Base Date 2026-11-03
Last Run 2026-11-03 21:00:25
S&P 500
7,569
NDX 100
28,180
VIX
20.82
US 10Y
4.95%
DXY
101.86
WTI
$96.60
Gold
$3,932
USD/KRW
1,344
SOX
10,599
Trump Approve
35.4%
S&P 500
7,569
NDX 100
28,180
VIX
20.82
US 10Y
4.95%
DXY
101.86
WTI
$96.60
Gold
$3,932
USD/KRW
1,344
SOX
10,599
Trump Approve
35.4%
AI LEADERS · MOVES & PAPERS

AI Leaders: Moves & Papers — latest job changes and research

LAST UPDATE 2026-11-03 08:00 KST PEOPLE 77 · PAPERS 76 · MOVES 0
Where the key people at AI labs, big tech and universities move, and what they publish, are early signals of where technology and competitive advantage are heading. Career moves and papers are collected automatically every day, checked by the editor and summarized in plain language with links to the originals.

Highlights

Career moves

Recent career records (from Wikidata; may include unverified information)

Papers

Paper2026-10-30

Development of a Hierarchical Model Surpassing Limitations in Complex Robot Control

Hierarchical World Models with Implicit Dynamics
Yann LeCun +6 more
What
Researchers developed ID-WM, a hierarchical world model capable of performing complex and long-horizon tasks by separating state accessibility and local control.
Key result
It recorded a 38.5 percent performance improvement compared to the previous state-of-the-art method, and a 135 percent performance improvement compared to single-level world models.
Why it matters
This research demonstrates the possibility for robots to plan long-term while reducing errors, even in complex and novel environments.AI summary
Paper2026-10-29 · arXiv (Cornell University)

A technology for jointly training robot data using diverse controllers

Cross-Controller Transfer via In-Context Behavior Adaptation
Chelsea Finn +2 more
What
Developed a technology to utilize data from robots using different control methods for learning without conflicts.
Key result
The average success rate improved by 33 percentage points compared to conventional methods.
Why it matters
Demonstrates the potential to more effectively utilize robot data collected from diverse environments.AI summary
Paper2026-10-29 · arXiv (Cornell University)

Proving that artificial intelligence models learn the same world even across different information modalities

Proving and Improving the Platonic Representation Hypothesis with JEPAs
Yann LeCun +1 more
What
Researchers proved the hypothesis that artificial intelligence models trained on different input modalities converge to the same world representation into a mathematical theorem and proposed a method to improve it.
Key result
It was confirmed that models using a good perspective can align representations much more efficiently with significantly less data than the next-token prediction method.
Why it matters
This study can contribute to improving how artificial intelligence efficiently understands data and connects different information modalities.AI summary
Paper2026-10-29 · arXiv (Cornell University)

Improving Psychiatric Intake AI Performance through Simulated Environment Training

Training language models in simulation for psychiatric intake
Mark Chen +55 more
What
Researchers demonstrated that language models trained through reinforcement learning in a simulated environment significantly improve structured text-based psychiatric initial intake capabilities.
Key result
In evaluations involving psychiatrists, the trained artificial intelligence outperformed the base model across multiple metrics, including clinical score, patient-centered communication, and differential diagnosis appropriateness.
Why it matters
This shows that simulated environments can serve as a useful testbed for training conversational artificial intelligence in psychiatry and verifying its safety.AI summary
Paper2026-10-28

Development of High-Efficiency AI Models Capable of Complex Tasks

Nemotron 3.5 Super and Lightning: Open, Efficient Mixture-of-Experts Hybrid Mamba-Transformer Models for Agentic AI
Ashish Vaswani +499 more
What
Researchers have developed Nemotron 3.5 Super and Lightning, open AI models well-suited for reasoning, long-context understanding, and autonomous task execution.
Key result
The Super model achieved 51.78% on software engineering evaluations while using fewer active parameters, outperforming previous models, and demonstrated high throughput performance through various optimizations.
Why it matters
This research demonstrates the possibility of simultaneously increasing the efficiency and performance of large-scale models so they can be practically utilized in diverse environments.AI summary
Paper2026-10-27 · SuperIntelligence - Robotics - Safety & Alignment

The US White House and AI companies have signed an agreement on superintelligence AI governance.

White House Accord on Super Intelligence
Dario Amodei +6 more
What
The US President and CEOs of major AI companies signed the White House Superintelligence Accord, a four-stage governance commitment for the development of superintelligent artificial intelligence.
Key result
Participating companies agreed to implement strong internal controls, operate independent internal verification teams, undergo external audits, and ensure oversight through an independent board committee.
Why it matters
This serves as a catalyst for advanced artificial intelligence developers to introduce systematic internal controls and external audits for security and safety.AI summary
Paper2026-10-22 · Science

Artificial intelligence achieved higher accuracy than doctors in iterative diagnosis processes

Sequential diagnosis with language models
What
Researchers developed a large language model based AI system and an interactive evaluation benchmark that mimic the clinical diagnosis process.
Key result
This AI system achieved accuracies of 83.2 percent and 80.3 percent respectively in medical case evaluations, showing superior performance compared to general practitioners.
Why it matters
Although in a controlled environment, this demonstrates the potential for language models to support more accurate and cost-effective diagnoses.AI summary
Paper2026-10-22

Discovered the laws for improving performance using artificial intelligence-generated synthetic data

