Skip to content
K = 0.731
All news · Today8 items

#Reasoning

All collected and published items, ordered by time. Picks are a selected subset.
10/12Mon
  1. The Decoder81

    Google DeepMind’s AlphaProof Nexus solves dozens of open math problems in peer-reviewed study

    Google DeepMind’s AlphaProof Nexus solved dozens of open math problems, according to a peer-reviewed study in Science, including nine of 353 open Erdős problems, two unsolved for 56 years. It also proved 44 OEIS conjectures, solved an open algebraic geometry question, and improved a min-max optimization bound. Built around Gemini 3.1 Pro, it drafts and revises proofs in Lean using checker feedback and can call AlphaProof for harder sub-problems, though DeepMind reports it solved the nine Erdős problems without that tool.

    Why it matters: DeepMind’s use of peer review provides a concrete comparison with OpenAI’s publication of AI proofs on GitHub amid a debate over how AI mathematics should be evaluated.

  2. 量子位47

    Zhipu’s GLM-5.3-Highspeed Model Reportedly Reaches 490 Tokens per Second

    Zhipu has added the GLM-5.3-Highspeed model to Zcode, where it is being gradually rolled out to subscribers at random. A user reported a generation speed of 490 tokens per second; the article says the model is currently available for testing in Zcode 3.15.1. The reported speed has not been independently confirmed, and the author says they could not access the model.

  3. 量子位75

    Huawei and UCL’s Memento 3 Clears 25 ARC-AGI-3 Games Without Updating Model Weights

    Huawei’s Noah’s Ark Lab, UCL, and collaborators report that Memento 3 cleared all 25 public ARC-AGI-3 games and 183 levels using 7518 actions, about 56% fewer than humans, achieving a 100 RHAE score. Built on Claude Opus 5 without weight updates or game-specific fine-tuning, the agent iteratively revises a Markdown rulebook and executable Python world model. In Atari Pong, its learned code-based strategy won three games 21:0 without further LLM calls; the experiments primarily cover game environments with relatively clear rules.

  4. 雷锋网57

    Microsoft releases Qwen-based Microsoft-Decision-1 and tests applications across domains

    Microsoft released Microsoft-Decision-1, a decision model post-trained from Qwen3.5-9B, and says it outperforms LLMs and decision models including Jev 1.13.0 on latency and accuracy in structured decision tasks. Its Xbox research team has used it to process more than 10,000 player feedback entries, while CEO Satya Nadella says Microsoft is testing applications across domains, including incident response, quality control, and scientific discovery. The report describes decision models as smaller, faster models that output predictions, probabilities, and confidence scores rather than natural language.

  5. Ethan Mollick · X40

    GPT-6 Pro finds supposed errors in academic papers all the time, just as part of other work. Confirming the errors & learning their importance would still take a lot of time today, but it will get faster. We don't have the right mechanisms in place to deal a flood of corrections.

    GPT-6 Pro finds supposed errors in academic papers all the time, just as part of other work. Confirming the errors & learning their importance would still take a lot of time today, but it will get faster. We don't have the right mechanisms in place to deal a flood of corrections.

  6. Ethan Mollick · X57

    Ethan Mollick argues advanced model use at labs remains concentrated in a few fields

    Ethan Mollick argues that many of the 5,000 people at labs with access to advanced models still haven’t used them to their full extent, with use concentrated in cybersecurity, math, and AI R&D. There are many other fields to advance and problems to solve, he adds, responding to @karpathy’s argument that the gap in shared understanding of LLM capabilities is widening.

  8. The Decoder62

    Agent teams cost up to 5.1x more for little quality gain, Vals AI finds

    Evals company Vals AI tested GPT-6 Sol and Claude Opus 5.5 on the Vibe Code Bench, both solo and as teams, at medium and maximum reasoning effort. Teams cost between 1.8x and 5.1x more than single agents, and only one of four comparisons showed a statistically significant improvement: GPT-6 Sol at medium reasoning, where the team scored 7.3 points higher. At maximum reasoning, teams gave neither model a real advantage.

10/11Sun
  1. Clément Delangue · X53

    OpenAI’s math problems become open-source RL environments on Hugging Face

    We turned @OpenAI's math problems into open-source RL environments on @huggingface! They are early and rough, and the verifier only accepts the exact original formalization, so an equivalent proof can still score 0. I find this an exciting direction for turning open research into executable environments anyone can build on to train better open models! https://huggingface.co/datasets/FineEnvs/openai-math

  2. 量子位71

    He Kaiming team releases VISTA visual harness for multimodal reasoning

    He Kaiming's team published VISTA, a visual-native harness that lets existing multimodal models observe environments, store raw frames outside context, and recall, zoom or inspect them during reasoning without retraining. Paired with VISTA, Claude Opus 5.0 completed all 25 public ARC-AGI-3 games with a perfect 100 action-efficiency score, using 57.4% fewer actions than the human first-play baseline, while GPT-5.6 Sol scored 99. The team also tested browser games, mazes and connect-the-dots tasks, and lists embodied tasks as future work.

