where delta is the real-valued gap in years since the previous position and lambda is a learned per-dimension decay rate, letting some career aspects persist while others fade. Around that cell, STEP layers FiLM modulation conditioned on a five-level education code, attention pooling over the sequence, and a learnable temperature that sharpens the softmax over the 3,007-class ESCO label space—the component with the largest single effect in the ablation. The supporting object is ROUTE, a two-stage embedding procedure (unsupervised denoising autoencoding on care

Load-bearing premise

The results depend on the automatically assigned occupation codes in the new benchmark being accurate enough that the reported margins over strong baselines reflect true predictive differences rather than systematic labeling artifacts.

What would settle it

Take a random sample of JobHop v2 test trajectories, have experienced annotators verify every occupation code and the true next job, and recompute the model comparison on only verified transitions. If STEP's margin over the strongest baseline disappears or reverses on the clean subset, the central claim is falsified. A cheaper diagnostic: compute each model's Recall@10 as a function of the confidence of the automatic code assignment and check whether the gap persists at high confidence.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • STEP achieves the top MRR, Recall@5, and Recall@10 on Decorte, Karrierewege, JobHop, and JobHop v2, outperforming MLP, BiLSTM, Markov, linear, and retrieval baselines on every metric.
  • The gains are largest where temporal and degree annotations exist: on JobHop v2, ROUTE+STEP beats the stronger BiLSTM baseline by 2.4 percentage points in R@10.
  • ROUTE embeddings improve direct cosine-similarity retrieval on all four datasets compared with the standard baseline, with the largest gain (+4.5 pp R@10) on JobHop v2, but the paper finds the embedding choice is a secondary lever once the STEP predictor is trained.
  • Autoregressive rollout degrades gradually: top-10 recall remains at 64.9% after five steps on Karrierewege and 28.7% on JobHop v2, so the model stays useful over a multi-step horizon.
  • Every learned model under-predicts how exploratory real career moves are; STEP's recommended occupations sit closer to the current one than the ground-truth transitions (novelty about 1.5 vs 2.9 on JobHop v2).

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the margins survive clean-label evaluation, the same time-decay and education-conditioning recipe should transfer to other irregularly spaced sequence-prediction settings—skills, courses, or internal mobility—where the gap between observations carries meaning; testing STEP on such taxonomies would be a direct extension.
  • The finding that retrieval-space improvements do not propagate to an end-to-end model suggests a caution for representation-learning benchmarks generally: gains measured by nearest-neighbour retrieval can overstate their value for downstream learners.
  • Because all models undershoot ground-truth novelty, a promising testable extension is a training objective with an explicit novelty bonus; one could check whether cross-sector recommendation improves without sacrificing top-10 accuracy.
  • Given the extraction pipeline's ~84% agreement with human annotations, the published numbers are likely an upper bound on cleanly labeled data; a model comparison on a human-verified subset would tell whether the structure of the claims changes.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper proposes STEP, a sequential model for next-job prediction over ESCO-encoded career trajectories, combining a time-decay GRU, FiLM-based education conditioning, attention pooling, and learnable temperature scaling; it also introduces ROUTE, a two-stage contrastive embedding procedure, and releases JobHop v2, a large-scale career-trajectory dataset derived from pseudonymized resumes. The system is evaluated on Decorte, Karrierewege, JobHop, and JobHop v2, with reported SOTA R@10 on all four datasets, accompanied by ablation studies, a diversity/novelty analysis, and a multi-step autoregressive evaluation. The empirical discipline is strong in places — five-seed runs, public code/dataset, multiple benchmarks, and component ablations — but the central SOTA claim rests on two load-bearing points that need further support: the comparison of STEP to baselines that do not appear to receive the dominant temperature-scaling component, and the reliability of the ESCO ground-truth labels in JobHop v2 and Karrierewege.

Significance. If the findings hold, the paper offers a useful public benchmark (JobHop v2) and a reproducible baseline for a practically relevant task. The paper is commendably transparent: it reports error bars, releases code and data, performs component ablations, includes multi-step rollout, and candidly discusses limitations such as fairness and the diversity-accuracy tradeoff. The retrieval-only comparison and the end-to-end ablations are informative. However, the headline claim of state-of-the-art performance on all four datasets is not yet fully secured: the ablation study shows that one implementation choice (learnable temperature scaling) accounts for 16.55 pp of the reported R@10, while the advertised architectural innovations contribute under 0.4 pp each, and the evaluation depends on ESCO labels whose noise is not quantified for the two datasets most central to the SOTA claim. The paper's significance would be strengthened by addressing these two issues directly.

