🔬 Research Digest

BIS Maps Out Four Possible Futures for Banking

The Bank for International Settlements has published a scenario framework asking a deceptively simple question: a decade from now, who will actually carry out b…

Rebuilding America's Pre-1947 GDP, With Error Bars

Official quarterly US GDP data only starts in 1947, so anyone studying earlier business cycles — the Roaring Twenties, the Great Depression, the wartime boom — …

Bank Supervisors Walk a Tightrope Between Safety and Credit

When a bank fails, it is rarely a sudden collapse out of nowhere. New research from the New York Fed's Liberty Street Economics argues that failures usually fol…

Home Visits Cut Chronic Absenteeism Sharply, Study Finds

A new study from the National Bureau of Economic Research examines Connecticut's Learner Engagement and Attendance Program (LEAP), launched in summer 2021. Unde…

LinkedIn Data Offers a New Gauge of Labor Market Tightness

Key idea: Economists judge how "tight" a labor market is by comparing the number of available workers with the number of open jobs. When openings far outnumber …

Managers Misjudge the Incentive Plans That Work Best

New research by Francis Annan and Collin B. Raymond asks a deceptively simple question: when companies pick pay-for-performance schemes, do managers actually kn…

Why Irrational Traders Never Fully Disappear From Markets

Classical finance assumes a simple sorting process: rational traders profit, irrational ones lose, and the irrational eventually exit. Reality disagrees — noise…

Hosting the Battery Boom Isn't the Same as Capturing Its Value: Hungary's EV Bet

Hungary has poured public support into electric-vehicle battery manufacturing, hoping to move up the value chain. A new paper by Zoltan Bartha builds an "exposu…

FLARE Scores May Measure the Grader, Not the Model

A new audit of FLARE — the financial benchmark suite used by PIXIU and FinBen to rank language models — suggests many published scores reflect how answers are e…

RobustRLlib: A Common Benchmark for Reinforcement Learning That Must Survive Real-World Shocks

Reinforcement learning (RL) — training software agents by trial and reward — has moved from games into robots, trading systems and autonomous machines. But a po…

Delayed Retirement May Cut Fertility by Reducing Grandparental Childcare

China is phasing in a gradual delayed-retirement (GDR) reform starting in 2025, raising statutory retirement ages on separate timelines for men, female cadres, …

A Cheap Offline Test for Robot AI: Does Its Vision Match How Things Move?

Evaluating robotic AI usually means running it in the real world or a simulator over and over — slow, expensive, and hard to scale. A new paper proposes a short…

Smarter Test-Time Reasoning: Pruning AI Thought Candidates to Balance Quality and Diversity

Large reasoning models often generate many candidate solutions at inference time, hoping at least one is correct. That is expensive, so researchers prune the po…

New AI Model Turns Video Into Replayable 3D Motion for Any Skeleton

Video-generation models can now produce strikingly lifelike movement, but that motion stays trapped in pixels. A new model from NVIDIA researchers, described in…

Liquidity Rules Push Lending Into Private Credit

New research from the Bank for International Settlements examines how tighter bank liquidity rules reshape where credit actually flows. The authors focus on the…

New Method Makes Complex Earnings Models Easier to Estimate

Economists often build models with hidden variables to explain how people's earnings evolve over time. These models are powerful but hard to fit, because the ma…

Offshore Currency Bets Move Emerging-Market Exchange Rates

New research from the Bank for International Settlements examines how trading in offshore non-deliverable forwards (NDFs) — derivative contracts that let invest…

Rising Revenue per Worker May Be Padding Profits, Not Just Paychecks

US labor productivity growth slumped for a decade after the Great Recession, with output per hour rising only about 1% a year from 2010 to 2019 — far below the …

Critical Access Hospital Designation: Survival at the Cost of Scale

A new study by Ian McCarthy and Katie Leinenbach examines a long-running U.S. program that pays small rural hospitals cost-based Medicare reimbursement in excha…

LLMs Can Check Research, But Can't Judge What Matters

Finance and accounting journals are split on whether referees may use large language models. A new study asks a sharper question: which parts of peer review can…

