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.
Global anchor method detects language shifts and domain adaptation.
problem Detecting corpus-level language shifts and domain adaptation.
method Global anchor method for comparing word embeddings.
result Global anchor method is superior to alignment method in applicability and implementation.
Research tackles distribution shift issues in ML to improve AI reliability.
problem Distribution shift limits ML reliability and trustworthiness.
method Study three distribution shifts (perturbation, domain, modality) and investigate robustness, explainability, adaptability.
result Proposes effective solutions and fundamental insights for enhancing ML robustness, adaptability, and safety.
Transformer models can solve complex math problems with less data.
problem Solving complex symbolic mathematics problems with limited data.
method Pretrain transformer models on language translation tasks and fine-tune for symbolic math.
result Pretrained transformer models achieve comparable accuracy to state-of-the-art models with less data.
New methods adapt conformal prediction to unknown subpopulation shifts.
problem Failure of conformal prediction under unknown subpopulation shifts.
method Proposes new methods that adapt conformal prediction to unknown subpopulation shifts without explicit subpopulation labels.
result Ensures valid coverage guarantees without explicit knowledge of subpopulation structure.
We describe a mechanism by which artificial neural networks can learn rapid adaptation - the ability to adapt on the fly, with little data, to new tasks - that we call conditionally shifted neurons. We apply this mechanism in the framework of metalearning, where the aim is to replicate some of the flexibility of human …
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.
Language models perform worse with implicit reward models than explicit ones.
problem Understanding why implicit reward models generalize worse than explicit ones.
method Investigated the root cause of the generalization gap between IM-RMs and EX-RMs.
result Implicit reward models rely more on superficial token-level cues, leading to worse generalization.
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.
This work aims to create a large-scale model for critical care time series data.
problem Lack of large-scale datasets and distribution shifts in critical care time series data.
method Harmonized dataset creation and transfer learning research.
result Established a foundation for large-scale multi-variate time series models in critical care.
Transformers show better in-context learning resilience under distribution shifts than simple MLPs.
problem Understanding in-context learning under varying distribution shifts.
method Comparing transformers and set-based MLPs on linear regression tasks.
result Transformers better emulate OLS performance and exhibit better resilience to mild distribution shifts.
Optimal selective classification using likelihood ratios improves model reliability.
problem Enhancing predictive model reliability by allowing uncertain predictions.
method Neyman--Pearson lemma applied to likelihood ratios for optimal selection.
result Neyman--Pearson-informed methods outperform existing baselines under covariate shifts.
Deep neural networks have excelled on a wide range of problems, from vision to language and game playing. Neural networks very gradually incorporate information into weights as they process data, requiring very low learning rates. If the training distribution shifts, the network is slow to adapt, and when it does adapt…
Enhanced regime shifts detection using unstructured text and financial data.
problem Detecting regime shifts in financial markets is challenging due to noisy and multicollinear data.
method Combines LLM reasoning on unstructured text and statistical validation on financial time series.
result Framework achieves F1 score of 0.82, outperforming pure data-driven methods.
FADE adapts machine learning models to evolving data efficiently.
problem Sequential covariate shift in dynamic environments.
method FADE uses Fisher information geometry for robust learning under SCS.
result FADE achieves up to 19% higher accuracy under severe shifts.
Extends conformal prediction for controlling expected risk of monotone loss functions.
problem Controlling expected risk of monotone loss functions.
method Generalizes split conformal prediction with coverage guarantee, extending to distribution shift, quantile risk, multiple, adversarial, and expectations of U-statistics.
result Tight up to an O(1/n) factor, with worked examples in computer vision and natural language processing. Every day, hundreds of millions of new Tweets containing over 40 languages of ever-shifting vernacular flow through Twitter. Models that attempt to extract insight from this firehose of information must face the torrential covariate shift that is endemic to the Twitter platform. While regularly-retrained algorithms can…
Paper proposes MAG to fine-tune BERT and XLNet for multimodal sentiment analysis.
problem Fine-tuning pre-trained models for multimodal language applications is challenging.
method Integrates vision and acoustic modalities into BERT and XLNet through Multimodal Adaptation Gate (MAG).
result Significant improvement in multimodal sentiment analysis performance over previous methods.