Rewriting Scaling Laws for Synthetic Data with Compute and Data Constraints
Percy Liang +6 more
What
When using synthetic rewrites where language models rewrite natural language documents into different styles for pre-training, we studied the laws for predicting the optimal ratio and performance under compute and data constraints.
Key result
As compute increases, the optimal proportion of synthetic data increases, and models trained on synthetic data generated by larger or equally sized models showed superior performance compared to models using only original data.
Why it matters
This suggests that AI training methods utilizing synthetic data are predictable, enabling the overcoming of data limitations and efficient performance improvement.AI summary
Paper2026-10-22 · BMJ digital health & AI

Definition and Evaluation Criteria for Medical Superintelligence in Artificial Intelligence Proposed

Medical superintelligence: foundations and aspirations
What
Discussed how artificial intelligence technology contributes to more effective medical services and introduced six interdependent dimensions of medical intelligence.
Key result
Revealed that improving healthcare quality requires not only outstanding performance in individual tasks, but also systems that collaborate with humans to perform integrated medical reasoning throughout the entire care process.
Why it matters
Proposed a direction to combine these capabilities with existing medical intelligence to enhance the quality and accessibility of medical services.AI summary
Paper2026-10-20 · Zenodo (CERN European Organization for Nuclear Research)

Release of Structural Data Used in AI-Based Enzyme Design Paper

Supplementary Data 2. Designed and predicted structures of all designs
What
The research team has released all structural data files covered in the paper on AI-driven enzyme design for novel chemical reactions.
Key result
The compressed file contains structures generated by the AI model, redesigned structures, and additional predicted structural data.
Why it matters
The released data allows related researchers to verify the results of AI-based enzyme design and conduct further research.AI summary
Paper2026-10-12 · arXiv (Cornell University)

Development of a Technology for Efficiently Discovering Diverse Antibody Candidates Using Artificial Intelligence

Multi-Objective Amortized CDRH3 Sequence Sampling Diversifies Promising Antibody Candidates
What
Researchers have developed a new artificial intelligence-based antibody sequence generation technology that can simultaneously consider multiple conditions and diversity during the antibody design process.
Key result
Compared to existing methods, it generated 10 times more diverse and acceptable antibody candidates at a 50 percent sequence identity.
Why it matters
This research can help reduce the probability of failure in the early stages of new drug development and explore various potential antibody candidates.AI summary
Paper2026-10-12 · Micromachines

Analysis of Construction and Medical Application Technologies for Bio-Integrated Micro-Robots

Construction and Functional Engineering of Bio-Integrated Micro/Nanorobots for Theranostic Applications
What
Reviewed the fabrication and theranostic applications of bio-integrated micro- and nanorobots that combine biological characteristics with technology.
Key result
Presented engineering strategies and manufacturing guidelines that implement mobility, in vivo imaging, and therapeutic functions using cell-based components.
Why it matters
Demonstrates the potential direction for utilizing bio-integrated micro-robots in the fields of precision medical intervention and personalized medicine.AI summary
Paper2026-10-11 · arXiv (Cornell University)

Reducing Prediction Errors in AI World Models by Applying Physical Laws

LeHami: Symplectic Latent Dynamics for World Models
Yann LeCun +2 more
What
Researchers developed LeHami, a new world model training method that combines the geometric properties of physical systems with artificial intelligence.
Key result
Unlike conventional approaches, this method prevents information from compressing and disappearing during long-term prediction and significantly improves velocity prediction performance.
Why it matters
This research can contribute to AI understanding and simulating physical spaces more accurately.AI summary
Paper2026-10-09 · DigitalCommons-IMSA (Illinois Mathematics and Science Academy)

Introduction to key achievements and career of AI researcher Jared Kaplan

Recipient of the 2026 Alumni Trailblazer Award
What
Introduced the research and career achievements in the field of artificial intelligence of Jared Kaplan, co-founder and Chief Scientist of Anthropic.
Key result
Researched laws predicting the performance of large-scale artificial intelligence systems and pioneered techniques ensuring artificial intelligence aligns with predetermined principles.
Why it matters
Contributed to providing theoretical frameworks and guidelines for the development of safe and beneficial artificial intelligence.AI summary
Paper2026-10-08 · Code Ocean

Technology to Interpret the Causes of Uncertainty in Artificial Intelligence Revealed

pce-uncertainty-attribution: code for "Interpretable attribution of decision-relevant uncertainty in AI"
What
Released code and methods to interpret which factors contribute to the decision-making uncertainty of trained artificial intelligence models.
Key result
Utilized polynomial surrogates to uncover the causes of model uncertainty in the fields of chemistry, experimental design, and language model alignment.
Why it matters
Can be utilized to better understand the decision-making process of artificial intelligence models across various fields and increase reliability.AI summary
Paper2026-10-08 · Science

Automatically proving unsolved math problems with AI

Advancing mathematics research with AI-driven formal proof search
What
We developed an AI agent that searches for mathematical proofs at scale by utilizing a language where a compiler verifies the proof steps.
Key result
This AI agent successfully and autonomously solved several unsolved math problems and conjectures.
Why it matters
It demonstrated the potential for AI to drive autonomous mathematical discoveries.AI summary
Paper2026-10-07 · arXiv (Cornell University)

Development of Long Visual Memory Processing Technology for Real-Time Robot Control