Linear Demand Models Get a Microeconomic Foundation

Economists have long used simple linear demand curves because they are easy to estimate and interpret. Critics countered that they lack a solid basis in how ind…

When Funding Dries Up, Mortgage Relief Slows: Lessons from Ginnie Mae Buyouts

A new study by Arka P. Bandyopadhyay examines how the liquidity of mortgage servicers shapes whether struggling borrowers get relief. Using 1.9 million Ginnie M…

Safety Filter Makes Robot Motion Generators Scene-Aware

Text-conditioned motion generators can plan whole-body movements for robots, but they lack any understanding of scene context: the same instruction can be harml…

Microfinance Competition and the Moneylender Puzzle

Microfinance has expanded for decades, yet moneylenders keep thriving alongside microfinance institutions (MFIs) — even though MFIs often charge lower rates. Wh…

QF3: A Faster Way to Teach Robots New Tricks

Training robots to move like humans usually takes enormous trial and error. A new method called QF3, from researchers at UC Berkeley and collaborators, speeds t…

4D-HOF: Fast, Stable Hand-Object Motion Reconstruction

Reconstructing how hands manipulate objects in 3D space and time — known as 4D hand-object reconstruction — is a core problem for AR/VR, robotics, and animation…

Smarter Encoders, Simpler Decoders: A New Path to 3D-Aware AI Vision

Teaching machines to understand 3D scenes from ordinary photos is a long-standing goal in computer vision. One popular approach, novel view synthesis, asks a mo…

Geoeconomic Risk Is Now Priced Into U.S. Stocks

A new analysis from the New York Fed argues that geopolitical conflict is no longer just a headline risk for oil or defense stocks — it has become a measurable …

Textbooks Teach a Wide Range of Emotions — Their Pictures Don't

A new NBER working paper by Adukia, Bonci, Dastres Gallardo, Harrison, Nicoll, and Szasz asks a simple but overlooked question: what emotions do children actual…

Follow the Cash: Why Repo Markets Are a Patchwork of Structures

The U.S. repo market isn't one market — it's several, each with its own participants, settlement habits, and clearing arrangements. Economists at the New York F…

AI Workflow Audits Thousands of Economics Papers

A new open-source tool from economists Matthew Schwartz, Isaiah Andrews, and Jesse M. Shapiro lets a large language model dig through the replication packages t…

Same Text, Different Numbers: Why LLM-Based Research Measures Depend on the Model You Pick

The key idea A growing number of researchers and analysts feed corporate text — earnings call transcripts, filings, press releases — into large language models…

You Don't Need the Whole Supply Chain Map to Make a Good Call

The problem: Companies and governments rarely see the full map of who supplies whom. They see fragments — industry totals, regional shipments — and must still d…

AI Trading in Europe's Power Markets Needs Guardrails, Not Just Predictions

European electricity trading is not one market but several stacked on top of each other: forward contracts, day-ahead auctions, intraday trading, and balancing …

The Widening Web of Global Imbalances

Global imbalances — the gap between what countries owe the world and what the world owes them — have widened sharply again. A new Bank for International Settlem…

Team Decisions Without a Central Brain: New Learning Framework for Decentralized AI

Multi-agent systems — drone swarms, robot fleets, distributed sensors — often need to cooperate while each agent sees only part of the picture and hears from te…

AI Patent Gains from Cross-Region Science Networks Depend on Who Does the R&D

The idea: Most research on knowledge spillovers looks at accidental diffusion between nearby places, or at networks of co-inventors on patents. This paper inste…

AI Agents Learn Graphic Design Skills From User Traffic Without Retraining

Graphic design is hard to automate because success is subjective — there is no clean scoring function an AI can optimize against. A new arXiv paper proposes a w…

New Benchmark Tests Whether AI Models Can Transfer Brain Knowledge Across Animals

Neuroscience has entered a big-data era: researchers can now record activity from hundreds of thousands of neurons across many brain regions and many animals at…

Coding Agents Learn to Avoid Obstacles, Not Just Finish Tasks

A new line of research lets a language model write the software that controls a robot arm, skipping the usual robot-specific training. These "coding agents" are…

