Synthetic data, or data generated by machine learning models, is increasingly emerging as a solution to the data access problem. However, its use introduces significant governance and accountability challenges, and potentially debases existing governance paradigms, such as compute and data governance. In this paper, we identify 3 key governance and accountability challenges that synthetic data poses - it can enable the increased emergence of malicious actors, spontaneous biases and value drift. We thus craft 3 technical mechanisms to address these specific challenges, finding applications for synthetic data towards adversarial training, bias mitigation and value reinforcement. These could not only counteract the risks of synthetic data, but serve as critical levers for governance of the frontier in the future.

Recent advancements in large language models (LLMs) have demonstrated remarkable reasoning capabilities through long chain-of-thought (CoT) reasoning. The R1 distillation scheme has emerged as a promising approach for training cost-effective models with enhanced reasoning abilities. However, the underlying mechanisms driving its effectiveness remain unclear. This study examines the universality of distillation data and identifies key components that enable the efficient transfer of long-chain reasoning capabilities in LLM distillation. Our findings reveal that the effectiveness of long CoT reasoning distillation from teacher models like Qwen-QwQ degrades significantly on nonhomologous models, challenging the assumed universality of current distillation methods. To gain deeper insights into the structure and patterns of long CoT reasoning, we propose DLCoT (Deconstructing Long Chain-of-Thought), a distillation data enhancement framework. DLCoT consists of three key steps: (1) data segmentation to decompose complex long CoT structures, (2) simplification by eliminating unsolvable and redundant solutions, and (3) optimization of intermediate error states. Our approach significantly improves model performance and token efficiency, facilitating the development of high-performance LLMs.

Language models (LMs) have exhibited exceptional versatility in reasoning and in-depth financial analysis through their proprietary information processing capabilities. Previous research focused on evaluating classification performance while often overlooking explainability or pre-conceived that refined explanation corresponds to higher classification accuracy. Using a public dataset in finance domain, we quantitatively evaluated self-explanations by LMs, focusing on their factuality and causality. We identified the statistically significant relationship between the accuracy of classifications and the factuality or causality of self-explanations. Our study built an empirical foundation for approximating classification confidence through self-explanations and for optimizing classification via proprietary reasoning.

Pre-trained Transformers, through in-context learning (ICL), have demonstrated exceptional capabilities to adapt to new tasks using example prompts without model update. Transformer-based wireless receivers, where prompts consist of the pilot data in the form of transmitted and received signal pairs, have shown high detection accuracy when pilot data are abundant. However, pilot information is often costly and limited in practice. In this work, we propose the DEcision Feedback INcontExt Detection (DEFINED) solution as a new wireless receiver design, which bypasses channel estimation and directly performs symbol detection using the (sometimes extremely) limited pilot data. The key innovation in DEFINED is the proposed decision feedback mechanism in ICL, where we sequentially incorporate the detected symbols into the prompts as pseudo-labels to improve the detection for subsequent symbols. Furthermore, we proposed another detection method where we combine ICL with Semi-Supervised Learning (SSL) to extract information from both labeled and unlabeled data during inference, thus avoiding the errors propagated during the decision feedback process of the original DEFINED. Extensive experiments across a broad range of wireless communication settings demonstrate that a small Transformer trained with DEFINED or IC-SSL achieves significant performance improvements over conventional methods, in some cases only needing a single pilot pair to achieve similar performance of the latter with more than 4 pilot pairs.

Isla Fulford, a researcher at OpenAI, had a hunch that Deep Research would be a hit even before it was released. Fulford had helped build the artificial intelligence agent, which autonomously explores the web, deciding for itself what links to click, what to read, and what to collate into an in-depth report. OpenAI first made Deep Research available internally; whenever it went down, Fulford says, she was inundated with queries from colleagues eager to have it back. "The number of people who were DMing me made us pretty excited," says Fulford. Since going live to the public on February 2, Deep Research has proven to be a hit with many users outside the company too.

If you don't follow science news much, you may only be aware of Artificial Intelligence as a new-fangled thingy that somehow makes TikToks of kittens singing versions of Don't You Want Me Like I Want You Baby in an ickle kid voice. If that's your level of engagement with AI, then this competent, fluent documentary offers a very approachable entry into the subject. As it happens, DeepMind is now officially called Google DeepMind and is part of a suite of divisions developing robotics and solving problems using AI – sorry, AGI because it's not just Artificial Intelligence we're talking about here, but Artificial General Intelligence (it gets explained). The fact that Google and its tech overlords are involved, however amiable they seem when seen in their office casual dress (former CEO Eric Schmidt at least), means this has more than a little flavour of corporate video, investor-fluffing, and self-congratulatory smugness about it. But DeepMind's British founder and CEO Demis Hassabis manages to come across as a pretty nice guy with a genuinely interesting backstory.

