High-risk industries like nuclear and aviation use real-time monitoring to detect dangerous system conditions. Similarly, Large Language Models (LLMs) need monitoring safeguards. We propose a real-time framework to predict harmful AI outputs before they occur by using an unsupervised approach that treats normal behavior as the baseline and harmful outputs as outliers. Our study focuses specifically on backdoor-triggered responses -- where specific input phrases activate hidden vulnerabilities causing the model to generate unsafe content like violence, pornography, or hate speech. We address two key challenges: (1) identifying true causal indicators rather than surface correlations, and (2) preventing advanced models from deception -- deliberately evading monitoring systems. Hence, we approach this problem from an unsupervised lens by drawing parallels to human deception: just as humans exhibit physical indicators while lying, we investigate whether LLMs display distinct internal behavioral signatures when generating harmful content. Our study addresses two critical challenges: 1) designing monitoring systems that capture true causal indicators rather than superficial correlations; and 2)preventing intentional evasion by increasingly capable "Future models''. Our findings show that models can produce harmful content through causal mechanisms and can become deceptive by: (a) alternating between linear and non-linear representations, and (b) modifying feature relationships. To counter this, we developed Safety-Net -- a multi-detector framework that monitors different representation dimensions, successfully detecting harmful behavior even when information is shifted across representational spaces to evade individual monitors. Our evaluation shows 96% accuracy in detecting harmful cases using our unsupervised ensemble approach.

Large Language Models (LLMs) are increasingly used not only to generate text but also to evaluate it, raising urgent questions about whether their judgments are consistent, unbiased, and robust to framing effects. In this study, we systematically examine inter - and intra - model agreement across four state - of - the - art LLMs - OpenAI o3 - mini, Deepseek Reasone r, xAI Grok 2, and Mistral - tasked with evaluating 4,800 narrative statements on 24 different topics of social, political, and public health relevance, for a total of 192,000 assessments. W e manipulate the disclosed source of each statement to assess how attribution to either another LLM or a human author of specified nationality affects evaluation outcomes. We find that, in the blind condition, different LLMs display a remarkably high degree of inter - and intra - model agreement across topics . However, this alignment breaks down when source framing is introduced. Here we show that attributing statements to Chinese individuals systematically lowers agreement scores across all models, and in particular for Deepseek Reasoner . Our findings reveal that framing effects can deeply affect text evaluation, with significant implications for the integrity, neutrality, and fairness of LLM - mediated information systems.

Existing AutoML systems have advanced the automation of machine learning (ML); however, they still require substantial manual configuration and expert input, particularly when handling multimodal data. We introduce MLZero, a novel multi-agent framework powered by Large Language Models (LLMs) that enables end-to-end ML automation across diverse data modalities with minimal human intervention. A cognitive perception module is first employed, transforming raw multimodal inputs into perceptual context that effectively guides the subsequent workflow. To address key limitations of LLMs, such as hallucinated code generation and outdated API knowledge, we enhance the iterative code generation process with semantic and episodic memory. MLZero demonstrates superior performance on MLE-Bench Lite, outperforming all competitors in both success rate and solution quality, securing six gold medals. Additionally, when evaluated on our Multimodal AutoML Agent Benchmark, which includes 25 more challenging tasks spanning diverse data modalities, MLZero outperforms the competing methods by a large margin with a success rate of 0.92 (+263.6\%) and an average rank of 2.28. Our approach maintains its robust effectiveness even with a compact 8B LLM, outperforming full-size systems from existing solutions.

Large language models (LLMs) have recently achieved remarkable results in complex multi-step tasks, such as mathematical reasoning and agentic software engineering. However, they often struggle to maintain consistent performance across multiple solution attempts. One effective approach to narrow the gap between average-case and best-case performance is guided test-time search, which explores multiple solution paths to identify the most promising one. Unfortunately, effective search techniques (e.g. MCTS) are often unsuitable for non-serializable RL environments, such as Docker containers, where intermediate environment states cannot be easily saved and restored. We investigate two complementary search strategies applicable to such environments: 1-step lookahead and trajectory selection, both guided by a learned action-value function estimator. On the SWE-bench Verified benchmark, a key testbed for agentic software engineering, we find these methods to double the average success rate of a fine-tuned Qwen-72B model, achieving 40.8%, the new state-of-the-art for open-weights models. Additionally, we show that these techniques are transferable to more advanced closed models, yielding similar improvements with GPT-4o.

Text-to-Image (T2I) has been prevalent in recent years, with most common condition tasks having been optimized nicely. Besides, counterfactual Text-to-Image is obstructing us from a more versatile AIGC experience. For those scenes that are impossible to happen in real world and anti-physics, we should spare no efforts in increasing the factual feel, which means synthesizing images that people think very likely to be happening, and concept alignment, which means all the required objects should be in the same frame. In this paper, we focus on concept alignment. As controllable T2I models have achieved satisfactory performance for real applications, we utilize this technology to replace the objects in a synthesized image in latent space step-by-step to change the image from a common scene to a counterfactual scene to meet the prompt. We propose a strategy to instruct this replacing process, which is called as Explicit Logical Narrative Prompt (ELNP), by using the newly SoTA language model DeepSeek to generate the instructions. Furthermore, to evaluate models' performance in counterfactual T2I, we design a metric to calculate how many required concepts in the prompt can be covered averagely in the synthesized images. The extensive experiments and qualitative comparisons demonstrate that our strategy can boost the concept alignment in counterfactual T2I.

