Recent findings suggest that misaligned models may exhibit deceptive behavior, raising concerns about output trustworthiness. Chain-of-thought (CoT) is a promising tool for alignment monitoring: when models articulate their reasoning faithfully, monitors can detect and mitigate harmful behaviors before undesirable outcomes occur. However, a key uncertainty is: Can models obfuscate their CoT in order to pursue hidden adversarial objectives while evading detection? To answer this question and thus stress-test CoT monitorability, we develop a composable and quantifiable taxonomy of prompts to elicit CoT obfuscation. We evaluate both internal CoT (reasoning traces) and external CoT (prompted reasoning in outputs) using toy tasks and more realistic environments in SHADE-Arena. We show that: (i) CoT monitoring performs accurately and efficiently without obfuscation pressure. (ii) Under strong obfuscation pressure, some models successfully complete adversarial tasks while evading detection. (iii) Models do not obfuscate their internal CoT as much as their external CoT (under prompt pressure). These results suggest that while CoT provides valuable oversight in benign settings, robust deployment requires model-specific stress-testing of monitorability.

Immersive telepresence, when a user views the video stream of a $360^\circ$ camera in a remote environment using a Head Mounted Display (HMD), has great potential to improve the sense of being in a remote environment. In most cases of immersive robotic telepresence, the camera is mounted on a mobile robot which increases the portion of the environment that the remote user can explore. However, robot motions can induce unpleasant symptoms associated with Virtual Reality (VR) sickness, degrading the overall user experience. Previous research has shown that unwinding the rotations of the robot, that is, decoupling the rotations that the camera undergoes due to robot motions from what is seen by the user, can increase user comfort and reduce VR sickness. However, that work considered a virtual environment and a simulated robot. In this work, to test whether the same hypotheses hold when the video stream from a real camera is used, we carried out a user study $(n=36)$ in which the unwinding rotations method was compared against coupled rotations in a task completed through a panoramic camera mounted on a robotic arm. Furthermore, within an inspection task which involved translations and rotations in three dimensions, we tested whether unwinding the robot rotations impacted the performance of users. The results show that the users found the unwinding rotations method to be more comfortable and preferable, and that a reduced level of VR sickness can be achieved without a significant impact on task performance.

We stress test monitoring systems for detecting covert misbehavior in autonomous LLM agents (e.g., secretly sharing private information). To this end, we systematize a monitor red teaming (MRT) workflow that incorporates: (1) varying levels of agent and monitor situational awareness; (2) distinct adversarial strategies to evade the monitor, such as prompt injection; and (3) two datasets and environments -- SHADE-Arena for tool-calling agents and our new CUA-SHADE-Arena, which extends TheAgentCompany, for computer-use agents. We run MRT on existing LLM monitor scaffoldings, which orchestrate LLMs and parse agent trajectories, alongside a new hybrid hierarchical-sequential scaffolding proposed in this work. Our empirical results yield three key findings. First, agent awareness dominates monitor awareness: an agent's knowledge that it is being monitored substantially degrades the monitor's reliability. On the contrary, providing the monitor with more information about the agent is less helpful than expected. Second, monitor scaffolding matters more than monitor awareness: the hybrid scaffolding consistently outperforms baseline monitor scaffolding, and can enable weaker models to reliably monitor stronger agents -- a weak-to-strong scaling effect. Third, in a human-in-the-loop setting where humans discuss with the LLM monitor to get an updated judgment for the agent's behavior, targeted human oversight is most effective; escalating only pre-flagged cases to human reviewers improved the TPR by approximately 15% at FPR = 0.01. Our work establishes a standard workflow for MRT, highlighting the lack of adversarial robustness for LLMs and humans when monitoring and detecting agent misbehavior. We release code, data, and logs to spur further research.

Level 3 automated driving systems allows drivers to engage in secondary tasks while diminishing their perception of risk. In the event of an emergency necessitating driver intervention, the system will alert the driver with a limited window for reaction and imposing a substantial cognitive burden. To address this challenge, this study employs a Large Language Model (LLM) to assist drivers in maintaining an appropriate attention on road conditions through a " humanized " persuasive advice. Our tool leverages the road conditions encountered by Level 3 systems as triggers, proactively steering driver behavior via both visual and auditory routes. Empirical study indicates that our tool is effective in sustaining driver attention with reduced cognitive load and coordinating secondary tasks with takeover behavior. Our work provides insights into the potential of using LLMs to support drivers during multi-task automated driving. I. INTRODUCTION Level 3 automated driving systems allow drivers to perform secondary tasks while driving, yet drivers still need to pay attention to the road conditions .

Given the ubiquity of multi-task in practical systems, Multi-Task Learning (MTL) has found widespread application across diverse domains. In real-world scenarios, these tasks often have different priorities. For instance, In web search, relevance is often prioritized over other metrics, such as click-through rates or user engagement. Existing frameworks pay insufficient attention to the prioritization among different tasks, which typically adjust task-specific loss function weights to differentiate task priorities. However, this approach encounters challenges as the number of tasks grows, leading to exponential increases in hyper-parameter tuning complexity. Furthermore, the simultaneous optimization of multiple objectives can negatively impact the performance of high-priority tasks due to interference from lower-priority tasks. In this paper, we introduce a novel multi-task learning framework employing Lagrangian Differential Multiplier Methods for step-wise multi-task optimization. It is designed to boost the performance of high-priority tasks without interference from other tasks. Its primary advantage lies in its ability to automatically optimize multiple objectives without requiring balancing hyper-parameters for different tasks, thereby eliminating the need for manual tuning. Additionally, we provide theoretical analysis demonstrating that our method ensures optimization guarantees, enhancing the reliability of the process. We demonstrate its effectiveness through experiments on multiple public datasets and its application in Taobao search, a large-scale industrial search ranking system, resulting in significant improvements across various business metrics.

