We explore how to enhance next-token prediction models to perform in-context imitation learning on a real robot, where the robot executes new tasks by interpreting contextual information provided during the input phase, without updating its underlying policy parameters. We propose In-Context Robot Transformer (ICRT), a causal transformer that performs autoregressive prediction on sensorimotor trajectories without relying on any linguistic data or reward function. This formulation enables flexible and training-free execution of new tasks at test time, achieved by prompting the model with sensorimotor trajectories of the new task composing of image observations, actions and states tuples, collected through human teleoperation. Experiments with a Franka Emika robot demonstrate that the ICRT can adapt to new tasks specified by prompts, even in environment configurations that differ from both the prompt and the training data. In a multitask environment setup, ICRT significantly outperforms current state-of-the-art next-token prediction models in robotics on generalizing to unseen tasks. Code, checkpoints and data are available on https://icrt.dev/

Cardinality estimation is crucial for enabling high query performance in relational databases. Recently learned cardinality estimation models have been proposed to improve accuracy but there is no systematic benchmark or datasets which allows researchers to evaluate the progress made by new learned approaches and even systematically develop new learned approaches. In this paper, we are releasing a benchmark, containing thousands of queries over 20 distinct real-world databases for learned cardinality estimation. In contrast to other initial benchmarks, our benchmark is much more diverse and can be used for training and testing learned models systematically. Using this benchmark, we explored whether learned cardinality estimation can be transferred to an unseen dataset in a zero-shot manner. We trained GNN-based and transformer-based models to study the problem in three setups: 1-) instance-based, 2-) zero-shot, and 3-) fine-tuned. Our results show that while we get promising results for zero-shot cardinality estimation on simple single table queries; as soon as we add joins, the accuracy drops. However, we show that with fine-tuning, we can still utilize pre-trained models for cardinality estimation, significantly reducing training overheads compared to instance specific models. We are open sourcing our scripts to collect statistics, generate queries and training datasets to foster more extensive research, also from the ML community on the important problem of cardinality estimation and in particular improve on recent directions such as pre-trained cardinality estimation.

This thesis provides methods and analysis of models which make progress on this goal. The techniques outlined are task agnostic, and should provide benefit when used with nearly any transformer LM. We introduce two new finetuning methods which add new capabilities to the models they are used on. The first adds a recurrence mechanism, which removes the fixed-window sized constraint and improves the efficiency of a transformer decoder. The second allows masked language models (MLMs) to be used for initialization of both the encoder and decoder of a non-autoregressive sequence-to-sequence transformer, opening up generative applications of models which were previously only used for natural language understanding tasks. We also introduce two new techniques for improving the quality of predictions of any transformer decoder without additional finetuning. One, hidden state optimization, can be applied to any transformer decoder to improve the quality of predictions at inference time, especially for few-shot classification. The other, conditional beam search, allows practitioners to search for natural language generation (NLG) model outputs with high likelihood while conditioning on the event that the output is not degenerate (e.g. empty, repetitive, etc.). Finally, we provide theoretical and empirical insights on the divergence of model-likelihood and output quality which has widely been observed in prior work. These insights apply to any model which represents a distribution over text, and apply to language models which are not transformers or even autoregressive. We argue that the NLP community has, to some extent, misunderstood the implications of these findings, and encourage a point of view which has more nuance.

The remarkable success of large language models (LLMs) across various multi-modality applications is well established. However, integrating large language models with humans, or brain dynamics, remains relatively unexplored. In this paper, we introduce BELT-2, a pioneering multi-task model designed to enhance both encoding and decoding performance from EEG signals. To bolster the quality of the EEG encoder, BELT-2 is the first work to innovatively 1) adopt byte-pair encoding (BPE)-level EEG-language alignment and 2) integrate multi-task training and decoding in the EEG domain. Inspired by the idea of \textbf{\textit{Bridging the Brain with GPT}}, we further connect the multi-task EEG encoder with LLMs by utilizing prefix-tuning on intermediary output from the EEG encoder. These innovative efforts make BELT-2 a pioneering breakthrough, making it the first work in the field capable of decoding coherent and readable sentences from non-invasive brain signals. Our experiments highlight significant advancements over prior techniques in both quantitative and qualitative measures, achieving a decoding performance with a BLEU-1 score of 52.2\% on the ZuCo dataset. Furthermore, BELT-2 shows a remarkable improvement ranging from 31\% to 162\% on other translation benchmarks. Codes can be accessed via the provided anonymous link~\footnote{https://anonymous.4open.science/r/BELT-2-0048}.