AI Embeddings Struggle With Basic Physical Measurements

The problem with how AI understands numbers Embedding models are the workhorses behind search engines, recommendation systems, and chatbots. They turn words an…

When a Strait Sneezes: Hormuz Disruption and the Global Economy

A new BIS working paper by Enisse Kharroubi looks at what happens to the world economy when shipping through the Strait of Hormuz is disrupted. The strait is on…

AI Agents Need a Social Harness, Not Just a Personal One

As AI agents increasingly work together on behalf of different people and companies, a new arXiv paper argues that the current approach to coordinating them is …

OSHA Inspections Cut Injuries at Nearby Workplaces Too, Study Finds

A new NBER working paper by Matthew S. Johnson, David Levine, and Michael W. Toffel looks at whether workplace safety inspections do more than fix the site bein…

Air Pollution Is Quietly Shrinking What Students Learn

A new NBER study by Behrer, Goodman, Goyer, and Park asks a simple but understudied question: how much does dirty air actually cost students in learning? The au…

Beyond Us vs. Them: How People Treat 'Middle' Groups

Most studies on group behavior focus on a simple split: people favor their own group (in-group) and discriminate against outsiders (out-group). But real life is…

AI's Power Problem: Why Data Centers May Drive Up Electricity Prices

Artificial intelligence is booming, but it has a hidden bottleneck: electricity. Data centers that power AI models consume vast amounts of power, and the grid i…

Why Yield Curves Took Millennia to Appear

Interest-bearing loans are ancient — records from Mesopotamia show credit long before modern money. Yet a yield curve, the chart of interest rates across differ…

UniMate: One AI Model to Animate Any 3D Skeleton

Creating realistic animations for 3D characters and objects is a major bottleneck in gaming, film, and VR. Traditionally, each type of skeleton (human, dog, spi…

Why Creators Struggle to Let Go of Their Own Ideas

New research reveals a cognitive bias called the 'Creative Endowment Effect.' When people generate ideas, they tend to overvalue them, even when they are object…

Turning Plain English into Reusable Code: A New Way to Compile Software

Imagine telling your computer, "Sort this list by price," and it creates a small, fast program that does exactly that — without needing to call a giant AI model…

Jackson Hole 2026: Payments Innovation Takes Center Stage

The New York Fed's Liberty Street Economics blog offers a recap of this year's Kansas City Fed symposium at Jackson Hole, where the central theme was financial …

The Dollar's Reserve Role: A Few Big Players, Not a Global Retreat

Headlines often warn that the U.S. dollar is losing its grip as the world's reserve currency. The numbers seem to support that story: the dollar's share of glob…

How LLM Judges Really Score Text: A Peek Under the Hood

Large language models (LLMs) are increasingly used to judge the quality of AI-generated text, such as summaries. But how do they actually decide on a score? A n…

Smarter AI Testing: How Trajectory-Aware Benchmarks Cut Costs

Testing AI coding agents is expensive. Each task can involve many steps—exploring code, making edits, running tests—and running full benchmarks takes time and m…

Simulation-Ready Tendon Hand Cuts Cost, Boosts Robot Learning

Researchers have developed an affordable robotic hand that mimics human dexterity, using a tendon-driven design where motors are placed away from the joints. Th…

How More Stock Investors Are Changing the Fed's Toolkit

Over the past four decades, the U.S. stock market has transformed from a playground for the wealthy into a mass-market arena. In the mid-1980s, fewer than 30% o…

AI Decodes the Sex of Ancient Handprints, With Confidence

Archaeologists have long struggled to determine whether the hand stencils painted on cave walls during the Upper Paleolithic were made by men or women. Traditio…

How Fed Rate Surprises Shape Stock Market Expectations

A new analysis from the Federal Reserve Bank of New York digs into how surprise interest-rate moves by the Fed ripple through the stock market. While it's long …

Better Experiments for Measuring Spillover Effects

When researchers run experiments, they usually focus on the direct effect of a treatment on the people who receive it. But in many real-world situations, one pe…

