Prompt injection poses a serious threat to the reliability and safety of LLM agents. Recent defenses against prompt injection, such as Instruction Hierarchy and SecAlign, have shown notable robustness against static attacks. However, to more thoroughly evaluate the robustness of these defenses, it is arguably necessary to employ strong attacks such as automated red-teaming. To this end, we introduce RL-Hammer, a simple recipe for training attacker models that automatically learn to perform strong prompt injections and jailbreaks via reinforcement learning. RL-Hammer requires no warm-up data and can be trained entirely from scratch. To achieve high ASRs against industrial-level models with defenses, we propose a set of practical techniques that enable highly effective, universal attacks. Using this pipeline, RL-Hammer reaches a 98% ASR against GPT -4o and a 72% ASR against GPT -5 with the Instruction Hierarchy defense. We further discuss the challenge of achieving high diversity in attacks, highlighting how attacker models tend to "reward-hack" diversity objectives. Finally, we show that RL-Hammer can evade multiple prompt injection detectors. We hope our work advances automatic red-teaming and motivates the development of stronger, more principled defenses. More recently, a new paradigm has emerged that allows LLMs to behave as autonomous agents in complex environments, including full-fledged operating systems, integrated software platforms, and multi-step tool pipelines. In these contexts, LLMs can function as coding assistants, system administrators, and even academic researchers. Notable examples include Microsoft Copilot (GitHub, 2025), Anthropic Claude Computer Use (Anthropic, 2024), OpenAI Operator (OpenAI, 2025), and Zochi (Intology, 2025), each demonstrating the potential to combine sophisticated reasoning with direct system control. As these capabilities continue to advance, LLM agents are expected to be integrated into an even broader range of systems, becoming indispensable in both consumer and enterprise applications. However, these capabilities also introduce significant security risks, most notably prompt injection.

Abstract--This paper presents a novel approach for enabling robust robotic perception in dark environments using infrared (IR) stream. IR stream is less susceptible to noise than RGB in low-light conditions. However, it is dominated by active emitter patterns that hinder high-level tasks such as object detection, tracking and localisation. T o address this, a U-Net-based architecture is proposed that reconstructs clean IR images from emitter-populated input, improving both image quality and downstream robotic performance. This approach outperforms existing enhancement techniques and enables reliable operation of vision-driven robotic systems across illumination conditions from well-lit to extreme low-light scenes. Lighting-invariant vision systems are desirable for enabling robots to operate robustly across diverse and unpredictable environments without requiring modifications to the underlying perception pipeline. In order to support high-level tasks such as object detection, semantic segmentation, and image classification, the vision system must remain reliable even in low light or completely dark scenes. Such capabilities are critical in domains like mine shaft exploration, post-disaster victim identification, nuclear facility inspection, and visual loop closure in feature-deprived environments using aruco markers.

Hierarchical Reasoning Model (HRM) is a novel approach using two small neural networks recursing at different frequencies. This biologically inspired method beats Large Language models (LLMs) on hard puzzle tasks such as Sudoku, Maze, and ARC-AGI while trained with small models (27M parameters) on small data (around 1000 examples). HRM holds great promise for solving hard problems with small networks, but it is not yet well understood and may be suboptimal. We propose Tiny Recursive Model (TRM), a much simpler recursive reasoning approach that achieves significantly higher generalization than HRM, while using a single tiny network with only 2 layers. With only 7M parameters, TRM obtains 45% test-accuracy on ARC-AGI-1 and 8% on ARC-AGI-2, higher than most LLMs (e.g., Deepseek R1, o3-mini, Gemini 2.5 Pro) with less than 0.01% of the parameters.

