Text-to-SQL simplifies database interactions by enabling non-experts to convert their natural language (NL) questions into Structured Query Language (SQL) queries. While recent advances in large language models (LLMs) have improved the zero-shot text-to-SQL paradigm, existing methods face scalability challenges when dealing with massive, dynamically changing databases. This paper introduces DBCopilot, a framework that addresses these challenges by employing a compact and flexible copilot model for routing across massive databases. Specifically, DBCopilot decouples the text-to-SQL process into schema routing and SQL generation, leveraging a lightweight sequence-to-sequence neural network-based router to formulate database connections and navigate natural language questions through databases and tables. The routed schemas and questions are then fed into LLMs for efficient SQL generation. Furthermore, DBCopilot also introduced a reverse schema-to-question generation paradigm, which can learn and adapt the router over massive databases automatically without requiring manual intervention. Experimental results demonstrate that DBCopilot is a scalable and effective solution for real-world text-to-SQL tasks, providing a significant advancement in handling large-scale schemas.

The proliferation of Large Language Models (LLMs) has spurred extensive research into LLM-related Prompt investigations, such as Instruction Learning (IL), In-context Learning (ICL), and Chain-of-Thought (CoT). These approaches aim to improve LLMs' responses by enabling them to provide concise statements or examples for deeper contemplation when addressing questions. However, independent thinking by LLMs can introduce variability in their thought processes, leading to potential inaccuracies. In response, our study seeks to bridge the gap between LLM and human-like thinking processes, recognizing that text comprehension begins with understanding individual words. To tackle this challenge, we have expanded the CoT method to cater to a specific domain. Our approach, known as "Think from Words" (TFW), initiates the comprehension process at the word level and then extends it to encompass the entire text. We also propose "TFW with Extra word-level information" (TFW Extra), augmenting comprehension with additional word-level data. To assess our methods, we employ text classification on six Japanese datasets comprising text-level and word-level elements. Our findings not only validate the effectiveness of TFW but also shed light on the impact of various word-level information types on LLMs' text comprehension, offering insights into their potential to cause misinterpretations and errors in the overall comprehension of the final text.

LoRA Hu et al. [2022] paper has introduced low-rank adapters to fine-tune large LLMs on downstream tasks. This approach quickly became popular due to reduced cost of the update. Different modifications of LoRA followed, for example, QLoRA Dettmers et al. [2023] utilizes quantization and further reduces fine-tuning costs, and ReLoRA Lialin et al. [2023] which showed that low-rank updates can also be used for full training. However, all variations of LoRA use the same chain of operations while calculating the output, which often leads to sub-optimal graph of computations. We propose RunLora: a framework which contains different variations of forward and backward pass through an adapter-induced linear layer and chooses the best pair for a given architecture. We evaluated our framework's performance on a series of Llama models and achieved up to 17% speedup only due to optimized chain of PyTorch operations. Additionally, we managed to save up to 4Gb of memory due to reduction in number of saved activations.

We argue that many general evaluation problems can be viewed through the lens of voting theory. Each task is interpreted as a separate voter, which requires only ordinal rankings or pairwise comparisons of agents to produce an overall evaluation. By viewing the aggregator as a social welfare function, we are able to leverage centuries of research in social choice theory to derive principled evaluation frameworks with axiomatic foundations. These evaluations are interpretable and flexible, while avoiding many of the problems currently facing cross-task evaluation. We apply this Voting-as-Evaluation (VasE) framework across multiple settings, including reinforcement learning, large language models, and humans. In practice, we observe that VasE can be more robust than popular evaluation frameworks (Elo and Nash averaging), discovers properties in the evaluation data not evident from scores alone, and can predict outcomes better than Elo in a complex seven-player game. We identify one particular approach, maximal lotteries, that satisfies important consistency properties relevant to evaluation, is computationally efficient (polynomial in the size of the evaluation data), and identifies game-theoretic cycles.

Recent advances in language models (LMs), have demonstrated significant efficacy in tasks related to the arts and humanities. While LMs have exhibited exceptional performance across a wide range of natural language processing tasks, there are notable challenges associated with their utilization on small datasets and their ability to replicate more creative human capacities. In this study, we aim to address these challenges by training a Persian classical poetry generation model using a transformer architecture on a specialized dataset with no pretraining. Additionally, we propose a novel decoding method to enhance coherence and meaningfulness in the generated poetry, effectively managing the tradeoff between diversity and quality. Furthermore, the results of our training approach and the proposed decoding method are evaluated through comprehensive set of automatic and human evaluations and showed its superior capability to generate coherent and meaningful poetry in compare to other decoding methods and an existing Persian large language model (LLM).