Long-WAM: Scaling the Context of World-Action Models
Song Han +15 more
What
Researchers have developed Long-WAM, a model that efficiently utilizes past visual information of robots under real-time constraints.
Key result
When pre-training the foundational model that processes visual information using an autoregressive (AR) method, the utilization of past memory is maximized, and robot task success rates are improved.
Why it matters
It demonstrated that real-time operation is possible in various hardware environments, which can contribute to future complex long-horizon robot control tasks.AI summary
Paper2026-10-07 · arXiv (Cornell University)

Development of a World Model Integrating Various Robot Behaviors into One

Cross-Embodiment Robot Foundation World Models with Latent Actions
Fei-Fei Li +11 more
What
Developed a robot world model utilizing an integrated latent action space that can be shared by various types of robots.
Key result
When applied to new robots, it showed up to a 46.7% improvement in precise manipulation tasks and an 11.7% improvement in LIBERO benchmark performance compared to conventional methods.
Why it matters
Enables efficient learning across diverse robots, contributing to solving major challenges in robotics.AI summary
Paper2026-10-07 · arXiv (Cornell University)

AI technology that transforms virtual robot videos into realistic ones

RoboRender: Robot-Oriented Video Generation for Visual Sim-to-Real Transfer
Fei-Fei Li +8 more
What
Developed a technology that enhances robot control capabilities by converting robot data from virtual environments into videos that look like the real world.
Key result
Robots trained with this technology achieved an average success rate of 71 percent in real-world environments, showing superior performance compared to existing methods.
Why it matters
This can help alleviate the visual problem where data created in virtual environments was difficult to directly apply to real robots.AI summary
Paper2026-10-07

Method developed to predict coding agent performance of base AI models in advance

Before They Can Solve: Predicting Post-Training Coding-Agent Performance from Base Models
What
Developed three evaluation screens utilizing the success paths of post-trained coding agents to evaluate the potential of base AI models.
Key result
All three developed evaluation methods closely predicted the rankings of the open model group to actual performance.
Why it matters
Presents the possibility of utilizing future coding agent benchmarks for base model evaluation.AI summary
Paper2026-10-06 · arXiv (Cornell University)

Development of an algorithm that accelerates AI robot behavior learning

QF3: Fast Flow RL with Filtered Q-Gradients
What
Researchers have developed QF3, a new reinforcement learning algorithm that efficiently learns robot behavior patterns.
Key result
It has been demonstrated that it can learn humanoid robot locomotion and motion tracking policies about 10 times faster than existing methods.
Why it matters
It shows the potential to learn and improve robot behavior control technology faster and more efficiently.AI summary
Paper2026-10-06 · arXiv (Cornell University)

Whole-Body Humanoid Loco-Manipulation Technology Utilizing Human Demonstrations

Workhorse: Learning Robust Whole-Body Humanoid Loco-Manipulation from Human Data
What
Developed a technology that enables humanoid robots to manipulate objects and move by utilizing human demonstration videos through a visual planner and a reinforcement learning tracker.
Key result
Successfully performed complex whole-body manipulation tasks such as box carrying, box grasping, and stepping over a suitcase on a real Unitree G1 robot.
Why it matters
Demonstrates that robots can improve their ability to handle objects while maintaining balance in complex environments based on human behavior data.AI summary
Paper2026-10-06 · arXiv (Cornell University)

Humanize technique to improve the reliability of AI coding

Humanize: Judgement Engineering for Agentic Coding
Song Han +8 more
What
Introduced Humanize, a multi-AI collaborative workflow that applies clear, mechanical decisions between the planning, implementation, and review stages.
Key result
Production AIs and review AIs from different providers worked in alternating turns, achieving high performance in competitive problem-solving and performance evaluation.
Why it matters
Demonstrates that independent reviews help catch errors in AI-assisted programming processes.AI summary
Paper2026-10-06 · arXiv (Cornell University)

Development of a new artificial neural network architecture separating short-term and long-term memory

Continuous Memory Machines
Llion Jones +5 more
What
Researchers have developed a new artificial neural network architecture called Continuous Memory Machines (CMM) that separates matrix-shaped short-term and long-term memory states.
Key result
It showed superior performance and generalization capabilities compared to existing models in algorithmic and reasoning tasks.
Why it matters
It can contribute to improving the memory and reasoning capabilities of artificial intelligence across various tasks.AI summary
Paper2026-10-06 · arXiv (Cornell University)

Clarifying the Relationship between Environmental Control and Structural Centrality in AI Agents

The Geometry of Empowerment
What
Researchers revealed the relationship between empowerment, which refers to an AI agent's ability to control its surrounding environment, and structural centralization, which grants broad access to future outcomes.
Key result
They proposed a new geometric structure capable of interpreting and analyzing this by connecting empowerment maximization with skill-learning methods.
Why it matters
Provides theoretical implications necessary for building scalable empowerment maximization methods.AI summary
Paper2026-10-06 · arXiv (Cornell University)

Open Framework Combining Robot Control and Future Prediction Released

OpenWAM: An Open Framework for Composable World-Action Models
Fei-Fei Li +9 more
What
Developed OpenWAM, an open-source framework for comparing and researching world-action models that handle robot control and future video prediction together.
Key result
The proposed method significantly increased the success rate of robotic manipulation tasks and improved the performance of dynamics models through counterfactual data training.
Why it matters
It provides a common foundation for comparing various robot control and video prediction methods, thereby increasing the efficiency of related technology research.AI summary
Paper2026-10-06 · arXiv (Cornell University)

