정렬: 최신순
30건의 논문
We present a retrieval-augmented translation pipeline tailored to low-resource technical Korean corpora. The method improves terminology consistency and reduces factual drift in domain-heavy documents.
PDF 열기This paper introduces sparse routing across expert groups for efficient long-context reasoning. We show lower inference cost while preserving performance on multi-document QA benchmarks.
PDF 열기We build a benchmark for evaluating trust calibration when multiple AI agents collaborate on scientific writing tasks. Results show confidence estimates remain poorly aligned under disagreement.
PDF 열기GraphDiffusion combines graph neural operators with diffusion priors to improve uncertainty estimates in molecular property prediction. The approach outperforms deterministic baselines on OOD compounds.
PDF 열기We propose an adaptive curriculum that reorders multimodal training data according to evolving confidence and disagreement signals. The strategy significantly stabilizes vision-language alignment with noisy web...
PDF 열기This work studies request-aware scheduling for heterogeneous edge TPU clusters. By combining queueing features and model cost profiles, we reduce p95 latency across bursty workloads.
PDF 열기OpenDataset-30K provides a cleaned and documented corpus for paper-level and section-level summarization. We detail filtering heuristics and release train-dev-test splits with quality annotations.
PDF 열기We introduce a layout-aware transformer architecture for parsing noisy PDF documents and recovering semantic structure. The model improves extraction quality on math-heavy articles and scanned proceedings.
PDF 열기Counterfactual regularization is used to constrain temporal causal graph discovery from observational sequences. We report stronger edge precision in partially observed healthcare data.
PDF 열기This paper proposes a privacy-preserving Bayesian optimization framework for site-distributed clinical trial tuning. The federated method improves sample efficiency while protecting participant-level data.
PDF 열기We revisit multilingual embedding evaluation with strict reproducibility controls, including tokenization checks and seed logging. Findings highlight large variance hidden by single-run reporting.
PDF 열기Operator caching is integrated into neural PDE pipelines to avoid repeated expensive transformations. We demonstrate notable speedups for inverse problems in fluid simulation.
PDF 열기This survey organizes practical hallucination detection methods by deployment constraints, including latency and annotation budget. We compare confidence-based, retrieval-based, and verifier-based pipelines.
PDF 열기We adapt consistency distillation to radar nowcasting and evaluate robustness to sensor drift. The model provides competitive forecasting quality at a fraction of diffusion sampling time.
PDF 열기A contrastive teacher-student objective is proposed for creating compact speech encoders under strict memory budgets. The method retains speaker and phoneme discrimination on mobile devices.
PDF 열기We design a hybrid ranking stack that combines lexical match, citation graph signals, and semantic embeddings for academic search. Offline experiments show better relevance at top-k for broad and narrow queries...
PDF 열기The paper introduces an uncertainty decomposition strategy for selecting pathology slides during active learning. Label efficiency improves while reducing false confidence under class imbalance.
PDF 열기Constraint-guided decoding applies static checks during generation to reduce invalid program outputs. We observe improved pass@k without expensive post-hoc reranking.
PDF 열기We study multi-objective reinforcement learning for power grid dispatch, balancing cost, carbon emissions, and reliability constraints. A Pareto-conditioned policy yields stable control across demand regimes.
PDF 열기This analysis characterizes token mixing behavior across depth using geometric probes on intermediate representations. We find distinct regimes that correlate with downstream transfer stability.
PDF 열기We propose governance guidelines and risk controls for synthetic datasets used in public-sector AI systems. The framework covers audit trails, disclosure levels, and bias escalation criteria.
PDF 열기A symbolic-neural retrieval index is introduced for large theorem libraries. The approach improves retrieval recall while preserving precise symbolic constraints during proof search.
PDF 열기We automate scientific taxonomy expansion by mining hypernym candidates from abstracts and refining them with contrastive validation. Human reviewers confirm substantial coverage gains.
PDF 열기This paper applies latent world models to long-horizon warehouse planning under dynamic obstacles. The proposed planner improves throughput and safety margins in multi-robot simulations.
PDF 열기We study continual domain adaptation with no replay buffers by combining parameter isolation and consistency regularization. Results show reduced forgetting on streaming benchmarks.
PDF 열기Dynamic precision windows adjust numeric precision throughout optimization to reduce energy use. The method preserves final quality while cutting accelerator power consumption in large runs.
PDF 열기Prompt tuning is evaluated for citation intent classification on a large multilingual corpus. We find robust zero-shot transfer for minority disciplines with lightweight adaptation.
PDF 열기We design a verification pipeline where annotators review high-impact scientific claim extractions suggested by language models. Targeted review significantly increases precision at minimal extra cost.
PDF 열기This empirical study compares feature families used in academic search re-ranking, including recency, citation velocity, and semantic agreement. We provide practical guidance for production tuning.
PDF 열기We propose a Bayesian guardrail framework for model updates in production systems, combining posterior risk thresholds with staged rollout triggers. The method reduces harmful regressions during fast iteration.
PDF 열기