major comments (3)
  1. [Section 5.3.1 and Table 4] The ablation in Table 4 shows that removing learnable temperature scaling degrades R@10 by 16.55 pp on JobHop v2, whereas every other advertised component changes R@10 by at most 0.38 pp. The MLP baseline is described as using cross-entropy over temperature-scaled cosine similarities, but it is not stated whether the learnable temperature parameter is included in the baselines. If the MLP and BiLSTM baselines use fixed or no temperature scaling, the reported SOTA margins on JobHop and JobHop v2 (2.2–2.9 pp R@10) may be attributable to this implementation choice rather than to STEP's sequential architecture. Please report baselines with the same temperature-scaling mechanism applied, or justify its exclusion.
  2. [Section 4.1, Appendix E, Appendix H] The SOTA claim depends on the accuracy of the ground-truth next-job ESCO codes. For Karrierewege, the paper itself reports a Berufenet-to-ESCO mapping accuracy of only 60.4% (Section 4.1). For JobHop v2, the ESCO occupation-code assignment is performed by a proprietary Nobl.ai classifier (Appendix E) whose assignment accuracy is never quantified; the 83.9% figure in Appendix H measures field-level extraction, not the final ESCO code assignment. Systematic label noise concentrated in certain occupations or transition types could plausibly create the observed 1.4–2.4 pp margins. The novelty analysis in Section 5.5.1, showing that true transitions are far more exploratory than any model predicts, is circumstantially consistent with mislabeling. A noise-sensitivity analysis, or a validation of the final ESCO codes on a labeled subsample, is needed to secure the central claim.
  3. [Section 5.4.2 and Table 5] With the STEP predictor held fixed, ROUTE embeddings are statistically indistinguishable from the STBase embeddings on three of four datasets (e.g., JobHop v2: 41.1±0.03 vs. 41.0±0.03; Karrierewege: 70.3±0.01 vs. 70.4±0.03). The retrieval-only gains in Table 1 are computed against the same next-job-matching objective that ROUTE was trained on, so they measure task alignment rather than independent representation quality. The paper's contribution statement that ROUTE 'improves the occupational embedding quality that STEP relies on' is therefore not supported in the end-to-end setting. Please either temper this claim or provide a downstream scenario in which ROUTE materially improves STEP.
minor comments (4)
  1. [Equation (3) and Appendix K] Appendix K states that λ=softplus(0)≈0.693 gives 'approximately no decay' for δ_t=1 year. In fact exp(-0.693·1)≈0.5, which is substantial decay. Please correct the description of the initialization's effect.
  2. [Table 2] The rows marked with † are attributed to Decorte et al. and Senger et al., but the values appear to be re-computed with the paper's own settings (e.g., cross-entropy training). Please clarify in the caption or text which numbers are direct re-implementations and which are copied from the original papers.
  3. [Section 5.5.2] The multi-step evaluation at M=1 is computed on a filtered subset (trajectories with length-T≥5) and is explicitly not comparable to Table 2. The text notes this, but the caption of Figure 3 should state it as well to prevent misreading.
  4. [Appendix H] The label-agreement row is described as an 'interpretive ceiling,' but the gap between the best extractor and the agreement ceiling could also reflect systematic errors shared across annotation sets. Please avoid over-interpreting this reference value.

Circularity Check

0 steps flagged

No significant circularity: main SOTA claim rests on held-out test splits; ROUTE evaluation is task-aligned but not circular.

full rationale

The derivation chain is self-contained. STEP and ROUTE are trained on training splits and evaluated on held-out test splits of four datasets; no parameter is fitted to test labels. The only self-referential aspect is that ROUTE's embedding-quality evaluation (Section 5.2, Table 1) measures cosine-similarity retrieval of the next job from a history prefix, which is exactly the objective used for supervised contrastive fine-tuning (Section 3.2.1, Eq. 1). However, the evaluation uses held-out prefixes and ground-truth next jobs, so it is a legitimate measure of task alignment rather than a reduction of the prediction to the fitted input. The self-citation of JobHop [26] is not load-bearing: JobHop is a publicly released benchmark, and JobHop v2 is a new dataset with its own construction pipeline, extraction evaluation, and held-out split. The skeptical concern about ESCO label noise in JobHop v2 and Karrierewege is a correctness/robustness risk, not a circularity, and the paper itself discloses extraction accuracy (83.9% on 200 hand-annotated resumes, Appendix H, Table 7) and Berufenet-to-ESCO mapping accuracy (60.4%, Section 4.1). No equation or fitted parameter is found to be equivalent to the claimed prediction by construction.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The central claims rest on the ESCO label space, the accuracy of LLM-extracted career data, and standard train/test split assumptions. The fitted parameters are dominated by a learnable temperature scalar; the time-decay rates and interval clamp are secondary. No exotic entities such as new particles, forces, or conserved quantities are introduced.

