We introduce Inference-Time Intervention (ITI), a technique designed to enhance the "truthfulness" of large language models (LLMs). ITI operates by shifting model activations during inference, following a set of directions across a limited number of attention heads. This intervention significantly improves the performance of LLaMA models on the TruthfulQA benchmark. On an instruction-finetuned LLaMA called Alpaca, ITI improves its truthfulness from 32.5% to 65.1%. We identify a trade-off between truthfulness and helpfulness and demonstrate how to balance it by tuning the intervention strength. ITI is minimally invasive and computationally inexpensive. Moreover, the technique is data efficient: while approaches like RLHF require extensive annotations, ITI locates truthful directions using only few hundred examples. Our findings suggest that LLMs may have an internal representation of the likelihood of something being true, even as they produce falsehoods on the surface.

The prevalence and strong capability of large language models (LLMs) present significant safety and ethical risks if exploited by malicious users. To prevent the potentially deceptive usage of LLMs, recent works have proposed algorithms to detect LLM-generated text and protect LLMs. In this paper, we investigate the robustness and reliability of these LLM detectors under adversarial attacks. We study two types of attack strategies: 1) replacing certain words in an LLM's output with their synonyms given the context; 2) automatically searching for an instructional prompt to alter the writing style of the generation. In both strategies, we leverage an auxiliary LLM to generate the word replacements or the instructional prompt. Different from previous works, we consider a challenging setting where the auxiliary LLM can also be protected by a detector. Experiments reveal that our attacks effectively compromise the performance of all detectors in the study with plausible generations, underscoring the urgent need to improve the robustness of LLM-generated text detection systems.

This paper introduces PMIndiaSum, a multilingual and massively parallel summarization corpus focused on languages in India. Our corpus provides a training and testing ground for four language families, 14 languages, and the largest to date with 196 language pairs. We detail our construction workflow including data acquisition, processing, and quality assurance. Furthermore, we publish benchmarks for monolingual, cross-lingual, and multilingual summarization by fine-tuning, prompting, as well as translate-and-summarize. Experimental results confirm the crucial role of our data in aiding summarization between Indian languages. Our dataset is publicly available and can be freely modified and re-distributed.

A salient characteristic of pre-trained language models (PTLMs) is a remarkable improvement in their generalization capability and emergence of new capabilities with increasing model capacity and pre-training dataset size. Consequently, we are witnessing the development of enormous models pushing the state-of-the-art. It is, however, imperative to realize that this inevitably leads to prohibitively long training times, extortionate computing costs, and a detrimental environmental impact. Significant efforts are underway to make PTLM training more efficient through innovations in model architectures, training pipelines, and loss function design, with scant attention being paid to optimizing the utility of training data. The key question that we ask is whether it is possible to train PTLMs by employing only highly informative subsets of the training data while maintaining downstream performance? Building upon the recent progress in informative data subset selection, we show how we can employ submodular optimization to select highly representative subsets of the training corpora and demonstrate that the proposed framework can be applied to efficiently train multiple PTLMs (BERT, BioBERT, GPT-2) using only a fraction of data. Further, we perform a rigorous empirical evaluation to show that the resulting models achieve up to $\sim99\%$ of the performance of the fully-trained models. We made our framework publicly available at https://github.com/Efficient-AI/ingenious.

Growing literature has shown that NLP systems may encode social biases; however, the political bias of summarization models remains relatively unknown. In this work, we use an entity replacement method to investigate the portrayal of politicians in automatically generated summaries of news articles. We develop an entity-based computational framework to assess the sensitivities of several extractive and abstractive summarizers to the politicians Donald Trump and Joe Biden. We find consistent differences in these summaries upon entity replacement, such as reduced emphasis of Trump's presence in the context of the same article and a more individualistic representation of Trump with respect to the collective US government (i.e., administration). These summary dissimilarities are most prominent when the entity is heavily featured in the source article. Our characterization provides a foundation for future studies of bias in summarization and for normative discussions on the ideal qualities of automatic summaries.

