This research introduces **INPAINTING**, a framework that treats finding adversarial "jailbreak" prompts as a simple inference task rather than a slow optimization problem. By using **Diffusion Large Language Models (DLLMs)**, which understand the joint relationship between prompts and responses, the researchers can directly generate prompts that trigger specific harmful outputs. This method effectively **inverts the standard generation process**, allowing a surrogate model to "sample" candidate attacks that are highly transferable to black-box targets like ChatGPT. The resulting prompts are **semantically natural and exhibit low perplexity**, making them difficult for traditional security filters to detect. Compared to existing gradient-based or iterative attacks, this approach is **significantly more efficient** and achieves higher success rates against robustly trained models. Ultimately, the paper highlights a critical security vulnerability: any model capable of modeling joint data distributions can be repurposed as a **powerful natural adversary**.

Language models face a significant partial token problem (PTP) when user prompts end in the middle of a multi-character token, causing the model to misinterpret the expected continuation. This research highlights that the issue is not just a theoretical glitch but a pervasive failure mode in natural language use, especially in Chinese, German, and programming code where word and token boundaries frequently misalign. Experiments reveal that even elite models suffer a dramatic drop in accuracy—between 60% and 95%—when encountering these "word-complete" but "token-incomplete" prompts. Surprisingly, this degradation does not improve with increased model scale, as larger models are often more strictly tuned to their specific tokenizers. To address these distortions, the authors evaluate several inference-time mitigations, finding that heuristic "token healing" offers inconsistent results. In contrast, the study validates that an exact solution called ByteSampler can completely eliminate the problem by reconstructing valid token paths. Ultimately, the paper provides practical recommendations for model providers to ensure more reliable text generation across diverse languages and technical domains.

This paper investigates whether large language models (LLMs) can effectively utilize novel information they learn during a single conversation. While models successfully create internal representations of new concepts from context, the study reveals these representations are often inert and cannot be applied to downstream tasks. Experiments on next-token prediction and adaptive world modeling show that both open-weights and closed-source reasoning models struggle to use this in-context knowledge reliably. Even when a model's latent states reflect a new logical structure, it frequently fails to deploy that structure to solve problems. These findings suggest a significant gap between encoding information and the flexible deployment required for truly adaptable artificial intelligence. Overall, the authors conclude that current architectures require new training methods to move beyond simple data recognition toward functional in-context reasoning.

This research explores the discrepancy between transformer in-context learning and Bayesian inference, arguing that models are Bayesian in expectation rather than through every individual realization. While previous studies used martingale diagnostics to question the Bayesian nature of these models, this paper identifies positional encodings as the primary factor that breaks the required exchangeability. By accounting for how architectural design prioritizes sequence order, the authors prove that transformers still achieve near-optimal compression and information-theoretic efficiency when performance is averaged across different orderings. Empirical tests on black-box LLMs and controlled ablations demonstrate that order-induced variance exists but predictably decays as context length increases. Ultimately, the study suggests permutation averaging as a practical and effective method for reducing uncertainty and improving the reliability of model outputs in tasks with exchangeable data.

This paper discusses a framework for reflective test-time planning designed to improve the performance of embodied Large Language Models (LLMs) during robotic tasks. This system utilizes double-loop learning, where agents re-evaluate their past decisions through hindsight assessments to correct underlying strategic errors. By incorporating internal reflection for immediate scoring and retrospective reflection for long-term credit assignment, the model adapts its policy at deployment without requiring additional pretraining data. Experimental results in household and cupboard fitting tasks demonstrate that this approach significantly reduces execution waste and improves success rates compared to standard methods. Furthermore, the researchers employ Low-Rank Adaptation (LoRA) to efficiently update the models, ensuring that the robots can learn from their own trials and errors in real-time environments.

This research introduces Optimal Advantage-based Policy Optimization with Lagged Inference policy (OAPL), a novel reinforcement learning algorithm designed to improve Large Language Model (LLM) reasoning. Traditional methods like GRPO often struggle with "off-policy" data caused by technical mismatches between training and inference engines. OAPL embraces these discrepancies by using a squared regression objective and KL-regularization, allowing the model to learn effectively even when data is significantly outdated. Empirical tests show that OAPL outperforms existing benchmarks in competition mathematics and matches top-tier coding models while using three times fewer training samples. Furthermore, the algorithm prevents entropy collapse, ensuring that the model maintains diverse and scalable problem-solving capabilities during test-time. Ultimately, the authors demonstrate that fully asynchronous, off-policy training is a more stable and efficient path for advancing machine reasoning.

This research paper argues that Test-Time Training (TTT) with key-value binding—previously understood as a way for models to "memorize" data during inference—is actually a form of linear attention. The authors identify a "memorization paradox" where improving the model's internal memory fitting actually degrades task performance, and even reversing the learning process can improve results. By mathematically unrolling the TTT update rules, they prove that complex inner-loop architectures are equivalent to learned linear attention operators. This theoretical shift allows for architectural simplifications, such as removing redundant normalization and momentum components. Furthermore, this new perspective enables fully parallel formulations of TTT, significantly increasing inference speed. Ultimately, the work reframes TTT as a dynamic feature mixer rather than a retrieval system, providing a more efficient framework for sequence modeling.

This paper introduces Unified Latents (UL), a novel framework designed by Google DeepMind researchers to improve the efficiency of generative diffusion models. By integrating the encoder, prior, and decoder into a single cohesive system, the method creates a more principled approach to managing latent representations. A key innovation involves using a fixed amount of Gaussian noise during encoding, which simplifies the training process and allows for a tight bound on latent information. This setup provides users with specific hyper-parameters, such as the loss factor and sigmoid bias, to precisely balance reconstruction accuracy against the complexity of the modeling task. Experimental results across image and video datasets demonstrate that this architecture achieves superior performance with lower computational costs compared to established baselines like Stable Diffusion. Overall, the research highlights how jointly optimizing all components of the latent space leads to higher-quality generative outputs and more effective scaling.

This paper introduces the Spectral Bellman Method (SBM), a novel framework designed to enhance value-based reinforcement learning by unifying representation learning and exploration. By leveraging the Inherent Bellman Error (IBE) condition, the authors demonstrate that optimal feature representations are intrinsically linked to the spectral properties of the Bellman operator. This theoretical connection allows the agent to learn state-action features whose covariance structure is naturally aligned with environment dynamics, facilitating more effective Thompson Sampling for exploration. Empirical evaluations on the Atari benchmark show that SBM significantly improves performance in hard-exploration and long-horizon tasks when integrated into standard algorithms like DQN and R2D2. Ultimately, the method offers a computationally tractable and principled approach to achieving Bellman consistency across a broad space of value functions.

This paper details the development of capability boundaries to predict the downstream performance of language models based on pre-training compute budgets. Researchers move beyond standard scaling laws by using high-quantile pinball loss and monotone splines to track the upper envelope of achievable results rather than average trends. This methodology addresses benchmark saturation, where static tests lose their ability to distinguish between top-tier models as capabilities grow. By focusing on a 98th percentile frontier, the framework creates a stable, probabilistic estimate of what competitive training pipelines can reliably achieve. Ultimately, the work offers a principled way for practitioners to manage resource allocation and evaluate model performance relative to an evolving global capability ceiling.