Tag
Georgetown linguistics researchers propose PropAlign, a multiple-summary-based method for graded proposition salience evaluation, and release PropAlignBank, a dataset with 138K alignments of 35K+ propositions. They show that summary-derived salience metrics are more consistent than direct model judgments, and that model ensembles with linguistic features approach human annotation quality.
提出 Weighted Ensemble (WE)，一个无需训练的框架，通过在生成过程中自适应重新分配采样预算来提升语言模型中稀有事件概率估计的效率与可靠性，并在安全相关任务上验证了其有效性。
MARGIN提出了一种运行时置信度校准方法，无需重训练或预留校准集即可从观测结果中学习模型特定的置信度修正，用于异构基础模型的多智能体协调决策，在代码生成和问答任务上相比未校准的置信度加权提升了答案选择准确率。
MITE reformulates biomedical NER as a structure-to-structure generation task by representing instructions and entity outputs in multiple programming-language formats (Python, C++, Java), using entity-level voting at inference to improve robustness. Experiments on six BioNER datasets show it outperforms BERT- and LLM-based baselines with strong cross-dataset generalization.
Researchers from KAIST, Kookmin University, and NYU propose a proper scoring rule-based diffusion framework for probabilistic weather forecasting that adds auxiliary conditional denoising tasks to CRPS-trained models, enabling improved forecast accuracy and calibration at longer horizons while keeping single-forward-pass ensemble generation.
A new paper introduces CAGE, a framework combining generative models, adversarial discrimination, and conformal prediction to weight time series ensemble forecasts, showing improved robustness against outliers and noise on milk collection and monkeypox datasets.
The paper proposes EnJoi, an ensemble joint score filter for generative data assimilation that learns the joint distribution of past and future states, enabling a modified En4DVar approach to balance confidence in the current state versus new observations. Experiments on fluid and traffic flow simulations show improved reconstruction, especially with sparse, non-homogeneous observations.
MIT researchers argue that algorithmic monoculture in AI may not always be harmful; its negative effects depend on details like domain and algorithm accuracy, and ensemble methods can mitigate issues such as systematic exclusion.
CUSP is a training-free framework for quantifying collective uncertainty in multi-agent multimodal systems using semantic opinion pooling, improving reliability and accuracy over baseline methods.
This paper introduces a probabilistic approach to end-to-end AI weather forecasting by enhancing the Aardvark Weather model with learned observation noise and Monte Carlo dropout, improving mean forecasts by 4.2% on average and separating aleatoric and epistemic uncertainty.
Introduces an efficient Bayesian deep ensemble method for predictive regression that combines low-dimensional ensemble representation, closed-form Bayesian aggregation, and independent ensemble training to achieve calibrated uncertainty estimates with computational efficiency.
提出一种基于AdaBoost的文本提示提升（TPB）框架，通过顺序集成弱文本提示分类器来改进视觉-语言模型的少样本分类性能，并在跨模型迁移中保持增益。
This paper investigates how different inference-time deployment rules (rollout strategies) impact multi-horizon volatility forecasting. It shows that non-default rollout rules often improve performance and that validation-based deployment policies provide a low-cost improvement over standard MIMO deployment, emphasizing the importance of deployment policy alongside model architecture.
This paper identifies a fundamental constraint on multi-model LLM systems: accuracy is capped by the rate at which all models fail on the same query. Across 67 frontier models, the all-wrong rate is significantly underestimated by common metrics, limiting gains from voting, routing, and ensemble strategies.
AgentCodec is a source-available library unifying 28 LLM reliability techniques (retries, ensembling, generator/critic refinement, etc.) under a single OpenAI-compatible API, with adaptive routers that can reduce inference costs by ~56% at matched quality. It adopts a communication-theory framing and supports drop-in replacement for OpenAI, Anthropic, and Ollama clients.
This paper introduces CAGE, a counterfactual graph-based method for calibrating multi-agent LLM systems, evaluating on benchmarks like TriviaQA and MMLU-Pro across various communication topologies. The method outperforms existing post-hoc and LLM-elicited calibration approaches.
This paper reveals a fundamental vulnerability in LLM watermarking: when users have access to multiple models, averaging their output distributions cancels watermark perturbations, enabling detection evasion. The authors propose WASH and demonstrate empirically that averaging 3-5 models suppresses detection z-scores below thresholds while improving text quality.
This paper presents a comparative evaluation of classical, ensemble, and neural machine learning approaches for predicting financial distress under severe class imbalance, using SMOTE for oversampling and SHAP for interpretability.
This paper investigates disagreement-based drift detection in ensembles of incremental decision trees, finding that while effective in neural networks, the method underperforms loss-based detectors for tree ensembles due to limited model plasticity.
OpenAI presents a novel exploration strategy for deep reinforcement learning using ensembles of Q-functions with upper-confidence bounds (UCB), demonstrating significant performance improvements on the Atari benchmark.