A new method for evaluating the self-improvement capabilities of AI agents through experience

Agent Plasticity: Measuring Self-Improvement Through Experience
Rob Fergus +10 more
What
Introduced the concept of agent plasticity and studied evaluation methods to measure the self-improvement efficiency of AI agents that evolve on their own through experience.
Key result
Despite similar learning opportunities, state-of-the-art models showed significant differences in their improvement trajectories, and the agent with the best final performance did not necessarily improve the most efficiently.
Why it matters
Demonstrates that when evaluating the potential of artificial intelligence, it is important to measure not only fixed-point capabilities but also how effectively they improve through experience.AI summary
Paper2026-10-06

OTel releases AI datasets and models dedicated to the telecommunications sector

OTel: Open Telco AI Datasets, Benchmarks, and Models
What
Researchers have released 'Open Telco (OTel)', a collection of resources including datasets, evaluation benchmarks, and models trained with them, necessary for AI training dedicated to the telecommunications sector.
Key result
The models based on the released OTel achieved performances of 93.1%, 0.947, and 87.8% in telecommunications data retrieval, document reranking, and language model accuracy evaluation, respectively.
Why it matters
This release provides a reproducible starting point that promotes artificial intelligence research and applications in the telecommunications field.AI summary
Paper2026-10-05 · arXiv (Cornell University)

Development of a new benchmark to study the phenomenon of AI behavioral cloning technology

Behavioral Cloning Mystery
What
Proposed OCBench, a robot manipulation evaluation tool with characteristics similar to human demonstrations.
Key result
Showed that this tool enables scientific research on various phenomena related to behavioral cloning and hypothesis testing in graphics processing unit (GPU) based environments.
Why it matters
Provides a foundation for analyzing previously unclear phenomena in robot learning technology within a controlled environment.AI summary
Paper2026-10-05 · arXiv (Cornell University)

A Performance-Guided Data Collection Method to Enhance Robot Learning Efficiency

Mulligan: Performance-Guided Data Collection for Efficient On-Robot Learning
Chelsea Finn +6 more
What
Researchers proposed Mulligan, a technique that improves efficiency in robot manipulation learning by collecting data centered around failed and untried states.
Key result
The Mulligan technique improved the robot's final task success rate from 10 percent to 34 percent compared to conventional methods.
Why it matters
It demonstrates a technological advancement that helps human trainers use time and effort more efficiently when teaching robots.AI summary
Paper2026-10-05 · arXiv (Cornell University)

Development of H-JEPA, an AI model that learns hierarchical structures for complex visual planning

H-JEPA: End-to-End Learning of Hierarchical World Models for Visual Planning
Yann LeCun +4 more
What
Developed H-JEPA, a hierarchical world model that learns by dividing abstraction levels according to time horizons.
Key result
In simulation environments and real robot video experiments, planning success rates and computational efficiency improved compared to existing single-level models.
Why it matters
It can contribute to increasing efficiency in robot control and visual tasks that require complex long-term planning.AI summary
Paper2026-10-04 · arXiv (Cornell University)

New method proposed to train AI models with a balance of multiple values

Which Preferences to Train On? End-to-End Multi-Objective Alignment with an Adversarial Preference Distribution
What
Researchers proposed a training method called MAESTRO that enables AI language models to adjust the balance among multiple values, such as helpfulness, harmlessness, and humor, according to user requests.
Key result
According to the abstract, MAESTRO achieved the best trade-off curves (Pareto fronts) in most tasks within a single training run in experiments handling up to three objectives, and had the lowest training cost among the compared methods.
Why it matters
While this shows the potential for companies to build AI services tailored to individual user settings more efficiently if a single model can handle multiple value combinations, the actual market impact is difficult to judge from the abstract alone.AI summary
Paper2026-10-04 · arXiv (Cornell University)

Measuring how easily AI comes up with future research ideas

Priced Guidance: Can Language Models Generate Future Research Ideas?
Percy Liang +2 more
What
The study proposes a method to indirectly estimate the likelihood that a language model generates novel research ideas without hints, using the amount of information (compression cost) required to reveal the correct answer.
Key result
Five models were evaluated on the core ideas of 87 recent deep learning papers, and the median compression cost of the best single model was 69.9 bits. This is significantly lower than the median of 5,712 bits for the general compression tool gzip.
Why it matters
This research demonstrates a new evaluation framework that can calculate the lower bound of the probability that a model generates ideas. However, the abstract does not address research productivity or market effects.AI summary
Paper2026-10-04 · arXiv (Cornell University)

AI Model 'SheetSage2' Released for Converting Music to Sheet Music

SheetSage2: Coherent Lead-Sheet Transcription with Synthetic Supervision
Yann LeCun +6 more
What
Presented SheetSage2, an integrated music transcription model that automatically converts audio recordings into human-readable sheet music (lead-sheets).
Key result
The single model SheetSage2-AR outperformed existing listed systems in 12 out of 15 benchmark and metric pairs evaluated, showed significant improvements over its predecessor SheetSage1, and even surpassed task-specific dedicated models in some benchmarks.
Why it matters
This research improves the accuracy of music transcription technology, and the model weights and inference code have been released to be utilized for follow-up research or applications. However, it is difficult to judge commercialization potential or market size based on the abstract alone.AI summary
Paper2026-10-04 · arXiv (Cornell University)

