This article evaluates the use of probing classifiers to modify the internal hidden state of a chess-playing transformer, which has been trained on sequences of chess moves and can generate new moves with prompted. Probing classifiers are a technique for understanding and modifying the operation of neural networks in which a smaller classifier is trained to use the model's internal representation to learn a probing task. The aim of this research is to discover whether the learned model possesses an editable internal representation of the chess game, despite being trained without explicit information about the rules of chess. We contrast the performance of standard linear probes against Sparse Autoencoders (SAEs), a latent space interpretability technique designed to decompose polysemantic concepts into atomic features via an overcomplete basis. Our experiments demonstrate that linear probes trained directly on the residual stream significantly outperform probes based on SAE latents. When quantifying the success of interventions via the probability of legal moves, linear probe edits achieved an 88% success rate, whereas SAE-based edits yielded only 41%. These findings suggest that while SAEs are valuable for specific interpretability tasks, they do not enhance the controllability of hidden states compared to raw vectors. Finally, we show that the residual stream respects the Markovian property of chess, validating the feasibility of applying consistent edits across different time steps for the same board state.

Probing classifiers are a technique for understanding and modifying the operation of neural networks in which a smaller classifier is trained to use the model’s internal representation to learn a probing task. Similar to a neural electrode array, probing classifiers help both discern and edit the internal representation of a neural network. This article evaluates the use of probing classifiers to modify the internal hidden state of a chess-playing transformer. We contrast the performance of standard linear probes against Sparse Autoencoders (SAEs), a latent space interpretability technique designed to decompose polysemantic concepts into atomic features via an overcomplete basis. Our experiments demonstrate that linear probes trained directly on the residual stream significantly outperform probes based on SAE latents. When quantifying the success of interventions via the probability of legal moves, linear probe edits achieved an 88% success rate, whereas SAE-based edits yielded only 41%. These findings suggest that while SAEs are valuable for specific interpretability tasks, they do not enhance the controllability of hidden states compared to raw vectors. Finally, we show that the residual stream respects the Markovian property of chess, validating the feasibility of applying consistent edits across different time steps for the same board state.

Traditional social navigation systems often treat perception and motion as decoupled tasks, leading to reactive behaviors and perceptual surprise due to limited field of view. While active vision—the ability to choose where to look—offers a solution, most existing frameworks decouple sensing from execution to simplify the learning process. This article introduces a novel joint reinforcement learning (RL) framework (Active Vision for Social Navigation) that unifies locomotion and discrete gaze control within a single, end-to-end policy. Unlike existing factored approaches, our method leverages a model-based RL architecture with a latent world model to explicitly address the credit assignment problem inherent in active sensing. Experimental results in cluttered, dynamic environments demonstrate that our joint policy outperforms factored sensing-action approaches by prioritizing viewpoints specifically relevant to social safety, such as checking blind spots and tracking human trajectories. Our findings suggest that tight sensorimotor coupling is essential for reducing perceptual surprise and ensuring safe, socially aware navigation in unstructured spaces.

Probing classifiers are a technique for understanding and modifying the operation of neural networks in which a smaller classifier is trained to use the model's internal representation to learn a related probing task. Similar to a neural electrode array, training probing classifiers can help researchers both discern and edit the internal representation of a neural network. This paper presents an evaluation of the use of probing classifiers to modify the internal hidden state of a chess-playing transformer. We demonstrate that intervention vector scaling should follow a negative exponential according to the length of the input to ensure model outputs remain semantically valid after editing the residual stream activations.

Public deliberation forums produce copious amounts of unorganized textual data reflecting diverse viewpoints. Visualization can serve as a valuable tool for understanding the relationships between policy preferences. We introduce a unique approach to visualizing opinion data gathered from the Polis online platform in which LLMs are used to generate positions and structure the data into argument maps. Each AI-generated position is supported by human-generated statements, providing a more meaningful organization of Polis's opinion data. We believe that these argument maps can provide policymakers with easy-to-understand visualizations that summarize public sentiment.

