Develops PromptShift-CRC for drift-aware conformal risk control in foundation models under prompt and domain shift.
problem Fixed calibration risk in foundation models due to prompt and domain shift.
method Embeds prompts and responses, measures drift, gives more weight to recent examples, and updates risk online.
result Develops method to control risk up to terms for distribution mismatch and weighted quantile uncertainty.
Emergent misalignment is influenced by training dynamics, model priors, and data.
problem Emergent misalignment in models
method Exploring training dynamics, model priors, and data
result Activation deltas before and after narrow fine-tuning correlate with their similarities when measured with the last prompt-token activations.
DS-CP improves reliability of uncertainty quantification for large language models under domain shift.
problem Overconfident and factually incorrect outputs (hallucinations) from large language models.
method Adapts conformal prediction to large language models under domain shift by reweighting calibration samples.
result DS-CP delivers more reliable coverage than standard conformal prediction, especially under substantial distribution shifts.
Transformers learn linear models in-context without updates.
problem Understanding how transformers mimic linear models in-context.
method Gradient flow on linear regression tasks with random initialization.
result Transformers achieve prediction error competitive with best linear predictors.
MIP framework improves urban flow prediction by adapting to distribution shifts.
problem Distribution shifts in urban flow data make prediction models unreliable.
method Memory-enhanced Invariant Prompt learning with learnable memory bank.
result MIP ensures robust predictions by focusing on invariant features.
Improved image synthesis with user scribbles and text prompts.
problem Inadequate details in generated images due to domain shift.
method Optimization problem formulation and cross-attention for control.
result Significant improvement in user satisfaction (85.32% higher).
Test-time training adapts a pretrained model to each prompt via parameter updates, improving accuracy under pretraining-to-test distribution shifts.
problem Improving accuracy of pretrained models under distribution shifts.
method Explaining TTT behavior through a decision-theoretic lens.
result TTT reduces prediction error when updates are spectrally matched to the prompt's signal-to-noise ratio and aligned with query-relevant eigen-directions.
M-FISHER detects and adapts to streaming data shifts with statistical validity and stability.
problem Detecting and adapting to distributional shifts in streaming data.
method Constructs an exponential martingale from non-conformity scores and applies Ville's inequality for detection. Fisher-preconditioned updates for adaptation.
result Establishes M-FISHER as a principled approach for robust, anytime-valid detection and geometrically stable adaptation.
Gradient-based framework for optimizing text prompts in diffusion models.
problem Efficiently optimizing prompts in text-to-image diffusion models with large domain space and non-differentiable embeddings.
method Formulated as discrete optimization over language space, designed compact subspaces, and introduced shortcut text gradient.
result Empirically discovered prompts that enhance or destroy image faithfulness.
Theory for RLHF generalization under reward shift and clipped KL.
problem Theoretical understanding of RLHF generalization, especially with reward shift and clipped KL.
method Developed generalization theory for RLHF, accounting for reward shift and clipped KL.
result Presented generalization bounds for RLHF, suggesting generalization error from sampling, reward shift, and KL clipping.
eva method improves RL models by adaptively creating prompts, boosting performance.
problem Fixed prompt distribution limits scalability of RL models post-training.
method Evolving Alignment via Asymmetric Self-Play (eva) approach.
result Significant performance boost on benchmarks, e.g. 51.6% to 60.1% win-rate.
Sequence probability predicts correctness in LLMs, but not for repeated prompts
problem Predicting correctness in large language models
method Quantifying sequence probability and correctness across different levels
result Higher sequence probability often predicts correctness across prompt-answer pairs
Unified Bayesian model explains in-context learning and activation steering in LLMs.
problem Understanding and controlling the behavior of large language models (LLMs) through prompts and activations.
method Developed a Bayesian model to explain and predict the effects of in-context learning and activation steering.
result Unified model predicts distinct phases and sudden shifts in LLM behavior, explaining prior empirical phenomena.
A new model designs molecules with desired properties.
problem Finding molecules with optimal chemical or biological properties.
method A latent prompt Transformer model with three components: latent vector, molecule generation, and property prediction.
result The model achieves state-of-the-art performance on molecule design tasks.
New dataset for industrial machine malfunction detection with domain shifts.
problem Challenges in detecting anomalies due to domain shifts in industrial sounds.
method Created a dataset with domain shifts for five types of industrial machines.
result Significant performance differences between source and target domains.
This work introduces RISE to explain LLMs more reliably by distinguishing essential context.
problem Identifying which context elements influence LLM outputs reliably.
method RISE (Redundancy-Insensitive Scoring of Explanation) method.
result RISE provides more robust explanations than traditional methods.
