Large language models display remarkable capabilities in logical and mathematical reasoning, allowing them to solve complex tasks. Interestingly, these abilities emerge in networks trained on the simple task of next-token prediction. In this work, we present a theoretical framework for studying auto-regressive next-token predictors. We demonstrate that even simple models such as linear next-token predictors, trained on Chain-of-Thought (CoT) data, can approximate any function efficiently computed by a Turing machine. We introduce a new complexity measure -- length complexity -- which measures the number of intermediate tokens in a CoT sequence required to approximate some target function, and analyze the interplay between length complexity and other notions of complexity. Finally, we show experimentally that simple next-token predictors, such as linear networks and shallow Multi-Layer Perceptrons (MLPs), display non-trivial performance on text generation and arithmetic tasks. Our results demonstrate that the power of language models can be attributed, to a great extent, to the auto-regressive next-token training scheme, and not necessarily to a particular choice of architecture.
相關內容
This paper proposes two nonlinear dynamics to solve constrained distributed optimization problem for resource allocation over a multi-agent network. In this setup, coupling constraint refers to resource-demand balance which is preserved at all-times. The proposed solutions can address various model nonlinearities, for example, due to quantization and/or saturation. Further, it allows to reach faster convergence or to robustify the solution against impulsive noise or uncertainties. We prove convergence over weakly connected networks using convex analysis and Lyapunov theory. Our findings show that convergence can be reached for general sign-preserving odd nonlinearity. We further propose delay-tolerant mechanisms to handle general bounded heterogeneous time-varying delays over the communication network of agents while preserving all-time feasibility. This work finds application in CPU scheduling and coverage control among others. This paper advances the state-of-the-art by addressing (i) possible nonlinearity on the agents/links, meanwhile handling (ii) resource-demand feasibility at all times, (iii) uniform-connectivity instead of all-time connectivity, and (iv) possible heterogeneous and time-varying delays. To our best knowledge, no existing work addresses contributions (i)-(iv) altogether. Simulations and comparative analysis are provided to corroborate our contributions.
In-context learning with large language models (LLMs) has recently caught increasing attention due to its superior few-shot performance on various tasks. However, its performance on text-to-SQL parsing still has much room for improvement. In this paper, we hypothesize that a crucial aspect of LLMs to improve for text-to-SQL parsing is their multi-step reasoning ability. Thus, we systematically study how to enhance LLMs' reasoning ability through chain of thought (CoT) style prompting, including the original chain-of-thought prompting (Wei et al., 2022b) and least-to-most prompting (Zhou et al., 2023). Our experiments demonstrate that iterative prompting as in Zhou et al. (2023) may be unnecessary for text-to-SQL parsing, and using detailed reasoning steps tends to have more error propagation issues. Based on these findings, we propose a new CoT-style prompting method for text-to-SQL parsing. It brings 5.2 and 6.5 point absolute gains on the Spider development set and the Spider Realistic set, respectively, compared to the standard prompting method without reasoning steps; 2.4 and 1.5 point absolute gains, compared to the least-to-most prompting method.
The goal of causal representation learning is to find a representation of data that consists of causally related latent variables. We consider a setup where one has access to data from multiple domains that potentially share a causal representation. Crucially, observations in different domains are assumed to be unpaired, that is, we only observe the marginal distribution in each domain but not their joint distribution. In this paper, we give sufficient conditions for identifiability of the joint distribution and the shared causal graph in a linear setup. Identifiability holds if we can uniquely recover the joint distribution and the shared causal representation from the marginal distributions in each domain. We transform our identifiability results into a practical method to recover the shared latent causal graph.
Large language models (LLMs) trained on huge corpora of text datasets demonstrate complex, emergent capabilities, achieving state-of-the-art performance on tasks they were not explicitly trained for. The precise nature of LLM capabilities is often mysterious, and different prompts can elicit different capabilities through in-context learning. We propose a Cognitive Interpretability framework that enables us to analyze in-context learning dynamics to understand latent concepts in LLMs underlying behavioral patterns. This provides a more nuanced understanding than success-or-failure evaluation benchmarks, but does not require observing internal activations as a mechanistic interpretation of circuits would. Inspired by the cognitive science of human randomness perception, we use random binary sequences as context and study dynamics of in-context learning by manipulating properties of context data, such as sequence length. In the latest GPT-3.5+ models, we find emergent abilities to generate pseudo-random numbers and learn basic formal languages, with striking in-context learning dynamics where model outputs transition sharply from pseudo-random behaviors to deterministic repetition.
We study a novel ensemble approach for feature selection based on hierarchical stacking in cases of non-stationarity and limited number of samples with large number of features. Our approach exploits the co-dependency between features using a hierarchical structure. Initially, a machine learning model is trained using a subset of features, and then the model's output is updated using another algorithm with the remaining features to minimize the target loss. This hierarchical structure allows for flexible depth and feature selection. By exploiting feature co-dependency hierarchically, our proposed approach overcomes the limitations of traditional feature selection methods and feature importance scores. The effectiveness of the approach is demonstrated on synthetic and real-life datasets, indicating improved performance with scalability and stability compared to the traditional methods and state-of-the-art approaches.
