Putin and Kim will be among 26 other heads of state who are expected to attend the parade. This is the first time a North Korean leader has attended a Chinese military parade since 1959. China is likely to display its latest weaponry, including hundreds of aircraft, tanks and anti-drone systems. This will be the first time its military's new force structure is being fully showcased in a parade. The highly choreographed event will see tens of thousands of military personnel march in formation through the historic Tiananmen Square, with troops from 45 of the so-called echelons of China's military as well as war veterans. The 70-minute parade, which will be surveyed by Xi, is expected to be closely watched by analysts and western powers.

An overnight Russian drone and missile attack on Ukraine's capital Kyiv has killed at least four people and wounded more than 20 others, officials said. Powerful explosions rocked the city into the early hours of Thursday morning, illuminating the sky and leaving behind columns of smoke as Russian projectiles damaged and destroyed buildings in several districts of the city. The attack was the first major combined Russian drone and missile attack to strike Kyiv since United States President Donald Trump met with Russian President Vladimir Putin in Alaska earlier this month to discuss ending the war in Ukraine. Tymur Tkachenko, the head of Kyiv's city military administration, said a 14-year-old girl was among those reported killed, citing preliminary information. A five-storey residential building in the city's Darnytskyi district was hit directly.

Chip giant Nvidia has set a new sales record, a sign that demand for artificial intelligence remains strong despite fearsthe technology may be overhyped. Nvidia, the world's most valuable company, on Wednesday reported revenue of 46.74bn for the three months that ended in July, a rise of 56 percent year-on-year. Profit for the quarter was 26.42bn, a yearly rise of 59 percent. Nvidia's latest earnings report had been hotly anticipated as the tech giant is widely seen as a barometer of the AI boom, which has lifted the US stock market from all-time high to all-time high. Nvidia CEO Jensen Huang said that production of Blackwell Ultra, Nvidia's latest platform using its most advanced chips, was ramping up "at full speed" and demand for the company's products was "extraordinary".

Chipmaker Nvidia set a fresh sales record in the second quarter, surpassing Wall Street expectations for its artificial intelligence chips. But shares of the chip giant still dropped 2.3% in after hours trading, in a sign that investors' worries of an AI bubble and the repercussions of Donald Trump's trade wars are not quelled. Nvidia's financial report was the first test of investor appetite since last week's mass AI-stock selloff, when several tech stocks saw shares tumble last week amid growing questions over whether AI-driven companies are being overvalued. On Wednesday, Nvidia reported an adjusted earnings per share of 1.08 on 46.74bn in revenue, surpassing Wall Street's projection of 1.01 in earnings per share on 46.05bn in revenue, according to Fact Set data. But investors had high expectations for the company.

We present an extension of the Levenberg-Marquardt algorithm for fitting multichannel nuclear cross section data. Our approach offers a practical and robust alternative to conventional trust-region methods for analyzing experimental data. The CoH$_3$ code, based on the Hauser-Feshbach statistical model, involves a large number of interdependent parameters, making optimization challenging due to the presence of "sloppy" directions in parameter space. To address the uneven distribution of experimental data across reaction channels, we construct a weighted Fisher Information Metric by integrating prior distributions over dataset weights. This framework enables a more balanced treatment of heterogeneous data, improving both parameter estimation and convergence robustness. We show that the resulting weighted Levenberg-Marquardt method yields more physically consistent fits for both raw and smoothed datasets, using experimental data for ${}^{148}$Sm as a representative example. Additionally, we introduce a geometric scaling strategy to accelerate convergence -- a method based on the local geometry of the manifold.

Given a trained neural network, can any specified output be generated by some input? Equivalently, does the network correspond to a function that is surjective? In generative models, surjectivity implies that any output, including harmful or undesirable content, can in principle be generated by the networks, raising concerns about model safety and jailbreak vulnerabilities. In this paper, we prove that many fundamental building blocks of modern neural architectures, such as networks with pre-layer normalization and linear-attention modules, are almost always surjective. As corollaries, widely used generative frameworks, including GPT-style transformers and diffusion models with deterministic ODE solvers, admit inverse mappings for arbitrary outputs. By studying surjectivity of these modern and commonly used neural architectures, we contribute a formalism that sheds light on their unavoidable vulnerability to a broad class of adversarial attacks.