Navigating uneven, unstructured terrain with dynamic obstacles remains a challenge for autonomous mobile robots. This article introduces Dynamic Unstructured Environment (DUnE) for evaluating the performance of off-road navigation systems in simulation. DUnE is a versatile software framework that implements the Gymnasium reinforcement learning (RL) interface for ROS 2, incorporating unstructured Gazebo simulation environments and dynamic obstacle integration to advance off-road navigation research. The testbed automates key performance metric logging and provides semi-automated trajectory generation for dynamic obstacles including simulated human actors. It supports multiple robot platforms and five distinct unstructured environments, ranging from forests to rocky terrains. A baseline reinforcement learning agent demonstrates the framework’s effectiveness by performing pointgoal navigation with obstacle avoidance across various terrains. By providing an RL interface, dynamic obstacle integration, specialized navigation tasks, and comprehensive metric tracking, DUnE addresses significant gaps in existing simulation tools.

In this study, we explore the future of work by examining differences in productivity when teams are composed of only humans or both humans and machine agents. Our objective was to characterize the similarities and differences between human and human-machine teams as they work to coordinate across their specialized roles. This form of research is increasingly important given that machine agents are becoming commonplace in sociotechnical systems and playing a more active role in collaborative work. One particular class of machine agents, bots, is being introduced to these systems to facilitate both taskwork and teamwork. We investigated the association between bots and productivity outcomes in open source software development through an analysis of hundreds of project teams. The presence of bots in teams was associated with higher levels of productivity and higher work centralization in addition to greater amounts of observed communication. The adoption of bots in software teams may have tradeoffs, in that doing so may increase productivity, but could also increase workload. We discuss the theoretical and practical implications of these findings for advancing human-machine teaming research.

Mechanistic interpretability (MI) studies aim to identify the specific neural pathways that underlie decision-making in neural networks. Here we analyze both the horizontal and vertical information flows of a chess-playing transformer. This paper introduces a new taxonomy of chessboard attention patterns that synchronize to guide move selection. Our findings show that the early layers of the chess transformer correctly identify moves that are highly ranked by the final layer. Experiments conducted on human chess players laid the foundation for much of our current understanding of human problem-solving, cognition, and visual memory. We believe that the study of chess language transformers may be an equally fruitful research area for AGI systems.

Due to the rise in video content creation targeted towards children, there is a need for robust content moderation schemes for video hosting platforms. A video that is visually benign may include audio content that is inappropriate for young children while being impossible to detect with a unimodal content moderation system. Popular video hosting platforms for children such as YouTube Kids still publish videos which contain audio content that is not conducive to a child's healthy behavioral and physical development. A robust classification of malicious videos requires audio representations in addition to video features. However, recent content moderation approaches rarely employ multimodal architectures that explicitly consider non-speech audio cues. To address this, we present an efficient adaptation of CLIP (Contrastive Language–Image Pre-training) that can leverage contextual audio cues for enhanced content moderation. We incorporate 1) the audio modality and 2) prompt learning, while keeping the backbone modules of each modality frozen. We conduct our experiments on a multimodal version of the MOB (Malicious or Benign) dataset in supervised and few-shot settings.

Tech-related online communities on GitHub, Reddit, and Stack Overflow are an invaluable resource for software engineers, allowing them to find solutions to problems and connect with other professionals. Much of the discourse on these platforms is conducted using commenting mechanisms in which one user responds to content posted by another user. Even though these communities lack formal organizational structures, these technologists are often followed by other software developers who monitor their posts; users who regularly post useful solutions are recognized using platform-specific mechanisms such as stars or karma points. This paper investigates the relationship between popularity and discourse in tech-related online communities. To do this, we create comment timelines from sequences of user interactions and extract commenting networks from comment response patterns. Although there are some commonalities, there are distinct differences between the commenting behavior of GitHub users vs. Reddit and Stack Overflow. By understanding how popularity affects user interactions, we can design communities that are more effective at supporting learning, collaboration, and knowledge sharing.

