Jailbreaks are adversarial attacks designed to bypass the built-in safety mechanisms of large language models. Automated jailbreaks typically optimize an adversarial suffix or adapt long prompt templates by forcing the model to generate the initial part of a restricted or harmful response. In this work, we show that existing jailbreak attacks that leverage such mechanisms to unlock the model response can be detected by a straightforward perplexity-based filtering on the input prompt. To overcome this issue, we propose LatentBreak, a white-box jailbreak attack that generates natural adversarial prompts with low perplexity capable of evading such defenses. LatentBreak substitutes words in the input prompt with semantically-equivalent ones, preserving the initial intent of the prompt, instead of adding high-perplexity adversarial suffixes or long templates. These words are chosen by minimizing the distance in the latent space between the representation of the adversarial prompt and that of harmless requests. Our extensive evaluation shows that LatentBreak leads to shorter and low-perplexity prompts, thus outperforming competing jailbreak algorithms against perplexity-based filters on multiple safety-aligned models.

Large scale robot learning has recently shown promise in enabling robots to perform complex tasks by integrating perception, control, and optionally, language understanding into a unified framework. However, they continue to struggle with long-horizon, contact-rich manipulation tasks, such as the handling of deformable objects, where supervision from demonstrations is often inconsistent in quality. In such settings, reward modeling offers a natural solution: by providing grounded progress signals, it can transform noisy demonstrations into stable supervision that generalizes across diverse trajectories. In this work, we introduce a stage-aware, video-based reward modeling framework that jointly predicts the high-level task stage and fine-grained progress within each stage. Reward labels are automatically derived from natural language subtask annotations, enabling consistent progress estimation across variable-length and heterogeneous demonstrations. This design overcomes the limitations of frame-index-based labeling, which collapses in long, variable-duration tasks such as folding a T -shirt. Our reward model demonstrates robustness to demonstration variability, generalization to out-of-distribution scenarios, and strong utility for downstream policy training. Building upon this reward model, we propose the Reward-Aligned Behavior Cloning (RA-BC) framework, which selectively filters high-quality data and reweights training samples according to reward estimates. Extensive experiments demonstrate that the reward model outperforms baselines on out-of-distribution real robot policy rollouts and human demonstration validation. Our approach achieves 83% success on folding T -shirts from the flattened state and 67% from the crumpled state--dramatically surpassing vanilla behavior cloning, which attains only 8% and 0% success under the same training dataset, respectively. Overall, our results highlight reward modeling as a key enabler for scalable, annotation-efficient, and robust imitation learning in long-horizon robotic manipulation. The long-standing vision of enabling robots to seamlessly assist humans in household chores has inspired decades of research in robotics. From tidying living spaces to preparing meals, such capabilities hold the promise of freeing up human time, and improving quality of life.

Reinforcement Learning with Human Feedback (RLHF) and Reinforcement Learning with Verifiable Rewards (RLVR) are the main RL paradigms used in LLM post-training, each offering distinct advantages. However, RLHF struggles with interpretability and reward hacking because it relies on human judgments that usually lack explicit criteria, whereas RLVR is limited in scope by its focus on correctness-based verifiers. We propose Reinforcement Learning with Binary Flexible Feedback (RLBFF), which combines the versatility of human-driven preferences with the precision of rule-based verification, enabling reward models to capture nuanced aspects of response quality beyond mere correctness. RLBFF extracts principles that can be answered in a binary fashion (e.g. accuracy of information: yes, or code readability: no) from natural language feedback. Such principles can then be used to ground Reward Model training as an entailment task (response satisfies or does not satisfy an arbitrary principle). We show that Reward Models trained in this manner can outperform Bradley-Terry models when matched for data and achieve top performance on RM-Bench (86.2%) and JudgeBench (81.4%, #1 on leaderboard as of September 24, 2025). Additionally, users can specify principles of interest at inference time to customize the focus of our reward models, in contrast to Bradley-Terry models. Finally, we present a fully open source recipe (including data) to align Qwen3-32B using RLBFF and our Reward Model, to match or exceed the performance of o3-mini and DeepSeek R1 on general alignment benchmarks of MT-Bench, WildBench, and Arena Hard v2 (at <5% of the inference cost). Models: https://huggingface.co/collections/nvidia/reward-models-10-2025

