Characterizing Truthfulness in Large Language Model Generations with Local Intrinsic Dimension

We study how to characterize and predict the truthfulness of texts generated from large language models (LLMs), which serves as a crucial step in building trust between humans and LLMs. Although several approaches based on entropy or verbalized uncertainty have been proposed to calibrate model predictions, these methods are often intractable, sensitive to hyperparameters, and less reliable when applied in generative tasks with LLMs. In this paper, we suggest investigating internal activations and quantifying LLM's truthfulness using the local intrinsic dimension (LID) of model activations. Through experiments on four question answering (QA) datasets, we demonstrate the effectiveness ohttps://info.arxiv.org/help/prep#abstractsf our proposed method. Additionally, we study intrinsic dimensions in LLMs and their relations with model layers, autoregressive language modeling, and the training of LLMs, revealing that intrinsic dimensions can be a powerful approach to understanding LLMs.

Contrastive Instruction Tuning

Instruction tuning has been used as a promising approach to improve the performance of large language models (LLMs) on unseen tasks. However, current LLMs exhibit limited robustness to unseen instructions, generating inconsistent outputs when the same instruction is phrased with slightly varied forms or language styles. This behavior indicates LLMs' lack of robustness to textual variations and generalizability to unseen instructions, potentially leading to trustworthiness issues. Accordingly, we propose Contrastive Instruction Tuning, which maximizes the similarity between the hidden representations of semantically equivalent instruction-instance pairs while minimizing the similarity between semantically different ones. To facilitate this approach, we augment the existing FLAN collection by paraphrasing task instructions. Experiments on the PromptBench benchmark show that CoIN consistently improves LLMs' robustness to unseen instructions with variations across character, word, sentence, and semantic levels by an average of +2.5% in accuracy.

Prompt-Driven LLM Safeguarding via Directed Representation Optimization

Prepending model inputs with safety prompts is a common practice of safeguarding large language models (LLMs) from complying with queries that contain harmful intents. However, the working mechanisms of safety prompts have not yet been fully understood, which hinders the potential for automatically optimizing them for improved LLM safety. Motivated by this problem, we investigate the impact of safety prompts from the perspective of model representations. We find that in models' representation space, harmful and harmless queries can be largely distinguished, but this is not noticeably enhanced by safety prompts. Instead, the queries' representations are moved by different safety prompts in similar directions, where models become more prone to refusal (i.e., refusing to provide assistance) even when the queries are harmless. Inspired by these findings, we propose a method called DRO (Directed Representation Optimization) for automatic safety prompt optimization. DRO treats safety prompts as continuous, trainable embeddings and learns to move the representations of harmful/harmless queries along/opposite the direction in which the model's refusal probability increases. We demonstrate that DRO remarkably improves the safeguarding performance of human-crafted safety prompts and outperforms strong baselines, as evaluated on out-of-domain benchmarks, without compromising the general model capability.

Active Instruction Tuning: Improving Cross-Task Generalization by Training on Prompt Sensitive Tasks

Instruction tuning (IT) achieves impressive zero-shot generalization results by training large language models (LLMs) on a massive amount of diverse tasks with instructions. However, how to select new tasks to improve the performance and generalizability of IT models remains an open question. Training on all existing tasks is impractical due to prohibiting computation requirements, and randomly selecting tasks can lead to suboptimal performance. In this work, we propose active instruction tuning based on prompt uncertainty, a novel framework to identify informative tasks, and then actively tune the models on the selected tasks. We represent the informativeness of new tasks with the disagreement of the current model outputs over perturbed prompts. Our experiments on NIV2 and Self-Instruct datasets demonstrate that our method consistently outperforms other baseline strategies for task selection, achieving better out-of-distribution generalization with fewer training tasks. Additionally, we introduce a task map that categorizes and diagnoses tasks based on prompt uncertainty and prediction probability. We discover that training on ambiguous (prompt-uncertain) tasks improves generalization while training on difficult (prompt-certain and low-probability) tasks offers no benefit, underscoring the importance of task selection for instruction tuning.