Probing classifiers are a technique for understanding and modifying the operation of neural networks in which a smaller classifier is trained to use the model's internal representation to learn a probing task. Similar to a neural electrode array, probing classifiers help both discern and edit the internal representation of a neural network. This paper evaluates the use of probing classifiers to modify the internal hidden state of a chess-playing transformer. The weights of the learned linear classifiers are very informative and can be used to reliably delete pieces from the board showing that the model internally maintains an editable emergent representation of game state.

Many social media platforms promote political polarization by creating online echo chambers where people are only exposed to information confirming their beliefs. Newer systems such as Polis and Kialo aim to foster constructive conversations and teach critical reasoning skills. However, these platforms rely heavily on human moderators to manage discussions effectively. This paper examines the effectiveness of large language models (LLMs) as moderators on Polis, an open-source, real-time system designed for democratic discourse. We evaluate the F1 score of various prompting techniques at classifying five Polis datasets labeled by human moderators. Our findings indicate that LLMs are robust to different prompting strategies and produce minimal false positives. While LLMs come with certain risks, we argue that they can be valuable tools to support human moderators, enabling broader participation in democratic discourse.

The advent of low cost sensors for measuring gaze, heart rate, EEG, and galvanic skin response have made it feasible to cheaply collect physiological data from human operators. However, leveraging this data for machine learning problems requires a good multimodal fusion architecture. When dealing with multimodal features, uncovering the correlations between different modalities is as crucial as identifying effective unimodal features. This paper proposes a hybrid multimodal tensor fusion network that is effective at learning both unimodal and bimodal dynamics for cognitive workload modeling. Our architecture comprises two parts: (1) intra-modality for learning high-level representations of each signal modality (2) inter-modality for modeling bimodal interactions using a tensor fusion layer created from the Cartesian product of modality embeddings. We compare this architecture to the usage of a cross-modal transformer fusion module that learns an inter-modality embedding. Experimental results conducted on the HP Omnicept Cognitive Load Database (HPO-CLD) show that both techniques outperform the most commonly used techniques used for multimodal fusion of physio-logical data and that the cross-modal transformer fusion module is especially effective.

This paper introduces SC-Phi2, a fine-tuned StarCraft II small language model for macromanagement tasks. Small language models, like Phi2, Gemma, and DistilBERT, are streamlined versions of large language models (LLMs) with fewer parameters that require less power and memory to run. To teach Microsoft's Phi2 model about StarCraft, we create a new SC2 text dataset with information about StarCraft races, roles, and actions and use it to fine-tune Phi-2 with self-supervised learning. We pair this language model with a Vision Transformer (ViT) from the pre-trained BLIP-2 (Bootstrapping Language Image Pre-training) model, fine-tuning it on the MSC replay dataset. This enables us to construct dynamic prompts that include visual game state information. Unlike the large models used in StarCraft LLMs such as GPT-3.5, Phi2 is trained primarily on textbook data and contains little inherent knowledge of StarCraft II beyond what is provided by our training process. By using LoRA (Low-rank Adaptation) and quantization, our model can be trained on a single GPU. We demonstrate that our model performs well at micromanagement tasks such as build order and global state prediction with a small number of parameters.

Background: This article introduces SC-Phi2, a fine-tuned StarCraft II small language model. Small language models, like Phi2, Gemma, and DistilBERT, are streamlined versions of large language models (LLMs) with fewer parameters that require less computational power and memory to run. Method: To teach Microsoft’s Phi2 model about StarCraft, we create a new SC2 text dataset with information about StarCraft races, roles, and actions and use it to fine-tune Phi-2 with self-supervised learning. We pair this language model with a Vision Transformer (ViT) from the pre-trained BLIP-2 (Bootstrapping Language Image Pre-training) model, fine-tuning it on the StarCraft replay dataset, MSC. This enables us to construct dynamic prompts that include visual game state information. Results: Unlike the large models used in StarCraft LLMs such as GPT-3.5, Phi2 is trained primarily on textbook data and contains little inherent knowledge of StarCraft II beyond what is provided by our training process. By using LoRA (Low-rank Adaptation) and quantization, our model can be trained on a single GPU. We demonstrate that our model performs well at build order prediction, an important StarCraft macromanagement task. Conclusions: Our research on the usage of small models is a step towards reducing the carbon footprint of AI agents.