Proposal of a method where multiple AI agents stack work in a Git repository to solve problems

GitSwarm: Decentralized Compounding Inference
Rob Fergus +13 more
What
Proposed and evaluated GitSwarm, a 'compounding inference' system where multiple identical AI agents work independently in a shared Git repository, continuously saving, referencing, and combining intermediate results.
Key result
Reported solving all 30 math proof problems in a single run, outperforming the strongest baseline under a fixed budget (65.1 points) with an average of 79.4 points on the programming benchmark ProgramBench, and improving starting architectures through iterative experiments across three neural network structure tasks.
Why it matters
Demonstrates that AI can accumulate intermediate achievements rather than discarding results after each attempt, reusing them for subsequent tasks, thereby suggesting a direction for long-horizon problem solving and AI design for scientific research. However, since these are results at the abstract level, actual industrial impact is difficult to judge from this material alone.AI summary
Paper2026-10-03 · arXiv (Cornell University)

Researchers Propose a Framework for Evaluating AI Answer Oversight via Code Execution

PyINE: A Framework for Scalable Elicitation and Oversight via Code Execution
What
Created the PyINE framework, which utilizes program execution results as ground truth, to examine whether supervisors can trust AI answers that feature plausible yet incomplete reasoning.
Key result
Models trained to follow shortcuts significantly improved their ability to predict execution outcomes, but still made systematic errors when misleading cues conflicted with actual behavior, and cheap learned supervisors frequently missed rare shortcut errors.
Why it matters
Significant for presenting an experimental environment that compares and reuses AI answer verification oversight methods under the same conditions, although the abstract does not mention actual service application effects.AI summary
Paper2026-10-02 · arXiv (Cornell University)

Proposing GOTT, a reusable object grasping technology for multiple robot hands

GOTT: Object-centric Dexterous Manipulation with a Reusable Cross-Embodiment Primitive
What
To enable multi-fingered robot hands, similar to human hands, to stably grasp and move objects in desired directions, the authors proposed GOTT, a contact technology that can be commonly used across different robots and hand shapes.
Key result
In simulations and real-world robot experiments, GOTT established stable contacts across various objects, arm-hand platforms, and unseen hand shapes, and consistently improved overall task success rates compared to pre-planned grasping methods.
Why it matters
This study demonstrates the potential to reuse the same technology even when robot hand types change by using a common motion component not dependent on a specific robot. However, it is difficult to judge the commercialization level or real industrial impact from the abstract alone.AI summary
Paper2026-10-02 · arXiv (Cornell University)

Modular memory design for quantum computer error correction significantly reduces resource consumption

Low-Overhead Quantum Error Correction with Boundary-Connected Planar Modules
What
Researchers divided the quantum processor into planar modules and demonstrated through simulations that utilizing inter-module connections can reduce the number of physical qubits required for quantum error correction by more than 10 times.
Key result
The new modular code design was shown to reduce physical qubit consumption by more than 10 times (and over 30 times in large-scale systems) compared to surface codes.
Why it matters
This study demonstrates the potential for high-density storage devices that maintain the manufacturing advantages of planar modules while leveraging the benefits of highly efficient codes.AI summary
Paper2026-10-02 · bioRxiv (Cold Spring Harbor Laboratory)

New enzyme chemistry technology combining artificial intelligence protein design and evolution

Design-Evolution Synergy Enables New-to-Nature Enzyme Chemistry
What
Developed an enzyme that performs carbon-hydrogen amination reactions not found in nature by combining artificial intelligence-based protein design and evolution techniques.
Key result
Succeeded in synthesizing a new enzyme with programmable positional and enantioselectivity, high catalytic efficiency, and superior enzyme stability.
Why it matters
Demonstrates that protein design technology can be utilized to create enzymes that promote demanding chemical reactions.AI summary
Paper2026-10-01 · arXiv (Cornell University)

VISTA: Research on improving multimodal AI agent performance with visual memory

VISTA: A Visual Harness for Reasoning in an Interactive World
Kaiming He +4 more
What
Researchers proposed 'VISTA', a visual harness designed to help multimodal AI models that handle both images and text solve tasks better in interactive visual environments.
Key result
They reported that applying VISTA on ARC-AGI-3 increased the relative human action efficiency score from 40.68 to 100.00, completing all 25 public games with 57.4% fewer actions than the initial participating human.
Why it matters
This demonstrates that adding visual memory and retrieval functions to existing models was effective in specific visual game tasks. However, since the abstract presents only a few benchmark results, it is premature to conclude that it can be applied broadly across actual industrial settings.AI summary
Paper2026-10-01 · arXiv (Cornell University)

An RPG framework where robots improve their own skills through simulation practice

Reconstruct, Practice, Go Real: Guided Self-Improvement for Embodied Agents
What
Proposed an RPG framework that autonomously improves robotic task systems without additional human intervention.
Key result
In new initial condition tests across 22 manipulation tasks, the success rate increased from 28.6% after the first practice to 95.0% after 15 rounds, and all 30 real-world robot trials succeeded.
Why it matters
This study demonstrates the potential to improve robot control software with less human effort, although actual industrial application or commercialization levels are difficult to judge from the abstract alone.AI summary
Paper2026-10-01 · arXiv (Cornell University)