ORCA calibrates LLMs for efficient, generalizable reasoning.
problem Miscalibration of large language models leading to inefficiencies.
method Online Reasoning Calibration (ORCA) using conformal prediction and test-time training.
result ORCA provides higher efficiency and generalization across different reasoning tasks.
Study evaluates how well question-answering models generalize to new data types.
problem Generalization of question-answering models to new data types.
method Constructed new test sets from different domains and evaluated models' performance.
result Models show significant performance drops when tested on new data types.
We propose a simple extension to the ReLU-family of activation functions that allows them to shift the mean activation across a layer towards zero. Combined with proper weight initialization, this alleviates the need for normalization layers. We explore the training of deep vanilla recurrent neural networks (RNNs) with…
Automates debiasing for large language model evaluations through Fisher random walk.
problem Rigorous and scalable evaluation of large language models.
method Semiparametric efficient estimator using Fisher random walk for weighted residual balancing.
result Efficient estimation of contextual preference scores for large language models.
Study uses LLMs for financial sentiment analysis without fine-tuning.
problem Lack of prescriptive knowledge to leverage generative models in FSA.
method Proposes a design framework with heterogeneous LLM agents based on Minsky's theory.
result Framework yields better accuracies, especially with substantial discussions.
ALMANACS benchmarks explainability methods on simulatability.
problem Evaluating the effectiveness of explainability methods for language models.
method ALMANACS is a simulatability benchmark that evaluates explainability methods on twelve safety-relevant topics.
result No explainability method outperforms the explanation-free control across all topics.
Study improves estimation of rare language model outputs.
problem Estimating probabilities of rare outputs in language models.
method Importance sampling vs. activation extrapolation for low probability estimation.
result Importance sampling outperforms activation extrapolation.
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.
Unified framework for ICL in causal and masked models.
problem Understanding ICL in masked language models and comparing it to causal models.
method Developed a statistical learning framework representing context by empirical measure and predicting using context and query.
result Upper bounds for masked and autoregressive objectives under Wasserstein-type regularity conditions.
Recent breakthroughs in computer vision and natural language processing have spurred interest in challenging multi-modal tasks such as visual question-answering and visual dialogue. For such tasks, one successful approach is to condition image-based convolutional network computation on language via Feature-wise Linear …
SAIL improves online alignment of large language models with minimal feedback.
problem Offline RLHF methods often lead to sub-optimal performance due to fixed preference datasets.
method SAIL uses bilevel optimization and a single-level first-order method to iteratively refine model alignment.
result SAIL significantly improves alignment performance on open-sourced datasets with minimal computational overhead.
The study analyzes numerical stability in large language models using mixed-precision arithmetic.
problem Numerical stability of large language models using low-precision arithmetic.
method Developed a mixed-precision analysis of transformer inference, deriving bounds for condition numbers and forward error.
result Established that numerical stability is determined by the interplay between weight magnitude and the growth of the residual stream.
LoCo-RLHF models diverse human feedback with contextual information.
problem Heterogeneous human feedback from diverse contexts and preferences.
method Low-rank contextual preference model, PRS policy.
result LoCo-RLHF achieves tighter sub-optimality gap than existing methods.
Study recommends PLM choices for minimizing calibration error in NLP tasks.
problem Minimizing calibration error in PLM-based NLP predictions.
method Compared various options for PLM encoding, size, uncertainty quantifier, and fine-tuning loss.
result Recommendations for a well-calibrated PLM-based prediction pipeline.
Statisticians contribute to LLMs for better trust and transparency.
problem Emerging statistical challenges in LLMs.
method Exploring statistical contributions to LLMs.
result Statisticians can enhance LLMs' trustworthiness and transparency.
Geostatistical learning faces unique challenges due to spatial correlation and covariate shifts.
problem Challenges in applying statistical learning to geospatial data.
method Assessing generalization error under covariate shift and spatial correlation.
result No classical learning methods are adequate for model selection in geospatial contexts.