free parameters (4)
  • Learnable temperature tau = initialized 0.07, clamped [1e-3, 1] during training
    Scalar that scales cosine-similarity logits; ablation shows removing it drops R@10 by 16.55 pp on JobHop v2, so the headline result depends critically on this fitted scalar.
  • Per-dimension time-decay rates lambda = lambda = softplus(l), l initialized at 0 (approx. 0.693 at start)
    Learned decay rates for the GRU hidden state; central to the time-decay mechanism, but the ablation impact is small (-0.16 pp R@10).
  • Inter-job interval clamp upper bound = 20.0 years
    Chosen after experimenting with ten different thresholds (Section 5.1); a hand-tuned preprocessing constant with no uncertainty propagated into metrics.
  • ROUTE training hyperparameters = TSDAE 3 epochs, contrastive 2 epochs, batch sizes 32/64, LR 3e-5/2e-5
    Selected via validation on a 20% subset of Karrierewege (Appendix I); these choices are empirical and not derived from theory.
axioms (4)
  • domain assumption The ESCO taxonomy with 3,007 leaf occupations is an adequate label space for career transitions.
    The entire prediction task is defined over ESCO codes (Section 3.1); if this taxonomy is not the right granularity or is internally inconsistent, the evaluation is only meaningful relative to that choice.
  • domain assumption LLM-extracted ESCO codes in JobHop v2 are accurate enough that residual label noise does not invalidate the benchmark.
    Extraction accuracy is 83.9% on 200 hand-annotated resumes (Appendix H); the paper does not analyze error propagation into next-job prediction metrics.
  • domain assumption Official train/validation/test splits are independent, with no leakage into ROUTE or STEP training.
    All results assume the released splits are respected; the paper states this, but code was not independently executed in this review.
  • domain assumption The frozen guide model all-MiniLM-L6-v2 provides reliable similarity signals for GISTEmbedLoss negative selection.
    ROUTE's contrastive stage relies on this guide model (Section 3.2.2); if its similarities are miscalibrated, false-negative filtering could distort the embedding space.

pith-pipeline@v1.3.0-alltime-deepseek · 37605 in / 13024 out tokens · 126800 ms · 2026-08-02T06:47:16.873768+00:00 · methodology

0 comments
read the original abstract

Career paths encode decades of skill acquisition, role transitions, and educational investment, and understanding them at scale underpins workforce planning, labor market policy, and job recommendation. Resumes are a rich source of information about career paths: they contain detailed descriptions of work experience, education, and skills. Yet their unstructured, heterogeneous, and multilingual nature has long prevented large-scale systematic analysis. With the advent of large language models (LLMs), it is now possible to source rich career trajectory data containing temporal and educational signals from unstructured resumes, enabling new opportunities for career-path recommendation. Exploiting this opportunity, we present STEP (Sequential Trajectory of Employment Prediction), a novel career-path recommendation system that leverages temporal and educational signals to predict the next job in a career trajectory. STEP integrates a time-decay Gated Recurrent Unit (GRU) cell to model temporal dynamics, Feature-wise Linear Modulation (FiLM) conditioned on educational attainment, and attention-based sequence pooling to select relevant features for next job prediction. To improve internal occupation representation for STEP, we introduce ROUTE, a two-stage contrastive procedure that first adapts a multilingual encoder to the career domain via unsupervised denoising autoencoding, then performs supervised contrastive fine-tuning with guided negative selection. We evaluate STEP on four datasets of career trajectories, including an improved version of our publicly available JobHop dataset, and show that it outperforms state-of-the-art baselines in next job prediction. The dataset and code are publicly released to support reproducible career-trajectory research.

Figures

Figures reproduced from arXiv: 2607.11722 by Alexandru C. Mara, Guillaume Bied, Iman Johary, Tijl De Bie.

Figure 1
Figure 1. Figure 1: Overview of the career path recommendation pipeline. Each occupation in a career trajectory is [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Architecture of STEP (Sequential Trajectory of Employment Prediction). The model processes input [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Multi-step autoregressive evaluation across all three datasets. Performance degrades monotonically [PITH_FULL_IMAGE:figures/full_fig_p021_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Full developer instruction used for resume extraction (Section 4.2). [PITH_FULL_IMAGE:figures/full_fig_p028_4.png] view at source ↗
Figure 4
Figure 4. Figure 4: In Figure 4, internal infrastructure details and model-specific encoding directives have [PITH_FULL_IMAGE:figures/full_fig_p029_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Extraction illustration using a synthetic resume representative of the VDAB corpus (Section 4.2). [PITH_FULL_IMAGE:figures/full_fig_p030_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Occupational diversity and career history lengths in JobHop v2. Professionals (ISCO 2), service & sales [PITH_FULL_IMAGE:figures/full_fig_p033_6.png] view at source ↗

Review history (2 revisions) →

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

67 extracted references · 11 linked inside Pith

  1. [1]

