Search and rescue (SAR) is a complex teamwork environment that requires efficient spatial communication between commanders and field teams with heterogeneous perspectives and asymmetric information. Maps are central artifacts in SAR, yet they are also a space of technological tension due to constantly changing situation at disaster sites. Sketch mapping is an effective method of externalizing and communicating spatial understanding, increasing situation awareness in spatial decision-making tasks including SAR. Current paper-based sketch mapping in SAR struggles to handle the three-dimensional nature of physical space and remote collaboration. We propose CoMap, a collaborative 3D sketch mapping system validated in a virtual reality fire-rescue game. In a within-subject study with 13 commander–field team pairs, CoMap enabled more accurate and efficient spatial communication than conventional 2D sketch mapping. Communication analysis further showed that CoMap fostered proactive descriptions. We distill three design implications for next-generation mapping tools to advance SAR training and real-world operations.

Large Language Models (LLMs) have demonstrated impressive reasoning abilities, but their generated rationales often suffer from issues such as reasoning inconsistency and factual errors, undermining their reliability.Prior work has explored improving rationale quality via multi-reward fine-tuning or reinforcement learning (RL), where models are optimized for diverse objectives.While effective, these approaches train the model in a fixed manner and do not have any inference-time adaptability, nor can they generalize reasoning requirements for new test-time inputs.Another approach is to train specialized reasoning experts using reward signals and use them to improve generation at inference time.Existing methods in this paradigm are limited to using only a single expert and cannot improve upon multiple reasoning aspects.To address this, we propose MIXIE, a novel inference-time expertmixing framework that dynamically determines mixing proportions for each expert, enabling contextualized and flexible fusion.We demonstrate the effectiveness of MIXIE on improving chain-of-thought reasoning in LLMs by merging commonsense and entailment reasoning experts finetuned on reward-filtered data.Our approach outperforms existing baselines on three question-answering datasets: StrategyQA, CommonsenseQA, and ARC, highlighting its potential to enhance LLM reasoning with efficient, adaptable expert integration.

An intuitive way to externalize spatial memory is to sketch it. Compared to traditional paper-based sketches, virtual reality (VR) creates new opportunities to investigate the 3D aspect of spatial memory as it empowers users to express 3D information on a 3D interface directly. The goal of this study is to design a 3D sketch mapping tool for researchers and non-expert users without sketching expertise that enables externalizing memories of spatial information after some 3D-critical tasks. There exist 3D sketching tools using VR, but there are two issues with the current mid-air 3D sketching approach: (1) distortion of sketches due to depth perception errors and (2) increased cognitive and sensorimotor demands due to an increased degree of freedom and absence of physical support. To address these problems, we implemented VResin, a novel sketching interface that synergizes 3D mid-air sketching with 2D surface sketching to scaffold 3D sketching into a layer-by-layer process. An experimental study with 48 participants on multi-layer building scenarios showed that VResin supports users in creating less distorted sketches while maintaining the level of completeness and generalization compared to mid-air sketching in VR. We also demonstrate the potential applications that can benefit from 3D sketch maps and the suitability of VResin for a variety of building shapes.

Though linguistic knowledge emerges during large-scale language model pretraining, recent work attempt to explicitly incorporate humandefined linguistic priors into task-specific finetuning.Infusing language models with syntactic or semantic knowledge from parsers has shown improvements on many language understanding tasks.To further investigate the effectiveness of structural linguistic priors, we conduct empirical study of replacing parsed graphs or trees with trivial ones (rarely carrying linguistic knowledge e.g., balanced tree) for tasks in the GLUE benchmark.Encoding with trivial graphs achieves competitive or even better performance in fully-supervised and few-shot settings.It reveals that the gains might not be significantly attributed to explicit linguistic priors but rather to more feature interactions brought by fusion layers.Hence we call for attention to using trivial graphs as necessary baselines to design advanced knowledge fusion methods in the future.