Towards a Multi-Agent Vision-Language System for Zero-Shot Novel Hazardous Object Detection for Autonomous Driving Safety Shashank Shriram 1, Srinivasa Perisetla 1, Aryan Keskar 1, Harsha Krishnaswamy 1, Tonko Emil Westerhof Bossen 1, 2, Andreas Møgelmose 2, Ross Greer 1 Machine Intelligence, Interaction, and Imagination (Mi 3) Laboratory 1 University of California, Merced 2 Aalborg Universitet Abstract --Detecting anomalous hazards in visual data, particularly in video streams, is a critical challenge in autonomous driving. Existing models often struggle with unpredictable, out-of-label hazards due to their reliance on predefined object categories. In this paper, we propose a multimodal approach that integrates vision-language reasoning with zero-shot object detection to improve hazard identification and explanation. Our pipeline consists of a Vision-Language Model (VLM), a Large Language Model (LLM), in order to detect hazardous objects within a traffic scene. We refine object detection by incorporating OpenAI's CLIP model to match predicted hazards with bounding box annotations, improving localization accuracy. T o assess model performance, we create a ground truth dataset by denoising and extending the foundational COOOL (Challenge-of-Out-of-Label) anomaly detection benchmark dataset with complete natural language descriptions for hazard annotations. We define a means of hazard detection and labeling evaluation on the extended dataset using cosine similarity. This evaluation considers the semantic similarity between the predicted hazard description and the annotated ground truth for each video. Additionally, we release a set of tools for structuring and managing large-scale hazard detection datasets.

One of the goals of language model unlearning is to reduce memorization of selected text instances while retaining the model's general abilities. Despite various proposed methods, reducing memorization of large datasets without noticeable degradation in model utility remains challenging. In this paper, we investigate the mean teacher algorithm (Tarvainen & Valpola, 2017), a simple proximal optimization method from continual learning literature that gradually modifies the teacher model. We show that the mean teacher can approximate a trajectory of a slow natural gradient descent (NGD), which inherently seeks low-curvature updates that are less likely to degrade the model utility. While slow NGD can suffer from vanishing gradients, we introduce a new unlearning loss called "negative log-unlikelihood" (NLUL) that avoids this problem. We show that the combination of mean teacher and NLUL improves some metrics on the MUSE benchmarks (Shi et al., 2024).

Reasoning models have demonstrated impressive performance on difficult tasks that traditional language models struggle at. However, many are plagued with the problem of overthinking--generating large amounts of unnecessary tokens which don't improve accuracy on a question. We introduce approximate measures of problem-level difficulty and demonstrate that a clear relationship between problem difficulty and optimal token spend exists, and evaluate how well calibrated a variety of reasoning models are in terms of efficiently allocating the optimal token count. We find that in general, reasoning models are poorly calibrated, particularly on easy problems. To evaluate calibration on easy questions we introduce DUMB500, a dataset of extremely easy math, reasoning, code, and task problems, and jointly evaluate reasoning model on these simple examples and extremely difficult examples from existing frontier benchmarks on the same task domain. Finally, we introduce THOUGHTTERMINATOR, a training-free black box decoding technique that significantly improves reasoning model calibration.

Unsurprisingly, mathematics is one of the main benchmarks f or current AI. There are even significant efforts to build AI systems dedicated to mathematics. O ne of them is o3-mini [ 6 ], developed by OpenAI. It is claimed that it solved 80% of the subject shee t at American Invitational Mathematics Examination (AIME) 2024 [ 5 ], a prestigious competition leading to the USA Mathematica l Olympiad. Another one is Grok-3 [ 11 ], developed by xAI (an Elon Musk company), which is also claimed to be very good at math and physics. These claims are i n stark contrast with the statistics put forward in [ 3 ] where, in the case of mathematical research, the AI can solv e only 2% of the problems. Our motivation for this AI experiment was to try to understan d, from the perspective of a non-specialist in machine learning, what can this kind of AI do for the working mathematicians, how can we use it to support our work, and what is behind the vas t gap (claimed in [ 3 ]) between what current AI capabilities and the prowess of the mathemat ical research community. The readership target of this paper consists mainly of the ma thematicians with a moderate degree of fluency with elementary category-theoretic think ing.

The widespread adoption of AI systems in the economy hinges on their ability to generate economic value that outweighs their inference costs. Evaluating this tradeoff requires metrics that account for both performance and costs. We propose a framework grounded in production theory for evaluating language models by combining accuracy and inference cost. We introduce "cost-of-pass", the expected monetary cost of generating a correct solution. We then define the "frontier cost-of-pass" as the minimum cost-of-pass achievable across available models or the "human-expert, using the approximate cost of hiring an expert. Our analysis reveals distinct economic insights. First, lightweight models are most cost-effective for basic quantitative tasks, large models for knowledge-intensive ones, and reasoning models for complex quantitative problems, despite higher per-token costs. Second, tracking this frontier cost-of-pass over the past year reveals significant progress, particularly for complex quantitative tasks where the cost has roughly halved every few months. Third, to trace key innovations driving this progress, we examine counterfactual frontiers: estimates of cost-efficiency without specific model classes. We find that innovations in lightweight, large, and reasoning models have been essential for pushing the frontier in basic quantitative, knowledge-intensive, and complex quantitative tasks, respectively. Finally, we assess the cost-reductions afforded by common inference-time techniques like majority voting and self-refinement, finding that their marginal accuracy gains rarely justify their costs. Our findings underscore that complementary model-level innovations are the primary drivers of cost-efficiency, and our economic framework provides a principled tool for measuring this progress and guiding deployment.