    Alibasic, H

    A. Alibasic, H. Upadhyay, M.C.E. Simsekler, T. Kurfess, W.L. Woon, and M.A. Omar. 2022. Evaluation of the trends in jobs and skill-sets using data analytics: a case study.Journal of Big Data9, 1 (2022). doi:10.1186/s40537-022-00576-5

    doi:10.1186/s40537-022-00576-5 2022

  2. [2]

    2022.The Direct and Indirect Effects of Online Job Search Advice

    Steffen Altmann, Anita Marie Glenny, Robert Mahlstedt, and Alexander Sebald. 2022.The Direct and Indirect Effects of Online Job Search Advice. IZA Discussion Paper 15830. Institute of Labor Economics (IZA)

    2022

  3. [3]

    C. H. Ayishathahira, C. Sreejith, and C. Raseek. 2018. Combination of Neural Networks and Conditional Random Fields for Efficient Resume Parsing. In2018 International CET Conference on Control, Communication, and Computing (IC4). IEEE, Thiruvananthapuram, India, 388–393

    2018

  4. [4]

    Baytas, Cao Xiao, Xi Zhang, Fei Wang, Anil K

    Inci M. Baytas, Cao Xiao, Xi Zhang, Fei Wang, Anil K. Jain, and Jiayu Zhou. 2017. Patient Subtyping via Time-Aware LSTM Networks. InProceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD ’17). Association for Computing Machinery, New York, NY, USA, 65–74. doi:10.1145/3097983.3097997

    arXiv 2017

  5. [5]

    2025.Advising Job Seekers in Occupations with Poor Prospects: A Field Experiment

    Michèle Belot, Bart de Koning, Didier Fouarge, Philipp Kircher, Paul Muller, and Sandra Phlippen. 2025.Advising Job Seekers in Occupations with Poor Prospects: A Field Experiment. NBER Working Paper 33819. National Bureau of Economic Research. doi:10.3386/w33819

    doi:10.3386/w33819 2025

  6. [6]

    Michèle Belot, Philipp Kircher, and Paul Muller. 2019. Providing Advice to Jobseekers at Low Cost: An Experimental Study on Online Advice.The Review of Economic Studies86, 4 (2019), 1411–1447. doi:10.1093/restud/rdy059

    doi:10.1093/restud/rdy059 2019

  7. [7]

    Michèle Belot, Philipp Kircher, and Paul Muller. 2026. Do the Long-term Unemployed Benefit from Automated Occupational Advice during Online Job Search?The Economic Journal136, 673 (2026), 184–206

    2026

  8. [8]

    Census Bureau and U.S

    U.S. Census Bureau and U.S. Bureau of Labor Statistics. 2023. Current Population Survey. U.S. Census Bureau. https://www.census.gov/programs-surveys/cps.html

    2023

  9. [9]

    Lilia Chang, Lisa Simon, Karthik Rajkumar, and Susan Athey. 2019. A bayesian approach to predicting occupational transitions. Working Paper

    2019

  10. [10]

    Ricky T. Q. Chen, Yulia Rubanova, Jesse Bettencourt, and David K. Duvenaud. 2018. Neural Ordinary Differential Equations. InAdvances in Neural Information Processing Systems (NeurIPS 2018), Vol. 31

    2018

  11. [11]

    Kyunghyun Cho, Bart van Merriënboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. 2014. Learning Phrase Representations using RNN Encoder–Decoder for Statistical Machine Translation. InProceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP). Association for Computational Lingu...

    doi:10.3115/v1/d14-1179 2014

  12. [12]

    Vachik S Dave, Baichuan Zhang, Mohammad Al Hasan, Khalifeh AlJadda, and Mohammed Korayem. 2018. A combined representation learning approach for better job and skill recommendation. InProceedings of the 27th ACM International Conference on Information and Knowledge Management. ACM, Turin, Italy, 1997–2005

    2018

  13. [13]

    Maria De-Arteaga, Alexey Romanov, Hanna Wallach, Jennifer Chayes, Christian Borgs, Alexandra Chouldechova, Sahin Geyik, Krishnaram Kenthapadi, and Adam Tauman Kalai. 2019. Bias in Bios: A Case Study of Semantic Representation Bias in a High-Stakes Setting. InProceedings of the Conference on Fairness, Accountability, and Transparency(Atlanta, GA, USA)(FAT*...

    arXiv 2019

  14. [14]

    Jens-Joris Decorte, Jeroen Van Hautte, Johannes Deleu, Chris Develder, and Thomas Demeester. 2023. Career path prediction using resume representation learning and skill-based matching.arXiv preprint arXiv:2310.15636(2023)

    Pith/arXiv arXiv 2023

  15. [15]

