"Grok is coming to Tesla vehicles very soon. Next week at the latest," Elon Musk stated Thursday in a post on X. The EV brand would catch up with the likes of Mercedes-Benz and Volkswagen, which have already integrated ChatGPT into their vehicles as voice assistants. The announcement follows a livestream late Wednesday that Musk held with xAI colleagues where the billionaire unveiled Grok 4, the latest AI model from xAI, his multibillion-dollar initiative to rival OpenAI and Google. The timing for both announcements is unfortunate for an embattled Musk. Over the past few days, a version of Grok built into Musk's X social media platform praised Adolf Hitler and provided antisemitic responses to multiple prompts from X users.

A poster on Reddit lived with a painful clicking jaw, the result of a boxing injury, for five years. They saw specialists, got MRIs, but no one could give them a solution to fix it, until they described the problem to ChatGPT. The AI chatbot suggested a specific jaw-alignment issue might be the problem and offered a technique involving tongue placement as a treatment. The individual tried it, and the clicking stopped. "After five years of just living with it," they wrote on Reddit in April, "this AI gave me a fix in a minute."

We propose a unified food-domain QA framework that combines a large-scale multimodal knowledge graph (MMKG) with generative AI. Our MMKG links 13,000 recipes, 3,000 ingredients, 140,000 relations, and 14,000 images. We generate 40,000 QA pairs using 40 templates and LLaVA/DeepSeek augmentation. Joint fine-tuning of Meta LLaMA 3.1-8B and Stable Diffusion 3.5-Large improves BERTScore by 16.2\%, reduces FID by 37.8\%, and boosts CLIP alignment by 31.1\%. Diagnostic analyses-CLIP-based mismatch detection (35.2\% to 7.3\%) and LLaVA-driven hallucination checks-ensure factual and visual fidelity. A hybrid retrieval-generation strategy achieves 94.1\% accurate image reuse and 85\% adequacy in synthesis. Our results demonstrate that structured knowledge and multimodal generation together enhance reliability and diversity in food QA.

ABSTRACT The recent surge in State Space Models (SSMs) [8, 9], particularly the emergence of Mamba, has established them as strong alternatives or complementary modules to Transformers across diverse domains. In this work, we aim to explore the potential of Mamba-based architectures for text-to-music generation. We adopt discrete tokens of Residual V ector Quantization (RVQ) as the modeling representation and empirically find that a single-layer code-book can capture semantic information in music. Motivated by this observation, we focus on modeling a single-codebook representation and adapt SiMBA, originally designed as a Mamba-based encoder, to function as a decoder for sequence modeling. We compare its performance against a standard Transformer-based decoder. Our results suggest that, under limited-resource settings, SiMBA achieves much faster convergence and generates outputs closer to the ground truth. This demonstrates the promise of SSMs for efficient and expressive text-to-music generation. We put audio examples on Github.

The fine-tuning of pretrained large language models (LLMs) has demonstrated state-of-the-art performance across a range of downstream applications. However, two main challenges hinder the wide adoption of these models: the substantial memory requirements of gradient-based optimizers used for fine-tuning and the critical need to protect the privacy of domain-specific fine-tuning data. As fine-tuning LLMs grows increasingly memory-intensive, a range of strategies has emerged to address this issue. In particular, zeroth-order (ZO) optimization methods recently have gained traction due to their memory efficiency, as they do not require explicit gradient computations. Instead, the zeroth-order gradients can be computed using forward step only, significantly reducing memory use compared to gradient computation. In a pioneering approach, Malladi et al. (2023) introduced a memory-efficient technique for fine-tuning LLMs using two-point Simultaneous Perturbation Stochastic Approximation (SPSA) estimators (Spall, 1992), enabling large model fine-tuning on memory-limited devices. Since then, zeroth-order methods have gained popularity in dealing with large machine learning models due to their memory efficiency and favorable upper bounds on gap from optimality under certain conditions on the Hessian of the objective function (Zhang et al., 2024b,a; Guo et al., 2024). Another major concern in training LLMs is privacy . As large parameterized models are increasingly used in sensitive data applications, these models must protect sensitive information, especially given privacy regulations like the E.U.

Diffusion models have emerged as a powerful class of generative models, achieving state-of-the-art results in continuous data domains such as image and video generation. Their core mechanism involves a forward diffusion process that gradually transforms structured data into a Gaussian-like distribution, followed by a learned reverse process to reconstruct the data. While successful in continuous modalities, applying this framework to discrete data-particularly natural language-remains challenging due to token dependency complexities and the lack of a defined generation order.This thesis investigates the feasibility and performance of discrete diffusion models for natural language generation. Specifically, we evaluate the Discrete Denoising Diffusion Probabilistic Model (D3PM) and compare it with traditional autoregressive (AR) language models. To assess generative performance, we use Bits Per Token (BPT), Negative Log-Likelihood (NLL), Perplexity (PPL), and Batch Processing Speed. Results show the best-performing D3PM model achieves a BPT of 5.72, with a mean of 8.05. The AR model outperforms in compression with a lower mean BPT of 4.59, but D3PM achieves higher processing speed, reaching up to 3.97 batches per sec., indicating potential for parallel generation.All evaluations were conducted under consistent conditions-generating 100,000 tokens per model with a fixed batch size of four-for fair comparison. This research presents a detailed analysis of diffusion-based vs. autoregressive models, highlighting trade-offs in generative quality and efficiency. Findings emphasize both the promise and limitations of diffusion models for discrete data, supporting future work in non-autoregressive language generation.

