biometric best-of-n sampling

To address low matching accuracy and inefficiency in multi-position concurrent recruitment, this paper proposes an end-to-end intelligent recommendation method that integrates large language model (LLM)-driven semantic understanding with graph-structured similarity computation. We innovatively construct a dual-perspective, multimodal embedding representation—jointly modeling candidates and positions—to unify resumes and job descriptions into a shared semantic space; further, we employ graph neural networks to capture cross-entity relational dependencies, enabling dynamic and interpretable multi-vacancy collaborative matching. Our approach is the first to deeply fuse LLM-powered fine-grained semantic parsing with graph-structural similarity measurement. Evaluated on a real-world recruitment dataset, it achieves an average 32.7% improvement in matching precision and recall, while reducing initial screening time by over 60%.

Early-stage candidate screening suffers from low efficiency due to the need to integrate heterogeneous information—resumes, interview videos, coding assignments, and public online data. This paper proposes a risk-aware, modular multi-agent system orchestrated by constraint-augmented large language models (LLMs), encompassing multimodal parsing (PDF/video), structured profile construction, knowledge-graph–driven public evidence verification, dual-dimension (technical/cultural) scoring with explicit risk penalization, and a human-in-the-loop review interface. Key contributions include: (i) the first explainable and traceable risk-aware scoring framework; (ii) a novel efficiency metric—“time per qualified candidate”; and (iii) component-level attribution and low-variance decision-making. Evaluated on real screening of 64 Python backend engineers, the system reduced time per qualified candidate from 3.33 to 1.70 hours, maintaining baseline precision and recall, while preserving final hiring authority exclusively with human recruiters.

This study addresses the inefficiency and training instability of large language models (LLMs) in multi-token candidate selection tasks—such as preference ranking, multiple-choice question answering, and clinical decision-making—arising from reliance on autoregressive decoding. We propose a decoding-free direct candidate selection paradigm. Through systematic evaluation of over ten logits-based probability estimation methods—including logsumexp, softmax truncation, and token-level weighted aggregation—we identify critical principles governing logits reweighting, normalization, and architecture-aware adaptation, and formulate task- and model-informed selection strategies. Evaluated across five multiple-choice QA benchmarks and four large-scale clinical decision benchmarks (with candidate sets up to >100K), our approach demonstrates consistent accuracy gains and improved gradient differentiability across LLaMA, Qwen, and Phi families. This work provides the first comprehensive, empirically grounded guide for efficient, trainable non-autoregressive generation in candidate selection.

Existing data selection methods introduce bias in heterogeneous data pools due to their reliance on reference trajectories misaligned with downstream task objectives. This work proposes a zeroth-order data selection approach that leverages a compact warm-up trajectory induced by the target validation set as an aligned and decoupled reference path, scoring candidate samples via normalized loss reduction at the trajectory endpoint. The method requires neither gradients nor Hessian approximations and enables reuse of the warm-up trajectory across diverse data pools, substantially reducing computational and storage overhead. Empirical evaluations across logistic regression, vision, and instruction fine-tuning tasks demonstrate performance comparable to strong dynamic attribution baselines while significantly decreasing warm-up time and storage costs.

This study investigates fairness degradation in AI-driven recruitment under dual bias—external discrimination and internal self-censorship—and evaluates the mitigating effect of resume anonymization. We propose the first systematic framework for synthesizing realistic, reproducible datasets that jointly encode both bias types. Using these data, we quantitatively assess fairness decay across five standard classifiers—logistic regression, decision trees, random forests, SVM, and XGBoost—under objective evaluation criteria. Results show an average 37% drop in top-candidate recall across all models on biased data; anonymization yields only marginal fairness improvement and fails to eliminate systemic bias rooted in historical inequities. Our core contributions are: (1) a transparent, reproducible dual-bias data generation framework; and (2) empirical evidence demonstrating the fundamental limitations of anonymization in algorithmic hiring, establishing a new benchmark for bias溯源 (origin tracing) and mitigation.

This work addresses key challenges in single-channel speech separation—namely, source permutation ambiguity, unstable sampling in generative models, and difficulties in aligning long audio segments across chunks—by proposing an ordered two-speaker separation method based on conditional flow matching. The approach fixes the source ordering during training by freezing the speaker encoder and introduces a biometric-guided Best-of-N sampling mechanism together with a chunk alignment strategy during inference to ensure stable and consistently ordered outputs. Built upon a Transformer U-Net architecture, the method achieves competitive performance on Libri2Mix in terms of SI-SDR, PESQ, and ESTOI metrics, while substantially reducing error rates in downstream tasks such as automatic speech recognition (cpWER) and speaker verification (EER).

This work addresses the challenge of robust speaker diarization (SD) in multilingual and code-switched scenarios, where low-resource conditions severely degrade SD performance. We propose a language-agnostic end-to-end SD–ASR–NMT joint pipeline. To enhance SD robustness, we introduce a novel multi-kernel consensus spectral clustering framework that integrates lightweight voice activity detection (VAD), fine-tuned ECAPA-TDNN speaker embeddings, multilingual ASR (Whisper/XLS-R), and neural machine translation, augmented by language identification and rule-based post-processing. To our knowledge, this is the first work to empirically validate the engineering feasibility of full-chain co-optimization of SD–ASR–NMT on real-world multilingual mixed audio. Evaluated on the NCIIPC challenge training set, our system reduces diarization error rate (DER) by 32% over baseline methods, supports Hindi, Tamil, English, and their code-switched combinations, achieves an end-to-end real-time factor <1.8×, and significantly improves cross-lingual generalization and system robustness.

This work addresses the challenge in vision-language tasks where full responses are difficult to verify, yet only partial information—such as regions, relations, or numerical values—is verifiable, rendering traditional Best-of-N (BoN) selection ineffective. The authors propose the Best-of-Evidence (BoE) framework, which formalizes candidate selection under partial verification for the first time. BoE models reusable claims via a signed candidate-factor graph and dynamically selects the most decision-influential evidence within a limited query budget. Theoretically, shared factor queries reduce query complexity from Θ(K) to O(log K), with BoE naturally degenerating to BoN under zero budget. Experiments demonstrate that BoE significantly improves selection performance across four medical VQA datasets, rectifies failure cases of BoN, and reveals that channel quality and candidate generation capacity critically constrain achievable performance.

This work addresses the challenge of model collapse and rapidly decaying generation diversity—following a power-law decay—under low-resource and data-isolated settings, where sample selection based on local biased reference distributions exacerbates instability. The study is the first to uncover this degradation mechanism and proposes a collaborative proxy reference construction method that requires no sharing of raw data. By aligning distributions across isolated data silos via Wasserstein distance, the approach constructs a globally consistent reference signal to guide recursive training of synthetic data. Theoretical analysis and distributed experiments demonstrate that, compared to local reference strategies, the proposed method effectively mitigates diversity degradation and substantially enhances model stability and generalization capability.