    Tianyu Du, Ayush Kanodia, Herman Brunborg, Keyon Vafa, and Susan Athey. 2024. LABOR-LLM: Language-Based Occupational Representations with Large Language Models

    2024

  16. [16]

    Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, et al. 2024. The Llama 3 Herd of Models. arXiv:2407.21783 [cs.AI]

    Pith/arXiv arXiv 2024

  17. [17]

    European Commission, Directorate-General for Employment, Social Affairs and Inclusion, Martin Le Vrang, Agis Papantoniou, Erika Pauwels, Pieter Fannes, Dominique Vandensteen, and Johan De Smedt. 2014. ESCO: Boosting Job ACM Trans. Inf. Syst., Vol. 1, No. 1, Article 1. Publication date: January 2026. STEP: A Career-Path Recommendation System 1:25 Matching ...

    doi:10.1109/mc.2014.283 2014

  18. [18]

    Tianyu Gao, Xingcheng Yao, and Danqi Chen. 2021. SimCSE: Simple Contrastive Learning of Sentence Embeddings. InProceedings of the 2021 Conference on Empirical Methods in Natural Language Processing (EMNLP). Association for Computational Linguistics, 6894–6910. doi:10.18653/v1/2021.emnlp-main.552

    doi:10.18653/v1/2021.emnlp-main.552 2021

  19. [19]

    Gemma Team, Morgane Riviere, Shreya Pathak, Pier Giuseppe Sessa, Cassidy Hardin, et al. 2024. Gemma 2: Improving Open Language Models at a Practical Size. arXiv:2408.00118 [cs.CL]

    Pith/arXiv arXiv 2024

  20. [20]

    Ann-Sophie Gnehm, Eva Bühlmann, and Simon Clematide. 2022. Evaluation of Transfer Learning and Domain Adaptation for Analyzing German-Speaking Job Advertisements. InProceedings of the Thirteenth Language Resources and Evaluation Conference (LREC). 3892–3901. Introduces jobGBERT: GBERT base + domain-adaptive pretraining on 4M Swiss job advertisements

    2022

  21. [21]

    Akshay Gugnani and Hemant Misra. 2020. Implicit Skills Extraction Using Document Embedding and Its Use in Job Recommendation. InProceedings of the AAAI Conference on Artificial Intelligence (AAAI 2020, IAAI Track), Vol. 34. 13286–13293. doi:10.1609/aaai.v34i08.7038

    doi:10.1609/aaai.v34i08.7038 2020

  22. [22]

    Matthew Henderson, Rami Al-Rfou, Brian Strope, Yun-hsuan Sung, László Lukács, Ruiqi Guo, Sanjiv Kumar, Balint Mik- los, and Ray Kurzweil. 2017. Efficient Natural Language Response Suggestion for Smart Reply. arXiv:1705.00652 [cs.CL] https://arxiv.org/abs/1705.00652

    Pith/arXiv arXiv 2017

  23. [23]

    Balázs Hidasi, Alexandros Karatzoglou, Linas Baltrunas, and Domonkos Tikk. 2016. Session-based Recommendations with Recurrent Neural Networks. InInternational Conference on Learning Representations (ICLR). https://arxiv.org/abs/ 1511.06939

    Pith/arXiv arXiv 2016

  24. [24]

    Sepp Hochreiter and Jürgen Schmidhuber. 1997. Long Short-Term Memory.Neural Computation9, 8 (1997), 1735–1780. doi:10.1162/neco.1997.9.8.1735

    doi:10.1162/neco.1997.9.8.1735 1997

  25. [25]

    Horton and Prasanna Tambe

    John J. Horton and Prasanna Tambe. 2015. Labor Economists Get Their Microscope: Big Data and Labor Market Analysis.Big Data3, 3 (2015), 130–137. arXiv:https://doi.org/10.1089/big.2015.0017 doi:10.1089/big.2015.0017 PMID: 27442956

    arXiv 2015

  26. [26]

    Mara, and Tijl De Bie

    Iman Johary, Raphaël Romero, Alexandru C. Mara, and Tijl De Bie. 2025. JobHop: A Large-Scale Dataset of Career Trajectories. In2025 IEEE International Conference on Big Data (BigData). 2184–2191. doi:10.1109/BigData66926.2025. 11402454

    arXiv 2025

  27. [27]

    Jeff Johnson, Matthijs Douze, and Hervé Jégou. 2019. Billion-Scale Similarity Search with GPUs.IEEE Transactions on Big Data7, 3 (2019), 535–547. doi:10.1109/TBDATA.2019.2921572 FAISS library

    arXiv 2019

  28. [28]

    Wang-Cheng Kang and Julian McAuley. 2018. Self-Attentive Sequential Recommendation. In2018 IEEE International Conference on Data Mining (ICDM). IEEE, 197–206. doi:10.1109/ICDM.2018.00035

    arXiv 2018

  29. [29]