Uncovering cause-and-effect mechanisms from data is fundamental to scientific progress. While large language models (LLMs) show promise for enhancing causal discovery (CD) from unstructured data, their application to the increasingly prevalent multimodal setting remains a critical challenge. Even with the advent of multimodal LLMs (MLLMs), their efficacy in multimodal CD is hindered by two primary limitations: (1) difficulty in exploring intra- and inter-modal interactions for comprehensive causal variable identification; and (2) insufficiency to handle structural ambiguities with purely observational data. To address these challenges, we propose MLLM-CD, a novel framework for multimodal causal discovery from unstructured data. It consists of three key components: (1) a novel contrastive factor discovery module to identify genuine multimodal factors based on the interactions explored from contrastive sample pairs; (2) a statistical causal structure discovery module to infer causal relationships among discovered factors; and (3) an iterative multimodal counterfactual reasoning module to refine the discovery outcomes iteratively by incorporating the world knowledge and reasoning capabilities of MLLMs. Extensive experiments on both synthetic and real-world datasets demonstrate the effectiveness of the proposed MLLM-CD in revealing genuine factors and causal relationships among them from multimodal unstructured data.

Modern signal processing (SP) pipelines, whether model-based or data-driven, often constrained by complex and fragmented workflow, rely heavily on expert knowledge and manual engineering, and struggle with adaptability and generalization under limited data. In contrast, Large Language Models (LLMs) offer strong reasoning capabilities, broad general-purpose knowledge, in-context learning, and cross-modal transfer abilities, positioning them as powerful tools for automating and generalizing SP workflows. Motivated by these potentials, we introduce SignalLLM, the first general-purpose LLM-based agent framework for general SP tasks. Unlike prior LLM-based SP approaches that are limited to narrow applications or tricky prompting, SignalLLM introduces a principled, modular architecture. It decomposes high-level SP goals into structured subtasks via in-context learning and domain-specific retrieval, followed by hierarchical planning through adaptive retrieval-augmented generation (RAG) and refinement; these subtasks are then executed through prompt-based reasoning, cross-modal reasoning, code synthesis, model invocation, or data-driven LLM-assisted modeling. Its generalizable design enables the flexible selection of problem solving strategies across different signal modalities, task types, and data conditions. We demonstrate the versatility and effectiveness of SignalLLM through five representative tasks in communication and sensing, such as radar target detection, human activity recognition, and text compression. Experimental results show superior performance over traditional and existing LLM-based methods, particularly in few-shot and zero-shot settings.

Speculative decoding accelerates large language model inference, but its reliance on a fixed speculation length is suboptimal in large-batch serving environments with diverse requests. This paper explores a new direction for dynamic adaptation by investigating a novel class of post-hoc, diagnostic signals. We propose Dynamic Speculative Decoding Engine (DSDE), a training-free framework built on two primary components: (1) a predictive signal based on the variance of the Kullback-Leibler (KLD) divergence, which diagnoses the generation's regional stability, and (2) an adaptive speculation length cap to mitigate the straggler problem in per-sequence decoding. Experiments demonstrate the potential of using KLD-based stability signals for dynamic adaptation. An algorithm guided by these signals achieves end-to-end latency competitive with leading baselines and exhibits superior robustness across diverse workloads. This robustness is particularly valuable in challenging low-acceptance-rate regimes, where the proposed signal maintains its diagnostic utility. Collectively, these findings validate post-hoc signals as a valuable component for building more robust and intelligent LLM inference systems, and highlight a promising direction for future research on dynamic speculation length adaptation.

