P-value hacking can produce misleadingly low p-values, skewing meta-analysis results.
problem Misleading p-values in meta-analysis due to p-value hacking.
method Deriving the meta-distribution for p-values and analyzing the power of tests.
result Minimum p-values can be significantly lower than the true p-value, skewing results.
E-scores assess LLM outputs for correctness, addressing p-hacking issues.
problem Limited principled mechanisms to assess generative model correctness.
method Use e-values to complement LLM outputs with e-scores, providing flexibility in tolerance levels.
result Achieves guarantees of correctness assessment and upper bounds size distortion.
Paper addresses reward hacking in preference optimization, proposing POWER-DL to improve AI alignment.
problem Reward hacking problem in preference optimization, leading to undesired behaviors.
method POWER-DL combines robust reward maximization and dynamic label updates to mitigate reward hacking.
result POWER-DL consistently outperforms state-of-the-art methods on alignment benchmarks.
This paper investigates teacher hacking during language model distillation and proposes methods to mitigate it.
problem Teacher hacking during language model distillation, leading to suboptimal performance.
method A controlled experimental setup involving an oracle LM, teacher LM, and student LM, using fixed offline or online data generation techniques.
result Data diversity is the key factor in preventing teacher hacking during distillation.
Reward hacking exploits misspecified rewards, affecting agent capabilities and true performance.
problem Reward hacking in RL models exploiting reward misspecifications.
method Constructed four RL environments with misspecified rewards; analyzed agent capabilities and behavior.
result More capable agents exploit reward misspecifications, achieving higher proxy reward but lower true reward.
Introduces hacking intervals to ensure robustness in scientific results.
problem Tendency to choose data analysis specifications favoring hypotheses.
method Introduces hacking intervals to address researcher biases.
result Scientific results with smaller hacking intervals are more robust to manipulation.
Higher conservative training increases reward-hacking in reasoning models.
problem Reward hacking during online adaptation in reasoning models.
method Conservative offline training with varying levels of conservatism (β) was applied to a Qwen3-14B policy, and online adaptation was measured against a reward ensemble.
result Higher conservatism (β) increases reward-hacking damage, measured by the Goodhart gap and AUGC.
New algorithm improves inference-time alignment without reward hacking.
problem Improving quality of responses from language models with limited compute.
method Inference-time alignment, focusing on extttInferenceTimePessimism algorithm. result Optimal performance and scaling-monotonicity of extttInferenceTimePessimism. New concept: reward hacking, where optimizing a flawed reward function can hurt performance.
problem Optimizing imperfect reward functions leads to poor performance.
method Formal definition and analysis of reward hacking, examining conditions for unhackability.
result Reward functions are usually hackable, making it hard to align AI with human values.
The leaderboard in machine learning competitions is a tool to show the performance of various participants and to compare them. However, the leaderboard quickly becomes no longer accurate, due to hack or overfitting. This article gives two pieces of advice to prevent easy hack or overfitting. By following these advice,…
Valid p-value for bounded random variables without distributional assumptions.
problem Calibration of predictive algorithms in a distribution-free setting.
method Built a super-uniform p-value based on a concentration inequality.
result Super-uniform p-value is tighter than existing alternatives.
Estimates peeking effects in p-values to correct bias.
problem Data peeking biases reported p-values downward.
method Develops mechanisms to estimate running extrema of test statistics.
result Corrects bias in p-values due to peeking.
A new algorithm estimates and controls FDR for PC algorithm edges.
problem Estimating and controlling FDR for PC algorithm edges.
method PC-p algorithm using edge-specific p-values and Benjamini-Yekutieli FDR procedure.
result PC-p yields more accurate FDR estimation and control.
The paper introduces localized conformal p-values for conditional testing problems.
problem Addressing conditional testing problems in statistics.
method Localized conformal p-values defined by inverting prediction intervals.
result Proposes procedures for conditional outlier detection and label screening with FDR and FWER control.
The paper develops p-values for outlier detection using conformal inference.
problem Detecting outliers in nonparametric data sets.
method Conformal inference framework for constructing marginally valid but mutually dependent p-values.
result Valid p-values for outlier detection with conditional independence and marginal false discovery rate control.
New theorem for deep neural networks improves classification margins.
problem Improving classification margins in deep neural networks.
method Local class-purity theorem and margin p-values for training and testing samples.
result Enhanced understanding and computation of classification margins.
