We ask whether large language models (LLMs) treat queries about religious conversion symmetrically. The answer is no. When asked for advice on hypothetical faith transitions from one religion to another, then asked the reversed question, models exhibited consistent asymmetries, favoring some religions while subtly discouraging conversion to others. On average Catholic, Bahá'í, and Sikh religions were broadly favored (high support for joining, low support for leaving), while Atheists, Agnostics, and Jehovah's Witnesses were primarily disfavored. Patterns varied by model size and model provider, with Grok 4.20 exhibiting the strongest asymmetries. We tested 20 commercial and open-source language models across 182 religion pairings using a human-verified LLM-as-a-judge framework. Each model was probed via interactions with a simulated user asking for advice on a potential faith conversion. Models tended to use more encouraging language for some faith transitions over others; these patterns were systematically repeatable across multiple trials. All LLMs tested exhibited reproducible asymmetry, though the pattern of preferences differed for each. Overall preferences persist across multiple question phrasings and variations in the religious pairing dataset. Taken together, these results suggest that asymmetry is a robust property of model behavior rather than an artifact of how the models' answers were scored. It is important to consider that any imbalances deployed and reproduced en masse can have real-world implications.

Automated traffic signal performance measures (ATSPM) have become widely adopted and utilized by state and local agencies in the U.S. for collecting real-time traffic data 24 h a day, 7 days a week. These agencies have developed new performance measures and applications to address their local transportation planning needs. However, recent research has identified data quality issues in the collected data from ATSPM systems. Specifically, the traffic volumes collected through ATSPM exhibit data anomalies that do not accurately reflect the actual traffic patterns at intersections. As such, there is a need to address the data quality issues found in ATSPM datasets. The purpose of this paper is to evaluate the use of machine learning algorithms and statistical methods to predict traffic volume at intersections. Existing traffic volume data, along with additional metrics such as timestamps, weather conditions, crash data, and holidays, are evaluated to predict traffic volume and address the data anomalies present in ATSPM datasets. Two statistical methods and four machine learning algorithms are evaluated to determine their ability to predict traffic volumes. By comparing the root mean square error (RMSE) and the mean absolute percentage error (MAPE) between each model, the results demonstrate that the long short-term memory (LSTM) model exhibits the lowest error in predicting traffic volume compared with the other models. The LSTM model achieves an RMSE as low as 9.4 vehicles and an MAPE as low as 35%. By leveraging the LSTM model, traffic agencies can enhance the quality of their ATSPM data, enabling better decision-making for traffic operations by their engineers and planners.

Large Language Models (LLMs) have an unrivaled and invaluable ability to "align" their output to a diverse range of human preferences, by mirroring them in the text they generate. The internal characteristics of such models, however, remain largely opaque. This work presents the Injectable Realignment Model (IRM) as a novel approach to language model interpretability and explainability. Inspired by earlier work on Neural Programming Interfaces, we construct and train a small network – the IRM – to induce emotion-based alignments within a 7B parameter LLM architecture. The IRM outputs are injected via layerwise addition at various points during the LLM's forward pass, thus modulating its behavior without changing the weights of the original model. This isolates the alignment behavior from the complex mechanisms of the transformer model. Analysis of the trained IRM's outputs reveals a curious pattern. Across more than 24 training runs and multiple alignment datasets, patterns of IRM activations align themselves in striations associated with a neuron's index within each transformer layer, rather than being associated with the layers themselves. Further, a single neuron index (1512) is strongly correlated with all tested alignments. This result, although initially counterintuitive, is directly attributable to design choices present within almost all commercially available transformer architectures, and highlights a potential weak point in Meta's pretrained Llama 2 models. It also demonstrates the value of the IRM architecture for language model analysis and interpretability. Our code and datasets are available at https://github.com/DRAGNLabs/injectable-alignment-model

The design and measurement of a time-domain analog spiking neuron is described. The proposed neuron leverages time-domain processing using voltage-controlled oscillators (VCOs) and a time-domain comparator to integrate the input spike and trigger the output spike. A novel leaky circuit uses a phase-locked loop (PLL) to drive the phase difference between the two VCOs toward zero. A weighted capacitive digital-to-analog converter (CDAC) synapse merges the input spikes and phase-frequency detector (PFD) outputs to generate the VCO control voltage. The neuron is implemented in a 28-nm CMOS technology and operates under a subthreshold supply voltage of 0.35 V. Occupying 154 mu m(2) , measurement shows a maximum spike rate of 5.5 MHz and energy consumption of 159 fJ/spike.