    Mesut Kaya and Toine Bogers. 2021. Effectiveness of job title based embeddings on résumé to job ad recommendation. InCEUR Workshop Proceedings, Vol. 2967. CEUR Workshop Proceedings

    2021

  30. [30]

    Aditya Kusupati, Gantavya Bhatt, Aniket Rege, Matthew Wallingford, Aditya Sinha, Vivek Ramanujan, William Howard-Snyder, Kaifeng Chen, Sham Kakade, Prateek Jain, and Ali Farhadi. 2022. Matryoshka Representation Learning. InAdvances in Neural Information Processing Systems (NeurIPS 2022), Vol. 35. 30233–30249

    2022

  31. [31]

    Gonzalez, Hao Zhang, and Ion Stoica

    Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph E. Gonzalez, Hao Zhang, and Ion Stoica. 2023. Efficient Memory Management for Large Language Model Serving with PagedAttention. InProceedings of the 29th Symposium on Operating Systems Principles (SOSP ’23). Association for Computing Machinery, New York, NY, USA. doi:10...

    arXiv 2023

  32. [32]

    Jaeyoung Lee, Michiharu Yamashita, Thanh Tran, Dongwon Lee, and Sangpil Kim. 2025. CAPER: Enhancing Career Trajectory Prediction with Query-Aware Graph Augmentation and Continual Learning. InProceedings of the ACM Web Conference 2025 (WWW ’25). To appear; arXiv preprint

    2025

  33. [33]

    Jiacheng Li, Yujie Wang, and Julian McAuley. 2020. Time Interval Aware Self-Attention for Sequential Recommendation. InProceedings of the 13th International Conference on Web Search and Data Mining (WSDM ’20). Association for Computing Machinery, New York, NY, USA, 322–330. doi:10.1145/3336191.3371786

    arXiv 2020

  34. [34]

    Liangyue Li, How Jing, Hanghang Tong, Jaewon Yang, Qi He, and Bee-Chung Chen. 2017. NEMO: Next Career Move Prediction with Contextual Embedding. InProceedings of the 26th International Conference on World Wide Web Companion(Perth, Australia)(WWW ’17 Companion). International World Wide Web Conferences Steering Committee, Republic and Canton of Geneva, CHE...

    arXiv 2017

  35. [35]

    Nan Li, Bo Kang, and Tijl De Bie. 2023. LLM4Jobs: unsupervised occupation extraction and standardization leveraging Large Language Models. doi:10.48550/arXiv.2309.09708 arXiv:2309.09708 [cs]

    Pith/arXiv arXiv doi:10.48550/arxiv.2309.09708 2023

  36. [36]

    Nan Li, Bo Kang, and Tijl De Bie. 2023. SkillGPT: a RESTful API service for skill extraction and standardization using a Large Language Model. doi:10.48550/arXiv.2304.11060 arXiv:2304.11060 [cs]

    Pith/arXiv arXiv doi:10.48550/arxiv.2304.11060 2023

  37. [37]

    Ilya Loshchilov and Frank Hutter. 2017. SGDR: Stochastic Gradient Descent with Warm Restarts. InInternational Conference on Learning Representations (ICLR). https://openreview.net/forum?id=Skq89Scxx ACM Trans. Inf. Syst., Vol. 1, No. 1, Article 1. Publication date: January 2026. 1:26 Johary et al

    2017

  38. [38]

    Ilya Loshchilov and Frank Hutter. 2019. Decoupled Weight Decay Regularization. InInternational Conference on Learning Representations (ICLR). https://openreview.net/forum?id=Bkg6RiCqY7

    2019

  39. [39]

    Qingxin Meng, Hengshu Zhu, Keli Xiao, Le Zhang, and Hui Xiong. 2019. A hierarchical career-path-aware neural network for job mobility prediction. InProceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining. 14–24

    2019

  40. [40]

    Whitney Moore, Steven Pedlow, Parvati Krishnamurty, Kirk Wolter, and IL Chicago. 2000. National longitudinal survey of youth 1997 (NLSY97).National Opinion Research Center, Chicago, IL254 (2000), 22

    2000

  41. [41]

    OpenAI. 2025. gpt-oss-120b and gpt-oss-20b Model Card. https://openai.com/index/introducing-gpt-oss/. Model release and accompanying technical report

    2025

  42. [42]

    Ethan Perez, Florian Strub, Harm de Vries, Vincent Dumoulin, and Aaron Courville. 2018. FiLM: Visual Reasoning with a General Conditioning Layer. InProceedings of the AAAI Conference on Artificial Intelligence (AAAI 2018), Vol. 32. doi:10.1609/aaai.v32i1.11671

    doi:10.1609/aaai.v32i1.11671 2018

  43. [43]