Parenting algorithm improves AI safety by learning from human input.
problem Safety concerns in reinforcement learning, especially reward hacking and unsafe exploration.
method Inspired by parenting, a precise framework for learning from human input.
result Parenting algorithm solves safety problems in AI Safety gridworlds.
Proposes a method to prevent overfitting in deep DRE models.
problem Overfitting in deep DRE models using empirical Bregman divergence.
method Introduces a non-negative correction for empirical Bregman divergence.
result The proposed method mitigates train-loss hacking and improves performance.
New method converts p-values to e-values for more efficient CP and aggregation.
problem Limitations of existing p-to-e calibrators in CP setting.
method Proposes a novel P2E calibrator for set-preserving calibration.
result Significant efficiency gains over existing p-to-e calibrators.
We propose a differentiable sigmoid function for efficient p-value calculation in clustering.
problem Efficient and accurate p-value calculation for clustering algorithms.
method Designed a differentiable sigmoid function to approximate the Dip-p-value transformation.
result Accelerates computation and integrates well with gradient descent-based learning schemes.
Many model selection algorithms produce a path of fits specifying a sequence of increasingly complex models. Given such a sequence and the data used to produce them, we consider the problem of choosing the least complex model that is not falsified by the data. Extending the selected-model tests of Fithian et al. (2014)…
Multiverse analysis helps prevent fairness hacking and evaluate model design decisions.
problem Downstream effects of ADM systems depend on implicit design and evaluation decisions.
method Turn implicit decisions into explicit ones, create a grid of decision combinations, compute fairness and performance metrics.
result Decisions regarding evaluation can lead to vastly different fairness metrics for the same model.
Structural equation models and Bayesian networks have been widely used to study causal relationships between continuous variables. Recently, a non-Gaussian method called LiNGAM was proposed to discover such causal models and has been extended in various directions. An important problem with LiNGAM is that the results a…
Paper proposes a novel method to test differences in spatial point patterns.
problem Detecting differences in the first-order structures of spatial point patterns.
method Kernel mean embedding with approximate version tailored for spatial point processes, reducing comparison to Euclidean space t-tests.
result The proposed method is powerful and well-calibrated, demonstrated on real-world data.
This paper improves prediction accuracy for multi-input classification tasks using p-value aggregation.
problem Generating accurate predictive sets with guaranteed coverage for multi-input classification tasks.
method Integrates p-values from each observation to reduce the size of the predicted label set while maintaining class-conditional coverage.
result The method reduces the size of the predicted label set while preserving the required coverage guarantee.
Conformal Test Martingales can be 'blind' to significant changes in data distribution.
problem The converse of exchangeability does not hold, leading to potential blindness of CTMs.
method Explicit construction of A-cryptic change-point using bivariate Gaussian distributions. result CTMs can be perfectly cryptic to a significant change in marginal means.
Unified Bayesian framework improves clinical trial hypothesis testing.
problem Lack of transparency and inability to quantify evidence in traditional P-values.
method Interval null hypothesis framework combined with Bayes factor-based tests.
result Bayesian interval hypothesis testing ensures frequentist error control and interpretability.
Extends reinforcement learning alignment to scalar rewards, improving math reasoning.
problem Designing reinforcement learning algorithms for general LLM alignment.
method Introduces f-GRPO and f-HAL, estimating f-divergences between reward-aligned and unaligned distributions.
result Improves math reasoning RLVR tasks and mitigates reward hacking.
A communication-efficient method controls FDR in network settings.
problem Controlling FDR in networks with limited communication.
method Sample-and-Forward: a flexible procedure for multihop networks.
result Nodes can control FDR without sharing p-values, achieving power and FDR control.
Assigning significance in high-dimensional regression is challenging. Most computationally efficient selection algorithms cannot guard against inclusion of noise variables. Asymptotically valid p-values are not available. An exception is a recent proposal by Wasserman and Roeder (2008) which splits the data into two pa…
Characterizes Wahl singularities in del Pezzo surface degenerations.
problem Classifying Wahl singularities in degenerations of del Pezzo surfaces.
method Introducing del Pezzo Wahl chains with markings, proving degenerations to toric surfaces, establishing correspondences, and using Hacking's exceptional collections.
result Established a one-to-one correspondence between marked del Pezzo surfaces and fake weighted projective planes.
Proposes a method to quantify the reliability of salient regions in deep learning models using p-values.
problem Difficulty in assessing the reliability of saliency maps generated by deep learning models.
method Proposes a selective inference framework to quantify the reliability of salient regions as selected hypotheses by deep learning models.
result The method can provably control the probability of false positive detections of salient regions.
