Large code datasets have become increasingly accessible for pre-training source code models. However, for the fine-tuning phase, obtaining representative training data that fully covers the code distribution for specific downstream tasks remains challenging due to the task-specific nature and limited labeling resources. Moreover, fine-tuning pretrained models can result in forgetting previously acquired pre-training knowledge. These lead to out-of-distribution (OOD) generalization issues with unexpected model inference behaviors that have not been systematically studied yet. In this paper, we contribute the first systematic approach that simulates various OOD scenarios along different dimensions of source code data properties and study the fine-tuned model behaviors in such scenarios. We investigate the behaviors of models under different fine-tuning methodologies, including full fine-tuning and Low-Rank Adaptation (LoRA) fine-tuning methods. Our comprehensive analysis, conducted on four state-of-the-art pretrained models and applied to two code generation tasks, exposes multiple failure modes attributed to OOD generalization issues. Additionally, our analysis uncovers that LoRA finetuning consistently exhibits significantly better OOD generalization performance than full fine-tuning across various scenarios. There has been increasing success in applying Large Language Models (LLMs) to various source code understanding and generation tasks. LLMs for codes such as CodeBERT Feng et al. (2020), GraphCodeBERT Guo et al. (2021), CodeT5+ Wang et al. (2023), CodeGen Nijkamp et al. (2023), and Code Llama Rozière et al. (2023) are pretrained using large-scale source code Figure 1: Our approach simulates out-ofdistribution datasets and serve as universal initialization for a (OOD) scenarios and analyzes the variety of downstream tasks. These tasks include corresponding behaviors of models. The emerging abilities of LLMs, such as in-context learning, demonstrate their potential to handle a wide range of tasks (Wei et al., 2022; Brown et al., 2020). However, it has been shown that not all tasks can be effectively addressed by relying only on the pretrained LLMs Anil et al. (2022). To adapt pretrained models for specific tasks, they can be fine-tuned with specific datasets for each downstream task. This fine-tuning process can involve optimizing all parameters or adopting a parameter-efficient approach (Houlsby et al., 2019; Hu et al., 2022), such as Low-Rank Adaptation (LoRA)Hu et al. (2022).

Abstract reasoning is a key ability for an intelligent system. Large language models (LMs) achieve above-chance performance on abstract reasoning tasks, but exhibit many imperfections. However, human abstract reasoning is also imperfect. For example, human reasoning is affected by our real-world knowledge and beliefs, and shows notable "content effects"; humans reason more reliably when the semantic content of a problem supports the correct logical inferences. These content-entangled reasoning patterns play a central role in debates about the fundamental nature of human intelligence. Here, we investigate whether language models $\unicode{x2014}$ whose prior expectations capture some aspects of human knowledge $\unicode{x2014}$ similarly mix content into their answers to logical problems. We explored this question across three logical reasoning tasks: natural language inference, judging the logical validity of syllogisms, and the Wason selection task. We evaluate state of the art large language models, as well as humans, and find that the language models reflect many of the same patterns observed in humans across these tasks $\unicode{x2014}$ like humans, models answer more accurately when the semantic content of a task supports the logical inferences. These parallels are reflected both in answer patterns, and in lower-level features like the relationship between model answer distributions and human response times. Our findings have implications for understanding both these cognitive effects in humans, and the factors that contribute to language model performance.

In computational social science (CSS), researchers analyze documents to explain social and political phenomena. In most scenarios, CSS researchers first obtain labels for documents and then explain labels using interpretable regression analyses in the second step. One increasingly common way to annotate documents cheaply at scale is through large language models (LLMs). However, like other scalable ways of producing annotations, such surrogate labels are often imperfect and biased. We present a new algorithm for using imperfect annotation surrogates for downstream statistical analyses while guaranteeing statistical properties -- like asymptotic unbiasedness and proper uncertainty quantification -- which are fundamental to CSS research. We show that direct use of surrogate labels in downstream statistical analyses leads to substantial bias and invalid confidence intervals, even with high surrogate accuracy of 80--90\%. To address this, we build on debiased machine learning to propose the design-based supervised learning (DSL) estimator. DSL employs a doubly-robust procedure to combine surrogate labels with a smaller number of high-quality, gold-standard labels. Our approach guarantees valid inference for downstream statistical analyses, even when surrogates are arbitrarily biased and without requiring stringent assumptions, by controlling the probability of sampling documents for gold-standard labeling. Both our theoretical analysis and experimental results show that DSL provides valid statistical inference while achieving root mean squared errors comparable to existing alternatives that focus only on prediction without inferential guarantees.

