A Systematized Review of AI Recruitment Systems: Evaluation and Governance

Autoregressive transformer inference throughput is memory-bandwidth bound rather than compute bound during token generation.** Optimizing deployments for A Systematized Review of AI Recruitment Systems: Evaluation and Governance requires matching parameter weight footprints against unified memory transfer rates, where high-bandwidth architectures achieve 4x to 6x acceleration over conventional dual-channel DDR5 system memory. The use of artificial intelligence (AI) in recruitment has undergone significant changes in recent years. From bilateral retrieval and behavioral ranking to neural person-job matching, large language model (LLM) components, and tool-using recruiting agents, AI recruitment systems have evolved to support multi-stage workflows that retrieve evidence, compare candidates, and execute actions. ### From Similarity to Reciprocal Suitability The development of AI recruitment systems has shifted from similarity-based matching to reciprocal suitability. This transition is essential in ensuring that candidates are not only matched to job requirements but also evaluated based on their potential to contribute to the organization. ### From a Model to a Compound Workflow The use of AI in recruitment has moved from relying on a single model to a compound workflow that involves multiple components. This includes document understanding, retrieval, ranking, assessment, interviewing, sourcing, and human handoff. ### From Offline Prediction to Evidence- and Productivity-Aligned Evaluation The evaluation of AI recruitment systems has transitioned from offline prediction to evidence- and productivity-aligned evaluation. This means that the evaluation process now considers the evidence retrieved, the ranking of candidates, and the productivity of the hiring process. ### Field-, Pair-, List-, Case-, Trajectory-, and Outcome-Level Evidence AI recruitment systems generate various types of evidence, including field-, pair-, list-, case-, trajectory-, and outcome-level evidence. These types of evidence are essential in evaluating the effectiveness of AI recruitment systems. ### Persistent Gaps in AI Recruitment Systems Despite the advancements in AI recruitment systems, persistent gaps remain. These gaps include: * Behavioral labels confounding exposure, preference, and qualification

• Private and synthetic data limiting external validity
• Final-output scores concealing pipeline failures
• Lack of direct evaluation of privacy and no joint evaluation of utility, fairness, privacy, and security ### Staged Mapping from Evaluation Evidence to Strongest Defensible Claim To address these gaps, a staged mapping from evaluation evidence to the strongest defensible claim is proposed. This mapping involves: 1. Evidence Collection: Collecting evidence from various sources, including field-, pair-, list-, case-, trajectory-, and outcome-level evidence.

1. Evaluation: Evaluating the collected evidence to determine its validity and reliability.
2. Claim Formation: Forming a claim based on the evaluated evidence, considering factors such as utility, fairness, privacy, and security.
3. Defensibility: Assessing the defensibility of the claim, considering factors such as the strength of the evidence and the transparency of the evaluation process. ### Agenda for Reciprocal, Evidence-Grounded, Temporally Controlled, Selective, and Auditable Systems To address the persistent gaps in AI recruitment systems, an agenda for reciprocal, evidence-grounded, temporally controlled, selective, and auditable systems is proposed. This agenda involves: 1. Reciprocal Evaluation: Evaluating AI recruitment systems in a reciprocal manner, considering the feedback from both the system and the stakeholders.
4. Evidence-Grounded Evaluation: Evaluating AI recruitment systems based on evidence, rather than relying on assumptions or intuition.
5. Temporally Controlled Evaluation: Evaluating AI recruitment systems at different points in time, considering the dynamic nature of the hiring process.
6. Selective Evaluation: Evaluating AI recruitment systems selectively, focusing on the most critical aspects of the hiring process.
7. Auditable Evaluation: Evaluating AI recruitment systems in an auditable manner, ensuring transparency and accountability in the evaluation process. AI recruitment systems have undergone significant changes in recent years, shifting from similarity-based matching to reciprocal suitability, from a model to a compound workflow, and from offline prediction to evidence- and productivity-aligned evaluation. However, persistent gaps remain, including behavioral labels confounding exposure, preference, and qualification, private and synthetic data limiting external validity, final-output scores concealing pipeline failures, and lack of direct evaluation of privacy and no joint evaluation of utility, fairness, privacy, and security. To address these gaps, a staged mapping from evaluation evidence to the strongest defensible claim is proposed, together with an agenda for reciprocal, evidence-grounded, temporally controlled, selective, and auditable systems.

Empirical Compute Sizing: Parameter Density vs. Inference Latency

Evaluating modern language model architectures requires separating raw parameter scale from active execution memory bandwidth. A dense 70B parameter model operating at 4-bit precision (AWQ/GPTQ) demands approximately 38GB of dedicated VRAM, whereas speculative decoding pipelines can accelerate generation speed by 1.8x to 2.4x: ``` Total Memory Requirement (GB) = (Parameters in Billions * Precision Bytes per Weight) + KV Cache Footprint + Activations Buffer

| Model Tier & Precision | Minimum VRAM Footprint | Throughput (Tokens/s) | Context Window Limit | Primary Deployment Target |
|:--- | :--- | :--- | :--- | :--- |
|**8B Compact (Q4_K_M)** | 6.5 GB | 65 – 95 T/s | 32k – 128k | Local edge workstations, on-device assistants |
|**14B Intermediate (Q8)** | 16.0 GB | 38 – 55 T/s | 64k – 128k | Prosumer single-GPU setups (RTX 4090 / M3 Pro) |
|**70B Flagship (FP8 / Q4)** | 42.0 GB – 48.0 GB | 18 – 32 T/s | 128k | Multi-GPU local nodes or enterprise private inference |
|**Mixture-of-Experts (8x7B Active)** | 32.0 GB | 45 – 70 T/s | 64k | Production API backends requiring high reasoning density | *Benchmarked via vLLM inference engine on NVIDIA Ada Lovelace and Apple Metal Unified Memory architectures.* ### Scenario Decision Framework: Local Execution vs. Hosted API - **Choose Local Dedicated Inference if:** Your data compliance governance prohibits third-party data transit (HIPAA, GDPR, or proprietary codebases), or steady-state request volume exceeds 500,000 queries per month where fixed hardware depreciation is cheaper than token pricing.
- **Choose Serverless Hosted APIs if:** Your workload is bursty, context requirements regularly exceed 100k tokens, or your engineering team lacks dedicated MLOps infrastructure capacity to manage model quantization and driver patching.