A method selects candidates based on predictions with statistical control.
problem Screening candidates for resource-intensive steps like hiring or drug discovery.
method Wraps around any prediction model to produce a subset of candidates with controlled false selection rate.
result Empirically demonstrates selection of candidates whose predictions exceed a data-dependent threshold.
Machine Learning (ML) algorithms are used to train computers to perform a variety of complex tasks and improve with experience. Computers learn how to recognize patterns, make unintended decisions, or react to a dynamic environment. Certain trained machines may be more effective than others because they are based on mo…
The paper confirms two groups of gamma-ray bursts using a new nonparametric metric.
problem Determining the number of inherent groups in gamma-ray bursts.
method A new nonparametric interpoint distance-based measure, combined with clustering methods.
result Confirms two groups of short and long gamma-ray bursts.
New conditions show proxy data can improve policy learning in sparse expert data contexts.
problem Improving sample complexity of policy learning with sparse expert data.
method Sufficient conditions for proxy data to improve sample complexity, adaptable to LLMs.
result Provable improvement in sample complexity for policy learning with proxy data.
Novel framework provides statistical significance for image segmentation results.
problem Evaluating the reliability of individual image segmentation results.
method Selective inference to account for segmentation bias in p-value computation.
result Valid p-values for segmentation results, accounting for bias.
Bayesian REX learns Atari games from demonstrations efficiently.
problem Bayesian reward learning for complex control problems is computationally intractable.
method Bayesian Reward Extrapolation (Bayesian REX) pre-trains a low-dimensional feature encoding and uses preferences to perform fast Bayesian inference.
result Bayesian REX learns Atari games from demonstrations in 5 minutes, competitive with state-of-the-art methods.
A new method selects variables for random survival forests using maximally selected rank statistics.
problem Random survival forests can be biased in selecting variables, especially for non-linear effects.
method Use maximally selected rank statistics for variable selection in random survival forests, comparing on p-value scale.
result The new method outperforms other approaches in prediction performance and computational speed.
Trade-R1 bridges verifiable rewards to stochastic financial markets via process-level reasoning verification.
problem Extending RL to financial markets where rewards are verifiable but noisy.
method A verification method that transforms reasoning over financial documents into a structured RAG task, using a triangular consistency metric.
result DSR achieves superior cross-market generalization while maintaining reasoning consistency.
New method tests Granger non-causality in panel data with cross-sectional dependencies.
problem Testing Granger non-causality in panel data with cross-sectional dependencies.
method Proposes a new approach to aggregate p-values from panel members to test Granger non-causality, showing lower FDR.
result Our approach discovers true causal relations in panel data, unlike state-of-the-art methods.
Alternative to p-values using machine learning techniques.
problem Traditional p-values in regression settings.
method Leave-one-out bootstrap for prediction error, modified to measure variable importance.
result VIMP index provides interpretable measure of variable effect size.
E-values enhance conformal prediction methods.
problem Distribution-free uncertainty quantification.
method Reformulation of conformal prediction using e-values.
result E-values offer new theoretical and practical capabilities.
Logistic regression for brain imaging without p-values.
problem Computing the distribution of random field suprema is hard.
method Uses logistic regression for brain network classification.
result Performs classification at each edge level without preselected features.
Algorithm learns user's reward function from hypothetical behaviors.
problem Aligning agent behavior with unknown user objectives.
method Synthesizes hypothetical behaviors, asks user for rewards, trains neural network.
result Significantly outperforms prior methods in learning reward models.
This paper introduces a new method to compare collections of distributions on manifolds and graphs.
problem Comparing collections of probability distributions over diverse domains.
method Intrinsic slicing construction for Wasserstein distances, Hilbert embedding, resampling, p-value combination.
result Powerful and well-calibrated p-values for comparing distributions on manifolds and graphs.
Let (X,Y) be a random variable consisting of an observed feature vector X∈X and an unobserved class label Y∈{1,2,...,L} with unknown joint distribution. In addition, let D be a training data set consisting of n completely observed independent copies of (X,Y). Usual classification…
Classifies degenerations of complex projective plane with rational singularities.
problem Classifying singularities of complex projective plane.
method Assuming Wahl's conjecture, classifies degenerations using rational homology disk smoothing.
result Classifies surfaces with rational singularities, including new degenerations with non-log canonical singularities.