In this study, we generate and maintain a database of 10 million virtual lipids through METiS's in-house de novo lipid generation algorithms and lipid virtual screening techniques. These virtual lipids serve as a corpus for pre-training, lipid representation learning, and downstream task knowledge transfer, culminating in state-of-the-art LNP property prediction performance. We propose LipidBERT, a BERT-like model pre-trained with the Masked Language Model (MLM) and various secondary tasks. Additionally, we compare the performance of embeddings generated by LipidBERT and PhatGPT, our GPT-like lipid generation model, on downstream tasks. The proposed bilingual LipidBERT model operates in two languages: the language of ionizable lipid pre-training, using in-house dry-lab lipid structures, and the language of LNP fine-tuning, utilizing in-house LNP wet-lab data. This dual capability positions LipidBERT as a key AI-based filter for future screening tasks, including new versions of METiS de novo lipid libraries and, more importantly, candidates for in vivo testing for orgran-targeting LNPs. To the best of our knowledge, this is the first successful demonstration of the capability of a pre-trained language model on virtual lipids and its effectiveness in downstream tasks using web-lab data. This work showcases the clever utilization of METiS's in-house de novo lipid library as well as the power of dry-wet lab integration.

Mobile robots are increasingly being used in noisy environments for social purposes, e.g. to provide support in healthcare or public spaces. Since these robots also operate beyond human sight, the question arises as to how different robot types, ambient noise or cognitive engagement impacts the detection of the robots by their sound. To address this research gap, we conducted a user study measuring auditory detection distances for a wheeled (Turtlebot 2i) and quadruped robot (Unitree Go 1), which emit different consequential sounds when moving. Additionally, we also manipulated background noise levels and participants' engagement in a secondary task during the study. Our results showed that the quadruped robot sound was detected significantly better (i.e., at a larger distance) than the wheeled one, which demonstrates that the movement mechanism has a meaningful impact on the auditory detectability. The detectability for both robots diminished significantly as background noise increased. But even in high background noise, participants detected the quadruped robot at a significantly larger distance. The engagement in a secondary task had hardly any impact. In essence, these findings highlight the critical role of distinguishing auditory characteristics of different robots to improve the smooth human-centered navigation of mobile robots in noisy environments.

Large Language Models (LLMs) are routinely used in retrieval-augmented applications to orchestrate tasks and process inputs from users and other sources. These inputs, even in a single LLM interaction, can come from a variety of sources, of varying trustworthiness and provenance. This opens the door to prompt injection attacks, where the LLM receives and acts upon instructions from supposedly data-only sources, thus deviating from the user's original instructions. We define this as task drift, and we propose to catch it by scanning and analyzing the LLM's activations. We compare the LLM's activations before and after processing the external input in order to detect whether this input caused instruction drift. We develop two probing methods and find that simply using a linear classifier can detect drift with near perfect ROC AUC on an out-of-distribution test set. We show that this approach generalizes surprisingly well to unseen task domains, such as prompt injections, jailbreaks, and malicious instructions, without being trained on any of these attacks. Our setup does not require any modification of the LLM (e.g., fine-tuning) or any text generation, thus maximizing deployability and cost efficiency and avoiding reliance on unreliable model output. To foster future research on activation-based task inspection, decoding, and interpretability, we will release our large-scale TaskTracker toolkit, comprising a dataset of over 500K instances, representations from 4 SoTA language models, and inspection tools.

Teleoperation is increasingly recognized as a viable solution for deploying robots in hazardous environments. Controlling a robot to perform a complex or demanding task may overload operators resulting in poor performance. To design a robot controller to assist the human in executing such challenging tasks, a comprehensive understanding of the interplay between the robot's autonomous behavior and the operator's internal state is essential. In this paper, we investigate the relationships between robot autonomy and both the human user's cognitive load and trust levels, and the potential existence of three-way interactions in the robot-assisted execution of the task. Our user study (N=24) results indicate that while autonomy level influences the teleoperator's perceived cognitive load and trust, there is no clear interaction between these factors. Instead, these elements appear to operate independently, thus highlighting the need to consider both cognitive load and trust as distinct but interrelated factors in varying the robot autonomy level in shared-control settings. This insight is crucial for the development of more effective and adaptable assistive robotic systems.

Data generation remains a bottleneck in training surrogate models to predict molecular properties. We demonstrate that multitask Gaussian process regression overcomes this limitation by leveraging both expensive and cheap data sources. In particular, we consider training sets constructed from coupled-cluster (CC) and density function theory (DFT) data. We report that multitask surrogates can predict at CC level accuracy with a reduction to data generation cost by over an order of magnitude. Of note, our approach allows the training set to include DFT data generated by a heterogeneous mix of exchange-correlation functionals without imposing any artificial hierarchy on functional accuracy. More generally, the multitask framework can accommodate a wider range of training set structures -- including full disparity between the different levels of fidelity -- than existing kernel approaches based on $\Delta$-learning, though we show that the accuracy of the two approaches can be similar. Consequently, multitask regression can be a tool for reducing data generation costs even further by opportunistically exploiting existing data sources.