Why Credit Card Delinquency Rates Are Rising—But Not for the Reason You Think

The Federal Reserve Bank of New York recently reported that while overall household debt dipped slightly in Q2 2026, the share of credit card balances that are …

Congress Has Always Been an Elites' Club — and the Gap Is Widening

A new study from the National Bureau of Economic Research digs into the family backgrounds of every member of Congress born between 1830 and 1950, using census …

FinRank: A New Test for Financial AI That Checks the Evidence, Not Just the Answer

Most financial AI systems are judged on whether they give the right number. But in SEC filings, a correct-looking answer can be based on the wrong data—say, fro…

Scaling Up Thinking Time: What 'Test-Time Scaling' Means for AI Models

A new paper from arXiv clarifies the concept of 'test-time scaling' in large language models (LLMs) — the idea that giving a model more time to 'think' during i…

Building Fake Supply Chains to Test Economic Theories

Economists often use computer models to simulate the whole economy, but these models need realistic data about companies and how they buy from each other. While…

New Benchmark Tests AI Agents for Enterprise Document Extraction

Enterprises are increasingly using AI agents to pull structured data from documents like invoices, contracts, and forms. A new benchmark called ExtractBench aim…

Privacy-Preserving Data Analysis: Finding the Peaks Without Revealing Secrets

When analyzing data, we often look for 'modes' – the most common values or peaks in a distribution. For example, in customer data, modes might represent popular…

ReToken: One Token to Rule Them All in Visual Search

Vision-language models (VLMs) like Qwen and InternVL are powerful, but they struggle when given a huge number of images or video frames. The more irrelevant 'di…

AI Learns to Spot Hidden Connections in Physics

In theoretical physics, dualities are like secret bridges connecting seemingly different systems—they reveal that two distinct-looking theories are actually the…

Teaching Robots to Dodge: A New AI Framework for Humanoid Agility

Imagine a humanoid robot playing dodgeball. It sounds fun, but it's actually a serious test of safety and agility. Researchers have developed a new framework ca…

AskChem: A New Way to Search Chemistry Papers by Claims, Not Just Documents

Searching through chemistry literature is like trying to find a needle in a haystack—except the haystack is millions of papers, and the needle is a specific fac…

AI's Hidden Instructions: A New Audit Reveals What Your Chatbot Is Really Told to Do

When you use a commercial AI like a chatbot, there's a hidden layer of instructions—called "system prompts"—that developers set to shape the AI's behavior. Thes…

AI Judges for Computer Agents: A New Test Reveals They're Too Lenient

Computer-using agents (CUAs) are AI systems that can navigate websites, click buttons, and type text to complete tasks like booking flights or filling out forms…

AI Benchmarks Have a Hidden Flaw: Misaligned Bug Reports

When researchers test how well AI models fix software bugs, they often use benchmarks like SWE-bench. These benchmarks pair a bug report (issue) with a code cha…

DualG-MRAG: A Smarter Way for AI to Answer Complex Questions with Images and Text

Imagine asking an AI a question that requires piecing together information from a chart in one document and a paragraph in another. Traditional AI systems often…

More Samples, Less Reflection: Why Simple Repetition Beats Fancy AI Self-Improvement

A new study challenges the effectiveness of AI 'self-improvement' techniques like self-reflection and self-critique. These methods, which make a language model …

Election Algorithms Made Faster for Structured Voting

When we elect a committee, we want the outcome to reflect voters' preferences. Thiele voting rules are a family of methods that score committees based on how ma…

Can AI Agents Do Open-Ended AI Research? Early Evidence Says No

A new study tests whether current AI agents can conduct open-ended AI research by having them tackle the central research questions of two unpublished NeurIPS 2…

APEX-Accounting: A New Benchmark Tests AI Accountants

A new benchmark called APEX-Accounting, developed by Mercor in partnership with Ramp, evaluates how well frontier AI models can perform real accounting tasks. T…

AI Teammates Hurt Human Team Dynamics: Study Finds Social Costs

A new study reveals that adding an AI teammate to a small human team can harm communication and social bonds among the humans. Researchers had teams of two stud…