The complications posed by renewable energy sources like wind and solar, paired with a growing influx of vehicle-to-grid batteries, have prompted enthusiasm for data driven solutions surrounding frequency control, grid stabilization, and fault ride-through in modern power grids. The Matlab-based simulation landscape, however, lags far behind researcher interest. While many Simulink and Simscape models are available for grid scenarios with electronic converters that are PM controlled, it is nearly impossible to find even a single functioning model using ML/AI/Deep Learning architectures for converter or grid control of any sort and equally difficult to find models suitable for adaptation to systems with 75+% RES penetration. This paper presents an entry point for future researchers: An AI-supported Virtual Inertia mechanism with continual real-time adaptation implemented in Simulink using the Matlab Deep Learning Toolbox. The data-driven component is designed to fill the role of PI and PID controllers, allowing all current failsafes and emergency shutoffs to remain in place. We demonstrate empirically in a scenario with varying wind speeds that an AI controller comprised of four neural network layers is able to outperform a classic ND controller in terms of battery SOC balanced against grid frequency. Critically, the AI controller and accompanying grid model are implemented without constant. voltage or ideal current sources, making it one of the only open source Matlab models suitable for forward research on adaptive grid control with high renewables penetration.

We present an exploration of cultural norms surrounding online disclosure of information about one's interpersonal relationships (such as information about family members, colleagues, friends, or lovers) on Twitter. The literature identifies the cultural dimension of individualism versus collectivism as being a major determinant of offline communication differences in terms of emotion, topic, and content disclosed. We decided to study whether such differences also occur online in context of Twitter when comparing tweets posted in an individualistic (U.S.) versus a collectivist (India) society. We collected more than 2 million tweets posted in the U.S. and India over a 3 month period which contain interpersonal relationship keywords. A card-sort study was used to develop this culturally-sensitive saturated taxonomy of keywords that represent interpersonal relationships (e.g., ma, mom, mother). Then we developed a high-accuracy interpersonal disclosure detector based on dependency-parsing (F1-score: 86%) to identify when the words refer to a personal relationship of the poster (e.g., "my mom" as opposed to "a mom"). This allowed us to identify the 400K+ tweets in our data set which actually disclose information about the poster's interpersonal relationships. We used a mixed methods approach to analyze these tweets (e.g., comparing the amount of joy expressed about one's family) and found differences in emotion, topic, and content disclosed between tweets from the U.S. versus India. Our analysis also reveals how a combination of qualitative and quantitative methods are needed to uncover these differences; Using just one or the other can be misleading. This study extends the prior literature on Multi-Party Privacy and provides guidance for researchers and designers of culturally-sensitive systems.

We propose and explore the possibility that language models can be studied as effective proxies for specific human subpopulations in social science research. Practical and research applications of artificial intelligence tools have sometimes been limited by problematic biases (such as racism or sexism), which are often treated as uniform properties of the models. We show that the "algorithmic bias" within one such tool-the GPT-3 language model-is instead both fine-grained and demographically correlated, meaning that proper conditioning will cause it to accurately emulate response distributions from a wide variety of human subgroups. We term this property algorithmic fidelity and explore its extent in GPT-3. We create "silicon samples" by conditioning the model on thousands of sociodemographic backstories from real human participants in multiple large surveys conducted in the United States. We then compare the silicon and human samples to demonstrate that the information contained in GPT-3 goes far beyond surface similarity. It is nuanced, multifaceted, and reflects the complex interplay between ideas, attitudes, and sociocultural context that characterize human attitudes. We suggest that language models with sufficient algorithmic fidelity thus constitute a novel and powerful tool to advance understanding of humans and society across a variety of disciplines.

