The study compares feed-forward and attention layers in language models.
problem Understanding the role of feed-forward and attention layers in language models.
method Empirical and theoretical analysis in a synthetic setting.
result Feed-forward layers learn simple distributional associations, while attention layers focus on in-context reasoning.
Study reveals how depth of reasoning affects generalization in models.
problem Understanding scaling behavior of generalization with CoT depth.
method Theoretical model of CoT in linear regression using random matrix theory.
result Sharp phase transition between exponential and polynomial improvement, saturation, and overthinking.
LLMs learn probability density functions in-context, showing distinct learning trajectories.
problem Density estimation of time series data in LLMs.
method Intensive Principal Component Analysis (InPCA) to visualize and analyze LLMs' learning dynamics.
result LLMs follow similar learning trajectories in a low-dimensional InPCA space, distinct from traditional methods.
TTT improves transformer models for in-context learning.
problem Improving transformer models for efficient in-context learning.
method Gradient-based TTT method for linear transformers, with theoretical and empirical analysis.
result TTT significantly reduces the sample size required for in-context learning.
LLMs learn new tasks from unstructured data, but it depends on word co-occurrence and positional information.
problem Understanding how LLMs can learn new tasks from unstructured data without explicit training.
method Examined the capabilities of LLMs trained on unstructured data, focusing on sequence model requirements and training data structure.
result Many ICL capabilities can emerge from word co-occurrence in unstructured data, but positional information is crucial for certain tasks.
Improves many-shot learning by optimizing and generating influential examples.
problem Limited performance in many-shot in-context learning (ICL).
method Iterative optimization and generation of influential examples.
result Significant improvements across various tasks using BRIDGE.
Theoretical work shows integrating coherent reasoning improves LLM performance and error correction.
problem Improving reasoning and error correction in large language models (LLMs) with few-shot prompting.
method Theoretical analysis and sensitivity experiments on transformer behavior with coherent reasoning and corrupted demonstrations.
result The transformer gains better error correction ability and more accurate predictions when coherent reasoning is integrated.
Generative AI can solve in-context learning problems using a martingale perspective.
problem Estimating when a conditional generative model can solve an in-context learning problem.
method Bayesian interpretation, ancestral sampling, generative predictive p-value.
result Developed a method to assess the suitability of CGMs for ICL problems using generative predictive p-values.
Looped Transformers learn to implement multi-step gradient descent for in-context learning.
problem Understanding the learnability of multi-step algorithms in multi-layer Transformers.
method Training weight-sharing looped Transformers for in-context linear regression, proving gradient dominance condition for convergence.
result Looped Transformers implement multi-step preconditioned gradient descent, converging to global minimizer.
This paper explores how Transformers predict next tokens in autoregressive tasks.
problem Understanding the success of Transformers in autoregressive learning.
method Trained a Transformer on a next-token prediction task, focusing on commuting orthogonal matrices.
result Trained Transformers can be seen as implementing gradient descent for a specific objective function.
A new method for inventory control using in-context learning and generative models.
problem Inventory control with decision-dependent censoring, focusing on the censored newsvendor problem.
method In-context generative posterior sampling (ICGPS) combining modern generative models and in-context autoregressive generation.
result ICGPS achieves sublinear Bayesian regret for the censored newsvendor problem, outperforming existing methods.
LLMs excel at summarizing and repairing complex models without needing full models.
problem Understanding and repairing complex models like GAMs.
method Hierarchical reasoning and extensive background knowledge.
result LLMs can detect anomalies, describe reasons, and suggest repairs.
MLPs can approximate any function in context, challenging the importance of in-context universality.
problem Understanding why transformers are more effective than classical models.
method Proved MLPs with trainable activation functions are universal in context.
result Transformer success is likely due to factors other than in-context universality.
Enhances neural processes to learn from multiple related datasets.
problem Improving predictions from datasets with shared similarities.
method Developed the in-context in-context learning pseudo-token TNP (ICICL-TNP) to condition on both sets of datapoints and sets of datasets.
result Demonstrated the importance and effectiveness of in-context in-context learning.
This work introduces a method to decompose uncertainty in in-context learning for large language models.
problem Understanding the sources of uncertainty in in-context learning for large language models.
method Variational uncertainty decomposition framework without sampling from latent parameter posterior.
result Quantitative and qualitative validation of decomposed epistemic and aleatoric uncertainties.
