Study finds phase transition in context-sensitive language model with short-range interactions.
problem Understanding phase transitions in language models with short-range interactions.
method Constructed a random language model with short-range interactions and investigated its statistical properties.
result Phase transition occurs in context-sensitive language models with constant context length.
Stable and consistent model alignment for language models without assuming human preference models.
problem Lack of statistical consistency in existing alignment methods.
method Relative density ratio optimization between preferred and mixture of preferred and non-preferred data distributions.
result Our approach achieves statistical consistency and stability, providing tighter convergence guarantees.
New model shows natural language exhibits phase transition similar to physics.
problem Understanding critical properties in natural language models.
method Created a context-sensitive random language model.
result Demonstrated a Berezinskii--Kosterlitz--Thouless phase transition.
We develop a statistical framework to benchmark and select large language models based on their risks.
problem Benchmarking and selecting large language models based on their associated risks.
method A distributional framework using first and second order stochastic dominance, linked to mean-risk models in finance.
result Formalizes a risk-aware approach for model selection, balancing risk and utility.
The paper explains how language models acquire complex skills through scaling laws and statistical analysis.
problem Understanding how language models acquire new skills as their size and training data increase.
method Statistical framework and mathematical analysis of scaling laws.
result Language models can learn complex skills efficiently due to a strong inductive bias.
Global watermark for diffusion language models decouples detection from local contexts.
problem Watermarking in diffusion language models is challenging due to joint sampling of distributions over many unresolved positions.
method Proposes a global vector-valued sketch representation to control watermarking in masked diffusion language models.
result The method decouples detection from local contexts, resulting in an order-agnostic statistic and robustness.
Theory predicts neural scaling exponents from language statistics.
problem No existing theory could quantitatively predict neural scaling exponents.
method Isolated two key statistical properties of language.
result Derives a simple formula predicting neural scaling exponents.
Improved language model for French clinical reports achieves state-of-the-art performance in medical NLP tasks.
problem Lack of specialized language models for French clinical reports.
method Adapted a general pre-trained language model (CamemBERT) to French clinical reports using a corpus of 21M reports.
result Pretrained and fine-tuned models improved F1-score by 3 percentage points on APMed task.
Improved language models by incorporating statistical discriminators.
problem Distinguishing model-generated text from real text reliably.
method Energy-Based Model framework to incorporate discriminators.
result Improves language model performance in perplexity and human evaluation.
Analyzes how BPE tokenisation affects corpus statistics and model entropy in transformer models.
problem Understanding how natural language properties relate to tokenisation schemes in transformer models.
method Analyzes Shannon entropy of corpora under Zipfian distribution, investigates BPE transformations, trains language models, and uses attention diagnostics.
result Transformer models trained on BPE-tokenised corpora increasingly agree with Zipfian predictions as BPE depth increases, indicating reduced local token dependencies.
We propose a statistical model for natural language that begins by considering language as a monoid, then representing it in complex matrices with a compatible translation invariant probability measure. We interpret the probability measure as arising via the Born rule from a translation invariant matrix product state.
We propose a new statistical model suitable for machine learning of systems with long distance correlations such as natural languages. The model is based on directed acyclic graph decorated by multi-linear tensor maps in the vertices and vector spaces in the edges, called tensor network. Such tensor networks have been …
New findings show language models can't simultaneously avoid hallucinations and capture all language richness.
problem Achieving both valid output and full language richness in language generation.
method Investigates language generation within a statistical setting, focusing on consistency and breadth.
result For most collections of candidate languages, a language model cannot simultaneously avoid hallucinations and capture all language richness.
Direct Density Ratio Optimization aligns LLMs with human preferences without assuming specific models.
problem Statistical inconsistency in aligning LLMs with human preferences.
method Direct Density Ratio Optimization (DDRO) estimates density ratio directly.
result DDRO is statistically consistent, converging to true human preferences as data grows.
Transformers encode latent distributions in text, improving performance in out-of-distribution cases.
problem What should embeddings from language models represent?
method Connecting autoregressive prediction to sufficient statistics, identifying three settings.
result Transformers encode latent generating distributions, improving performance.
Develops a statistical framework to measure uncertainty in model rankings based on human preferences.
problem Uncertainty in model rankings based on human preferences due to mismatch between human and model preferences.
method Statistical framework using pairwise comparisons by humans and models to provide rank-sets for each model.
result Rank-sets constructed using only pairwise comparisons by strong models often do not cover the true ranking of human preferences.