In humans, intrinsic motivation is an important mechanism for open-ended cognitive development; in robots, it has been shown to be valuable for exploration. An important aspect of human cognitive development is episodic memory which enables both the recollection of events from the past and the projection of subjective future. This paper explores the use of visual episodic memory as a source of intrinsic motivation for robotic exploration problems. Using a convolutional recurrent neural network autoencoder, the agent learns an efficient representation for spatiotemporal features such that accurate sequence prediction can only happen once spatiotemporal features have been learned. Structural similarity between ground truth and autoencoder generated images is used as an intrinsic motivation signal to guide exploration. Our proposed episodic memory model also implicitly accounts for the agent's actions, motivating the robot to seek new interactive experiences rather than just areas that are visually dissimilar. When guiding robotic exploration, our proposed method outperforms the Curiosity-driven Variational Autoencoder (CVAE) at finding dynamic anomalies.

We present a novel method aimed at enhancing the sample efficiency of ensemble Q learning. Our proposed approach integrates multi-head self-attention into the ensembled Q networks while bootstrapping the state-action pairs ingested by the ensemble. This not only results in performance improvements over the original REDQ and its variant DroQ, thereby enhancing Q predictions, but also effectively reduces both the average normalized bias and standard deviation of normalized bias within Q-function ensembles. Importantly, our method also performs well even in scenarios with a low update-to-data (UTD) ratio. Notably, the implementation of our proposed method is straightforward, requiring minimal modifications to the base model.

Human-centric visual understanding is an important desideratum for effective human-robot interaction. In order to navigate crowded public places, social robots must be able to interpret the activity of the surrounding humans. This paper addresses one key aspect of human-centric visual understanding, multi-person pose estimation. Achieving good performance on multi-person pose estimation in crowded scenes is difficult due to the challenges of occluded joints and instance separation. In order to tackle these challenges and overcome the limitations of image features in representing invisible body parts, we propose a novel prompt-based pose inference strategy called LAMP (Language Assisted Multi-person Pose estimation). By utilizing the text representations generated by a well-trained language model (CLIP), LAMP can facilitate the understanding of poses on the instance and joint levels, and learn more robust visual representations that are less susceptible to occlusion. This paper demonstrates that language-supervised training boosts the performance of single-stage multi-person pose estimation, and both instance-level and joint-level prompts are valuable for training. The code is available at https://github.com/shengnanh20/LAMP.

Natural language supervision has been shown to be effective for zero-shot learning in many computer vision tasks, such as object detection and activity recognition. However, generating informative prompts can be challenging for more subtle tasks, such as video content moderation. This can be difficult, as there are many reasons why a video might be inappropriate, beyond violence and obscenity. For example, scammers may attempt to create junk content that is similar to popular educational videos but with no meaningful information. This paper evaluates the performance of several CLIP variations for content moderation of children's cartoons in both the supervised and zero-shot setting. We show that our proposed model (Vanilla CLIP with Projection Layer) outperforms previous work conducted on the Malicious or Benign (MOB) benchmark for video content moderation. This paper presents an in depth analysis of how context-specific language prompts affect content moderation performance. Our results indicate that it is important to include more context in content moderation prompts, particularly for cartoon videos as they are not well represented in the CLIP training data.

We present an agent development and data collection framework for Generals.io (GIO)--a real-time strategy game with imperfect information in which players attempt to gain control of opponents' starting positions within a 2D grid world. The framework provides event-based communication amongst several modules implemented as microservices, enabling real-time data collection from GIO's streaming data. Its modular design facilitates rapid bot development and testing, while the emphasis on data collection makes it easy to analyze agent performance. We use this framework in a case study of a top-performing GIO agent called Flobot. Our analysis demonstrates that Flobot's performance varies based on its starting position. Based on the analysis performed with our framework, we propose a modification to Flobot's pathfinding algorithm. Statistical tests show that the new algorithm results in a significant reduction in performance variance.