As large language models (LLMs) continue to advance, aligning these models with human preferences has emerged as a critical challenge. Traditional alignment methods, relying on human or LLM annotated datasets, are limited by their resource-intensive nature, inherent subjectivity, and the risk of feedback loops that amplify model biases. To overcome these limitations, we introduce WildFeedback, a novel framework that leverages real-time, in-situ user interactions to create preference datasets that more accurately reflect authentic human values. WildFeedback operates through a three-step process: feedback signal identification, preference data construction, and user-guided evaluation. We applied this framework to a large corpus of user-LLM conversations, resulting in a rich preference dataset that reflects genuine user preferences. This dataset captures the nuances of user preferences by identifying and classifying feedback signals within natural conversations, thereby enabling the construction of more representative and context-sensitive alignment data. Our extensive experiments demonstrate that LLMs fine-tuned on WildFeedback exhibit significantly improved alignment with user preferences, as evidenced by both traditional benchmarks and our proposed user-guided evaluation. By incorporating real-time feedback from actual users, WildFeedback addresses the scalability, subjectivity, and bias challenges that plague existing approaches, marking a significant step toward developing LLMs that are more responsive to the diverse and evolving needs of their users. In summary, WildFeedback offers a robust, scalable solution for aligning LLMs with true human values, setting a new standard for the development and evaluation of user-centric language models.

Detecting semantic types of columns in data lake tables is an important application. A key bottleneck in semantic type detection is the availability of human annotation due to the inherent complexity of data lakes. In this paper, we propose using programmatic weak supervision to assist in annotating the training data for semantic type detection by leveraging labeling functions. One challenge in this process is the difficulty of manually writing labeling functions due to the large volume and low quality of the data lake table datasets. To address this issue, we explore employing Large Language Models (LLMs) for labeling function generation and introduce several prompt engineering strategies for this purpose. We conduct experiments on real-world web table datasets. Based on the initial results, we perform extensive analysis and provide empirical insights and future directions for researchers in this field.

A text-to-speech (TTS) model trained to reconstruct speech given text tends towards predictions that are close to the average characteristics of a dataset, failing to model the variations that make human speech sound natural. This problem is magnified for zero-shot voice cloning, a task that requires training data with high variance in speaking styles. We build off of recent works which have used Generative Advsarial Networks (GAN) by proposing a Transformer encoder-decoder architecture to conditionally discriminates between real and generated speech features. The discriminator is used in a training pipeline that improves both the acoustic and prosodic features of a TTS model. We introduce our novel adversarial training technique by applying it to a FastSpeech2 acoustic model and training on Libriheavy, a large multi-speaker dataset, for the task of zero-shot voice cloning. Our model achieves improvements over the baseline in terms of speech quality and speaker similarity. Audio examples from our system are available online.

Recently, large language and vision models have shown strong performance, but due to high pre-training and fine-tuning costs, research has shifted towards faster training via dataset pruning. Previous methods used sample loss as an evaluation criterion, aiming to select the most "difficult" samples for training. However, when the pruning rate increases, the number of times each sample is trained becomes more evenly distributed, which causes many critical or general samples to not be effectively fitted. We refer to this as Low-Frequency Learning (LFL). In other words, LFL prevents the model from remembering most samples. In our work, we decompose the scoring function of LFL, provide a theoretical explanation for the inefficiency of LFL, and propose adding a memory term to the scoring function to enhance the model's memory capability, along with an approximation of this memory term. Similarly, we explore memory in Self-Supervised Learning (SSL), marking the first discussion on SSL memory. Using contrastive learning, we derive the memory term both theoretically and experimentally. Finally, we propose Enhance Memory Pruning (EMP), which addresses the issue of insufficient memory under high pruning rates by enhancing the model's memory of data, thereby improving its performance. We evaluated the performance of EMP in tasks such as image classification, natural language understanding, and model pre-training. The results show that EMP can improve model performance under extreme pruning rates. For example, in the CIFAR100-ResNet50 pre-training task, with 70\% pruning, EMP outperforms current methods by 2.2\%.

While large language models (LLMs) are known to have strong language understanding and generation abilities, they can also generate harmful, biased, and toxic content [1][2]. Alignment of LLMs ensures that they generate content that is "desirable" for the user, typically meaning content that is safe and ethical. Various approaches for LLM alignment have been presented ([1], [2], [3] and reference therein). The major approach to the alignment is reinforcement learning from human feedback (RLHF) [4], where a reward model is constructed by human feedback and used for the training of LLMs. Variants of RLHF architectures are also proposed, such as Safe-RLHF [5], SENSEI [6], and f-DPG [7], and their implementations are presented, such as training pre-trained LLMs [8][9], and applications like information-seeking chatbot [10].

There is a growing trend of teaching large language models (LLMs) to solve mathematical problems through coding. Existing studies primarily focus on prompting powerful, closed-source models to generate seed training data followed by in-domain data augmentation, equipping LLMs with considerable capabilities for code-aided mathematical reasoning. However, continually training these models on augmented data derived from a few datasets such as GSM8K may impair their generalization abilities and restrict their effectiveness to a narrow range of question types. Conversely, the potential of improving such LLMs by leveraging large-scale, expert-written, diverse math question-answer pairs remains unexplored. To utilize these resources and tackle unique challenges such as code response assessment, we propose a novel paradigm that uses a code-based critic model to guide steps including question-code data construction, quality control, and complementary evaluation. We also explore different alignment algorithms with self-generated instruction/preference data to foster continuous improvement. Experiments across both in-domain (up to +5.7%) and out-of-domain (+4.4%) benchmarks in English and Chinese demonstrate the effectiveness of the proposed paradigm.