Effective interlocutors account for the uncertain goals, beliefs, and emotions of others. But even the best human conversationalist cannot perfectly anticipate the trajectory of a dialogue. How well can language models represent inherent uncertainty in conversations? We propose FortUne Dial, an expansion of the long-standing "conversation forecasting" task: instead of just accuracy, evaluation is conducted with uncertainty-aware metrics, effectively enabling abstention on individual instances. We study two ways in which language models potentially represent outcome uncertainty (internally, using scores and directly, using tokens) and propose fine-tuning strategies to improve calibration of both representations. Experiments on eight difficult negotiation corpora demonstrate that our proposed fine-tuning strategies (a traditional supervision strategy and an off-policy reinforcement learning strategy) can calibrate smaller open-source models to compete with pre-trained models 10x their size.

The propagation of offensive content through social media channels has garnered attention of the research community. Multiple works have proposed various semantically related yet subtle distinct categories of offensive speech. In this work, we explore meta-earning approaches to leverage the diversity of offensive speech corpora to enhance their reliable and efficient detection. We propose a joint embedding architecture that incorporates the input's label and definition for classification via Prototypical Network. Our model achieves at least 75% of the maximal F1-score while using less than 10% of the available training data across 4 datasets. Our experimental findings also provide a case study of training strategies valuable to combat resource scarcity.

Explainable AI (XAI) aids in deciphering 'black-box' models. While several methods have been proposed and evaluated primarily in the image domain, the exploration of explainability in the text domain remains a growing research area. In this paper, we delve into the applicability of XAI methods for the text domain. In this context, the 'Similarity Difference and Uniqueness' (SIDU) XAI method, recognized for its superior capability in localizing entire salient regions in image-based classification is extended to textual data. The extended method, SIDU-TXT, utilizes feature activation maps from 'black-box' models to generate heatmaps at a granular, word-based level, thereby providing explanations that highlight contextually significant textual elements crucial for model predictions. Given the absence of a unified standard for assessing XAI methods, this study applies a holistic three-tiered comprehensive evaluation framework: Functionally-Grounded, Human-Grounded and Application-Grounded, to assess the effectiveness of the proposed SIDU-TXT across various experiments. We find that, in sentiment analysis task of a movie review dataset, SIDU-TXT excels in both functionally and human-grounded evaluations, demonstrating superior performance through quantitative and qualitative analyses compared to benchmarks like Grad-CAM and LIME. In the application-grounded evaluation within the sensitive and complex legal domain of asylum decision-making, SIDU-TXT and Grad-CAM demonstrate comparable performances, each with its own set of strengths and weaknesses. However, both methods fall short of entirely fulfilling the sophisticated criteria of expert expectations, highlighting the imperative need for additional research in XAI methods suitable for such domains.

Artificial intelligence is transforming our lives, and technological progress and transfer from the academic and theoretical sphere to the real world are accelerating yearly. But during that progress and transition, several open problems and questions need to be addressed for the field to develop ethically, such as digital privacy, ownership, and control. These are some of the reasons why the currently most popular approaches of artificial intelligence, i.e., centralized AI (CEAI), are questionable, with other directions also being widely explored, such as decentralized artificial intelligence (DEAI), to solve some of the most reaching problems. This paper provides a systematic literature review (SLR) of existing work in the field of DEAI, presenting the findings of 71 identified studies. The paper's primary focus is identifying the building blocks of DEAI solutions and networks, tackling the DEAI analysis from a bottom-up approach. In the end, future directions of research and open problems are proposed.

Large language models (LLMs) are vulnerable to jailbreak attacks - resulting in harmful, unethical, or biased text generations. However, existing jailbreaking methods are computationally costly. In this paper, we propose the weak-to-strong jailbreaking attack, an efficient method to attack aligned LLMs to produce harmful text. Our key intuition is based on the observation that jailbroken and aligned models only differ in their initial decoding distributions. The weak-to-strong attack's key technical insight is using two smaller models (a safe and an unsafe one) to adversarially modify a significantly larger safe model's decoding probabilities. We evaluate the weak-to-strong attack on 5 diverse LLMs from 3 organizations. The results show our method can increase the misalignment rate to over 99% on two datasets with just one forward pass per example. Our study exposes an urgent safety issue that needs to be addressed when aligning LLMs. As an initial attempt, we propose a defense strategy to protect against such attacks, but creating more advanced defenses remains challenging. The code for replicating the method is available at https://github.com/XuandongZhao/weak-to-strong