DCASE 2022 Task 2 tackles domain shifts in ASD for machine condition monitoring.
problem Domain shifts change acoustic characteristics, affecting ASD performance.
method Domain generalization techniques to detect anomalies across unknown domains.
result Two types of domain generalization techniques were identified and analyzed.
New method reduces label and data shifts between domains using optimal transport.
problem Label shift between source and target domains in domain adaptation.
method Developed theory and LDROT method to mitigate label and data shifts.
result Theoretical and experimental validation of LDROT's effectiveness.
CoDAG combines domain adaptation and generalization for unsupervised continual domain shift learning.
problem Acquiring knowledge in unsupervised continual domain shift learning.
method Complementary Domain Adaptation and Generalization (CoDAG) framework.
result CoDAG outperforms state-of-the-art models in all datasets and evaluation metrics.
HySRL improves RL sample efficiency with shifted-dynamics data.
problem Leveraging historical data with shifted dynamics to improve sample efficiency in RL.
method HySRL, a hybrid transfer RL algorithm that uses prior information on dynamics shift to achieve better sample complexity.
result HySRL achieves problem-dependent sample complexity and outperforms pure online RL.
Generative model learns to create coherent slates from prompts.
problem Generating coherent slates from natural language prompts.
method DMSG, a generative framework based on diffusion models.
result DMSG outperforms baselines in relevance and diversity.
RLSbench benchmarks domain adaptation under label proportion shifts, revealing widespread failures and proposing a two-step meta-algorithm.
problem Domain adaptation under label proportion shifts is poorly understood and inconsistent across methods.
method RLSbench introduces a large-scale benchmark with 500 distribution shift pairs. It proposes a two-step meta-algorithm to improve domain adaptation methods under label proportion shifts.
result The two-step meta-algorithm improves domain adaptation methods by 2-10% accuracy points under large label proportion shifts.
New method learns models to adapt to domain shifts at test time.
problem Learning models robust to distribution shifts in practical applications.
method Adaptive Risk Minimization (ARM) framework.
result Performance gains of 1-4% on image classification problems.
Discrepancy between training and testing domains is a fundamental problem in the generalization of machine learning techniques. Recently, several approaches have been proposed to learn domain invariant feature representations through adversarial deep learning. However, label shift, where the percentage of data in each …
Proposes a new method for handling domain shift in samples with biases in both covariates and labels.
problem Domain shift in samples with biases in both covariates and labels.
method Factorizable Joint Shift (FJS) and Joint Importance Aligning (JIA).
result Our method can handle co-existence of sampling bias in covariates and labels.
Paper tackles robust classification under class-dependent domain shift.
problem Class-dependent domain shift in machine learning.
method Defined a simple optimization problem with an information theoretic constraint and solved it using neural networks.
result Demonstrated that the proposed method can learn robust classifiers that generalize well to unseen domains.
Proposes SGShift to identify shifted features causing model performance degradation under concept shift.
problem Concept shift leading to miscalibration in ML models across domains.
method SGShift method for identifying sparse set of shifted features using feature selection and statistical tools.
result SGShift identifies shifted features more accurately than baseline methods, requires few samples in the shifted domain, and is robust to complex cases.
Proposes a method to adapt to new classes in a domain shift.
problem Learning new classes in a domain shift without labeled supervision.
method Inspired by prototypical networks, the method classifies target samples into shared and novel classes.
result Superior performance compared to DA and CI methods in the CIDA paradigm.
Detects domain shifts in datasets using interpretable feature subspaces.
problem Detecting subtle differences in dataset probability distributions.
method Localised density anomaly detection in high-dimensional feature spaces.
result Extracts interpretable feature subspaces for domain shifts.
Adapts to shifts in latent subgroup distributions without labeled target data.
problem Adapting to domain shifts when latent subgroup distributions differ.
method Uses concept and proxy variables from source domain, and unlabeled target data.
result Optimal target predictor can be identified and estimated.
EQD model improves domain-specific QA by 0.6% to 10.5%.
problem Challenges in domain-specific quantitative reasoning for LLMs.
method Two-step fine-tuning framework guided by a reward function.
result EQD outperforms state-of-the-art models and prompting strategies.
Reasoning models generate differently based on problem difficulty, not just length.
problem Understanding how reasoning models handle different problem difficulties.
method Examined hidden-state trajectories across competitive programming, mathematics, and Boolean satisfiability.
result Corrected trajectory geometry shows difficulty-dependent differences in reasoning models, with stronger effects in the code domain.
