An Efficient Long-Context Ranking Architecture With Calibrated LLM Distillation: Application to Person-Job Fit

TL;DR

This work tackles real-time, cross-language person–project matching with long, structured resumes. It introduces a calibrated distillation framework where a generative LLM serves as a teacher to generate semantically grounded relevance scores, which are then mimicked by a lightweight two-branch reranker that processes utterance-level representations and uses late cross-attention to form detailed similarity distributions. Training leverages pointwise MSE, pairwise Calibrated Margin MSE Distillation (CMMD), and listwise CalID distillation to align both relative rankings and absolute score magnitudes with the teacher, yielding semantically calibrated scores in $[0,1]$ for robust ranking and interpretability. The approach delivers strong relevance, ranking, and calibration performance while maintaining real-time efficiency (e.g., processing 1,000 pairs in under a minute with precomputed embeddings) and demonstrates resilience to multilingual and out-of-distribution inputs. This has practical impact for scalable, transparent HR recommender systems across languages and domains.

Abstract

Finding the most relevant person for a job proposal in real time is challenging, especially when resumes are long, structured, and multilingual. In this paper, we propose a re-ranking model based on a new generation of late cross-attention architecture, that decomposes both resumes and project briefs to efficiently handle long-context inputs with minimal computational overhead. To mitigate historical data biases, we use a generative large language model (LLM) as a teacher, generating fine-grained, semantically grounded supervision. This signal is distilled into our student model via an enriched distillation loss function. The resulting model produces skill-fit scores that enable consistent and interpretable person-job matching. Experiments on relevance, ranking, and calibration metrics demonstrate that our approach outperforms state-of-the-art baselines.

Figures (4)

• Figure 1: Overview of the semantic score distillation pipeline: a generative LLM assigns semantic labels that are mapped to reference scores to form a calibrated teacher model $M_{\text{teacher}}$, whose outputs are used to train a lightweight student model $M_{\text{student}}$.
• Figure 2: Illustration of the proposed brief encoder branch (top) and profile encoder branch (bottom). Similar sections across freelancer profiles are first projected into a latent space (shown in the first three blocks). Positional encodings are then added, and the resulting embeddings are combined with learned section-type embeddings to retain structural information. These enriched vectors are finally passed through a linear layer.
• Figure 3: Overview of the proposed comparison architecture. The first two blocks (“Profiles to Brief” and “Brief to Profiles”) depict the multi-head cross-attention mechanism. The following stages show the computation of cosine similarity distributions, statistical pooling into fixed-size features, and final concatenation before scoring via an MLP.
• Figure 4: Box plots of prediction errors grouped by discrete ground-truth scores (left) and joint distribution between ground-truth scores and model predictions (right) for each method.