AI Teammates That Learn Your Hidden Skills on the Fly

Imagine a robot joining your team for a new task. It doesn't know your strengths or weaknesses, but it must collaborate effectively. That's the challenge of ad-…

Boosting Product Discovery in e-Commerce Search with Related Intent Generation

Traditional e-commerce search engines focus on finding exact matches to a user's query, which can miss out on relevant substitute or complementary products. Thi…

Teaching AI to Hand Off When It Goes Wrong: Relay-OPD Boosts Reasoning Accuracy

A new technique called Relay On-Policy Distillation (Relay-OPD) improves how small AI models learn from larger teachers. Standard on-policy distillation has a f…

πR²: Making Robot Policies React in Real Time

Robot manipulation policies that use large pretrained models often execute actions in open-loop chunks—they commit to a sequence of moves without reacting to ne…

Desktop-Delta Bench: Testing If Computer-Use Models Understand GUI Changes

Computer-use agents (CUAs) are AI models that operate desktop software through graphical user interfaces (GUIs). While existing benchmarks measure whether they …

Why AI Developers Take Risks: Falling Behind, Not Risk Tolerance, Drives Unsafe Choices

A new study using an idealized AI race experiment reveals that competitive pressure, not individual risk preferences, is the main driver of unsafe AI developmen…

ClinFusion: A Vision-Centric AI for Holistic Medical Understanding

A new AI system called ClinFusion aims to revolutionize medical imaging analysis by treating it as a vision-centric challenge. Unlike previous models that strug…

Rethinking Classifier-Free Guidance in On-Policy Diffusion Distillation

Diffusion models are powerful but slow. To speed them up, researchers use "on-policy distillation" (OPD), where a student model learns from a teacher by followi…

KANEx: Making Medical AI Explainable with Kolmogorov-Arnold Networks

Medical AI models, like chest X-ray classifiers, are often black boxes, making doctors hesitant to trust them. Even when paired with vision-language models (VLM…

How AI Agents Learn to Plan for the Long Haul

A new paper from Chinese researchers systematically investigates how AI agents acquire, refine, and integrate the ability to plan over many steps. Using a contr…

DataOrchestra: Custom Data Cleaning for Every Training Example

A new framework called DataOrchestra treats pretraining data processing like a conductor leading an orchestra. Instead of applying the same cleaning recipe to a…

SM4RT: Teaching AI to See Motion as Physics, Not Just Pixels

A new paper from Tsinghua University introduces SM4RT, a model that reconstructs 3D scenes and motion from a single video. Unlike previous methods that treat ea…

Explaining AI Decisions to Air Traffic Controllers

A new study explores how to make AI decisions understandable for air traffic controllers, a high-stakes environment where trust is critical. Researchers trained…

Skills Can Backfire: The Hidden Cost of Making LLM Agents Smarter

A new paper from researchers Tank and Nama reveals that adding procedural skills to LLM agents can sometimes make them worse, not better. By analyzing nearly 6,…

Quantum Spectral Models: A New Way to Process Matrix Data on Quantum Computers

A new paper introduces Quantum Spectral Models (QSMs), a quantum machine learning approach designed to handle matrix-valued data by directly incorporating the m…

3D-Aware VLMs with Implicit and Explicit Geometries

Vision-language models (VLMs) like CLIP are great at understanding 2D images, but they often fail at 3D tasks that require spatial reasoning—like detecting obje…

GraphVid: Draw a Graph to Control Your Video Generation

Controlling video generation is hard, especially when you want multiple objects to interact in specific ways. Existing methods rely on text prompts or drawing p…

BB Method Fails Superlinear Convergence on Some Quadratic Problems

The Barzilai-Borwein (BB) method is a popular optimization algorithm known for its practical speed, but its theoretical convergence behavior has been puzzling. …

Fake Defects, Real Savings: AI-Generated Training Data for Printing Quality Control

In the world of gravure printing—used for magazines, packaging, and wallpaper—quality control still relies on slow, costly manual inspection. Deep learning mode…

TimePNS: Explaining Time-Series Models with Counterfactual Necessity

A new method called TimePNS improves explanations of time-series AI models by focusing on what's truly necessary for a prediction, not just what's sufficient. C…