Proxy methods adapt to distribution shifts without explicitly modeling latent confounders.
problem Adapting to distribution shifts under latent variable confounding.
method Proximal causal learning, two-stage kernel estimation.
result Proxy methods outperform other methods in adapting to complex distribution shifts.
Paper develops a new method to improve model calibration under distribution shifts.
problem Challenges in uncertainty quantification with different training and test distributions.
method Develops multi-domain temperature scaling to handle distribution shifts.
result Outperforms existing methods on in-distribution and out-of-distribution test sets.
Paper tackles backwards-compatible data adaptation for confounded covariate and label shifts.
problem Adapt covariates to predict labels confounded with covariate shifts.
method Proposes confounded shift framework based on minimizing divergence between source and target conditional distributions, conditioning on confounders.
result Demonstrates approach on synthetic and real datasets, achieving backwards-compatible data adaptation.
Paper shows fairness and domain adaptation can work together.
problem Algorithmic bias and distributional shifts in ML models.
method Leveraging fairness and distribution shifts, the paper shows how domain adaptation methods can mitigate bias.
result Enforcing individual fairness can improve out-of-distribution accuracy under covariate shift.
Algorithm improves model performance on shifted concepts without retraining.
problem Improving model performance on shifted concepts with limited source data.
method Model consolidation of intermediate internal distributions after adaptation.
result Effective improvement in model performance on shifted concepts.
Proposes methods to aggregate prediction intervals for domain shift uncertainty.
problem Uncertainty quantification in distribution shifts.
method Aggregates prediction intervals for minimal width and adequate coverage.
result Effective methodologies for unsupervised domain shift under labeled source and unlabeled target.
Self-training improves gradual domain adaptation with unlabeled data.
problem Improving machine learning models' adaptability to gradually shifting data distributions.
method Proved upper bounds on self-training error, highlighted the importance of regularization and label sharpening, and demonstrated algorithmic insights.
result Self-training works well for gradual shifts, especially with small Wasserstein-infinity distance.
Estimates model performance under distribution shift using domain-invariant predictors.
problem Poor performance of models on test distributions different from training distributions.
method Uses domain-invariant predictors as a proxy for unknown target labels.
result Shows that the complexity of latent representations influences target risk.
Algorithm adapts to shifting domains with minimal label queries.
problem Adaptive learning in online machine learning systems with domain shifts.
method Adaptive algorithm balancing regret and label queries for hidden domains.
result Achieves lower regret compared to uniform and greedy queries.
Method distinguishes between failures and domain shifts in industrial data streams.
problem Confusing domain shifts with failures in industrial data.
method Modified Page-Hinkley changepoint detector and supervised domain-adaptation-based anomaly detection.
result Allows differentiation between failures and domain shifts.
Proposes a new approach to MSDA by introducing latent covariate shift to handle varying label distributions.
problem Challenges of conventional MSDA approaches in real-world settings where label distributions vary across domains.
method Introduces latent covariate shift (LCS) and a causal generative model with latent noises, latent content variable, and latent style variable.
result Identifies latent content variable up to block identifiability, enabling more nuanced label distribution recovery.
Self-training avoids spurious features in domain adaptation.
problem Domain shift with large differences between source and target domains.
method Entropy minimization on unlabeled target data, initialized with a source classifier.
result Entropy minimization avoids using spurious features in large domain shifts.
Bayesian approach quantifies uncertainty in LLM-based systems.
problem Uncertainty quantification in LLM-based systems, especially for high-stakes applications.
method Interpreting prompts as parameters in a Bayesian model, using MHLP for inference.
result Improvements in predictive accuracy and uncertainty quantification on various benchmarks.
REx tackles distributional shift by reducing risk differences across domains.
problem Tackling distributional shift when transferring machine learning systems to real-world applications.
method Risk Extrapolation (REx) assumes training domains represent test-time variations and uses extrapolated domains to minimize risk variance.
result REx reduces sensitivity to extreme distributional shifts, including causal and anti-causal inputs.
Post-hoc calibration improves uncertainty under domain shift.
problem Improving uncertainty calibration under domain shift.
method Apply perturbations to validation set before post-hoc calibration.
result Perturbation step results in better calibration under domain shift.
This work tackles robust multi-source domain adaptation under label shift.
problem Label shift and data contamination in multi-source domain adaptation.
method Domain-weighted empirical risk minimization framework with refinement procedure.
result The proposed method achieves superior performance in multi-category classification problems.