  3. 量子位74

    TypeSafe AI raises $870M Series A at $7.5B valuation for Jev decision model

    TypeSafe AI, the San Francisco company behind the decision-only model Jev, announced an $870 million Series A at a $7.5 billion post-money valuation, up from a $200 million valuation at its $40 million seed round 24 days earlier. a16z, Sequoia and DCVC participated, and a16z partner Martin Casado joined the board. The company says about one third of Fortune 500 companies now use or integrate Jev, which outputs decisions rather than chat.

  4. 量子位88

    OpenAI Posts Two Kakeya Conjecture Manuscripts, Extending Work to Four Dimensions

    On October 7, OpenAI released 722 mathematics manuscripts, including two on the Kakeya conjecture: a 97-page paper claiming a three-dimensional maximal-function result stronger than Wang and Zahl's, and a 175-page paper claiming the four-dimensional Hausdorff dimension case. Both are attributed to OpenAI, are not Lean-formalized or peer-reviewed, and OpenAI's README says some unformalized results may be flawed.

    Why it matters: The piece traces how two OpenAI manuscripts extend Wang and Zahl's three-dimensional Kakeya work to a stronger maximal-function version and to four dimensions, while noting the proofs lack peer review.

  5. François Chollet · X23

    How AI shifted from a transductive paradigm (token-by-token answer completion) to an inductive paradigm (on-the-fly reasoning chain synthesis), in one chart. This is the big story of 2025 and 2026. This is what made coding agents work.

    How AI shifted from a transductive paradigm (token-by-token answer completion) to an inductive paradigm (on-the-fly reasoning chain synthesis), in one chart. This is the big story of 2025 and 2026. This is what made coding agents work.

10/10Sat
  1. The Decoder66

    Microsoft launches Decision-1 decision model based on Qwen3.5-9B

    Microsoft released Decision-1, a decision model for classifications, evaluations and routing decisions, built on Qwen3.5-9B. Microsoft says it tops 36 benchmarks covering nearly 150,000 questions with 83.5% accuracy and 85 ms latency, and is 2.5 times faster than runner-up H2O-Lightning-4B. It is available through Microsoft Foundry and OpenRouter at $0.042 per million input tokens, with output tokens free.

  2. 雷锋网28

    IROS 2026 keynote: more data does not mean better robot decisions

    At IROS 2026, University of Queensland researcher Jen Jen Chung argued in a keynote that data volume is not the same as information, and that blindly piling up data can degrade robot policies. Drawing on drone soaring, multi-robot communication and contact-rich assembly work, she said data value is unevenly distributed and that sharing or collecting only decision-changing information beats full broadcasts or uniform sampling.

  3. 雷锋网36

    IROS 2026 Signals Embodied AI's Shift From Concepts to Engineering

    At IROS 2026, running September 27 to October 1 in Pittsburgh, Leiphone reports embodied AI research is converging on four trends: foundation models internalizing physics, "brain-cerebellum" architectures becoming standard, touch as core to dexterous manipulation, and cheaper data collection. Papers cited include ULTRA, LITHE, VOFA, KineFuse, VTAP Gripper, ViHaTeleop, and HALO, alongside workshops on physical world models and uncertainty.

  4. Ethan Mollick · X55

    Astra and Opus 5 beat Montezuma's Revenge, but Metaculus AGI bet remains unresolved

    Ethan Mollick says Astra and Opus 5 both beat Montezuma's Revenge, but resolving the AGI bet also requires AI to win a discontinued prize resembling a weak Turing Test. He says Metaculus has decided to recreate the test to confirm that criterion is met. In the quoted post, Metaculus says GPT-6 Astra beat the game last month, but its “weakly general AI” question, launched in 2020, remains unresolved.

  6. Satya Nadella · X49

    Introducing Microsoft-Decision-1, our new model for fast decision-making. It delivers top performance on structured decision tasks, outperforming both LLMs and other decision models in latency and quality. We’re already testing it across Microsoft for everything from incident response and quality control to scientific discovery.

    Introducing Microsoft-Decision-1, our new model for fast decision-making. It delivers top performance on structured decision tasks, outperforming both LLMs and other decision models in latency and quality. We’re already testing it across Microsoft for everything from incident response and quality control to scientific discovery.

10/9Fri
  2. 量子位62

    ByteDance Seed finds why DeepSeek-V4 can be inconsistent: chunked KV-cache compression creates position sensitivity

    The ByteDance Seed team found that the DeepSeek-V4 series shows periodic fluctuations in long-context retrieval depending on where information appears, with a cycle of every 4 tokens. In the 128K needle-in-a-haystack test, moving the same information to a different position changed accuracy by as much as 40.2 percentage points.