Transformers can generalize to a large task family with only a few demonstrations.
problem Can learning from a small set of tasks generalize to a large task family?
method Investigating autoregressive compositional structure where each task is a composition of T T T operations, each from a finite family of D D D subtasks. result Transformers can generalize to D T D^T D T tasks with only O ~ ( D ) \widetilde{O}(D) O ( D ) demonstrations. Transformers learn to perform logistic regression in-context.
problem Understanding how transformers learn to perform specific tasks in-context.
method Constructed multi-layer transformers that perform in-context logistic regression through normalized gradient descent.
result Transformers can be trained to perform in-context logistic regression effectively.
Fine-tuning harms in-context learning, but restricting updates to the value matrix improves zero-shot performance.
problem Fine-tuning harms in-context learning, reducing zero-shot performance on unseen tasks.
method Theoretical analysis of linear attention models, identifying conditions for degraded few-shot performance.
result Restricting updates to the value matrix improves zero-shot performance while preserving in-context learning.
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.
Transformers learn to play games in-context, proving Nash equilibrium.
problem Understanding in-context game-playing capabilities of pre-trained transformers.
method Theoretical guarantees and constructional results for transformer architecture in multi-agent games.
result Pre-trained transformers can learn Nash equilibrium in-context for two-player zero-sum games.
Transformers mimic Bayesian reasoning in controlled settings, revealing geometric mechanisms.
problem Verifying if transformers perform Bayesian reasoning rigorously in natural data.
method Constructing Bayesian wind tunnels with known posteriors and proving memorization impossibility.
result Transformers achieve 10 − 3 10^{-3} 1 0 − 3 - 10 − 4 10^{-4} 1 0 − 4 bit accuracy in Bayesian posteriors, while MLPs fail. Transformers forecast time series in-context, improving efficiency and performance.
problem Overfitting and limited performance in time series forecasting.
method Reformulate time series forecasting as input tokens, aligning with in-context learning mechanisms.
result Consistently better performance across various settings (full-data, few-shot, zero-shot).
Transformers can adaptively select and perform various machine learning tasks in context.
problem Transformers' ability to perform multiple machine learning tasks without explicit prompting.
method Statistical theory and efficient implementation of in-context gradient descent, enabling transformers to implement various algorithms and tasks.
result Transformers can implement a broad class of standard machine learning algorithms in context, including algorithm selection.
Continuum transformers learn operators in context via gradient descent.
problem Generalizing transformers to handle infinite-dimensional inputs for in-context learning.
method Gradient descent in an operator RKHS, leveraging generalized representer theorems and gradient flows.
result Operator learned in context is Bayes Optimal Predictor in infinite depth limit.
LLMs perform well in financial sentiment analysis without fine-tuning.
problem Challenges in financial terminology, emotions, and ambiguous expressions.
method In-context learning methods for financial document-sentiment pairs.
result LLMs can generalize in-context demonstrations to new financial documents.
Bayesian RL enhances LLMs to reflectively explore and correct errors.
problem LLMs trained via RL lack reflective behaviors like rethinking and error correction.
method Bayesian RL framework that optimizes expected return under posterior distribution over Markov decision processes.
result BARL algorithm improves LLM performance in reasoning tasks.
New framework shows cross-attention improves multi-modal in-context learning.
problem Understanding multi-modal in-context learning in neural networks.
method Mathematical framework and linearized cross-attention mechanism.
result Cross-attention mechanism is provably optimal for multi-modal in-context learning.
New model uses attention for in-context learning of categorical data.
problem Learning from categorical data in context.
method Attention-based network with self-attention and cross-attention layers, using functional gradient descent.
result Model can perform multi-step inference for categorical observations.
Orion-Bix combines biaxial attention and meta-learning for tabular few-shot learning.
problem Scaling and generalizing tabular models with mixed numeric and categorical fields, weak feature structure, and limited labeled data.
method Orion-Bix uses biaxial attention and meta-learned in-context reasoning to efficiently capture local and global dependencies.
result Orion-Bix outperforms gradient-boosting baselines and state-of-the-art tabular models on public benchmarks.
ICEE learns new RL tasks in less time with a Transformer model.
problem Efficient in-context policy learning for reinforcement learning.
method In-context Exploration-Exploitation (ICEE) algorithm that optimizes efficiency without explicit Bayesian inference.
result ICEE solves Bayesian optimization problems as efficiently as Gaussian process biased methods but in significantly less time.
New analysis shows transformer models can't learn effectively.
problem Transformer models learning in context can't achieve general predictive accuracy.
method Empirical evidence and mathematical analysis of transformer architecture limitations.
result Transformers cannot achieve general predictive accuracy due to architectural limitations.