We often attribute human characteristics to large language models (LLMs) and claim that they "know" certain things. LLMs have an internal probabilistic knowledge that represents information retained during training. How can we assess the veracity of this knowledge? We examine two common methods for probing the veracity of LLMs and discover several assumptions that are flawed. To address these flawed assumptions, we introduce sAwMIL (short for Sparse Aware Multiple-Instance Learning), a probing method that utilizes the internal activations of LLMs to separate statements into true, false, and neither. sAwMIL is based on multiple-instance learning and conformal prediction. We evaluate sAwMIL on 5 validity criteria across 16 open-source LLMs, including both default and chat-based variants, as well as on 3 new datasets. Among the insights we provide are: (1) the veracity signal is often concentrated in the third quarter of an LLM's depth; (2) truth and falsehood signals are not always symmetric; (3) linear probes perform better on chat models than on default models; (4) nonlinear probes may be required to capture veracity signals for some LLMs with reinforcement learning from human feedback or knowledge distillation; and (5) LLMs capture a third type of signal that is distinct from true and false and is neither true nor false. These findings provide a reliable method for verifying what LLMs "know" and how certain they are of their probabilistic internal knowledge.

The evaluation of large language model (LLM) outputs is increasingly performed by other LLMs, a setup commonly known as "LLM-as-a-judge", or autograders. While autograders offer a scalable alternative to human evaluation, they have shown mixed reliability and may exhibit systematic biases, depending on response type, scoring methodology, domain specificity, or other factors. Here we propose a statistical framework based on Bayesian generalised linear models (GLMs) that enables researchers to simultaneously assess their autograders while addressing their primary research questions (e.g., LLM evaluation). Our approach models evaluation outcomes (e.g., scores or pairwise preferences) as a function of properties of the grader (e.g., human vs. autograder) and the evaluated item (e.g., response length or the LLM that generated it), allowing for explicit quantification of scoring differences and potential biases within a unified framework. In addition, our method can be used to augment traditional metrics such as inter-rater agreement, by providing uncertainty estimates and clarifying sources of disagreement. Overall, this approach contributes to more robust and interpretable use of autograders in LLM evaluation, enabling both performance analysis and bias detection.

Decentralized learning offers a promising approach to crowdsource data consumptions and computational workloads across geographically distributed compute interconnected through peer-to-peer networks, accommodating the exponentially increasing demands. However, proper incentives are still in absence, considerably discouraging participation. Our vision is that a fair incentive mechanism relies on fair attribution of contributions to participating nodes, which faces non-trivial challenges arising from the localized connections making influence ``cascade'' in a decentralized network. To overcome this, we design the first method to estimate \textbf{D}ata \textbf{I}nfluence \textbf{C}ascad\textbf{E} (DICE) in a decentralized environment. Theoretically, the framework derives tractable approximations of influence cascade over arbitrary neighbor hops, suggesting the influence cascade is determined by an interplay of data, communication topology, and the curvature of loss landscape. DICE also lays the foundations for applications including selecting suitable collaborators and identifying malicious behaviors. Project page is available at https://raiden-zhu.github.io/blog/2025/DICE/.

Prior work on jailbreak detection has established the importance of adversarial robustness for LLMs but has largely focused on the model ability to resist adversarial inputs and to output safe content, rather than the effectiveness of external supervision systems. The only public and independent benchmark of these guardrails to date evaluates a narrow set of supervisors on limited scenarios. Consequently, no comprehensive public benchmark yet verifies how well supervision systems from the market perform under realistic, diverse attacks. To address this, we introduce BELLS, a Benchmark for the Evaluation of LLM Supervision Systems. The framework is two dimensional: harm severity (benign, borderline, harmful) and adversarial sophistication (direct vs. jailbreak) and provides a rich dataset covering 3 jailbreak families and 11 harm categories. Our evaluations reveal drastic limitations of specialized supervision systems. While they recognize some known jailbreak patterns, their semantic understanding and generalization capabilities are very limited, sometimes with detection rates close to zero when asking a harmful question directly or with a new jailbreak technique such as base64 encoding. Simply asking generalist LLMs if the user question is "harmful or not" largely outperforms these supervisors from the market according to our BELLS score. But frontier LLMs still suffer from metacognitive incoherence, often responding to queries they correctly identify as harmful (up to 30 percent for Claude 3.7 and greater than 50 percent for Mistral Large). These results suggest that simple scaffolding could significantly improve misuse detection robustness, but more research is needed to assess the tradeoffs of such techniques. Our results support the "bitter lesson" of misuse detection: general capabilities of LLMs are necessary to detect a diverse array of misuses and jailbreaks.