  3. vLLM · X40

    vLLM Semantic Router team ships Decision 2.0 for single-pass routing and classification

    Great work from the vLLM Semantic Router team on Decision 2.0. Multiple questions about one input, answered in a single forward pass with per-option probabilities. Useful for routing and classification. Quoted post by @XunzhuoLiu: In every size. Decision 2.0: six open decision models, 0.6B to 27B. On the #Jev Decision Index 0.3, each one tops the same-size open models we lined it up against. #Trending collections on Hugging Face 🤗 Decision 2.0: https://huggingface.co/collections/vllm-sr/decision-20 Vela 2.0: https://huggingface.co/collections/vllm-sr/vela-20

  4. 腾讯混元 · X58

    Tencent Hy and universities release ExplorationBench for AI exploration

    Researchers from Tencent Hy, Fudan University, and Tsinghua University released ExplorationBench, a benchmark for measuring how AI systems explore. It uses two verifiable sandboxes, AlienCode with 31 hidden rule changes and 70 tasks, and AlienLogic with 24 patched inference rules and 70 theorems, graded by an interpreter or proof checker rather than an LLM judge. Across 10 frontier systems, feedback beat thinking alone, with the best AlienCode run reaching 89.0% after four rounds.

  5. NVIDIA Blog59

    Into the Omniverse: How Developers Turn Ideas Into Simulations With Frontier AI Agents

    NVIDIA explains how its developers use frontier AI agents, including GPT-6 Astra, with Omniverse libraries to build and refine simulations through natural-language instructions, human review and iterative testing. Examples span warehouse scaling, cable routing, robot asset preparation, autonomous-driving workflows and sensor-based digital twin refinement. One warehouse simulation sustained 30 frames per second with 80 conveyors, 40 arms and 800 boxes during a two-minute test; a robot CAD example used estimated physical properties that were not calibrated against the real robot.

  7. OpenAI Developers · X81

    OpenAI rolls out Ultrafast for GPT-6.1 Sol across API, Codex, and ChatGPT Work

    Ultrafast is rolling out today for GPT-6.1 Sol in the API, Codex, and ChatGPT Work. OpenAI describes it as offering near-Astra intelligence at up to 8x faster speeds than Sol Standard. https://t.co/Cyy9k2hifV

    Why it matters: The comparison with Sol Standard gives developers a concrete speed reference for the rollout, alongside OpenAI’s stated positioning of its intelligence near Astra.

  8. Latent Space36

    Periodic Labs' Liam Fedus and Ekin Dogus Cubuk on "synthesis superintelligence": from semiconductors to superconductors

    In an interview, Periodic Labs' Liam Fedus and Ekin Dogus Cubuk proposed a "synthesis superintelligence" approach: reinforcement learning in physical experiments, AI-driven materials characterization, density functional theory (DFT), and high-throughput labs. The aim is for models to learn the full process of doing science, rather than only studying published results.

10/8Thu
  1. vLLM · X62

    vLLM v0.31.0 ships with DeepSeek-V4.1-Flash support and fast restart

    vLLM v0.31.0 is live, spanning 717 commits from 307 contributors, including 96 first-timers. Highlights include DeepSeek-V4.1-Flash support with FlashMLA mega attention and NVFP4 KV caching, a vllm preload fast-restart path that keeps weights in GPU memory, Model Runner V2 speculative decoding, and large-scale serving features such as MoonEP balanced all-to-all and DeepEPv2. Full release notes are on GitHub.

  7. Noam Brown · X62

    Noam Brown says LLMs surpass top human experts on some research problems

    Noam Brown argues that LLMs have crossed an important threshold by surpassing top human experts on some research problems, with the gap between slightly below and slightly above expert performance helping explain the seemingly sudden surge in math results. He emphasizes that capabilities remain jagged and models are still worse than humans in many ways, but expects continued improvements to bring breakthroughs in other domains and a new era of scientific discovery.

  8. NVIDIA AI · X38

    An AI agent makes a mistake early in a task, then keeps going in the wrong direction. Our researchers built PivotOPD to teach agents how to avoid those mistakes and recover when they happen. During training, a teacher model shows the agent a better action and how to get back on https://t.co/7hMLMdExEH

    An AI agent makes a mistake early in a task, then keeps going in the wrong direction. Our researchers built PivotOPD to teach agents how to avoid those mistakes and recover when they happen. During training, a teacher model shows the agent a better action and how to get back on https://t.co/7hMLMdExEH

  9. The Decoder82

    Anthropic releases Claude Haiku 5.5 with prices cut by up to 90%

    Anthropic released Claude Haiku 5.5, its fastest and least expensive small model so far. Average prices are about 75% lower than Haiku 4.5; for prompts under 100,000 tokens, prices are up to 90% lower, while prompts above that threshold cost five times as much.

    Why it matters: The source provides Haiku 5.5's pricing table and comparisons across several benchmarks, helping readers assess the tradeoffs for cost-sensitive tasks.

10/7Wed
  2. Hugging Face Blog62

    Nemotron fine-tuning reaches gold-medal level at both IOI 2026 and IMO 2026

    The NVIDIA team started with Nemotron 3 and used supervised fine-tuning, reinforcement learning, and a generate-verify-improve reasoning process to build specialized models that reached gold-medal thresholds at both IOI 2026 and IMO 2026.

    Why it matters: The source gives a complete recipe for fine-tuning Nemotron 3 to gold-medal-level results at IOI and IMO, which readers can use to reproduce its data and reasoning workflow.