Large transformers are powerful architectures used for self-supervised data analysis across various data types, including protein sequences, images, and text. In these models, the semantic structure of the dataset emerges from a sequence of transformations between one representation and the next. We characterize the geometric and statistical properties of these representations and how they change as we move through the layers. By analyzing the intrinsic dimension (ID) and neighbor composition, we find that the representations evolve similarly in transformers trained on protein language tasks and image reconstruction tasks. In the first layers, the data manifold expands, becoming high-dimensional, and then contracts significantly in the intermediate layers. In the last part of the model, the ID remains approximately constant or forms a second shallow peak. We show that the semantic information of the dataset is better expressed at the end of the first peak, and this phenomenon can be observed across many models trained on diverse datasets. Based on our findings, we point out an explicit strategy to identify, without supervision, the layers that maximize semantic content: representations at intermediate layers corresponding to a relative minimum of the ID profile are more suitable for downstream learning tasks.

This paper presents LightLM, a lightweight Transformer-based language model for generative recommendation. While Transformer-based generative modeling has gained importance in various AI sub-fields such as NLP and vision, generative recommendation is still in its infancy due to its unique demand on personalized generative modeling. Existing works on generative recommendation often use NLP-oriented Transformer architectures such as T5, GPT, LLaMA and M6, which are heavy-weight and are not specifically designed for recommendation tasks. LightLM tackles the issue by introducing a light-weight deep and narrow Transformer architecture, which is specifically tailored for direct generation of recommendation items. This structure is especially apt for straightforward generative recommendation and stems from the observation that language model does not have to be too wide for this task, as the input predominantly consists of short tokens that are well-suited for the model's capacity. We also show that our devised user and item ID indexing methods, i.e., Spectral Collaborative Indexing (SCI) and Graph Collaborative Indexing (GCI), enables the deep and narrow Transformer architecture to outperform large-scale language models for recommendation. Besides, to address the hallucination problem of generating items as output, we propose the constrained generation process for generative recommenders. Experiments on real-world datasets show that LightLM outperforms various competitive baselines in terms of both recommendation accuracy and efficiency. The code can be found at https://github.com/dongyuanjushi/LightLM.

Current gesture recognition systems primarily focus on identifying gestures within a predefined set, leaving a gap in connecting these gestures to interactive GUI elements or system functions (e.g., linking a 'thumb-up' gesture to a 'like' button). We introduce GestureGPT, a novel zero-shot gesture understanding and grounding framework leveraging large language models (LLMs). Gesture descriptions are formulated based on hand landmark coordinates from gesture videos and fed into our dual-agent dialogue system. A gesture agent deciphers these descriptions and queries about the interaction context (e.g., interface, history, gaze data), which a context agent organizes and provides. Following iterative exchanges, the gesture agent discerns user intent, grounding it to an interactive function. We validated the gesture description module using public first-view and third-view gesture datasets and tested the whole system in two real-world settings: video streaming and smart home IoT control. The highest zero-shot Top-5 grounding accuracies are 80.11% for video streaming and 90.78% for smart home tasks, showing potential of the new gesture understanding paradigm.