Learning to perform manipulation tasks from human videos is a promising approach for teaching robots. However, many manipulation tasks require changing control parameters during task execution, such as force, which visual data alone cannot capture. In this work, we leverage sensing devices such as armbands that measure human muscle activities and microphones that record sound, to capture the details in the human manipulation process, and enable robots to extract task plans and control parameters to perform the same task. To achieve this, we introduce Chain-of-Modality (CoM), a prompting strategy that enables Vision Language Models to reason about multimodal human demonstration data -- videos coupled with muscle or audio signals. By progressively integrating information from each modality, CoM refines a task plan and generates detailed control parameters, enabling robots to perform manipulation tasks based on a single multimodal human video prompt. Our experiments show that CoM delivers a threefold improvement in accuracy for extracting task plans and control parameters compared to baselines, with strong generalization to new task setups and objects in real-world robot experiments. Videos and code are available at https://chain-of-modality.github.io

From 2000 to 2015, the UN's Millennium Development Goals guided global priorities. The subsequent Sustainable Development Goals (SDGs) adopted a more dynamic approach, with annual indicator updates. As 2030 nears and progress lags, innovative acceleration strategies are critical. This study develops an AI-powered knowledge graph system to analyze SDG interconnections, discover potential new goals, and visualize them online. Using official SDG texts, Elsevier's keyword dataset, and 1,127 TED Talk transcripts (2020.01-2024.04), a pilot on 269 talks from 2023 applies AI-speculative design, large language models, and retrieval-augmented generation. Key findings include: (1) Heatmap analysis reveals strong associations between Goal 10 and Goal 16, and minimal coverage of Goal 6. (2) In the knowledge graph, simulated dialogue over time reveals new central nodes, showing how richer data supports divergent thinking and goal clarity. (3) Six potential new goals are proposed, centered on equity, resilience, and technology-driven inclusion. This speculative-AI framework offers fresh insights for policymakers and lays groundwork for future multimodal and cross-system SDG applications.

Stance detection has become an essential tool for analyzing public discussions on social media. Current methods face significant challenges, particularly in Chinese language processing and multi-turn conversational analysis. To address these limitations, we introduce C-MTCSD, the largest Chinese multi-turn conversational stance detection dataset, comprising 24,264 carefully annotated instances from Sina Weibo, which is 4.2 times larger than the only prior Chinese conversational stance detection dataset. Our comprehensive evaluation using both traditional approaches and large language models reveals the complexity of C-MTCSD: even state-of-the-art models achieve only 64.07% F1 score in the challenging zero-shot setting, while performance consistently degrades with increasing conversation depth. Traditional models particularly struggle with implicit stance detection, achieving below 50% F1 score. This work establishes a challenging new benchmark for Chinese stance detection research, highlighting significant opportunities for future improvements.

Large Reasoning Models (LRMs) like DeepSeek-R1 and OpenAI-o1 have demonstrated remarkable reasoning capabilities, raising important questions about their biases in LLM-as-a-judge settings. We present a comprehensive benchmark comparing judging biases between LLMs and LRMs across both subjective preference-alignment datasets and objective fact-based datasets. Through investigation of bandwagon, authority, position, and distraction biases, we uncover four key findings: (1) despite their advanced reasoning capabilities, LRMs remain susceptible to the above biases; (2) LRMs demonstrate better robustness than LLMs specifically on fact-related datasets; (3) LRMs exhibit notable position bias, preferring options in later positions; and (4) we identify a novel "superficial reflection bias" where phrases mimicking reasoning (e.g., "wait, let me think...") significantly influence model judgments. To address these biases, we design and evaluate three mitigation strategies: specialized system prompts that reduce judging biases by up to 19\% in preference alignment datasets and 14\% in fact-related datasets, in-context learning that provides up to 27\% improvement on preference tasks but shows inconsistent results on factual tasks, and a self-reflection mechanism that reduces biases by up to 10\% in preference datasets and 16\% in fact-related datasets, with self-reflection proving particularly effective for LRMs. Our work provides crucial insights for developing more reliable LLM-as-a-Judge frameworks, especially as LRMs become increasingly deployed as automated judges.

Given the growing use of generative artificial intelligence as a tool for creating multilingual content and bypassing both machine and traditional translation methods, this study explores the ability of linguistically trained individuals to discern machine-generated output from human-written text (HT). After brief training sessions on the textual anomalies typically found in synthetic text (ST), twenty-three postgraduate translation students analysed excerpts of Italian prose and assigned likelihood scores to indicate whether they believed they were human-written or AI-generated (ChatGPT-4o). The results show that, on average, the students struggled to distinguish between HT and ST, with only two participants achieving notable accuracy. Closer analysis revealed that the students often identified the same textual anomalies in both HT and ST, although features such as low burstiness and self-contradiction were more frequently associated with ST. These findings suggest the need for improvements in the preparatory training. Moreover, the study raises questions about the necessity of editing synthetic text to make it sound more human-like and recommends further research to determine whether AI-generated text is already sufficiently natural-sounding not to require further refinement.