--T extbook question answering (TQA) is a complex task, requiring the interpretation of complex multimodal context. Although recent advances have improved overall performance, they often encounter difficulties in educational settings where accurate semantic alignment and task-specific document retrieval are essential. In this paper, we propose a novel approach to multimodal textbook question answering by introducing a mechanism for enhancing semantic representations through multi-objective joint training. Our model, Joint Embedding Training With Ranking Supervision for T extbook Question Answering (JETRTQA), is a multimodal learning framework built on a retriever-generator architecture that uses a retrieval-augmented generation setup, in which a multimodal large language model generates answers. JETRTQA is designed to improve the relevance of retrieved documents in complex educational contexts. Unlike traditional direct scoring approaches, JETRTQA learns to refine the semantic representations of questions and documents through a supervised signal that combines pairwise ranking and implicit supervision derived from answers. We evaluate our method on the CK12-QA dataset and demonstrate that it significantly improves the discrimination between informative and irrelevant documents, even when they are long, complex, and multimodal. JETRTQA outperforms the previous state of the art, achieving a 2.4% gain in accuracy on the validation set and 11.1% on the test set. EXTBOOK question answering (TQA) has emerged as a central challenge in natural language processing because the complexity of educational content requires deep semantic reasoning. TQA involves the analysis of structured, often lengthy, educational documents that are frequently multimodal, incorporating elements such as diagrams, tables, or explanatory images. The retrieved information is then used to generate answers. This process is not a simple fusion; it demands a strategic approach to overcome the fundamental limitations of traditional question-answering (QA) models, which are often unable to effectively handle long, complex, or out-of-domain contexts [1], [2].

Memory-based approaches have shown strong performance in Continual Relation Extraction (CRE). However, storing examples from previous tasks increases memory usage and raises privacy concerns. Recently, prompt-based methods have emerged as a promising alternative, as they do not rely on storing past samples. Despite this progress, current prompt-based techniques face several core challenges in CRE, particularly in accurately identifying task identities and mitigating catastrophic forgetting. Existing prompt selection strategies often suffer from inaccuracies, lack robust mechanisms to prevent forgetting in shared parameters, and struggle to handle both cross-task and within-task variations. In this paper, we propose WAVE++, a novel approach inspired by the connection between prefix-tuning and mixture of experts. Specifically, we introduce task-specific prompt pools that enhance flexibility and adaptability across diverse tasks while avoiding boundary-spanning risks; this design more effectively captures variations within each task and across tasks. To further refine relation classification, we incorporate label descriptions that provide richer, more global context, enabling the model to better distinguish among different relations. We also propose a training-free mechanism to improve task prediction during inference. Moreover, we integrate a generative model to consolidate prior knowledge within the shared parameters, thereby removing the need for explicit data storage. Extensive experiments demonstrate that WAVE++ outperforms state-of-the-art prompt-based and rehearsal-based methods, offering a more robust solution for continual relation extraction. Our code is publicly available at https://github.com/PiDinosauR2804/WAVE-CRE-PLUS-PLUS.

Scaling up high-quality trajectory data has long been a critical bottleneck for developing human-like computer use agents. We introduce PC Agent-E, an efficient agent training framework that significantly reduces reliance on large-scale human demonstrations. Starting with just 312 human-annotated computer use trajectories, we further improved data quality by synthesizing diverse action decisions with Claude 3.7 Sonnet. Trained on these enriched trajectories, our PC Agent-E model achieved a remarkable 141% relative improvement, surpassing the strong Claude 3.7 Sonnet with extended thinking on WindowsAgentArena-V2, an improved benchmark we also released. Furthermore, PC Agent-E demonstrates strong generalizability to different operating systems on OSWorld. Our findings suggest that strong computer use capabilities can be stimulated from a small amount of high-quality trajectory data.

In the era of foundation models and Large Language Models (LLMs), Euclidean space is the de facto geometric setting of our machine learning architectures. However, recent literature has demonstrated that this choice comes with fundamental limitations. To that end, non-Euclidean learning is quickly gaining traction, particularly in web-related applications where complex relationships and structures are prevalent. Non-Euclidean spaces, such as hyperbolic, spherical, and mixed-curvature spaces, have been shown to provide more efficient and effective representations for data with intrinsic geometric properties, including web-related data like social network topology, query-document relationships, and user-item interactions. Integrating foundation models with non-Euclidean geometries has great potential to enhance their ability to capture and model the underlying structures, leading to better performance in search, recommendations, and content understanding. This workshop focuses on the intersection of Non-Euclidean Foundation Models and Geometric Learning (NEGEL), exploring its potential benefits, including the potential benefits for advancing web-related technologies, challenges, and future directions. Workshop page: [https://hyperboliclearning.github.io/events/www2025workshop](https://hyperboliclearning.github.io/events/www2025workshop)

This paper investigates the extent to which pretrained German BERT encodes knowledge of noun compound semantics. We comprehensively vary combinations of target tokens, layers, and cased vs. uncased models, and evaluate them by predicting the compositionality of 868 gold standard compounds. Looking at representational patterns within the transformer architecture, we observe trends comparable to equivalent prior work on English, with compositionality information most easily recoverable in the early layers. However, our strongest results clearly lag behind those reported for English, suggesting an inherently more difficult task in German. This may be due to the higher productivity of compounding in German than in English and the associated increase in constituent-level ambiguity, including in our target compound set.