Quantization enables efficient deployment of large language models (LLMs) in resource-constrained environments by significantly reducing memory and computation costs. While quantized LLMs often maintain performance on perplexity and zero-shot tasks, their impact on truthfulness-whether generating truthful or deceptive responses-remains largely unexplored. In this work, we introduce TruthfulnessEval, a comprehensive evaluation framework for assessing the truthfulness of quantized LLMs across three dimensions: (1) Truthfulness on Logical Reasoning; (2) Truthfulness on Common Sense; and (3) Truthfulness on Imitative Falsehoods. Using this framework, we examine mainstream quantization techniques (ranging from 4-bit to extreme 2-bit) across several open-source LLMs. Surprisingly, we find that while quantized models retain internally truthful representations, they are more susceptible to producing false outputs under misleading prompts. To probe this vulnerability, we test 15 rephrased variants of "honest", "neutral" and "deceptive" prompts and observe that "deceptive" prompts can override truth-consistent behavior, whereas "honest" and "neutral" prompts maintain stable outputs. Further, we reveal that quantized models "know" the truth internally yet still produce false outputs when guided by "deceptive" prompts via layer-wise probing and PCA visualizations. Our findings provide insights into future designs of quantization-aware alignment and truthfulness interventions.

Fine-tuning the large language models (LLMs) are prevented by the deficiency of centralized control and the massive computing and communication overhead on the decentralized schemes. While the typical standard federated learning (FL) supports data privacy, the central server requirement creates a single point of attack and vulnerability to poisoning attacks. Generalizing the result in this direction to 70B-parameter models in the heterogeneous, trustless environments has turned out to be a huge, yet unbroken bottleneck. This paper introduces FLock, a decentralized framework for secure and efficient collaborative LLM fine-tuning. Integrating a blockchain-based trust layer with economic incentives, FLock replaces the central aggregator with a secure, auditable protocol for cooperation among untrusted parties. We present the first empirical validation of fine-tuning a 70B LLM in a secure, multi-domain, decentralized setting. Our experiments show the FLock framework defends against backdoor poisoning attacks that compromise standard FL optimizers and fosters synergistic knowledge transfer. The resulting models show a >68% reduction in adversarial attack success rates. The global model also demonstrates superior cross-domain generalization, outperforming models trained in isolation on their own specialized data.

Large Language Models (LLMs) have demonstrated strong capabilities in transforming text descriptions or tables to data visualizations via instruction-tuning methods. However, it is not straightforward to apply these methods directly for a more real-world use case of visualizing data from long documents based on user-given intents, as opposed to the user pre-selecting the relevant content manually. We introduce the task of intent-based chart generation from documents: given a user-specified intent and document(s), the goal is to generate a chart adhering to the intent and grounded on the document(s) in a zero-shot setting. We propose an unsupervised, two-staged framework in which an LLM first extracts relevant information from the document(s) by decomposing the intent and iteratively validates and refines this data. Next, a heuristic-guided module selects an appropriate chart type before final code generation. To assess the data accuracy of the generated charts, we propose an attribution-based metric that uses a structured textual representation of charts, instead of relying on visual decoding metrics that often fail to capture the chart data effectively. To validate our approach, we curate a dataset comprising of 1,242 $<$intent, document, charts$>$ tuples from two domains, finance and scientific, in contrast to the existing datasets that are largely limited to parallel text descriptions/ tables and their corresponding charts. We compare our approach with baselines using single-shot chart generation using LLMs and query-based retrieval methods; our method outperforms by upto $9$ points and $17$ points in terms of chart data accuracy and chart type respectively over the best baselines.

Vision-language models (VLMs), such as CLIP, have gained significant popularity as foundation models, with numerous fine-tuning methods developed to enhance performance on downstream tasks. However, due to their inherent vulnerability and the common practice of selecting from a limited set of open-source models, VLMs suffer from a higher risk of adversarial attacks than traditional vision models. Existing defense techniques typically rely on adversarial fine-tuning during training, which requires labeled data and lacks of flexibility for downstream tasks. T o address these limitations, we propose robust test-time prompt tuning (R-TPT), which mitigates the impact of adversarial attacks during the inference stage. W e first reformulate the classic marginal entropy objective by eliminating the term that introduces conflicts under adversarial conditions, retaining only the pointwise entropy minimization. Furthermore, we introduce a plug-and-play reliability-based weighted ensembling strategy, which aggregates useful information from reliable augmented views to strengthen the defense. R-TPT enhances defense against adversarial attacks without requiring labeled training data while offering high flexibility for inference tasks. Extensive experiments on widely used benchmarks with various attacks demonstrate the effectiveness of R-TPT.