As Large Language Model (LLM) agents increasingly gain self-evolutionary capabilities to adapt and refine their strategies through real-world interaction, their long-term reliability becomes a critical concern. We identify the Alignment Tipping Process (ATP), a critical post-deployment risk unique to self-evolving LLM agents. Unlike training-time failures, ATP arises when continual interaction drives agents to abandon alignment constraints established during training in favor of reinforced, self-interested strategies. We formalize and analyze ATP through two complementary paradigms: Self-Interested Exploration, where repeated high-reward deviations induce individual behavioral drift, and Imitative Strategy Diffusion, where deviant behaviors spread across multi-agent systems. Building on these paradigms, we construct controllable testbeds and benchmark Qwen3-8B and Llama-3.1-8B-Instruct. Our experiments show that alignment benefits erode rapidly under self-evolution, with initially aligned models converging toward unaligned states. In multi-agent settings, successful violations diffuse quickly, leading to collective misalignment. Moreover, current reinforcement learning-based alignment methods provide only fragile defenses against alignment tipping. Together, these findings demonstrate that alignment of LLM agents is not a static property but a fragile and dynamic one, vulnerable to feedback-driven decay during deployment. Our data and code are available at https://github.com/aiming-lab/ATP.

Although deep neural networks and in particular Convolu-tional Neural Networks have demonstrated state-of-the-art performance in image classification with relatively high efficiency, they still exhibit high computational costs, often rendering them impractical for real-time and edge applications. Therefore, a multitude of compression techniques have been developed to reduce these costs while maintaining accuracy. In addition, dynamic architectures have been introduced to modulate the level of compression at execution time, which is a desirable property in many resource-limited application scenarios. The proposed method effectively integrates two well-established optimization techniques: early exits and knowledge distillation, where a reduced student early-exit model is trained from a more complex teacher early-exit model. The primary contribution of this research lies in the approach for training the student early-exit model. In comparison to the conventional Knowledge Distillation loss, our approach incorporates a new entropy-based loss for images where the teacher's classification was incorrect. The proposed method optimizes the trade-off between accuracy and efficiency, thereby achieving significant reductions in computational complexity without compromising classification performance. The validity of this approach is substantiated by experimental results on image classification datasets CIF AR10, CIF AR100 and SVHN, which further opens new research perspectives for Knowledge Distillation in other contexts.

Hallucination detection remains a fundamental challenge for the safe and reliable deployment of large language models (LLMs), especially in applications requiring factual accuracy. Existing hallucination benchmarks often operate at the sequence level and are limited to English, lacking the fine-grained, multilingual supervision needed for a comprehensive evaluation. In this work, we introduce PsiloQA, a large-scale, multilingual dataset annotated with span-level hallucinations across 14 languages. PsiloQA is constructed through an automated three-stage pipeline: generating question-answer pairs from Wikipedia using GPT-4o, eliciting potentially hallucinated answers from diverse LLMs in a no-context setting, and automatically annotating hallucinated spans using GPT-4o by comparing against golden answers and retrieved context. We evaluate a wide range of hallucination detection methods -- including uncertainty quantification, LLM-based tagging, and fine-tuned encoder models -- and show that encoder-based models achieve the strongest performance across languages. Furthermore, PsiloQA demonstrates effective cross-lingual generalization and supports robust knowledge transfer to other benchmarks, all while being significantly more cost-efficient than human-annotated datasets. Our dataset and results advance the development of scalable, fine-grained hallucination detection in multilingual settings.

The growing demand for efficient deep learning has positioned dataset distillation as a pivotal technique for compressing training dataset while preserving model performance. However, existing inner-loop optimization methods for dataset distillation typically rely on random truncation strategies, which lack flexibility and often yield suboptimal results. In this work, we observe that neural networks exhibit distinct learning dynamics across different training stages-early, middle, and late-making random truncation ineffective. To address this limitation, we propose Automatic Truncated Backpropagation Through Time (AT-BPTT), a novel framework that dynamically adapts both truncation positions and window sizes according to intrinsic gradient behavior. AT-BPTT introduces three key components: (1) a probabilistic mechanism for stage-aware timestep selection, (2) an adaptive window sizing strategy based on gradient variation, and (3) a low-rank Hessian approximation to reduce computational overhead. Extensive experiments on CIFAR-10, CIFAR-100, Tiny-ImageNet, and ImageNet-1K show that AT-BPTT achieves state-of-the-art performance, improving accuracy by an average of 6.16% over baseline methods. Moreover, our approach accelerates inner-loop optimization by 3.9x while saving 63% memory cost.