Activation patching is a standard method in mechanistic interpretability for localizing the components of a model responsible for specific behaviors, but it is computationally expensive to apply at scale. Attribution patching offers a faster, gradient-based approximation, yet suffers from noise and reduced reliability in deep, highly non-linear networks. In this work, we introduce Relevance Patching (RelP), which replaces the local gradients in attribution patching with propagation coefficients derived from Layer-wise Relevance Propagation (LRP). LRP propagates the network's output backward through the layers, redistributing relevance to lower-level components according to local propagation rules that ensure properties such as relevance conservation or improved signal-to-noise ratio. Like attribution patching, RelP requires only two forward passes and one backward pass, maintaining computational efficiency while improving faithfulness. We validate RelP across a range of models and tasks, showing that it more accurately approximates activation patching than standard attribution patching, particularly when analyzing residual stream and MLP outputs in the Indirect Object Identification (IOI) task. For instance, for MLP outputs in GPT-2 Large, attribution patching achieves a Pearson correlation of 0.006, whereas RelP reaches 0.956, highlighting the improvement offered by RelP. Additionally, we compare the faithfulness of sparse feature circuits identified by RelP and Integrated Gradients (IG), showing that RelP achieves comparable faithfulness without the extra computational cost associated with IG.

We study how prompt-level inductive biases influence the cognitive behavior of large language models (LLMs) in instructional dialogue. We introduce a symbolic scaffolding method paired with a short-term memory schema designed to promote adaptive, structured reasoning in Socratic tutoring. Using controlled ablation across five system variants, we evaluate model outputs via expert-designed rubrics covering scaffolding, responsiveness, symbolic reasoning, and conversational memory. We present preliminary results using an LLM-based evaluation framework aligned to a cognitively grounded rubric. This enables scalable, systematic comparisons across architectural variants in early-stage experimentation. The preliminary results show that our full system consistently outperforms baseline variants. Analysis reveals that removing memory or symbolic structure degrades key cognitive behaviors, including abstraction, adaptive probing, and conceptual continuity. These findings support a processing-level account in which prompt-level cognitive scaffolds can reliably shape emergent instructional strategies in LLMs.

We propose Collab-REC, a multi-agent framework designed to counteract popularity bias and enhance diversity in tourism recommendations. In our setting, three LLM-based agents -- Personalization, Popularity, and Sustainability generate city suggestions from complementary perspectives. A non-LLM moderator then merges and refines these proposals via multi-round negotiation, ensuring each agent's viewpoint is incorporated while penalizing spurious or repeated responses. Experiments on European city queries show that Collab-REC improves diversity and overall relevance compared to a single-agent baseline, surfacing lesser-visited locales that often remain overlooked. This balanced, context-aware approach addresses over-tourism and better aligns with constraints provided by the user, highlighting the promise of multi-stakeholder collaboration in LLM-driven recommender systems.

In the search for artificial general intelligence, model development and training has focused primarily on vast datasets of known problems and their accepted solutions. This process necessarily produces convergent systems which are fundamentally incapable of the conceptual reframing that is required for genuine creative breakthroughs. Inspired by the divergent cognitive processes that allow humans to make such creative leaps, our work introduces a family of language models, TinyTim, to serve as sources of divergent generation within broader systems. These models have been created by fine-tuning on the anti-parsimonious text of James Joyce's `Finnegans Wake'. Quantitative analysis of both an unsupervised fine-tuned model (TinyTim-V1) and a new instruction-tuned variant (TinyTim-V2) demonstrates a profound capacity for lexical invention; the foundational V1 model exhibits a Yule's K score for lexical richness over twenty times greater than that of convergent baselines. This trait is a stable property of the family, as the instruction-tuned V2 maintains a statistically distinct profile and resists factual convergence, sacrificing benchmark performance to preserve its core generative style. This work establishes a methodology for engineering specialized divergent models that, when paired with convergent systems, can reframe problems and force breakthroughs beyond the reach of statistical optimization alone.