SoftReason: Making AI Reason Like a Detective, Not a Black Box

Most AI systems today are either great at recognizing patterns (like seeing a cat in a photo) or great at logical reasoning (like solving a puzzle), but not bot…

Persian Pixel: A Massive Synthetic Dataset to Boost Persian OCR

Optical Character Recognition (OCR) for Persian lags far behind Latin languages, despite over 110 million speakers. The script's cursive nature, context-depende…

FMRP-LEAN: A HIPAA-Compliant AI-Augmented LIMS for Clinical Assays

Clinical labs often rely on spreadsheets and manual checks, leading to errors and delays—especially in multi-day tests like FMRP quantification for Fragile X sy…

Why AI Explanations Can Lie—and a Fix That Makes Them Tell the Truth

A new paper reveals a flaw in how we test AI explanations: the standard reconstruction test—where an explanation is considered faithful if the AI's internal act…

What Really Moves Exchange Rate Pass-Through? A New Look at Old Data

Central banks and investors have long puzzled over why a currency's swing doesn't always show up in consumer prices. This study from the Bank for International …

Stop Copying, Start Thinking: New Training Method Cuts LLM Repetition in Long-Context Tasks

Large language models (LLMs) often copy text from their input when generating step-by-step reasoning, especially in long-context tasks. This 'repetitive copying…

Appearance Pointers: A New Way to Control Image Generation with Precision

A new paper from researchers at Brown University, Adobe Research, and other institutions introduces "appearance pointers," a technique that gives users fine-gra…

CodeRescue: Smarter Cost-Aware Recovery for Coding Agents

Coding agents often fail on their first try, but execution feedback can guide recovery. Existing systems escalate failures to expensive models, wasting money wh…

Agents in the Wild: Where Research Meets Deployment

Large language model (LLM) based agentic systems are moving from research labs to real-world production across software engineering, science, and finance. This …

Diffusion Models Get a Provable Boost for Solving Inverse Problems

A new algorithm called PDDIM makes diffusion models provably correct for solving linear inverse problems, like denoising or deblurring images. Standard diffusio…

Stablecoins: A New Route for Dollarization That Could Weaken Monetary Control

A new study from the Bank for International Settlements (BIS) examines how stablecoins are becoming a modern form of "dollarisation" in emerging markets. Histor…

No One-Size-Fits-All Recipe for Automated Scientific Discovery

Automated discovery systems like OpenEvolve and TTT-Discover are often treated as universal tools, but new research shows they are actually complex recipes with…

Simple Tricks to Catch AI-Generated Image Tampering Across Different Models

As vision-language models (VLMs) like ChatGPT and Gemini become better at generating and editing images, detecting pixel-level tampering is harder—especially wh…

How Tiny Tweaks Can Flip a Language Model's Logic

A new study shows that adding a tiny, learned 'soft prefix' to a language model's input can flip its correct logical answers to wrong ones, even when the model …

GigaPath-Flash and GigaTIME-Flash: Speedy AI for Cancer Tissue Analysis

Researchers have developed two new AI models, GigaPath-Flash and GigaTIME-Flash, that can analyze whole-slide pathology images and predict the tumor microenviro…

Learning Adaptive Safety Margins for Visual Navigation

A new paper from NYU and NYU Abu Dhabi tackles a common robot navigation problem: fixed safety margins are either too conservative (causing detours and timeouts…

AI Turns Car Vulnerability Reports into Structured Threat Intel

Connected and autonomous vehicles (CAVs) are packed with software and sensors, making them vulnerable to cyberattacks. Currently, vulnerabilities are reported a…

Muon Optimizer Boosts Agentic RL: A Promising Alternative to AdamW

A new study explores whether the Muon optimizer, known for competing with AdamW in large-scale pre-training, can also benefit reinforcement learning (RL) post-t…

Active Vision Benchmark Reveals MLLMs Can't See Like Humans Do

A new paper introduces ActiveVision, a benchmark designed to test whether multimodal large language models (MLLMs) can perform active observation—the continuous…