Should I Stop or Should I Go: Early Stopping with Heterogeneous Populations

Randomized experiments often need to be stopped prematurely due to the treatment having an unintended harmful effect. Existing methods that determine when to stop an experiment early are typically applied to the data in aggregate and do not account for treatment effect heterogeneity. In this paper, we study the early stopping of experiments for harm on heterogeneous populations. We first establish that current methods often fail to stop experiments when the treatment harms a minority group of participants. We then use causal machine learning to develop CLASH, the first broadly-applicable method for heterogeneous early stopping. We demonstrate CLASH's performance on simulated and real data and show that it yields effective early stopping for both clinical trials and A/B tests.

Did You Read the Instructions? Rethinking the Effectiveness of Task Definitions in Instruction Learning

Large language models (LLMs) have shown impressive performance in following natural language instructions to solve unseen tasks. However, it remains unclear whether models truly understand task definitions and whether the human-written definitions are optimal. In this paper, we systematically study the role of task definitions in instruction learning. We first conduct an ablation analysis informed by human annotations to understand which parts of a task definition are most important, and find that model performance only drops substantially when removing contents describing the task output, in particular label information. Next, we propose an automatic algorithm to compress task definitions to a minimal supporting set of tokens, and find that 60\% of tokens can be removed while maintaining or even improving model performance. Based on these results, we propose two strategies to help models better leverage task instructions: (1) providing only key information for tasks in a common structured format, and (2) adding a meta-tuning stage to help the model better understand the definitions. With these two strategies, we achieve a 4.2 Rouge-L improvement over 119 unseen test tasks.

Efficient Shapley Values Estimation by Amortization for Text Classification

Despite the popularity of Shapley Values in explaining neural text classification models, computing them is prohibitive for large pretrained models due to a large number of model evaluations. In practice, Shapley Values are often estimated with a small number of stochastic model evaluations. However, we show that the estimated Shapley Values are sensitive to random seed choices -- the top-ranked features often have little overlap across different seeds, especially on examples with longer input texts. This can only be mitigated by aggregating thousands of model evaluations, which on the other hand, induces substantial computational overheads. To mitigate the trade-off between stability and efficiency, we develop an amortized model that directly predicts each input feature's Shapley Value without additional model evaluations. It is trained on a set of examples whose Shapley Values are estimated from a large number of model evaluations to ensure stability. Experimental results on two text classification datasets demonstrate that our amortized model estimates Shapley Values accurately with up to 60 times speedup compared to traditional methods. Furthermore, the estimated values are stable as the inference is deterministic. We release our code at https://github.com/yangalan123/Amortized-Interpretability.

Red Teaming Language Model Detectors with Language Models

The prevalence and high capacity of large language models (LLMs) present significant safety and ethical risks when malicious users exploit them for automated content generation. To prevent the potentially deceptive usage of LLMs, recent works have proposed several algorithms to detect machine-generated text. In this paper, we systematically test the reliability of the existing detectors, by designing two types of attack strategies to fool the detectors: 1) replacing words with their synonyms based on the context; 2) altering the writing style of generated text. These strategies are implemented by instructing LLMs to generate synonymous word substitutions or writing directives that modify the style without human involvement, and the LLMs leveraged in the attack can also be protected by detectors. Our research reveals that our attacks effectively compromise the performance of all tested detectors, thereby underscoring the urgent need for the development of more robust machine-generated text detection systems.

Dynosaur: A Dynamic Growth Paradigm for Instruction-Tuning Data Curation

Instruction tuning has emerged to enhance the capabilities of large language models (LLMs) in providing appropriate outputs based on input instructions. However, existing methods for collecting instruction-tuning data suffer from limitations in scalability and affordability. In this paper, we propose Dynosaur, a dynamic growth paradigm for instruction-tuning data curation. Built upon the metadata of existing NLP datasets, we generate multiple task instructions applicable to various NLP datasets and determine the relevant data fields for constructing instruction-tuning data with LLMs. Dynosaur offers several advantages: 1) lower generation costs (less than $12 for generating 800K instruction-tuning data), 2) good quality of instruction-tuning data (better performance than Alpaca and Instruction GPT-4 on Super-NI with comparable data sizes), and 3) the ability to grow dynamically by incorporating new datasets from Huggingface Datasets Platform. We further investigate continual learning as an approach to learning with the ever-growing instruction-tuning dataset. We demonstrate that replay methods not only help mitigate forgetting issues but help generalize to unseen tasks better. As a novel continual learning scenario for instruction tuning, selecting tasks based on instruction representations can be an effective replaying strategy. Code and data are released at \url{https://github.com/WadeYin9712/Dynosaur}.