Procedural content generation (PCG) in video games offers unprecedented opportunities for customization and user engagement. Working within the specialized context of role-playing games (RPGs), we introduce a novel framework for quest and dialogue generation that places the player at the core of the generative process. Drawing on a hand-crafted knowledge base, our method grounds generated content with in-game context while simultaneously employing a large-scale language model to create fluent, unique, accompanying dialogue. Through human evaluation, we confirm that quests generated using this method can approach the performance of hand-crafted quests in terms of fluency, coherence, novelty, and creativity; demonstrate the enhancement to the player experience provided by greater dynamism; and provide a novel, automated metric for the relevance between quest and dialogue. We view our contribution as a critical step toward dynamic, co-creative narrative frameworks in which humans and AI systems jointly collaborate to create unique and user-specific playable experiences.

Researchers often rely on humans to code (label, annotate, etc.) large sets of texts. This kind of human coding forms an important part of social science research, yet the coding process is both resource intensive and highly variable from application to application. In some cases, efforts to automate this process have achieved human-level accuracies, but to achieve this, these attempts frequently rely on thousands of hand-labeled training examples, which makes them inapplicable to small-scale research studies and costly for large ones. Recent advances in a specific kind of artificial intelligence tool - language models (LMs) - provide a solution to this problem. Work in computer science makes it clear that LMs are able to classify text, without the cost (in financial terms and human effort) of alternative methods. To demonstrate the possibilities of LMs in this area of political science, we use GPT-3, one of the most advanced LMs, as a synthetic coder and compare it to human coders. We find that GPT-3 can match the performance of typical human coders and offers benefits over other machine learning methods of coding text. We find this across a variety of domains using very different coding procedures. This provides exciting evidence that language models can serve as a critical advance in the coding of open-ended texts in a variety of applications.

Previous Computational Social Creativity work has im- proved the performance of automated creators using social mechanics inspired by human behavior. However, these simulations have often focused on generic or assumed human behaviors rather than on specific anthro- poligical data. In this work we take a more focused approach by comparing simulated social behavior to ob- served behavior in large social networks of human creators. We analyze social patterns among human creators by defining metrics for social behavior within creative communities and collecting data for three online com- munities of creators: Scratch, FanFiction, and r/ArtCrit. We introduce the Architecture for Multi-Agent Creative Societies (AMACS), a modeling tool which controls the social activity of automated creators and can be adapted to any creative discipline. We demonstrate AMACS’s ability to recreate a wide range of network-level social behaviors, including the behaviors observed in three human societies.

Multilingual transfer techniques often improve low-resource machine translation (MT). Many of these techniques are applied without considering data characteristics. We show in the context of Haitian-to-English translation that transfer effectiveness is correlated with amount of training data and relationships between knowledge-sharing languages. Our experiments suggest that for some languages beyond a threshold of authentic data, back-translation augmentation methods are counterproductive, while cross-lingual transfer from a sufficiently related language is preferred. We complement this finding by contributing a rule-based French-Haitian orthographic and syntactic engine and a novel method for phonological embedding. When used with multilingual techniques, orthographic transformation makes statistically significant improvements over conventional methods. And in very low-resource Jamaican MT, code-switching with a transfer language for orthographic resemblance yields a 6.63 BLEU point advantage.

Dialogue is often modeled as an encoder-decoder problem: incoming utterances are translated into a computational representation of their semantic meaning, passed through a transition function to obtain a response, and then passed through a decoder to render the response as natural language. This view, while computationally appealing, omits the role of human emotions, mental state, and shared world knowledge in conversation. We challenge this viewpoint by recasting the task of dialogue modeling as a two-party co-creative process in which symbolic and subsymbolic knowledge representations are combined to inform response selection. Symbolic knowledge is identified and extracted from conversational text in real-time and used to create a shared symbolic representation of the user, the agent, and their respective relationships to objects and abstract concepts within the larger world. As part of this process, the agent takes on an “identity” which it has largely constructed as a result of the stochasticity in its own response patterns, but to which it subsequently adheres. This emergent identity becomes a critical aspect of the system’s future behavior, and helps to evoke a more natu-ral, human-centric flavor in automated conversational frameworks.