A 'Finding Inspiring Papers' Benchmark Created by 184 Researchers

ScholarCatalyst: A Benchmark for Retrieving Papers That Inspire New Research
Chelsea Finn +13 more
What
We created ScholarCatalyst, a dataset where first authors of 207 recent computer science papers indicated which prior papers actually helped their research, and presented a retrieval task to find them.
Key result
Agentic search (AI that iteratively performs tasks on its own) did not show better performance than basic search methods (recall of 0.42 versus 0.48 based on the top 20 results). Agents based on Claude Fable 5.1 only reached 0.51.
Why it matters
This indicates that current AI still lacks the expert intuition needed to identify necessary prior ideas from vast research literature, pointing to the need for new training methods. While this can be viewed as a step forward for AI assisting scientific research, actual commercialization or industrial effects cannot be confirmed from the abstract.AI summary
Paper2026-09-30 · arXiv (Cornell University)

Researchers Propose TacEx to Enhance Robot Learning Efficiency Through Tactile Curiosity

Tactile Curiosity Drives Robot Interaction
What
The researchers proposed TacEx, an exploration framework that allows robots to develop curiosity about tactile sensations (the sense of touching objects) and learn how to touch and pick up objects on their own.
Key result
According to the abstract, TacEx collects interaction data rich in contact without task rewards or expert demonstrations, enables learning of pick-and-place policies using this data without further environment interactions, and improves the post-training performance of vision-language-action models.
Why it matters
Traditional robot learning spends a lot of time moving randomly in the air, demonstrating that exploring with a focus on contact can improve learning efficiency. However, it is difficult to judge the level of practical application or commercialization potential based on the abstract alone.AI summary
Paper2026-09-30 · Nature

Development of Technology to Invisibly Mark the Origin of AI-Generated Proteins

Function-preserving watermarking of AI-generated proteins.
What
Researchers have developed 'SynthIDBio', a technology that leaves marks on AI-generated protein sequences and structures without loss of function.
Key result
It has been demonstrated that watermarks can be detected with near-perfect accuracy while maintaining protein functions similar to existing ones.
Why it matters
It demonstrates the potential of provenance tracking tools in the field of AI-driven bioengineering to address biosecurity and information authenticity issues.AI summary
Paper2026-09-29 · arXiv (Cornell University)

Scaling humanoid object carrying learning by expanding a few real videos with AI

Counterfactual Video Generation Enables Scalable Humanoid Loco-Manipulation
What
Proposed a method (PRISM) where AI generates various virtual videos based on a few real videos of a person carrying objects, and uses them for robot learning.
Key result
Reports that a single control policy trained this way successfully enabled a real humanoid robot to pick up, transport, and place boxes, drums, kegs, and balls without additional real-world learning.
Why it matters
Significant in that it demonstrates the potential to increase training data using a few real videos in situations where collecting large amounts of real robot demonstration data is difficult.AI summary
Paper2026-09-29 · arXiv (Cornell University)

SoL-Refiner Upscales Low-Resolution Videos to 4K Quality in a Single Step

SoL-Refiner: Speed-of-Light One-Step Refinement for High-Resolution Video
Song Han +11 more
What
The researchers proposed SoL-Refiner, a refiner model that transforms low-resolution videos into high-resolution ones through a single denoising step.
Key result
At 2K resolution, it achieved higher average scores on VBench and UniPercept than comparable external refiners. At 3840x2176 resolution, it improved upon both metrics compared to the three-step LTX-2.3 refiner. It was reported that under 2K latency settings, refining latency became 8.91 times faster compared to the same baseline model.
Why it matters
High-resolution video generation significantly increases computational load and cost burdens. This approach shows that increasing quality in a single step instead of multiple steps can be a direction to reduce this burden. However, since the results presented in the abstract were measured under conditions set by the authors, they cannot be generalized as overall effects in actual services.AI summary
Paper2026-09-29 · arXiv (Cornell University)

LeapQuant: Accelerating AI Inference by Compressing Linear Attention States to 8 Bits

LeapQuant: Efficient Linear Attention with Accurate Recurrent State Quantization
Song Han +12 more
What
The paper proposes LeapQuant, a training-free method that compresses the recurrent states, which large language models repeatedly read and write when processing long contexts, into low precision.
Key result
According to the abstract, 8-bit compression maintains accuracy comparable to the FP32 baseline while achieving an average speedup of 2.05 to 3.70 times at the GPU kernel level and 1.47 times in overall inference.
Why it matters
This research can contribute to improving operational efficiency for AI services handling long contexts by reducing memory and computational burdens during inference. However, since the abstract only presents results from specific models and GPUs, it is difficult to conclude that the same effects will appear across all actual commercial services.AI summary
Paper2026-09-29 · arXiv (Cornell University)

T2Mem: A Technique for Training Robots to Remember Past Scenes and Perform Tasks

T$^2$Mem: Learning Test-Time Memory for Robotics
Fei-Fei Li +6 more
What
For robot manipulation tasks that require remembering past observation information, we propose a framework (T2Mem) that adds memory capabilities through test-time training to pretrained robot policies that previously lacked memory functions.
Key result
The average success rate across 16 RoboMME tasks increased from 17.93% for the baseline policy without memory to 56.83%, outperforming recurrent memory methods reported in the benchmark.
Why it matters
This is significant in showing that robots' ability to utilize past information not currently visible on screen can be fostered using only demonstration data without external memory devices, though the abstract only covers research-stage results, so commercialization or market effects cannot be confirmed.AI summary
Paper2026-09-29 · arXiv (Cornell University)