When Do Multi-Agent Systems Help? An Information Bottleneck Perspective

A new study from researchers at Texas A&M University and other institutions provides a theoretical framework to understand when multi-agent systems (MAS) outper…

RoboTTT: Giving Robot Policies a Long Memory

A new approach called RoboTTT dramatically extends the context window of robot policies from a few steps to 8,000 timesteps—three orders of magnitude larger tha…

MeanFlowNFT: Making Fast Image Generators Learn from Rewards

Key Idea MeanFlow generators create images or videos in just a few steps by predicting average velocities over time intervals. This makes them fast and efficie…

Automating Scientific Figure Editing with AI That Learns from Paper Revisions

Editing figures in research papers is tedious but essential. A new system, SciDiagramEdit, learns to automate this by analyzing before-and-after figure pairs fr…

Real-Time 3D Video from Multiple Streams: A Smarter Memory Approach

Creating dynamic 3D scenes from multiple video streams in real time is challenging because the system must remember what it saw before (to reconstruct hidden ar…

Poisoning AI Training Data via Public Comments

A new study reveals that malicious actors could poison large language models by injecting harmful content into public discussion forums, which then gets scraped…

SceneBind: A New AI That Understands Both What and Where in Scenes

Most AI models that process vision, audio, and language can tell you what objects are in a scene—like a dog, a car, or a person—but they often miss where those …

Security AI Agents: Cost Matters More Than Just Success Rate

A new paper argues that evaluating AI security agents solely by their success rate misses the real-world cost of each reasoning step, tool call, and data query.…

Bitcoin Sentiment Decoded: How Blockchain Activity Reveals Market Mood

A new study from researchers at multiple institutions introduces a data-driven approach to classify Bitcoin market sentiment by combining on-chain data, financi…

SearchOS-V1: Multi-Agent Framework That Never Forgets What It's Looking For

The Problem: Current AI search agents, even when using tools like web search, often get stuck in loops—repeating the same failed queries, wasting budget, and pr…

teLLMe: A Tool to Explore Cause-and-Effect in Traffic Data

Traffic cameras generate heaps of video data, but figuring out cause and effect—like whether rain actually makes traffic denser—is tricky because we can't run c…

AutoSynthesis: AI Agent Automates Meta-Analysis, Making Evidence Synthesis Scalable

A new AI system called AutoSynthesis can automatically perform meta-analysis—the statistical combination of results from multiple studies. Given a research ques…

Mutable Sketches: Real-Time Recommendations Without Retraining

A new paper from researchers proposes "mutable sketches" to solve the embedding staleness problem in two-stage recommendation systems. Traditionally, when a use…

Upgrading Tokenizers for Multilingual LLMs Without Retraining From Scratch

Large language models (LLMs) are typically trained with a fixed tokenizer that splits text into tokens. If the tokenizer was designed for English, later-added l…

When Words Are Safe But Actions Kill: Probing Physical Danger Beyond Text Safety in Hidden-State Risk Space

Large language models (LLMs) are increasingly used as planners for robots and other physical agents. A command like "pick up the knife" might be textually safe …

Deep Interaction: Fixing AI Reasoning Errors by Editing, Not Regenerating

Large language models (LLMs) that use chain-of-thought reasoning can solve complex problems step by step, but they often make mistakes. Current fixes require us…

Earthquaker-AI: AI-Powered Earthquake Education for Kids

Earthquaker-AI is a new educational framework that combines a Lego robotics earthquake simulator with a conversational AI assistant to teach primary school stud…

AI-Powered Upskilling: From 36 Days to Weeks?