    Chuan Qin, Hengshu Zhu, Tong Xu, Chen Zhu, Chao Ma, Enhong Chen, and Hui Xiong. 2020. An Enhanced Neural Network Approach to Person-Job Fit in Talent Recruitment.ACM Transactions on Information Systems38, 2, Article 15 (2020), 33 pages. doi:10.1145/3376927

    doi:10.1145/3376927 2020

  44. [44]

    Nils Reimers and Iryna Gurevych. 2019. Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). Association for Computational Linguistics, Hong Kong, China, 3982–3992. do...

    doi:10.18653/v1/d19-1410 2019

  45. [45]

    Julian Rosenberger, Nicolas Bougie, Ann-Sophie Gnehm, Linda Dietz, Simon Clematide, and Matthias Gramsch. 2025. CareerBERT: Matching resumes to ESCO jobs in a shared embedding space for generic job recommendations.Expert Systems with Applications(2025). doi:10.1016/j.eswa.2025.127043 arXiv:2503.02056

    arXiv 2025

  46. [46]

    Oscar Sainz, Iker García-Ferrero, Rodrigo Agerri, Oier Lopez de Lacalle, German Rigau, and Eneko Agirre. 2024. GoLLIE: Annotation Guidelines improve Zero-Shot Information-Extraction. InInternational Conference on Learning Representations (ICLR). https://openreview.net/forum?id=Y3wpuxd7u9

    2024

  47. [47]

    Abel Salinas, Parth Shah, Yuzhong Huang, Robert McCormack, and Fred Morstatter. 2023. The Unequal Opportunities of Large Language Models: Examining Demographic Biases in Job Recommendations by ChatGPT and LLaMA. In Proceedings of the 3rd ACM Conference on Equity and Access in Algorithms, Mechanisms, and Optimization (EAAMO ’23). Association for Computing ...

    arXiv 2023

  48. [48]

    Roan Schellingerhout, Volodymyr Medentsiy, and Maarten de Rijke. 2022. Explainable Career Path Predictions using Neural Models. InProceedings of the RecSys in HR 2022 Workshop (co-located with RecSys 2022) (CEUR Workshop Proceedings). https://ceur-ws.org/Vol-3218/RecSysHR2022-paper_6.pdf

    2022

  49. [49]

    Elena Senger, Yuri Campbell, Rob van der Goot, and Barbara Plank. 2024. KARRIEREWEGE: A Large Scale Career Path Prediction Dataset. arXiv:2412.14612 [cs.CL] https://arxiv.org/abs/2412.14612

    Pith/arXiv arXiv 2024

  50. [50]

    Elena Senger, Yuri Campbell, Rob van der Goot, and Barbara Plank. 2025. Deep Learning Methods for Multilingual Career Path Prediction. arXiv:2503.17276 [cs.CL] https://arxiv.org/abs/2503.17276

    Pith/arXiv arXiv 2025

  51. [51]

    Solatorio

    Aivin V. Solatorio. 2024. GISTEmbed: Guided In-sample Selection of Training Negatives for Text Embedding Fine-tuning. arXiv:2402.16829 [cs.CL] https://arxiv.org/abs/2402.16829

    Pith/arXiv arXiv 2024

  52. [52]

    Kaitao Song, Xu Tan, Tao Qin, Jianfeng Lu, and Tie-Yan Liu. 2020. MPNet: Masked and Permuted Pre-training for Language Understanding. InAdvances in Neural Information Processing Systems (NeurIPS 2020), Vol. 33. 16857–16867

    2020

  53. [54]

    Keyon Vafa, Emil Palikot, Tianyu Du, Ayush Kanodia, Susan Athey, and David Blei. 2024. CAREER: A Foundation Model for Labor Sequence Data. https://openreview.net/forum?id=4i1MXH8Sle

    2024

  54. [55]

    Cerilla, Aaron A

    Carl John Vinas, Micaela T. Cerilla, Aaron A. Santillan, and Michael B. Dela Fuente. 2023. Career Path Modeling and Recommendations with Linkedin Career Data and Predicted Salary Estimations. https://openreview.net/forum?id= R5NNAThG0i

    2023

  55. [56]

    Chao Wang, Hengshu Zhu, Qiming Hao, Keli Xiao, and Hui Xiong. 2021. Variable Interval Time Sequence Modeling for Career Trajectory Prediction: Deep Collaborative Perspective. InProceedings of the Web Conference 2021(Ljubljana, Slovenia)(WWW ’21). Association for Computing Machinery, New York, NY, USA, 612–623. doi:10.1145/3442381. 3449959

    doi:10.1145/3442381 2021

  56. [57]