Proposed meta-reasoning technique for AI agents to determine 'how to proceed' before performing tasks

Thinking Before Thinking: Scaling Agentic Inference Through Meta-Reasoning
Rob Fergus +11 more
What
Proposed an 'agentic meta-reasoning' approach that systematically treats execution control decisions—such as how much to reuse previous task results, whether to start fresh, and when to stop—for AI agents performing long-running tasks.
Key result
On the ProgramBench program reconstruction benchmark, the GPT-5.5-based version recorded 71.5%, which is higher than Codex's 58.0%, and the Opus 4.8-based version recorded 67.2%, higher than Claude Code's 65.5%.
Why it matters
According to the abstract, this suggests that as agents perform longer tasks, allocating computational resources to structured execution control may become more important. However, since the results in the abstract were obtained under specific benchmarks and experimental conditions, it is difficult to conclude that they directly translate to real-world product performance.AI summary
Paper2026-09-28 · The International Journal of Robotics Research

A comprehensive review summarizing 'world model' research for robot learning

World model for robot learning: A comprehensive survey
Pieter Abbeel +17 more
What
This is a survey paper that systematically organizes 'world models'—models that predict how the environment will change in response to actions—from the perspective of robot learning.
Key result
According to the abstract, it outlines how world models are used in robot policies, simulators, evaluation, and data generation, summarizes the evolutionary trajectory of robot video-based world models, and highlights representative datasets, benchmarks, and evaluation methods.
Why it matters
This serves as a resource that unifies scattered robot learning research approaches into a single framework, helping to grasp the research trends. However, the abstract does not prove the performance or commercial achievements of individual technologies.AI summary
Paper2026-09-28 · arXiv (Cornell University)

Proposal of a Technique to Reduce Speed and Memory Usage in Video Generation AI

Sol-H3: Recursive Self-Improvement for MiniMax-H3 Inference Acceleration on Sol-Engine across Cloud and Edge
Song Han +8 more
What
Created a complete execution pipeline to run the 33-billion parameter video generation AI model MiniMax-H3 faster and with less memory on both cloud servers and small devices.
Key result
According to the abstract, up to a 30x end-to-end speedup and about 20% memory reduction were achieved, generating a 5-second, 1344x768 resolution video with audio 3.5x faster than real-time on 8 GB200 server nodes.
Why it matters
This is significant as it demonstrates a way to lower the generation latency and memory limits that are the biggest obstacles when applying video generation AI to real services, though the commercialization effect and scale of cost reduction are difficult to judge from the abstract alone.AI summary
Paper2026-09-28 · arXiv (Cornell University)

Research improving block diffusion language model training by incorporating future context

d-OPD: Future-Aware On-Policy Distillation for Block Diffusion Language Models
Song Han +2 more
What
The authors propose d-OPD, a distillation training method that converts conventional autoregressive language models (which generate one token at a time) into block diffusion language models that generate multiple tokens in parallel.
Key result
Experiments with Qwen3 models (0.6B to 8B) showed that d-OPD improved the average score across six benchmarks by up to 4.0 points compared to the existing OPDLM method, while reducing training time by 1.35 to 1.58 times.
Why it matters
The abstract suggests that generating multiple tokens simultaneously in blocks has the potential to increase text generation speed, and this study demonstrates that the efficiency of the training process for creating such models can be improved.AI summary
Paper2026-09-28 · arXiv (Cornell University)

DexAgent automatically generates robot hand motion training data from human demonstration videos

DexAgent: An Agentic Human2Sim2Robot Framework for Dexterous Manipulation with Self-Evolving Tool Library
Fei-Fei Li +4 more
What
We propose an agentic framework that takes a single first-person human video and a task description to generate physically feasible robot motion trajectories for robot learning data.
Key result
Policies trained on data generated by DexAgent across 11 real-world tasks showed a 3.5 times higher success rate compared to baseline methods.
Why it matters
It demonstrates the potential to automate the process of converting human demonstration videos into robot learning data, though it is difficult to judge the industrial application level or commercialization timeline from the abstract alone.AI summary
Paper2026-09-28 · arXiv (Cornell University)

JEPA research makes autonomous driving world models 100 times faster

AD-E2E-JEPA: A Joint-Embedding Predictive Architecture For End-to-End Autonomous Driving
Yann LeCun +3 more
What
Evaluated predictive world models (JEPA family) used in autonomous driving and proposed an improved architecture, AD-E2E-JEPA, that significantly reduces computational load.
Key result
The proposed architecture reduced the number of patches used for planning by 1/16 and the embedding dimension by 1/4, increasing inference speed by about 100 times while maintaining planning performance.
Why it matters
This research shows a direction for reducing the computational cost of autonomous driving AI, though the actual in-vehicle application effect and commercialization potential are difficult to judge from the abstract alone.AI summary
Paper2026-09-28 · arXiv (Cornell University)