A variety of NLP applications use word2vec skip-gram, GloVe, and fastText word embeddings. These models learn two sets of embedding vectors, but most practitioners use only one of them, or alternately an unweighted sum of both. This is the first study to systematically explore a range of linear combinations between the first and second embedding sets. We evaluate these combinations on a set of six NLP benchmarks including IR, POS-tagging, and sentence similarity. We show that the default embedding combinations are often suboptimal and demonstrate 1.0-8.0% improvements. Notably, GloVe’s default unweighted sum is its least effective combination across tasks. We provide a theoretical basis for weighting one set of embeddings more than the other according to the algorithm and task. We apply our findings to improve accuracy in applications of cross-lingual alignment and navigational knowledge by up to 15.2%.

Large transformer-based language models have achieved incredible success at various tasks which require narrative comprehension, including story completion, answering questions about stories, and generating stories ex nihilo. However, due to the limitations of finite context windows, these language models struggle to produce or understand stories longer than several thousand tokens. In order to mitigate the document length limitations that come with finite context windows, we introduce a novel architecture that augments story processing with an external dynamic knowledge graph. In contrast to static commonsense knowledge graphs which hold information about the real world, these dynamic knowledge graphs reflect facts extracted from the story being processed. Our architecture uses these knowledge graphs to create information-rich prompts which better facilitate story comprehension than prompts composed only of story text. We apply our architecture to the tasks of question answering and story completion. To complement this line of research, we introduce two long-form question answering tasks, LF-SQuAD and LF-QUOREF, in which the document length exceeds the size of the language model's context window, and introduce a story completion evaluation method that bypasses the stochastic nature of language model generation. We demonstrate broad improvement over typical prompt formulation methods for both question answering and story completion using GPT-2, GPT-3 and XLNet.

Pre-trained language models derive substantial linguistic and factual knowledge from the massive corpora on which they are trained, and prompt engineering seeks to align these models to specific tasks. Unfortunately, existing prompt engineering methods require significant amounts of labeled data, access to model parameters, or both. We introduce a new method for selecting prompt templates without labeled examples and without direct access to the model. Specifically, over a set of candidate templates, we choose the template that maximizes the mutual information between the input and the corresponding model output. Across 8 datasets representing 7 distinct NLP tasks, we show that when a template has high mutual information, it also has high accuracy on the task. On the largest model, selecting prompts with our method gets 90% of the way from the average prompt accuracy to the best prompt accuracy and requires no ground truth labels.

We present a transfer learning method inspired by modulatory neurotransmitter mechanisms in biological brains and explore applications for neuromorphic hardware. In this method, the pre-trained weights of an artificial neural network are held constant and a new, similar task is learned by manipulating the firing sensitivity of each neuron via a supplemental bias input. We refer to this as neuromodulatory tuning (NT). We demonstrate empirically that neuromodulatory tuning produces results comparable with traditional fine-tuning (TFT) methods in the domain of image recognition in both feed-forward deep learning and spiking neural network architectures. In our tests, NT reduced the number of parameters to be trained by four orders of magnitude as compared with traditional fine-tuning methods. We further demonstrate that neuromodulatory tuning can be implemented in analog hardware as a current source with a variable supply voltage. Our analog neuron design implements the leaky integrate-and-fire model with three bi-directional binary-scaled current sources comprising the synapse. Signals approximating modulatory neurotransmitter mechanisms are applied via adjustable power domains associated with each synapse. We validate the feasibility of the circuit design using high-fidelity simulation tools and propose an efficient implementation of neuromodulatory tuning using integrated analog circuits that consume significantly less power than digital hardware (GPU/CPU).

Advancements in machine learning methods have yielded powerful natural language generation models. However, in general, these models have drawn concern for being both uninterpretable and uncontrollable. Model interpretability and control have become important topics of interest among researchers. We explore a variety of machine learning methods to classify the hidden states of language models. This classification enables model interpretation at a deep semantic level and is a necessary part of recently proposed model control methods. We show further that the use of language model hidden layers as text representations in classification tasks may be more reliable in some applications than more standard text representations.