Managing long sequences has become an important and necessary feature for large language models (LLMs). However, it is still an open question of how to comprehensively and systematically evaluate the long-sequence capability of LLMs. One of the reasons is that conventional and widely-used benchmarks mainly consist of short sequences. In this paper, we propose M4LE, a Multi-ability, Multi-range, Multi-task, Multi-domain benchmark for Long-context Evaluation. M4LE is based on a diverse NLP task pool comprising 36 NLP datasets, 11 task types and 12 domains. To alleviate the scarcity of tasks with naturally long sequences and incorporate multiple-ability assessment, we propose an automatic approach (but with negligible human annotations) to convert short-sequence tasks into a unified long-sequence scenario where LLMs have to identify single or multiple relevant spans in long contexts based on explicit or semantic hints. Specifically, the scenario includes five different types of abilities: (1) explicit single-span; (2) semantic single-span; (3) explicit multiple-span; (4) semantic multiple-span; and (5) global context understanding. The resulting samples in M4LE are evenly distributed from 1k to 8k input length. We conducted a systematic evaluation on 11 well-established LLMs, especially those optimized for long-sequence inputs. Our results reveal that: 1) Current LLMs struggle to understand long context, particularly when tasks require multiple-span attention. 2) Semantic retrieval task is more difficult for competent LLMs. 3) Models fine-tuned on longer text with position interpolation have comparable performance to those using Neural Tangent Kernel (NTK) aware scaling methods without fine-tuning. We make our benchmark publicly available to encourage future research in this challenging area.

Large language models have shown success as a tutor in education in various fields. Educating patients about their clinical visits plays a pivotal role in patients' adherence to their treatment plans post-discharge. This paper presents EHRTutor, an innovative multi-component framework leveraging the Large Language Model (LLM) for patient education through conversational question-answering. EHRTutor first formulates questions pertaining to the electronic health record discharge instructions. It then educates the patient through conversation by administering each question as a test. Finally, it generates a summary at the end of the conversation. Evaluation results using LLMs and domain experts have shown a clear preference for EHRTutor over the baseline. Moreover, EHRTutor also offers a framework for generating synthetic patient education dialogues that can be used for future in-house system training.

A model is considered well-calibrated when its probability estimate aligns with the actual likelihood of the output being correct. Calibrating language models (LMs) is crucial, as it plays a vital role in detecting and mitigating hallucinations, a common issue of LMs, as well as building more trustworthy models. Yet, popular neural model calibration techniques are not well-suited for LMs due to their lack of flexibility in discerning answer correctness and their high computational costs. For instance, post-processing methods like temperature scaling are often unable to reorder the candidate generations. Moreover, training-based methods require finetuning the entire model, which is impractical due to the increasing sizes of modern LMs. In this paper, we present LitCab, a lightweight calibration mechanism consisting of a single linear layer taking the input text representation and manipulateing the LM output logits. LitCab improves model calibration by only adding < 2% of the original model parameters. For evaluation, we construct CaT, a benchmark consisting of 7 text generation tasks, covering responses ranging from short phrases to paragraphs. We test LitCab with Llama2-7B, where it improves calibration across all tasks, by reducing the average ECE score by 20%. We further conduct a comprehensive evaluation with 7 popular open-sourced LMs from GPT and LLaMA families, yielding the following key findings: (1) Larger models within the same family exhibit better calibration on tasks with short generation tasks, but not necessarily for longer ones. (2) GPT-family models show superior calibration compared to LLaMA, Llama2 and Vicuna models despite having much fewer parameters. (3) Finetuning pretrained model (e.g., LLaMA) with samples of limited purpose (e.g., conversations) may lead to worse calibration, highlighting the importance of finetuning setups for calibrating LMs.

Student simulation presents a transformative approach to enhance learning outcomes, advance educational research, and ultimately shape the future of effective pedagogy. We explore the feasibility of using large language models (LLMs), a remarkable achievement in AI, to simulate student learning behaviors. Unlike conventional machine learning based prediction, we leverage LLMs to instantiate virtual students with specific demographics and uncover intricate correlations among learning experiences, course materials, understanding levels, and engagement. Our objective is not merely to predict learning outcomes but to replicate learning behaviors and patterns of real students. We validate this hypothesis through three experiments. The first experiment, based on a dataset of N = 145, simulates student learning outcomes from demographic data, revealing parallels with actual students concerning various demographic factors. The second experiment (N = 4524) results in increasingly realistic simulated behaviors with more assessment history for virtual students modelling. The third experiment (N = 27), incorporating prior knowledge and course interactions, indicates a strong link between virtual students' learning behaviors and fine-grained mappings from test questions, course materials, engagement and understanding levels. Collectively, these findings deepen our understanding of LLMs and demonstrate its viability for student simulation, empowering more adaptable curricula design to enhance inclusivity and educational effectiveness.