A Framework for Evaluating AI Consciousness Across Five Levels Proposed

From cacophony to hierarchy: a principled framework for assessing AI consciousness
Shane Legg +13 more
What
The authors organized various theories dealing with AI consciousness into a five-tier hierarchical framework and proposed a method to evaluate current Large Language Models (LLMs) using indicators for each level.
Key result
Current LLM consciousness assessments varied widely from under 0.01 to about 0.8, depending not only on how evidence is interpreted but also on how much weight is placed on which theory.
Why it matters
Given that consciousness-related indicators heavily overlap with structures necessary for general intelligence, this implies that as AI becomes more capable, it may become a stronger candidate for consciousness. However, this is not a definitive conclusion, and the abstract makes no direct mention of investment.AI summary
Paper2026-09-28

Analysis of the Impact and Policy Responses Triggered by Automating AI R&D

What if automating AI R&D triggers an intelligence explosion?
What
Researchers evaluated the possibility of AI R&D automation triggering an intelligence explosion, along with its potential impacts and policy response measures.
Key result
AI systems are expected to automate most AI R&D tasks within a few years, which could lead to rapid capability growth at a level society struggle to handle.
Why it matters
It warns of the extreme risks such as loss of control alongside the benefits that AI R&D automation may bring, urging policy preparedness.AI summary
Paper2026-09-12

Cohere Unveils Cost-Effective Large Language-Based Translation Model

North Small Translate: Advanced Cost-Effective Translation (Cohere CAT+)
What
Researchers have introduced 'North Small Translate', a cost-effective large language model-based machine translation model that delivers outstanding performance across 50 languages.
Key result
This model achieved state-of-the-art machine translation performance across 50 languages among model groups with fewer than 1 trillion parameters.
Why it matters
It demonstrates the potential to efficiently perform diverse language translation and post-editing tasks without expensive computations during the inference process.AI summary
Paper2026-09-10

Analysis of Overfitting in Mixture-of-Experts Models When Repeating AI Training Data

Data Scarcity and Model Sparsity: Mixtures-of-Experts Overfit More to Repeated Data
Percy Liang +4 more
What
It was revealed that when AI training data is repeated, the recently widely used Mixture-of-Experts architecture degrades in performance faster than conventional models.
Key result
When data is repeated, Mixture-of-Experts models easily overfit, and while existing regularization techniques can partially mitigate this, they fall short of the performance achieved by training on unique data.
Why it matters
This demonstrates the necessity of establishing training strategies that consider the characteristics of each AI model architecture in situations where iterative learning is required due to data scarcity.AI summary
Paper2026-08-24

Learning to Act While the Robot Thinks

Learning to Act While Waiting: RL Finetuning of Generalist Robot Policies Under Inference Latency
Sergey Levine +19 more
What
Researchers have developed ARLI, a latency-aware reinforcement learning framework that solves problems caused by the thinking time (inference latency) of robot models.
Key result
This new method enables robots to successfully continue learning even in environments with latency, showing performance equal to or better than an ideal latency-free environment.
Why it matters
It opens up the possibility for complex and massive robot AI to autonomously improve performance in real-world fields without freezing or stuttering.AI summary
Paper2026-08-17

Release of Open Source Tools and Performance Comparison Platform for Relational Learning

Advancing Open and Reproducible Relational Learning: RelArena-$\alpha$, TabPFN-Rel and RPI
Yann LeCun +46 more
What
Researchers have open-sourced an integrated comparison platform and two related softwares to promote research in the relational learning field.
Key result
Among the newly released tools, TabPFN-Rel, a model for relational databases, recorded the highest rank on the comparison platform.
Why it matters
This release can serve as a foundation to establish reliable comparison standards in relational learning and assist its application in research and industrial settings.AI summary

People

Why follow AI leaders' moves and papers?

In AI, people are the technology. Which company a key researcher joins, or which lab they found, often turns into products, funding rounds and acquisitions months later. This page tracks job changes across frontier labs, new labs, academia, AI chips, and people in Korea and China.

Papers show what a person is focused on right now. We sort papers into 10 investable fields such as HBM & memory, GPUs & accelerators, inference efficiency, agents and robotics, and give a three-line summary — what, key result, why it matters — based on the abstract, with links to the original and PDF. This is information for finding ideas, not a recommendation to buy or sell any security.

Frequently Asked Questions

Q. Who do you track?

Key researchers and executives at frontier labs (OpenAI, Google DeepMind, Anthropic, Meta and others), new labs, academia, AI chip companies, and in Korea and China. Executives who rarely publish are tracked for career moves only.

Q. Where does the data come from, and how often is it updated?

Papers come from Semantic Scholar and arXiv, and careers from Wikidata (CC0) plus official announcements and press reports, collected automatically every day. Only items checked and approved by the editor are published.

Q. Who writes the paper summaries?

An AI (Google Gemini) drafts each summary from the paper's abstract, and the editor checks it before publishing. Other languages are AI translations of the checked Korean summary. Always confirm exact figures and context in the original paper.

This page compiles published papers, official announcements and press reports for information only. Paper summaries are written by AI from the abstract and reviewed by the editor; they may differ from the original, so check the paper for anything important. Nothing here is a recommendation to buy or sell any security or investment advice. Career moves cite official announcements or major press reports; some career records come from Wikidata (CC0). Summaries on this page were translated by AI from the editor-reviewed Korean version.

🐩 Support Kobong's Treats
코봉이
If this page was helpful, feel free to support Kobong's treat fund too 🦴
💬 Community Comments
Anonymous nicknames based on IP · Abuse may result in restrictions
Loading...