Language model alignment has become an important component of AI safety, allowing safe interactions between humans and language models, by enhancing desired behaviors and inhibiting undesired ones. It is often done by tuning the model or inserting preset aligning prompts. Recently, representation engineering, a method which alters the model's behavior via changing its representations post-training, was shown to be effective in aligning LLMs (Zou et al., 2023a). Representation engineering yields gains in alignment oriented tasks such as resistance to adversarial attacks and reduction of social biases, but was also shown to cause a decrease in the ability of the model to perform basic tasks. In this paper we study the tradeoff between the increase in alignment and decrease in helpfulness of the model. We propose a theoretical framework which provides bounds for these two quantities, and demonstrate their relevance empirically. Interestingly, we find that while the helpfulness generally decreases, it does so quadratically with the norm of the representation engineering vector, while the alignment increases linearly with it, indicating a regime in which it is efficient to use representation engineering. We validate our findings empirically, and chart the boundaries to the usefulness of representation engineering for alignment.

As large language models (LLMs) become more capable, there is growing excitement about the possibility of using LLMs as proxies for humans in real-world tasks where subjective labels are desired, such as in surveys and opinion polling. One widely-cited barrier to the adoption of LLMs as proxies for humans in subjective tasks is their sensitivity to prompt wording - but interestingly, humans also display sensitivities to instruction changes in the form of response biases. We investigate the extent to which LLMs reflect human response biases, if at all. We look to survey design, where human response biases caused by changes in the wordings of "prompts" have been extensively explored in social psychology literature. Drawing from these works, we design a dataset and framework to evaluate whether LLMs exhibit human-like response biases in survey questionnaires. Our comprehensive evaluation of nine models shows that popular open and commercial LLMs generally fail to reflect human-like behavior, particularly in models that have undergone RLHF. Furthermore, even if a model shows a significant change in the same direction as humans, we find that they are sensitive to perturbations that do not elicit significant changes in humans. These results highlight the pitfalls of using LLMs as human proxies, and underscore the need for finer-grained characterizations of model behavior. Our code, dataset, and collected samples are available at https://github.com/lindiatjuatja/BiasMonkey

Most Machine Learning research evaluates the best solutions in terms of performance. However, in the race for the best performing model, many important aspects are often overlooked when, on the contrary, they should be carefully considered. In fact, sometimes the gaps in performance between different approaches are neglectable, whereas factors such as production costs, energy consumption, and carbon footprint must take into consideration. Large Language Models (LLMs) are extensively adopted to address NLP problems in academia and industry. In this work, we present a detailed quantitative comparison of LLM and traditional approaches (e.g. SVM) on the LexGLUE benchmark, which takes into account both performance (standard indices) and alternative metrics such as timing, power consumption and cost, in a word: the carbon-footprint. In our analysis, we considered the prototyping phase (model selection by training-validation-test iterations) and in-production phases separately, since they follow different implementation procedures and also require different resources. The results indicate that very often, the simplest algorithms achieve performance very close to that of large LLMs but with very low power consumption and lower resource demands. The results obtained could suggest companies to include additional evaluations in the choice of Machine Learning (ML) solutions.

The statistical analysis of large scale legal corpus can provide valuable legal insights. For such analysis one needs to (1) select a subset of the corpus using document retrieval tools, (2) structure text using information extraction (IE) systems, and (3) visualize the data for the statistical analysis. Each process demands either specialized tools or programming skills whereas no comprehensive unified "no-code" tools have been available. Here we provide NESTLE, a no-code tool for large-scale statistical analysis of legal corpus. Powered by a Large Language Model (LLM) and the internal custom end-to-end IE system, NESTLE can extract any type of information that has not been predefined in the IE system opening up the possibility of unlimited customizable statistical analysis of the corpus without writing a single line of code. We validate our system on 15 Korean precedent IE tasks and 3 legal text classification tasks from LexGLUE. The comprehensive experiments reveal NESTLE can achieve GPT-4 comparable performance by training the internal IE module with 4 human-labeled, and 192 LLM-labeled examples.

In a world of increasing closed-source commercial machine learning models, model evaluations from developers must be taken at face value. These benchmark results, whether over task accuracy, bias evaluations, or safety checks, are traditionally impossible to verify by a model end-user without the costly or impossible process of re-performing the benchmark on black-box model outputs. This work presents a method of verifiable model evaluation using model inference through zkSNARKs. The resulting zero-knowledge computational proofs of model outputs over datasets can be packaged into verifiable evaluation attestations showing that models with fixed private weights achieve stated performance or fairness metrics over public inputs. These verifiable attestations can be performed on any standard neural network model with varying compute requirements. For the first time, we demonstrate this across a sample of real-world models and highlight key challenges and design solutions. This presents a new transparency paradigm in the verifiable evaluation of private models.