A strong and fast baseline in molecular property prediction is a Random Forest (RF) trained on ECFP4/ECFP6 descriptors. In practice, the count-based variant of ECFP generally outperforms the binary variant, especially for classification. Recent deep-learning approaches can match or exceed these baselines, including pretrained transformer-CNN models (5) and graph neural networks such as ChemProp or AttentiveFP(6). Chemprop's key architectural choice is directed, bond-centered message passing, in contrast to the more common atom-centered formulations used by many MPNNs. Because much of the remaining architecture is comparable across message-passing GNNs, this raises a focused question: what concrete advantage does the bond-centered formulation confer over atom-centered approaches? To isolate this representational factor, we introduce a static Bond-Centered Fingerprint (BCFP) that mirrors Chemprop's bond-centric view, and we compare it directly against ECFP using a lightweight Random Forest or XGBoost pipeline on the Blood-Brain Barrier Penetration (BBBP) classification task. To our knowledge, this is the first study to propose BCFP and analyze its complementarity with ECFP (7) . Our results indicate that concatenating atom-and bond-centered fingerprints yields efficient and effective models for BBBP prediction, clarifying why bond-centric message passing often appears among top-k performers while offering a simple, fast alternative to full neural architectures.

Despite interpretability work analyzing VIT encoders and transformer activations, we don't yet understand why Multimodal Language Models (MLMs) struggle on perception-heavy tasks. We offer an under-studied perspective by examining how popular MLMs (LLaVA-OneVision, Qwen2.5-VL, and Llama-3-LLaVA-NeXT) process their visual key-value tokens. We first study the flow of visual information through the language model, finding that image value tokens encode sufficient information to perform several perception-heavy tasks zero-shot: segmentation, semantic correspondence, temporal correspondence, and referring expression detection. We find that while the language model does augment the visual information received from the projection of input visual encodings-which we reveal correlates with overall MLM perception capability-it contains less visual information on several tasks than the equivalent visual encoder (SigLIP) that has not undergone MLM finetuning. Further, we find that the visual information corresponding to input-agnostic image key tokens in later layers of language models contains artifacts which reduce perception capability of the overall MLM. Next, we discuss controlling visual information in the language model, showing that adding a text prefix to the image input improves perception capabilities of visual representations. Finally, we reveal that if language models were able to better control their visual information, their perception would significantly improve; e.g., in 33.3% of Art Style questions in the BLINK benchmark, perception information present in the language model is not surfaced to the output! Our findings reveal insights into the role of key-value tokens in multimodal systems, paving the way for deeper mechanistic interpretability of MLMs and suggesting new directions for training their visual encoder and language model components.

Recent progress in large language models demonstrates that hybrid architectures--combining self-attention mechanisms with structured state space models like Mamba--can achieve a compelling balance between modeling quality and computational efficiency, particularly for long-context tasks. While these hybrid models show promising performance, systematic comparisons of hybridization strategies and analyses on the key factors behind their effectiveness have not been clearly shared to the community. In this work, we present a holistic evaluation of hybrid architectures based on inter-layer (sequential) or intra-layer (parallel) fusion. We evaluate these designs from a variety of perspectives: language modeling performance, long-context capabilities, scaling analysis, and training and inference efficiency. By investigating the core characteristics of their computational primitive, we identify the most critical elements for each hybridization strategy and further propose optimal design recipes for both hybrid models. Our comprehensive analysis provides practical guidance and valuable insights for developing hybrid language models, facilitating the optimization of architectural configurations.