    Kexin Wang, Nils Reimers, and Iryna Gurevych. 2021. TSDAE: Using Transformer-based Sequential Denoising Auto-Encoder for Unsupervised Sentence Embedding Learning. InFindings of the Association for Computational Linguistics: EMNLP 2021. Association for Computational Linguistics, Punta Cana, Dominican Republic, 671–688. doi:10.18653/v1/2021.findings-emnlp.5...

    doi:10.18653/v1/2021.findings-emnlp.59 2021

  57. [58]

    Liang Wang, Nan Yang, Xiaolong Huang, Linjun Yang, Rangan Majumder, and Furu Wei. 2024. Multilingual E5 Text Embeddings: A Technical Report. arXiv:2402.05672 [cs.CL] https://arxiv.org/abs/2402.05672

    Pith/arXiv arXiv 2024

  58. [59]

    Shitao Xiao, Zheng Liu, Peitian Zhang, Niklas Muennighoff, Defu Lian, and Jian-Yun Nie. 2024. C-Pack: Packed Resources For General Chinese Embeddings. InProceedings of the 47th International ACM SIGIR Conference on Research and Develop- ment in Information Retrieval (SIGIR ’24). Association for Computing Machinery, 641–649. doi:10.1145/3626772.3657878 Int...

    arXiv 2024

  59. [60]

    Michiharu Yamashita, Yunqi Li, Thanh Tran, Yongfeng Zhang, and Dongwon Lee. 2022. Looking further into the future: Career pathway prediction.WSDM Computational Jobs Marketplace 2022(2022)

    2022

  60. [61]

    Michiharu Yamashita, Thanh Tran, and Dongwon Lee. 2024. OpenResume: Advancing Career Trajectory Modeling with Anonymized and Synthetic Resume Datasets. In2024 IEEE International Conference on Big Data (BigData). 6697–6706. doi:10.1109/BigData62323.2024.10825519

    arXiv 2024

  61. [62]

    Rui Zha, Ying Sun, Chuan Qin, Le Zhang, Tong Xu, Hengshu Zhu, and Enhong Chen. 2024. Toward Unified Represen- tation Learning for Career Mobility Analysis with Trajectory Hypergraph.ACM Transactions on Information Systems 42, 4, Article 110 (2024), 28 pages. doi:10.1145/3651158

    doi:10.1145/3651158 2024

  62. [63]

    Denghui Zhang, Junming Liu, Hengshu Zhu, Yanchi Liu, Lichen Wang, Pengyang Wang, and Hui Xiong. 2019. Job2Vec: Job Title Benchmarking with Collective Multi-View Representation Learning. InProceedings of the 28th ACM International Conference on Information and Knowledge Management (CIKM ’19). Association for Computing Machinery, New York, NY, USA, 2763–277...

    arXiv 2019

  63. [64]

    Le Zhang, Ding Zhou, Hengshu Zhu, Tong Xu, Rui Zha, Enhong Chen, and Hui Xiong. 2021. Attentive heterogeneous graph embedding for job mobility prediction. InProceedings of the 27th ACM SIGKDD conference on knowledge discovery & data mining. 2192–2201. Overview of the Appendices The appendices collect the supplementary technical material supporting the mai...

    2021

  64. [65]

    -” for missing/masked values. Drop parts containing only <MASK>. -Current Roles:Replace markers like “Present

    GENERAL CONSTRAINTS -Input:Resume text captured at timestamp: {timestamp}. -Output:A single JSON object. No markdown, no explanations, no <MASK> tokens. -Language:Do NOT translate descriptions or titles. Keep original language. -Missing Data:Use “-” for missing/masked values. Drop parts containing only <MASK>. -Current Roles:Replace markers like “Present”...

  65. [66]

    2008, 2009 and 2015

    LOGIC & PARSING RULES -MERGE:Consecutive entries for the same role/company with overlapping or adjacent dates must be combined into one entry. -SPLIT:Single entries listing disjoint dates (e.g., “2008, 2009 and 2015”) must be split into separate entries for each period. -GROUPS:Unpack grouped sections (e.g., “Interims”, “Stages”) into individual experienc...

    2008

  66. [67]

    Education

    EXTRACTION SCOPE -Skills:Extract ONLY technical skills/tools. Ignore soft skills. -Experience: CRITICAL: Do NOT extract items from “Education” or “Profile” sections as work experience. Be careful with multi-column layouts. For the title field, extract the job title exactly as written. For the standard_title field, infer a standardized title (preferably ES...

  67. [68]

    work_experiences

    OUTPUT SCHEMA { "work_experiences": [ { "title": "Original Job Title", "standard_title": "Standardized Job Title", "company_name": "Organization Name (or'-')", "location": "City only (no country)", "description": "Original text description", "start_date": "Month Year | Year | -", "end_date": "Month Year | Year | -", "work_schedule_type": "full-time | part...

    2026