Spatio-temporal graph learning is a fundamental problem in the Web of Things era, which enables a plethora of Web applications such as smart cities, human mobility and climate analysis. Existing approaches tackle different learning tasks independently, tailoring their models to unique task characteristics. These methods, however, fall short of modeling intrinsic uncertainties in the spatio-temporal data. Meanwhile, their specialized designs limit their universality as general spatio-temporal learning solutions. In this paper, we propose to model the learning tasks in a unified perspective, viewing them as predictions based on conditional information with shared spatio-temporal patterns. Based on this proposal, we introduce Unified Spatio-Temporal Diffusion Models (USTD) to address the tasks uniformly within the uncertainty-aware diffusion framework. USTD is holistically designed, comprising a shared spatio-temporal encoder and attention-based denoising networks that are task-specific. The shared encoder, optimized by a pre-training strategy, effectively captures conditional spatio-temporal patterns. The denoising networks, utilizing both cross- and self-attention, integrate conditional dependencies and generate predictions. Opting for forecasting and kriging as downstream tasks, we design Gated Attention (SGA) and Temporal Gated Attention (TGA) for each task, with different emphases on the spatial and temporal dimensions, respectively. By combining the advantages of deterministic encoders and probabilistic diffusion models, USTD achieves state-of-the-art performances compared to deterministic and probabilistic baselines in both tasks, while also providing valuable uncertainty estimates.
In applying reinforcement learning (RL) to high-stakes domains, quantitative and qualitative evaluation using observational data can help practitioners understand the generalization performance of new policies. However, this type of off-policy evaluation (OPE) is inherently limited since offline data may not reflect the distribution shifts resulting from the application of new policies. On the other hand, online evaluation by collecting rollouts according to the new policy is often infeasible, as deploying new policies in these domains can be unsafe. In this work, we propose a semi-offline evaluation framework as an intermediate step between offline and online evaluation, where human users provide annotations of unobserved counterfactual trajectories. While tempting to simply augment existing data with such annotations, we show that this naive approach can lead to biased results. Instead, we design a new family of OPE estimators based on importance sampling (IS) and a novel weighting scheme that incorporate counterfactual annotations without introducing additional bias. We analyze the theoretical properties of our approach, showing its potential to reduce both bias and variance compared to standard IS estimators. Our analyses reveal important practical considerations for handling biased, noisy, or missing annotations. In a series of proof-of-concept experiments involving bandits and a healthcare-inspired simulator, we demonstrate that our approach outperforms purely offline IS estimators and is robust to imperfect annotations. Our framework, combined with principled human-centered design of annotation solicitation, can enable the application of RL in high-stakes domains.
We propose LLM-FP4 for quantizing both weights and activations in large language models (LLMs) down to 4-bit floating-point values, in a post-training manner. Existing post-training quantization (PTQ) solutions are primarily integer-based and struggle with bit widths below 8 bits. Compared to integer quantization, floating-point (FP) quantization is more flexible and can better handle long-tail or bell-shaped distributions, and it has emerged as a default choice in many hardware platforms. One characteristic of FP quantization is that its performance largely depends on the choice of exponent bits and clipping range. In this regard, we construct a strong FP-PTQ baseline by searching for the optimal quantization parameters. Furthermore, we observe a high inter-channel variance and low intra-channel variance pattern in activation distributions, which adds activation quantization difficulty. We recognize this pattern to be consistent across a spectrum of transformer models designed for diverse tasks, such as LLMs, BERT, and Vision Transformer models. To tackle this, we propose per-channel activation quantization and show that these additional scaling factors can be reparameterized as exponential biases of weights, incurring a negligible cost. Our method, for the first time, can quantize both weights and activations in the LLaMA-13B to only 4-bit and achieves an average score of 63.1 on the common sense zero-shot reasoning tasks, which is only 5.8 lower than the full-precision model, significantly outperforming the previous state-of-the-art by 12.7 points. Code is available at: //github.com/nbasyl/LLM-FP4.
In predictive modeling with simulation or machine learning, it is critical to accurately assess the quality of estimated values through output analysis. In recent decades output analysis has become enriched with methods that quantify the impact of input data uncertainty in the model outputs to increase robustness. However, most developments are applicable assuming that the input data adheres to a parametric family of distributions. We propose a unified output analysis framework for simulation and machine learning outputs through the lens of Monte Carlo sampling. This framework provides nonparametric quantification of the variance and bias induced in the outputs with higher-order accuracy. Our new bias-corrected estimation from the model outputs leverages the extension of fast iterative bootstrap sampling and higher-order influence functions. For the scalability of the proposed estimation methods, we devise budget-optimal rules and leverage control variates for variance reduction. Our theoretical and numerical results demonstrate a clear advantage in building more robust confidence intervals from the model outputs with higher coverage probability.
We propose GAN-Supervised Learning, a framework for learning discriminative models and their GAN-generated training data jointly end-to-end. We apply our framework to the dense visual alignment problem. Inspired by the classic Congealing method, our GANgealing algorithm trains a Spatial Transformer to map random samples from a GAN trained on unaligned data to a common, jointly-learned target mode. We show results on eight datasets, all of which demonstrate our method successfully aligns complex data and discovers dense correspondences. GANgealing significantly outperforms past self-supervised correspondence algorithms and performs on-par with (and sometimes exceeds) state-of-the-art supervised correspondence algorithms on several datasets -- without making use of any correspondence supervision or data augmentation and despite being trained exclusively on GAN-generated data. For precise correspondence, we improve upon state-of-the-art supervised methods by as much as $3\times$. We show applications of our method for augmented reality, image editing and automated pre-processing of image datasets for downstream GAN training.
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191