Synapse: Evolving Job-Person Fit with Explainable Two-phase Retrieval and LLM-guided Genetic Resume Optimization

Abstract: Modern recruitment platforms operate under severe information imbalance: job seekers must search over massive, rapidly changing collections of postings, while employers are overwhelmed by high-volume, low-relevance applicant pools. Existing recruitment recommender systems typically rely on keyword matching or single-stage semantic retrieval, which struggle to capture fine-grained alignment between candidate experience and job requirements under real-world scale and cost constraints. We present Synapse, a multi-stage semantic recruitment system that separates high-recall candidate generation from high-precision semantic reranking, combining efficient dense retrieval using FAISS with an ensemble of contrastive learning and LLM reasoning. To improve transparency, Synapse incorporates a retrieval-augmented explanation layer that grounds recommendations in explicit evidence. Beyond retrieval, we introduce a novel evolutionary resume optimization framework that treats resume refinement as a black-box optimization problem. Using Differential Evolution with LLM-guided mutation operators, the system iteratively modifies candidate representations to improve alignment with screening objectives, without any labeled data. Evaluation shows that the proposed ensemble improves nDCG@10 by 22% over embedding-only retrieval baselines, while the evolutionary optimization loop consistently yields monotonic improvements in recommender scores, exceeding 60% relative gain across evaluated profiles. We plan to release code and data upon publication.

• The paper introduces Synapse, a two-phase retrieval system that integrates LLM-guided genetic resume optimization to enhance candidate-job alignment.
• It employs a scalable ensemble of rerankers and retrieval-augmented generation for explainable recommendations, achieving up to 31.9% nDCG@10 improvement.
• The evolutionary resume optimization process yields median fitness gains of 62% through iterative, label-free population-based refinement.

Synapse: A Multi-stage, Explainable Recruitment Recommender with LLM-guided Evolutionary Resume Optimization

Introduction and Motivation

The paper presents Synapse, a recruitment recommender system targeting the longstanding challenges of semantic candidate-job alignment, retrieval precision, and transparency in digital hiring flows (2604.02539). The proliferation of job postings and applicant volume has rendered existing keyword-based and embedding-only systems inadequate due to their inability to capture fine-grained, latent alignments and enforce explicit constraints, especially within highly dynamic and data-limited recruitment environments.

Synapse addresses these limitations with a multi-stage architecture combining high-recall dense retrieval with high-precision semantic reranking and contextual explainability. A novel contribution is the LLM-guided, evolutionary resume optimization loop, which iteratively refines candidate resumes towards better alignment with recruiter screening objectives, operationalized entirely in a label-free, black-box setting.

System Architecture: Two-phase Retrieval and Explainability

The Synapse pipeline is structured as a scalable two-phase recommender, carefully designed to partition efficiency-critical broad search from computationally expensive, context-sensitive reranking.

In Phase I, both resumes and job postings are embedded into a shared semantic space using a lightweight Sentence-BERT variant. Fast approximate nearest-neighbor search over a FAISS index retrieves top-K candidates, aggressively prioritizing recall and throughput over precision.

Figure 1: The Synapse recommender pipeline combines high-recall dense retrieval with precise semantic reranking.

Phase II then applies multiple more expressive reranking models to the retrieved subset: (1) deeper embedding similarity with higher-dimensional latent spaces, (2) a contrastive learning model trained on augmented resume-job pairings with a RoBERTa backbone and a token-aligned triplet loss to maximize semantic robustness, and (3) pairwise preference modeling using LLMs, which proved superior to direct scalar scoring. A variety of ensemble strategies (including weighted averaging, Borda count, and Reciprocal Rank Fusion) were explored for optimal aggregation, with weighted averaging yielding the highest nDCG improvements.

Complementing this retrieval stack, Synapse introduces an explainability module grounded in Retrieval-Augmented Generation (RAG). Here, relevant evidence snippets are extracted from both the resume and job description, and provided to an LLM, which generates textual rationales for recommendations. This mitigates user trust issues and hallucination risk inherent to opaque neural models, and provides actionable feedback for applicants.

LLM-guided Evolutionary Resume Refinement

A key innovation of Synapse is modeling resume optimization as a black-box, population-based search. The system operationalizes fitness as the weighted sum of embedding distance, contrastive model similarity, and LLM preference—all computed against a target posting set—with empirical weighting heavily favoring embedding alignment ($0.7$), and smaller contributions from the higher-precision rerankers.

The evolutionary process initializes a population of mutated resumes via temperature-controlled section rewriting (mutation), employs LLM-driven semantic merging (crossover), and ensures elitism for monotonic fitness improvement. Aggressive mutations become more probable in later generations to promote escape from local optima. Fitness-based selection iteratively advances the population, yielding resumes with progressively better system alignment, with no need for labeled data or human intervention.

Experimental Evaluation

The Synapse system was rigorously evaluated for computational efficiency, retrieval quality, and the effectiveness of the evolutionary refinement loop. Key findings include:

• System latency: GPU acceleration notably reduced embedding and scoring time (from 0.57s to 0.026s per resume-job pair). LLM-based RAG explanations remain the main bottleneck, averaging 7 seconds due to API overhead.
• Ranking quality: The ensemble of reranking methods achieved notable gains, with the best weighted-average rank ensemble improving nDCG@10 by 31.9% over the baseline embedding-only system. Contrastive learning and pairwise LLM ranking individually produced 17.4% and 10.2% gains, respectively, over the baseline.
• Resume optimization: Differential Evolution—using LLM-guided mutation and crossover—consistently improved fitness scores across diverse resume sets.

  Figure 2: Relative fitness improvement across generations for ten resumes, the blue line representing the mean.

Quantitatively, the evolutionary loop achieved monotonic fitness improvement, yielding median, mean, and upper-quartile gains of 62%, 68%, and 92% respectively, as measured by the composite objective function across top-ranked job postings.

Implications and Future Directions

The Synapse system offers several immediate and longer-term implications:

• Recruitment System Design: The demonstrated architecture provides a practical blueprint for scalable deployment in high-volume recruitment settings. The strong gains from ensemble reranking and explainability mechanisms underscore the limits of single-model semantic similarity and highlight the value of integrating orthogonal signals.
• Interactive Career Optimization: The LLM-guided evolutionary loop offers a data-agnostic, modular framework for automated, system-aligned resume refinement, which can be generalized to broader document optimization and recommendation tasks.
• Bias and Data Scarcity: While the system circumvents the need for large-scale labeled data, its reliance on automated fitness metrics introduces potential for encoding and perpetuating algorithmic biases inherent in underlying embedding or LLM models. As such, research on augmenting supervision via synthetic labeling, adversarial debiasing, and expanded transparency auditing is warranted.
• LLM Cost Tradeoffs: The latency and cost overhead of LLM-based evaluation remains the limiting factor for real-time, large-scale deployment. Continued advances in efficient LLMs or distillation approaches will be required for practical adoption.

Conclusion

Synapse advances recruitment recommendation by tightly integrating scalable two-phase retrieval, transparent RAG-style explainability, and a novel LLM-driven evolutionary optimization for resumes. The system realizes strong empirical improvements—up to 31.9% nDCG@10 over basic embeddings, and 62–92% evolutionary fitness gains—without requiring extensive labeled data. Ongoing work may focus on further efficiency optimizations, robustifying fitness objectives, and expanding explainability to address trust and bias in human-centric AI systems.