Building a generalist agent that can interact with the world is the intriguing target of AI systems, thus spurring the research for embodied navigation, where an agent is required to navigate according to instructions or respond to queries. Despite the major progress attained, previous works primarily focus on task-specific agents and lack generalizability to unseen scenarios. Recently, LLMs have presented remarkable capabilities across various fields, and provided a promising opportunity for embodied navigation. Drawing on this, we propose the first generalist model for embodied navigation, NaviLLM. It adapts LLMs to embodied navigation by introducing schema-based instruction. The schema-based instruction flexibly casts various tasks into generation problems, thereby unifying a wide range of tasks. This approach allows us to integrate diverse data sources from various datasets into the training, equipping NaviLLM with a wide range of capabilities required by embodied navigation. We conduct extensive experiments to evaluate the performance and generalizability of our model. The experimental results demonstrate that our unified model achieves state-of-the-art performance on CVDN, SOON, and ScanQA. Specifically, it surpasses the previous stats-of-the-art method by a significant margin of 29% in goal progress on CVDN. Moreover, our model also demonstrates strong generalizability and presents impressive results on unseen tasks, e.g., embodied question answering and 3D captioning.

The paper introduces LEMR, a framework that reduces annotation costs for model selection tasks. Our approach leverages ensemble methods to generate pseudo-labels, employs uncertainty sampling for target acquisition, and utilizes a Z-score mechanism for iterative committee reelection to refine model ranks. We present a systematic study across various selection metrics, demonstrating that LEMR achieves comparable results to fully labeled datasets with a fraction of the labeling budget. Our findings indicate that LEMR not only economizes the labeling effort in weak supervision and semi-supervised learning settings but also effectively guides prompt selection for large language models. With extensive experiments across 23 tasks, we reveal that our framework can dramatically decrease the labeling cost without compromising the accuracy of model selection, thereby offering a cost-effective alternative to traditional practices.

The capabilities of large language models have grown significantly in recent years and so too have concerns about their misuse. In this context, the ability to distinguish machine-generated text from human-authored content becomes important. Prior works have proposed numerous schemes to watermark text, which would benefit from a systematic evaluation framework. This work focuses on text watermarking techniques - as opposed to image watermarks - and proposes MARKMYWORDS, a comprehensive benchmark for them under different tasks as well as practical attacks. We focus on three main metrics: quality, size (e.g. the number of tokens needed to detect a watermark), and tamper-resistance. Current watermarking techniques are good enough to be deployed: Kirchenbauer et al. [1] can watermark Llama2-7B-chat with no perceivable loss in quality, the watermark can be detected with fewer than 100 tokens, and the scheme offers good tamper-resistance to simple attacks. We argue that watermark indistinguishability, a criteria emphasized in some prior works, is too strong a requirement: schemes that slightly modify logit distributions outperform their indistinguishable counterparts with no noticeable loss in generation quality. We publicly release our benchmark (https://github.com/wagner-group/MarkMyWords)

Mechanistic interpretability aims to understand model behaviors in terms of specific, interpretable features, often hypothesized to manifest as low-dimensional subspaces of activations. Specifically, recent studies have explored subspace interventions (such as activation patching) as a way to simultaneously manipulate model behavior and attribute the features behind it to given subspaces. In this work, we demonstrate that these two aims diverge, potentially leading to an illusory sense of interpretability. Counterintuitively, even if a subspace intervention makes the model's output behave as if the value of a feature was changed, this effect may be achieved by activating a dormant parallel pathway leveraging another subspace that is causally disconnected from model outputs. We demonstrate this phenomenon in a distilled mathematical example, in two real-world domains (the indirect object identification task and factual recall), and present evidence for its prevalence in practice. In the context of factual recall, we further show a link to rank-1 fact editing, providing a mechanistic explanation for previous work observing an inconsistency between fact editing performance and fact localization. However, this does not imply that activation patching of subspaces is intrinsically unfit for interpretability. To contextualize our findings, we also show what a success case looks like in a task (indirect object identification) where prior manual circuit analysis informs an understanding of the location of a feature. We explore the additional evidence needed to argue that a patched subspace is faithful.

Reasoning is a fundamental capability of AI agents. Recently, large language models (LLMs) have shown remarkable abilities to perform reasoning tasks. However, numerous evaluations of the reasoning capabilities of LLMs have also showed some limitations. An outstanding limitation is length generalization, meaning that when trained on reasoning problems of smaller lengths or sizes, the resulting models struggle with problems of larger sizes or lengths. This potentially indicates some theoretical limitations of generalization in learning reasoning skills. These evaluations and their observations motivated us to perform a theoretical study of the length generalization problem. This work focuses on reasoning tasks that can be formulated as Markov dynamic processes (MDPs) and/or directed acyclic graphs (DAGs). It identifies and proves conditions that decide whether the length generalization problem can be solved or not for a reasoning task in a particular representation. Experiments are also conducted to verify the theoretical results.