This research explores the effects of various training settings on a Polish to English Statistical Machine Translation system for spoken language. Various elements of the TED, Europarl, and OPUS parallel text corpora were used as the basis for training of language models, for development, tuning and testing of the tran…
Method detects if text is generated by a language model with watermarks.
problem Detecting if text is generated by a language model with watermarks.
method Randomization tests and change point detection techniques.
result Method ensures Type I and Type II error control and accurately identifies watermarked sub-strings.
Language models are at the heart of numerous works, notably in the text mining and information retrieval communities. These statistical models aim at extracting word distributions, from simple unigram models to recurrent approaches with latent variables that capture subtle dependencies in texts. However, those models a…
Deep models learn to parse complex language structures from local data patterns.
problem Understanding how deep models parse and represent language structures.
method Introduced tunable probabilistic context-free grammars and a learning algorithm inspired by deep networks.
result Data correlations across scales enable hierarchical language representations.
Analyzes word vectors and co-occurrence statistics in NLP models.
problem Understanding biases in NLP models through co-occurrence statistics.
method Developed an analytic model of statistics learned by Word2Vec and GloVe, derived the first solution to Word2Vec's algorithm, and analyzed independence in co-occurrence models.
result Demonstrated a universal property of word vectors that can reveal biases in data before they are absorbed by DL models.
The paper tackles data misappropriation in LLMs by embedding watermarks and testing for their presence.
problem Detecting data misappropriation in LLMs trained on copyrighted data.
method Embedding watermarks, formulating as hypothesis testing, developing statistical framework, constructing test statistics, determining optimal thresholds, controlling errors, establishing asymptotic optimality.
result The proposed statistical testing framework effectively detects data misappropriation in LLMs.
Framework ensures alignment between humans and machines in LLMs.
problem Human-machine misalignment in LLMs scoring mechanisms.
method Lightweight calibration framework for blackbox models.
result Provably guarantees alignment between humans and machines.
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.
LLMs struggle to generate random numbers from statistical distributions, leading to biased results in applications.
problem LLMs' inability to generate random numbers accurately from specified distributions.
method Dual-protocol design: Batch Generation and Independent Requests, benchmarking 11 models across 15 distributions.
result Sampling fidelity degrades with distributional complexity and horizon, leading to systematic biases in downstream applications.
Paper proposes a new RLHF framework for human preference learning.
problem Handling dependent online human preference outcomes with dynamic contexts.
method Two-stage algorithm with ε-greedy followed by exploitation; anti-concentration inequalities and matrix martingale concentration techniques. result Our method achieves optimal regret bound and asymptotic normality of estimators.
We propose a new statistical model for computational linguistics. Rather than trying to estimate directly the probability distribution of a random sentence of the language, we define a Markov chain on finite sets of sentences with many finite recurrent communicating classes and define our language model as the invarian…
Modern LLMs fail at authorship attribution without fine-tuning, but topic embeddings outperform them.
problem Authorship attribution of the Federalist Papers using modern LLMs.
method Examined popular LLMs, compared word/phrase embeddings, and used Bayesian analysis with topic embeddings.
result Topic embeddings trained on 'function words' outperform default LLM embeddings in authorship attribution.
This research investigates the Statistical Machine Translation approaches to translate speech in real time automatically. Such systems can be used in a pipeline with speech recognition and synthesis software in order to produce a real-time voice communication system between foreigners. We obtained three main data sets …
Method tackles uncertainty in reward models for LLMs from heterogeneous human feedback.
problem Uncertainty in reward models for LLMs from heterogeneous human feedback.
method Heterogeneous preference framework and alternating gradient descent algorithm.
result Established theoretical guarantees for estimator convergence and asymptotic distribution.
Stan is a probabilistic programming language that is popular in the statistics community, with a high-level syntax for expressing probabilistic models. Stan differs by nature from generative probabilistic programming languages like Church, Anglican, or Pyro. This paper presents a comprehensive compilation scheme to com…
The paper introduces Relative Bias to quantify LLM bias systematically.
problem Quantifying bias in LLMs is challenging due to ambiguity and rapid model emergence.
method Relative Bias framework using Embedding Transformation and LLM-as-a-Judge methodologies.
result The two scoring methods show strong alignment, providing a systematic approach.
Bidirectional attention is shown to be equivalent to a continuous bag of words model with mixture-of-experts.
problem Understanding the statistical underpinnings of bidirectional attention.
method Exploring bidirectional attention as a mixture-of-experts model and reparameterizing it.
result Bidirectional attention can be viewed as a continuous bag of words model with mixture-of-experts weights.