We define extensions of CTL and TCTL with strategic operators, called Strategic CTL (SCTL) and Strategic TCTL (STCTL), respectively. For each of the above logics we give a synchronous and asynchronous semantics, i.e., STCTL is interpreted over networks of extended Timed Automata (TA) that either make synchronous moves or synchronise via joint actions. We consider several semantics regarding information: imperfect (i) and perfect (I), and recall: imperfect (r) and perfect (R). We prove that SCTL is more expressive than ATL for all semantics, and this holds for the timed versions as well. Moreover, the model checking problem for SCTL[ir] is of the same complexity as for ATL[ir], the model checking problem for STCTL[ir] is of the same complexity as for TCTL, while for STCTL[iR] it is undecidable as for ATL[iR]. The above results suggest to use SCTL[ir] and STCTL[ir] in practical applications. Therefore, we use the tool IMITATOR to support model checking of STCTL[ir].

Differential Privacy (DP) is a well-established framework to quantify privacy loss incurred by any algorithm. Traditional formulations impose a uniform privacy requirement for all users, which is often inconsistent with real-world scenarios in which users dictate their privacy preferences individually. This work considers the problem of mean estimation, where each user can impose their own distinct privacy level. The algorithm we propose is shown to be minimax optimal and has a near-linear run-time. Our results elicit an interesting saturation phenomenon that occurs. Namely, the privacy requirements of the most stringent users dictate the overall error rates. As a consequence, users with less but differing privacy requirements are all given more privacy than they require, in equal amounts. In other words, these privacy-indifferent users are given a nontrivial degree of privacy for free, without any sacrifice in the performance of the estimator.

Gibbs posteriors are proportional to a prior distribution multiplied by an exponentiated loss function, with a key tuning parameter weighting information in the loss relative to the prior and providing a control of posterior uncertainty. Gibbs posteriors provide a principled framework for likelihood-free Bayesian inference, but in many situations, including a single tuning parameter inevitably leads to poor uncertainty quantification. In particular, regardless of the value of the parameter, credible regions have far from the nominal frequentist coverage even in large samples. We propose a sequential extension to Gibbs posteriors to address this problem. We prove the proposed sequential posterior exhibits concentration and a Bernstein-von Mises theorem, which holds under easy to verify conditions in Euclidean space and on manifolds. As a byproduct, we obtain the first Bernstein-von Mises theorem for traditional likelihood-based Bayesian posteriors on manifolds. All methods are illustrated with an application to principal component analysis.

If machine learning models were to achieve superhuman abilities at various reasoning or decision-making tasks, how would we go about evaluating such models, given that humans would necessarily be poor proxies for ground truth? In this paper, we propose a framework for evaluating superhuman models via consistency checks. Our premise is that while the correctness of superhuman decisions may be impossible to evaluate, we can still surface mistakes if the model's decisions fail to satisfy certain logical, human-interpretable rules. We instantiate our framework on three tasks where correctness of decisions is hard to evaluate due to either superhuman model abilities, or to otherwise missing ground truth: evaluating chess positions, forecasting future events, and making legal judgments. We show that regardless of a model's (possibly superhuman) performance on these tasks, we can discover logical inconsistencies in decision making. For example: a chess engine assigning opposing valuations to semantically identical boards; GPT-4 forecasting that sports records will evolve non-monotonically over time; or an AI judge assigning bail to a defendant only after we add a felony to their criminal record.

Former CIA director Gen. David Petraeus spoke with Fox News Digital about the "breathtaking" pace of AI development and what responsibility the US has to keep allies on the "cutting edge." The U.S. has no responsibility to keep its allies on the "cutting edge" of artificial intelligence (AI) development – unless it is a matter of national security, a former CIA director and retired Army officer tells Fox News Digital. "First and foremost is to ensure that we are on the cutting edge," retired four-star Gen. David Petraeus said during a recent Zoom interview. "Certainly, we should share to varying degrees with our closest partners, and they share with us – again, no one of us is smarter than all of us together in these kinds of endeavors – but it's not our job entirely to ensure that they are proceeding along these lines, unless it's, of course, in our national interest to do so, and it is in a number of different cases," he explained. The pace of AI development has dominated conversation since public access to ChatGPT in November 2022, particularly with concerns over who will stay at the top of the game – a race that drove countries to reassess their investments in the burgeoning field.