Tag
A Tsinghua/Delft paper presents a fast empirical-Bayes Beta-Bernoulli predictor for sparse AI benchmark evaluation, achieving a 24.6% Brier score reduction over a constant baseline with a 7.6 KiB program and microsecond-scale predictions.
DACE introduces a diversity-driven adversarial co-evolution framework for LLM safety alignment that jointly addresses attacker strategy collapse and defender adversarial forgetting, using a two-dimensional strategy space with normalized marginal coverage gain rewards and a Bayesian adversarial replay pool with Thompson sampling. Experiments show improved robustness to out-of-distribution attacks while preserving general capabilities.
A tutorial paper presenting a clear workflow for applying neural simulation-based inference (neural SBI) to computational cognitive models, bypassing intractable likelihood computations and offering practical use cases and recommendations to reduce computational burden for cognitive scientists.
This paper proposes a Bayesian framework for interactive program induction where a language model serves as a prior over programs, deriving a greedy query-selection algorithm that provably converges after O((1/γ) log 1/p⁰(f*)) informative oracle queries. On the CodeARC benchmark, it improves accuracy over LM-only agents by up to 22 percentage points while using up to 10.5 fewer input-output examples.
CONGA is a compiler that detects Gaussian latent blocks in structured Bayesian models and builds affine transport maps (Brenier and Knothe-Rosenblatt) to speed up posterior inference, improving NUTS minimum ESS per second by 12.5x on the Stock-Watson inflation model and restoring convergence on a 6760-dimensional electricity load model.
提出了一种基于条件归一化流的摊销变分推断方法,构建加权似然的局部后验来处理模型校准中的模型偏差与上下文不确定性,并在Branin-Hoo孪生实验及单柱海洋模型的k-ε湍流闭合校准中验证了其相较于全局后验的误差降低效果。
LEGO-SCD is an online scene change detection method that learns scene-specific statistical models from automatic edits to a reference 3D Gaussian Splatting scene without manual annotations, enabling closed-form Bayesian evidence accumulation with object-guided mask refinement; it achieves 67.64% mIoU on PASLCD and 75.82% on CCS3D at 17.2 FPS.
该论文提出层级分数后验恢复方法(HSPR),通过结合历史响应模式与自适应任务选择,仅用小子集即可高效估计软件代理在完整基准测试上的分数,在SWE-bench系列任务库上取得最低的平均与最大绝对误差。
该论文提出一个两参数贝叶斯线索组合模型,解释视觉-语言模型在文本与图像证据冲突时如何整合信息,并能准确预测模型在未见过的输入退化组合下的偏好,统一了此前关于文本主导和视觉主导的矛盾报告。
This paper presents a family of conditional generative models over phylogenetic trees trained with a discrete diffusion framework, enabling simulation-based inference under complex evolutionary models. The approach recovers MCMC posteriors more accurately than prior methods like Phyloformer and outperforms the conventional method IQ-Tree on realistic data.
该论文提出将序贯贝叶斯实验设计(BED)通过信息增益奖励训练摊销进 LLM 的推理过程中,使 LLM 学会自适应信息收集,并在 Twenty Questions 与临床对话基准 Medi-Q 上超越基线且泛化良好。
This paper proposes robust, adaptive Langevin Monte Carlo methods (R-LMC, A-LMC, RA-LMC) that handle generalized smoothness and state-dependent heavy-tailed noise, establishing Fisher information convergence guarantees without dissipativity assumptions, and demonstrates effective Bayesian fine-tuning of LLMs and vision-language models.
本文提出 OSIS(one-step implicit sampler),利用短时间 MCMC 动力学为隐式建模分布定义局部目标,实现单步摊销采样。作者分析了收敛性质,并表明该方法在采样效率上优于多步摊销采样器和非摊销 MCMC 基线。
本文提出"空间感应头"(spatial induction heads),一种两层 gather-and-match 电路,解释 Transformer 如何在多维细胞自动机的序列化数据中进行上下文学习。研究表明,该电路可逼近贝叶斯后验并在确定性规则上实现精确预测,训练后的两层 Transformer 能泛化到未见过的规则。
This paper introduces the Bayesian Mirror Architecture (BMA), a self-referential generative framework where circular self-recursion and hybrid event–self binding make consciousness an architectural property rather than an objective, with belief dynamics analyzed via optimal-transport geometry on probability measures. It defines a Causal Learning Regime as a diagnostic of learnable structure and introduces self-manifolds, Wasserstein ε-necks, and basin selection interpreted as internally determined choice.
This paper shows that Large Language Bayes (LLB), which averages posteriors from language-model-generated probabilistic programs using weights from an exponentiated evidence bound, is not reparameterisation-invariant. Equivalent program writings can receive substantially different weights (up to 31.9× discrepancy that can even reverse sign), inverting Bayes factors and inducing error in model posteriors controlled by the spread of bound shortfalls via a sharp Hilbert-distance bound.
This paper adds an informed adversary to a forced-choice signalling game, showing the optimal signal shifts from posterior-maximizing to margin-maximizing as persuasion budget grows, and that the robust optimum coincides with salience on all 108 confirmatory items — a structural limit on what evaluations of adversary-awareness can learn.
BaRe-Mem introduces a Bayesian online reliability memory for multi-agent consultation, enhancing robustness against misleading advisor information and guiding the choice between consultation and autonomous reasoning.
FB-GDM introduces a fully-Bayesian guided diffusion method for high-dimensional linear inverse problems, eliminating per-task hyperparameter tuning and demonstrating robust performance improvements over existing techniques.
The paper introduces Odds-Ratio Thompson Sampling (OR-TS), a method for batched multi-armed bandits that uses contrast-based updates to handle varying common levels, showing improved regret over absolute-rate memory in simulations and real experiments.