The paper distinguishes between conditional and marginal processes in language models and discusses conditions for usefulness.
problem The conditional nature of language models trained on observed sequences and the need for marginal text-only processes.
method Distinguishing between full conditional language process, marginal text-only process, and model-induced distribution; analyzing assumptions of stationarity and ergodicity.
result The marginal text-only law is useful only when the observed prefix is an approximately sufficient statistic for the latent circumstances relevant to continuation.
I introduce a new geometrical approach to thermo--statistical mechanics. Here I highlight the main physical ideas, and how do they translate into geometrical language. I contrast the present approach with previous thermo--statistical--geometrical formalisms, (pseudo-)Riemannian [Weinhold 1975; Ruppeiner 1979] as well a…
New method uses imperfect LLM annotations for valid statistical inference in social science.
problem Inaccurate large language model annotations in social science research.
method Design-based supervised learning (DSL) combining imperfect LLM surrogates with gold-standard labels.
result DSL provides valid statistical inference with comparable predictive accuracy to existing methods.
New estimator improves statistical validity of synthetic data integration.
problem Combining synthetic data generated by large language models with real data for valid inference.
method Generalized method of moments estimator with theoretical guarantees.
result Improves estimates of target parameter through interactions between synthetic and real data.
EFS uses LLMs to optimize sparse portfolios by evolving alpha factors.
problem Sparse portfolio optimization in dynamic market regimes.
method Evolutionary feedback loop with LLM-generated alpha factors.
result Significantly outperforms baselines in diverse datasets.
Paper introduces a statistical framework for watermarking LLM-generated text.
problem Detecting LLM-generated text from human-written text with statistical signals.
method Hypothesis testing formulation, pivotal statistic selection, secret key, closed-form expression of false negative rate, minimax optimization.
result Derives optimal detection rules for LLM-generated text, demonstrating higher power than existing methods.
Neural network framework for language recognition considers sequence information and improves accuracy.
problem Challenging task of automatic language identification in noisy conditions.
method Proposes a neural network framework with bidirectional LSTM and attention modeling for relevance weighting.
result Significant improvements over conventional methods in noisy conditions and multi-speaker speech.
This paper presents the beginnings of an automatic statistician, focusing on regression problems. Our system explores an open-ended space of statistical models to discover a good explanation of a data set, and then produces a detailed report with figures and natural-language text. Our approach treats unknown regression…
New attack reveals memorization patterns in pre-trained LLMs.
problem Determining if a data point was part of a pre-trained LLM's training set.
method Adapts MIA statistical tests to LLM's perplexity dynamics of subsequences.
result Significantly outperforms prior approaches in membership inference attacks.
Study bandwidth-limited training and inference of language models.
problem Training and inference of language models on scattered data with limited bandwidth.
method Analyzed two protocols: Federated Probe-Logit Distillation (FPLD) for training and Federated Conformal RAG (FC-RAG) for inference.
result Explicit high-probability KL-consistency rate and distribution-free marginal-coverage bound for Federated Conformal RAG.
LMs inevitably generate hallucinations, but can be made statistically negligible.
problem Inevitable hallucinations in language models.
method Probabilistic perspective and information theory.
result Hallucinations can be statistically negligible with sufficient training data and algorithm improvements.
DeepSeekMoE improves language model efficiency with shared experts and normalized gating.
problem Improving sample efficiency in language model architectures.
method Theoretical and empirical analysis of shared experts and normalized sigmoid gating.
result Theoretical and empirical evidence of improved sample efficiency with shared experts and normalized gating.
Transformers' self-attention mechanism is mapped to a generalized Potts model.
problem Uncertainty in what type of data distribution self-attention can efficiently learn.
method Decouple word positions and embeddings, then show self-attention learns a generalized Potts model.
result Training self-attention is equivalent to solving the inverse Potts problem.
New method assesses multivariate stochastic dominance using Optimal Transport.
problem Benchmarking models across multiple metrics considering dependencies.
method Characterization of multivariate first stochastic dominance via couplings, entropic regularization, and Optimal Transport.
result Established CLT and consistency for the empirical statistic, enabling hypothesis testing.
New research reveals how the pretraining distribution affects in-context learning in large language models.
problem Understanding how the pretraining distribution influences in-context learning in large language models.
method Developed a theoretical framework to characterize the relationship between pretraining distribution properties and in-context learning performance.
result Characterized a fundamental trade-off between robust task selection and generalization in ICL due to the pretraining distribution's statistical properties.