RACER optimizes LLM-as-judge accuracy with dynamic reasoning selection.
problem Balancing reasoning accuracy with computational cost in LLM-as-judge settings.
method Formulates routing as a constrained distributionally robust optimization problem, accounting for distribution shift via KL-divergence uncertainty set.
result RACER achieves superior accuracy-cost trade-offs under distribution shift.
TRIBE model uses LLMs to simulate human trading behavior in bond markets.
problem Complexities in decentralized bond market transactions.
method Agent-based model augmented with LLMs to simulate human-like decision-making.
result Slight trade aversion in LLMs can lead to complete market collapse.
Generative model evaluates text emotion intensity, outperforming classification.
problem Limitations of discrete emotion classification in applied domains.
method Fine-tuning generative language models to output continuous emotion intensity scores.
result Generative model outperforms classification baselines and reveals generalization capabilities.
Improved crypto market forecasting using historical price reactions to tweets.
problem Challenges in inferring market impact from human sentiment labels.
method Market-derived labeling approach to assign tweet sentiment labels based on historical price trends. Fine-tuned language model with context-aware prompt-tuning.
result 89.6% accuracy on Bitcoin news events, outperforming traditional fusion models.
New method samples from LLM posterior for coherent, useful responses.
problem Hallucinations in large language models.
method Posterior sampling for conditional generation, with calibration.
result Achieves statistical guarantees with higher downstream utility.
New method certifies risks of LLM outputs, improving accuracy and reliability.
problem Uncertain and incorrect outputs from large language models.
method Information-lift certificates using PAC-Bayes bounds and skeleton design.
result Achieves 77.0% coverage at 2% risk, outperforming baselines.
Study uses LLMs to create personalized treatment plans for rare gynecological tumors.
problem Suboptimal management and poor prognosis due to low incidence and heterogeneity of rare gynecological tumors.
method Developed a digital twin system using LLMs to integrate clinical and biomarker data.
result LLM-enabled digital twins efficiently model individual patient trajectories and identify potential treatment options.
Paper proposes AXE loss for non-autoregressive machine translation, improving performance.
problem Challenges in training non-autoregressive models due to lack of autoregressive factors and cross entropy loss penalties.
method Proposes aligned cross entropy (AXE) loss function using a differentiable dynamic program for better word order alignment.
result AXE-based training improves performance on major WMT benchmarks and sets a new state of the art for non-autoregressive models.
New benchmark protocol evaluates neural network optimizers for efficiency and data shift sensitivity.
problem Benchmarking neural network optimizers with hyperparameter complexity and data shift sensitivity.
method Proposed a new evaluation protocol combining end-to-end and data-addition training efficiency, using bandit hyperparameter tuning and human study validation.
result No clear winner across all tasks, highlighting the complexity of optimizer performance.
LLMs prefer Bitcoin under crisis frames, affecting financial decisions.
problem Testing whether LLMs have built-in biases towards specific financial assets.
method Developed a three-level audit protocol to examine Bitcoin's representation and influence in LLMs.
result An identifiable internal feature in LLMs can be perturbed to move financial choices, but only within measurable limits.
Corrects bias in LLM-as-a-judge evaluations using adaptive calibration.
problem Bias in LLM evaluations due to imperfect sensitivity and specificity.
method Plug-in framework with confidence intervals accounting for test and calibration dataset uncertainties.
result LML-based evaluation yields more reliable estimates than human-only evaluation.
Graph Weighted Models (GWMs) have recently been proposed as a natural generalization of weighted automata over strings and trees to arbitrary families of labeled graphs (and hypergraphs). A GWM generically associates a labeled graph with a tensor network and computes a value by successive contractions directed by its e…
New principles needed for scaling large language models, challenging traditional regularization methods.
problem The shift from generalization to scaling in machine learning requires new guiding principles.
method Examining the effectiveness of traditional regularization methods in the scaling-centric era.
result Traditional principles of regularization may not generalize to larger scales, highlighting new phenomena like scaling law crossover.
CaTs use DAGs with transformers to enforce causal constraints, improving neural network robustness.
problem Neural networks lack inherent causal structure respect, leading to reliability issues.
method Introducing Causal Transformers (CaTs) that operate under predefined causal constraints specified by DAGs.
result CaTs improve robustness and interpretability of neural networks under causal constraints.