Transformers can handle endogeneity in linear regression using IV methods.
problem Endogeneity in in-context linear regression models.
method Transformer architecture with gradient-based bi-level optimization and in-context pretraining.
result Transformers provide more robust predictions and estimates than 2SLS in endogenous scenarios.
Transformers learn new tasks from few examples via optimal approximation.
problem Learning new tasks from limited examples using large language models.
method Developed approximation and generalization error bounds for transformers trained on nonparametric regression tasks.
result Transformers achieve minimax optimal estimation risk in context.
Paper presents a Transformer model for automatic domain adaptation.
problem Challenges in selecting or designing domain adaptation algorithms.
method Transformer model approximates and selects domain adaptation algorithms.
result Transformers can approximate and automatically select domain adaptation algorithms.
Transformers converge linearly to optimal models for Gaussian mixtures classification.
problem Theoretical understanding of transformers' in-context classification.
method Gradient descent training of a single-layer transformer for Gaussian mixtures classification.
result Transformers converge linearly to globally optimal models for Gaussian mixtures classification.
Study on neural scaling laws for solving linear systems in-context.
problem Theoretical guarantees for solving linear systems using a linear transformer architecture.
method Neural scaling laws and task diversity for in-domain and out-of-domain generalization.
result Novel notion of task diversity for necessary and sufficient condition of generalization under task shifts.
Bayesian analysis reveals epistemic uncertainty as a key diagnostic for delayed generalization in in-context learning.
problem Delayed generalization in in-context learning from few examples.
method Bayesian perspective, modular arithmetic tasks, approximate Bayesian techniques, spectral mechanism analysis.
result Epistemic uncertainty collapses sharply when the model groks, indicating a practical diagnostic of generalization.
Transformers trained on random classification tasks generalize well and can overfit without error.
problem Understanding how transformers generalize and overfit in-context.
method Analysis of implicit regularization during gradient descent training.
result Transformers can overfit without error and still generalize well.
Transformers enable in-context learning with guarantees for a wide range of tasks.
problem How to enable in-context learning with transformers for various tasks.
method Developed a universal approximation theory integrating Barron's function approximation with transformer capabilities.
result Transformers can approximate any target function with vanishingly small risk using a few in-context examples.
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.
Transformers learn to generalize unseen tasks by composing self-attention layers.
problem How Transformers generalize to unseen, out-of-distribution tasks.
method Examined synthetic examples and pretrained LLMs, focusing on induction heads and latent subspace.
result Transformers can learn hidden rules for unseen tasks by composing self-attention layers, achieving OOD generalization.
Transformers learn low-dimensional target functions efficiently in-context.
problem Efficiently learning nonlinear target functions in-context using transformers.
method Nonlinear MLP layer in transformers optimized by gradient descent, focusing on single-index target functions.
result Transformers can learn target functions with low-dimensional structures efficiently in-context.
Transformers learn rich in-context dependencies efficiently.
problem Understanding how transformers learn long-range dependencies efficiently.
method Approximation and dynamics analysis of induction head mechanisms.
result Abrupt transition from lazy to rich mechanisms during training.
LLM4Causal democratizes causal reasoning via fine-tuned LLMs.
problem Limited capability of LLMs in causal inference and interpretation.
method Fine-tuning an open-source LLM for causal tasks, proposing datasets for instruction tuning.
result LLM4Causal delivers end-to-end solutions for causal problems and interprets results easily.
Looped Transformers improve robustness and expressivity in in-context learning for diverse tasks.
problem Improving robustness and expressivity in in-context learning for diverse tasks.
method Study in-context linear regression with diverse tasks, focusing on depth and looping.
result Looped Transformers exhibit similar expressive power and are provably robust under mild assumptions.
Multi-head attention outperforms single-head in in-context linear regression tasks.
problem Comparing performance of transformer with single-/multi-head attention in in-context learning.
method Theoretical analysis of performance of transformers with different attention mechanisms in linear regression tasks.
result Multi-head attention with a substantial embedding dimension outperforms single-head attention in in-context linear regression tasks.
Transformers learn to solve modular arithmetic tasks by in-context learning and skill composition.
problem Understanding how large language models generalize to unseen tasks in modular arithmetic.
method Pre-training on a set of modular arithmetic tasks and evaluating out-of-distribution performance.
result Transformers require two transformer blocks for out-of-distribution generalization, and deeper models exhibit transient out-of-distribution performance.
Transformers can implement reinforcement learning algorithms from data without updates.
problem Training reinforcement learning algorithms from data without parameter updates.
method Design a teacher-mimicking training procedure for transformers to implement policy-improvement methods.
result Gradient flow converges to an optimal parameter manifold corresponding to the desired RL update.