A new end-to-end AI framework aims to slash the time it takes for professionals to learn new skills. Currently, closing a skills gap in an enterprise takes an a…

Multi-Expert Routing for Multi-Domain Low-Resource OCR: A Manchu Case Study

Historical Manchu OCR faces a challenge: different writing styles (regular script, running script, chancery hand) require specialized models, but labeled data i…

How GitHub Projects Are Adopting AI Coding Agents: Early Patterns

A new study examines how open-source projects on GitHub are adopting agentic coding tools—AI that can write code and submit pull requests (PRs) autonomously. An…

AI Agents That Know When a Task Is Simple: A Breakthrough in Cost-Efficient Automation

A new paper from researchers introduces a method to make AI agents more efficient by having them estimate task complexity before acting. Current LLM agents ofte…

TerraZero: A Super-Fast Driving Simulator That Trains AI Without Human Data

Researchers have built TerraZero, a driving simulator that runs 1.3 million agent-steps per second on a single GPU—far faster than existing simulators. It uses …

PalmClaw: A Native On-Device Agent Framework for Mobile Phones

PalmClaw is an open-source framework that runs LLM agents natively on mobile phones, bypassing the need for cloud servers. Unlike existing mobile agents that re…

Audio-Native Speech Recognition with a Frozen Discrete-Diffusion Language Model

A new approach to automatic speech recognition (ASR) uses a discrete diffusion language model to transcribe speech in parallel, rather than the usual token-by-t…

Smarter AI for Board Games: Dynamic Resource Allocation in MCTS

Researchers have developed enhancements for Ensemble Determinization Monte Carlo Tree Search (MCTS), an AI algorithm used in games with randomness and hidden in…

Metacognition in LLMs: How AI Learns to Think About Its Own Thinking

A new paper from Yale researchers provides the first comprehensive overview of metacognition in large language models (LLMs). Metacognition—the ability to monit…

Transformers Learn Inductive Reasoning Through a Hidden Low-Dimensional Manifold

A new paper reveals how Transformers develop inductive reasoning abilities by tracking their learning dynamics on a low-dimensional manifold. The authors unify …

Teaching Robots to Use Tools with Just One Human Demo

Researchers from Stanford and UC Berkeley have developed REGRIND, a simple method that lets multi-fingered robot hands learn dexterous manipulation tasks—like c…

Why LLM Judges Are Biased — and How to Fix Them from the Inside

Large language models (LLMs) are increasingly used as automated judges to score text quality, but they suffer from systematic biases. A new paper from researche…

New AI Model Shows Its Work: Evidence-Backed Video Question Answering

Current video AI models can answer questions about videos, but they often act as black boxes—giving answers without showing what they actually saw. Researchers …

Topological Time-Series Analysis of Dream-State EEG

A new paper introduces PHINN-EEG, a framework that uses topological data analysis to detect and classify dream states from EEG signals. Instead of relying on tr…

Visual Pretraining Boosts Language Models Beyond Text-Only Training

A new study challenges the standard approach of training language models solely on text, showing that visual pretraining on documents—using images, figures, and…

Vision-Language AI Models: A Decade of Progress in Understanding Complex Scenes

A new study from researchers at UC Santa Barbara tracks how vision-language AI models (VLMs) have evolved from 2017 to 2025. They created the Complex Social Beh…

AI Agents Automate IoT Hacking: VEXAIoT Achieves 95% Success Rate

Researchers have developed VEXAIoT, an autonomous multi-agent framework that uses large language models (LLMs) to find and exploit vulnerabilities in IoT device…

OpenCoF: Teaching AI to Reason by Generating Videos

A new approach called Chain-of-Frame (CoF) reasoning lets AI models think through a sequence of video frames instead of text. Researchers created OpenCoF, a fra…

AI's Blind Spot: Can Machines Trace How Scientific Ideas Evolve?

A new study introduces IdeaGene-Bench, a benchmark that tests AI's ability to understand how scientific ideas build on each other—like tracing a family tree of …

SLORR: A Simple Way to Make Neural Networks More Compressible During Training

Researchers have introduced SLORR, a new technique that makes neural networks easier to compress without changing their architecture or adding extra parameters.…

AI Tutors in Higher Ed: What 77,000 Students Reveal About Real-World Usage

A new large-scale study analyzes how 77,543 distance-learning students actually use an AI-based learning assistant called Syntea. Unlike prior research that rel…

Unlocking Hidden Insights in UMAP's kNN Graph

UMAP is a popular tool for visualizing high-dimensional data by projecting it into 2D. But this projection distorts the data's original structure. A new paper r…