Large language models (LLMs) can generate survey responses at low cost, but their reliability varies substantially across questions and is unknown before data collection. Deploying LLMs in surveys still requires costly human responses for verification and correction. How should a limited human-labeling budget be allocated across questions in real time? We propose an adaptive allocation algorithm that learns which questions are hardest for the LLM while simultaneously collecting human responses. Each human label serves a dual role: it improves the estimate for that question and reveals how well the LLM predicts human responses on it. The algorithm directs more budget to questions where the LLM is least reliable, without requiring any prior knowledge of question-level LLM accuracy. We prove that the allocation gap relative to the best possible allocation vanishes as the budget grows, and validate the approach on both synthetic data and a real survey dataset with 68 questions and over 2000 respondents. On real survey data, the standard practice of allocating human labels uniformly across questions wastes 10--12% of the budget relative to the optimal; our algorithm reduces this waste to 2--6%, and the advantage grows as questions become more heterogeneous in LLM prediction quality. The algorithm achieves the same estimation quality as traditional uniform sampling with fewer human samples, requires no pilot study, and is backed by formal performance guarantees validated on real survey data. More broadly, the framework applies whenever scarce human oversight must be allocated across tasks where LLM reliability is unknown.

Thismakesinferenceveryfast,andmaynot even need to use all input featuresto make a prediction (with axis-aligned trees).

Neural Information Processing Systems http://nips.cc/

Mathematicians say Google's AI tools are supercharging their research AI tools developed by Google DeepMind are surprisingly effective at assisting mathematical research and could usher in a wave of AI-powered mathematical discovery at a previously unseen scale, say mathematicians who have tested the technology. In May, Google announced an AI system called AlphaEvolve that could find new algorithms and mathematical formulae. The system works by exploring many possible solutions, produced by Google's AI chatbot Gemini. Crucially, though, these are fed to a separate AI evaluator that can filter out the nonsensical solutions that a chatbot inevitably generates . At the time, Google researchers tested AlphaEvolve on more than 50 open mathematical problems and found that, in three-quarters of cases, the system could rediscover the best-known solutions found by humans.

This paper investigates whether Bittensor can be considered the Bitcoin of decentralized Artificial Intelligence by directly comparing its tokenomics, decentralization properties, consensus mechanism, and incentive structure against those of Bitcoin. Leveraging on-chain data from all 64 active Bittensor subnets, we first document considerable concentration in both stake and rewards. We further show that rewards are overwhelmingly driven by stake, highlighting a clear misalignment between quality and compensation. As a remedy, we put forward a series of two-pronged protocol-level interventions. For incentive realignment, our proposed solutions include performance-weighted emission split, composite scoring, and a trust-bonus multiplier. As for mitigating security vulnerability due to stake concentration, we propose and empirically validate stake cap at the 88th percentile, which elevates the median coalition size required for a 51-percent attack and remains robust across daily, weekly, and monthly snapshots.

Terence Tao, a mathematics professor at UCLA, is a real-life superintelligence. The "Mozart of Math," as he is sometimes called, is widely considered the world's greatest living mathematician. He has won numerous awards, including the equivalent of a Nobel Prize for mathematics, for his advances and proofs. Right now, AI is nowhere close to his level. But technology companies are trying to get it there.

Microarchitecture simulators are indispensable tools for microarchitecture designers to validate, estimate, and optimize new hardware that meets specific design requirements. While the quest for a fast, accurate and detailed microarchitecture simulation has been ongoing for decades, existing simulators excel and fall short at different aspects: (i) Although execution-driven simulation is accurate and detailed, it is extremely slow and requires expert-level experience to design. (ii) Trace-driven simulation reuses the execution traces in pursuit of fast simulation but faces accuracy concerns and fails to achieve significant speedup. (iii) Emerging deep learning (DL)-based simulations are remarkably fast and have acceptable accuracy but fail to provide adequate low-level microarchitectural performance metrics crucial for microarchitectural bottleneck analysis. Additionally, they introduce substantial overheads from trace regeneration and model re-training when simulating a new microarchitecture. Re-thinking the advantages and limitations of the aforementioned simulation paradigms, this paper introduces TAO that redesigns the DL-based simulation with three primary contributions: First, we propose a new training dataset design such that the subsequent simulation only needs functional trace as inputs, which can be rapidly generated and reused across microarchitectures. Second, we redesign the input features and the DL model using self-attention to support predicting various performance metrics. Third, we propose techniques to train a microarchitecture agnostic embedding layer that enables fast transfer learning between different microarchitectural configurations and reduces the re-training overhead of conventional DL-based simulators. Our extensive evaluation shows TAO can reduce the overall training and simulation time by 18.06x over the state-of-the-art DL-based endeavors.

Hello Nature readers, would you like to get this Briefing in your inbox free every day? Zhurong (pictured, centre) spent a year exploring Mars before it went into hibernation last May.Credit: Xinhua/Shutterstock Is something amiss with Zhurong, China's first Mars rover? The vehicle has been in induced hibernation since May. It was supposed to awaken last month, but the Chinese space agency has been tight-lipped about its status, leading some researchers to speculate that it might not have survived the harsh Martian winter and dust storms. "There's a long history of solar-powered landers and rovers on Mars running out of power," says astrobiologist David Flannery.

In this series, we are going to write a neural net completely from scratch (down to rudimentary tensor operations) with Nvidia's CUDA, the GPU parallel computing platform behind modern deep learning libraries. You can find the GitHub repository here. Audience: Familiarity with core concepts of C, such as pointers and object-oriented programming, is needed, and you must be thoroughly comfortable with neural networks and their various aspects like backpropagation. However, you need not know any CUDA or parallel programming, and everything will be covered in the upcoming articles. Without further ado, let's get coding!

"Whenever you compete, you have to accept simple rules – someone wins, someone loses, and usually the winner takes it all." For this week's ML practitioner's series, Analytics India Magazine got in touch with Oleg Yaroshevskiy from Ukraine. In this interview, he shares his experiences from his journey to the top 20 in one of the toughest data science competitions in the world. Oleg majored in maths and statistics from Cybernetics Faculty of Taras Shevchenko National University of Kyiv, which was co-founded by Victor Glushkov, one of the cybernetics pioneers who played a key role in the advancement of theoretical computer science, including artificial intelligence. Oleg had a formal introduction to machine learning (ML) during his graduation days where he had studied neural networks along with the popular Andrew NG's course on Coursera back in 2013.