PPP, a prediction-market tracking tool, showed that Polymarket contracts tied to four mathematicians winning the 2026 Fields Medal have all surged over the past week and now trade above 98%. The names cited are Deng Yu, Wang Hong, John Pardon and Jacob Tsimerman. The 2026 International Congress of Mathematicians is scheduled to open on July 23 in Philadelphia, where the Fields Medal and several other prizes are set to be announced. Recent online discussion has pointed to the ICM 2026 website schedule’s front-end code as a possible source of clues about the winners, but the reported list has not been officially confirmed. If Deng Yu and Wang Hong do ultimately receive the prize, it would mark the first time mathematicians of Chinese nationality have won the Fields Medal.

A paper accepted to ICML 2026 has stirred a heated debate after a Reddit post highlighted its central claim: large language models may become more diverse at inference time through prompt design alone. The paper introduces Verbalized Sampling, or VS, a method that asks a model to verbalize a probability distribution while generating outputs. According to the authors, that simple change can reduce mode collapse, the tendency of models to produce repetitive, high-probability answers across tasks such as creative writing, question answering, and code generation. The paper argues that the deeper source of the problem is not only decoding or reward-model design, but a “typicality bias” in human preference data. In that view, annotators systematically favor familiar, fluent, and conventional responses, which pushes aligned models toward safer and more homogeneous outputs. The authors report consistent findings across five preference datasets and multiple base models. In creative writing tasks, they say VS raised diversity to 1.6x to 2.1x that of standard prompting, without reducing factual accuracy or model safety. Critics on Reddit questioned whether a prompt-based method belongs at a top-tier ML conference, citing limited scale, uncertain generalization, and similarities to common prompting practices. Supporters argued the paper’s main contribution is its explanation of mode collapse and its framing of prompt design as a way to study model behavior rather than a superficial trick.

A new benchmark introduced by researchers from the Suzhou Institute for Advanced Research of the University of Science and Technology of China, the University of New South Wales and other institutions argues that large language models hit a clear limit in atom-level manipulation tasks. Presented at ICML 2026, AtomWorld focuses on practical crystal-structure operations rather than text understanding or theory questions, and tests whether a model can follow instructions to modify atomic arrangements correctly. The study reports that scaling still helps on some rule-based tasks, such as atom replacement, deletion and movement. But the gains become unstable when tasks require 3D spatial reasoning and geometric planning, including rotating around an atom, deleting atoms in a spatial region and supercell expansion. In one cited example, Claude Opus 4.6 reached only about a 12% success rate on rotation tasks. The benchmark covers Claude Opus 4.6, GPT-5.4, Gemini 3.1 Pro, Gemini 2.5 Pro, Qwen3-32B, GPT o3, GPT-4o-mini, DeepSeek Chat and Llama3-70B. The paper does not reject Scaling Law outright. Instead, it argues that language scaling alone cannot fill the gap between knowing materials science and carrying out physically valid operations. The authors say AI for Science now needs “Action Scaling,” built around executable actions, simulator feedback, physical-constraint checks and error correction.

LazAI's research paper on QoS-aware token scheduling and private data valuation for multi-modal agentic networks has been accepted at IEEE ICME 2026, which had only a 28.89% acceptance rate out of 3,810 submissions.

Bitmain unveils the Antminer L9, a next-gen ASIC miner for Scrypt coins (LTC, DOGE) with 16 GH/s hashrate and 210 J/G efficiency. The L9 outperforms its predecessor L7 by 78% in hashrate, promising higher profitability for miners.

Internet Capital Markets (ICM) turn any internet idea into a tradable token. Solana-based Believe platform has created 9,845 tokens, $400M volume, and Launch Coin soared 100x to $200M market cap.

Internet Capital Markets (ICM) are reshaping crypto fundraising. Believe platform enables instant tokenization of ideas via Launch Coin, with 9,845+ tokens created, 190K+ traders, and $400M+ volume. Launch Coin surged 100x to $200M market cap. This article explores ICM mechanics, its difference from traditional markets, and why it matters for the crypto ecosystem.

The Internet Capital Markets (ICM) narrative is transforming fundraising by tokenizing any idea. Believe platform has created over 9,800 tokens with $400M trading volume; its token Launch Coin surged 100x. This article breaks down ICM, its mechanics, and its impact on crypto.