Materials synthesis is vital for innovations such as energy storage, catalysis, electronics, and biomedical devices. Yet, the process relies heavily on empirical, trial-and-error methods guided by expert intuition. Our work aims to support the materials science community by providing a practical, data-driven resource. We have curated a comprehensive dataset of 17K expert-verified synthesis recipes from open-access literature, which forms the basis of our newly developed benchmark, AlchemyBench. AlchemyBench offers an end-to-end framework that supports research in large language models applied to synthesis prediction. It encompasses key tasks, including raw materials and equipment prediction, synthesis procedure generation, and characterization outcome forecasting. We propose an LLM-as-a-Judge framework that leverages large language models for automated evaluation, demonstrating strong statistical agreement with expert assessments. Overall, our contributions offer a supportive foundation for exploring the capabilities of LLMs in predicting and guiding materials synthesis, ultimately paving the way for more efficient experimental design and accelerated innovation in materials science.

The impressive capabilities of Large Language Models (LLMs) across diverse tasks are now well-established, yet their effective deployment necessitates careful hyperparameter optimization. Through extensive empirical studies involving grid searches across diverse configurations, we discover universal scaling laws governing these hyperparameters: optimal learning rate follows a power-law relationship with both model parameters and data sizes, while optimal batch size scales primarily with data sizes. Our analysis reveals a convex optimization landscape for hyperparameters under fixed models and data size conditions. This convexity implies an optimal hyperparameter plateau. We contribute a universal, plug-and-play optimal hyperparameter tool for the community. Its estimated values on the test set are merely 0.09% away from the globally optimal LLM performance found via an exhaustive search. These laws demonstrate remarkable robustness across variations in model sparsity, training data distribution, and model shape. To our best known, this is the first work that unifies different model shapes and structures, such as Mixture-of-Experts models and dense transformers, as well as establishes optimal hyperparameter scaling laws across diverse data distributions. This exhaustive optimization process demands substantial computational resources, utilizing nearly one million NVIDIA H800 GPU hours to train 3,700 LLMs of varying sizes and hyperparameters from scratch and consuming approximately 100 trillion tokens in total. To facilitate reproducibility and further research, we will progressively release all loss measurements and model checkpoints through our designated repository https://step-law.github.io/

Large Language Models (LLMs) have substantially improved the conversational capabilities of social robots. Nevertheless, for an intuitive and fluent human-robot interaction, robots should be able to ground the conversation by relating ambiguous or underspecified spoken utterances to the current physical situation and to the intents expressed non verbally by the user, for example by using referential gaze. Here we propose a representation integrating speech and gaze to enable LLMs to obtain higher situated awareness and correctly resolve ambiguous requests. Our approach relies on a text-based semantic translation of the scanpath produced by the user along with the verbal requests and demonstrates LLM's capabilities to reason about gaze behavior, robustly ignoring spurious glances or irrelevant objects. We validate the system across multiple tasks and two scenarios, showing its generality and accuracy, and demonstrate its implementation on a robotic platform, closing the loop from request interpretation to execution.

Despite the recent success of image-text contrastive models like CLIP and SigLIP, these models often struggle with vision-centric tasks that demand high-fidelity image understanding, such as counting, depth estimation, and fine-grained object recognition. These models, by performing language alignment, tend to prioritize high-level semantics over visual understanding, weakening their image understanding. On the other hand, vision-focused models are great at processing visual information but struggle to understand language, limiting their flexibility for language-driven tasks. In this work, we introduce TULIP, an open-source, drop-in replacement for existing CLIP-like models. Our method leverages generative data augmentation, enhanced image-image and text-text contrastive learning, and image/text reconstruction regularization to learn fine-grained visual features while preserving global semantic alignment. Our approach, scaling to over 1B parameters, outperforms existing state-of-the-art (SOTA) models across multiple benchmarks, establishing a new SOTA zero-shot performance on ImageNet-1K, delivering up to a $2\times$ enhancement over SigLIP on RxRx1 in linear probing for few-shot classification, and improving vision-language models, achieving over $3\times$ higher scores than SigLIP on MMVP. Our code/checkpoints are available at https://tulip-berkeley.github.io