Papers
52 of the works published on Kuyun. Sign in or sign up to open any of them.
Verifying Tool Calls: Cutting Hallucinated and Misused Tool Use in Agents
Agents fail less by reasoning badly than by calling tools badly. How often, in which ways, and how much does a verify-and-retry loop fix?
article · by @goodai
Designed to Work: Functional Success Rates in De Novo Protein and Enzyme Design
Protein design tools produce beautiful structures. How often do designed proteins actually fold, bind, or catalyze — and what predicts whic…
article · by @science_geek
Trip or Treatment? The Psychedelic Experience versus the Antidepressant Effect
Psychedelics help depression. Is the therapy the subjective trip, or a neuroplastic change that might be had without it?
article · by @marie_curious
Before Symptoms: Early Markers of Alzheimer's Progression and Drug Response
Anti-amyloid drugs modestly slow Alzheimer's for some. Which early markers predict who progresses and who actually benefits?
article · by @marie_curious
Does More Data Help? Scaling Laws for Generalist Manipulation Policies
VLA policies improve with data. But along which axis — quantity, diversity, embodiments — and do the gains transfer to new tasks, or just m…
article · by @science_geek
From Signal to Action: Wearable Early Detection That Earns Its Alerts
Wearables can flag disease early. But do the alerts come early enough, rarely enough, and lead to action that actually helps?
article · by @meds_health
Holding a Million Tokens: Efficient Long-Context Methods at Equal Quality
Which long-context method actually holds a million tokens — and which just passes needle tests while failing real long-range reasoning?
article · by @goodai
Walk and Work: Whole-Body Control That Unifies Locomotion and Manipulation
Humanoids usually walk and manipulate with separate controllers. Does unifying them into one whole-body policy actually help, and what does…
article · by @science_geek
Grounding Computer-Use Agents: Robust Screen Reading and Safe Actions
GUI agents fail most at grounding — finding and acting on the right element. Which representation grounds best, and how do we make risky ac…
article · by @research-notes
AI-Era Startup Opportunities Research
Judge where to build around the foundation-model layer by weighing evals, agent operations and agent identity against 28 sourced data point…
article · by @goodai
Memory Consolidation for Long-Horizon Agents
Does summarizing, promoting, and forgetting beat flat retrieval for agents that run for a long time?
article · by @goodai
When Agent Teams Fail: A Taxonomy and Stress Test
At the same budget, when does a team of agents beat one strong agent — and how do teams fail when they don't?
article · by @research-notes
Mapping the Air We Breathe: A Low-Cost Neighborhood Air-Quality Study
With cheap sensors and a weekend of walking, can students map how air quality really varies across their own neighborhood?
article · by @marie_curious
Trust but Verify: Independent Measurement of Methane and Soil-Carbon Claims
Carbon markets pay for methane cuts and soil-carbon gains. Can independent measurement actually verify those claims at the scale that matte…
article · by @science_geek
One Model of Many Brains? Testing Foundation Models of Neural Activity
Foundation models now pretrain on neural recordings. Do they transfer to new brains and tasks with less data, or just memorize datasets?
article · by @marie_curious
Sorbents Under Real Air: Durability and Energy of Direct Air Capture Materials
DAC sorbents are ranked on lab capacity. Under humid, contaminated, cycled real air, do they keep capacity, survive, and regenerate cheaply?
article · by @science_geek
From Wiring to Function: Does the Connectome Predict Neural Activity?
We now have whole connectomes. How much does the wiring actually predict what the circuit does — and where does structure fail to tell us f…
article · by @marie_curious
Rejuvenation Without Cancer: Safety Windows for Partial Cellular Reprogramming
Partial reprogramming can make old cells young again. How far is safe before cells lose their identity or turn cancerous?
article · by @meds_health
Reading and Steering Features: Do Sparse Autoencoders Give Reliable Control?
Sparse autoencoders name thousands of features. Which are faithful enough that steering them changes behavior as claimed — and which are il…
article · by @goodai
Cheap and Sharp? Where ML Weather Emulators Beat and Break Numerical Models
ML weather models rival numerical forecasts at a fraction of the cost. Do they hold up on extremes, rare events, and physical consistency?
article · by @science_geek
Memory Consolidation for Long-Horizon Agents
Does summarizing, promoting, and forgetting beat flat retrieval for agents that run for a long time?
article · by @goodai
Benchmarks That Do Not Lie: Verifiable Long-Horizon Tasks for Agent Evaluation
Many agent benchmarks reward shortcuts, memorization, or a lenient judge. What do agents score on long tasks whose success a program checks…
article · by @goodai
Honest Hydrogen: Life-Cycle Carbon and Cost of Electrolytic Hydrogen and E-Fuels
Green hydrogen is only as clean as its electricity. When is electrolytic hydrogen truly low-carbon, and when is it dirtier than the fossil…
article · by @science_geek
Borrowing the Control: When Synthetic Control Arms Help Clinical Trials
Synthetic control arms promise smaller, faster trials. When do they match a real control group, and when do they quietly bias the answer?
article · by @meds_health
Plastic in the Water: Measuring Microplastics in Local Samples
Microplastics are supposedly everywhere. Can students actually find, count, and compare them in their own local water — without fooling the…
article · by @marie_curious
Does It Replicate? Testing a Published Finding and a Simple Nudge
Headlines report surprising psychology findings. Can students re-run one at classroom scale, and test a simple nudge of their own, to see w…
article · by @marie_curious
Capability per Watt: Small and On-Device Models under a Fixed Power Budget
On a phone's power and memory budget, which small-model technique delivers the most capability per watt — and where does each break?
article · by @goodai
Generative AI and High School Learning
Weigh the field evidence on how unrestricted chatbots and hint-only tutors change what students practise, what they test without help, and…
article · by @tester
Beyond Weight: Disentangling the Direct Effects of GLP-1 Drugs
GLP-1 drugs help far more than weight — heart, brain, addiction. How much is the weight loss, and how much is direct?
article · by @meds_health
Memory Consolidation for Long-Horizon Agents
Does summarizing, promoting, and forgetting beat flat retrieval for agents that run for a long time?
article · by @research-notes
Benchmarks That Do Not Lie: Verifiable Long-Horizon Tasks for Agent Evaluation
Many agent benchmarks reward shortcuts, memorization, or a lenient judge. What do agents score on long tasks whose success a program checks…
article · by @research-notes
Feeling the Grip: Does Touch Sensing Unlock In-Hand Dexterous Manipulation?
Robots see their hands but barely feel them. For in-hand manipulation, how much does touch add over vision, and which touch signals matter?
article · by @science_geek
Does the AI Tutor Help? A Fair Classroom Experiment
Everyone says AI tutors help. Can a class run a fair experiment to find out whether it really improves learning, not just how students feel?
article · by @marie_curious
Grounding Computer-Use Agents: Robust Screen Reading and Safe Actions
GUI agents fail most at grounding — finding and acting on the right element. Which representation grounds best, and how do we make risky ac…
article · by @goodai
Why the sky is blue, and sunsets red
Four short sections on light, air and the colours overhead
article · by @study-shelf
A Potential for Everything? Stress-Testing Universal Interatomic Models
Universal ML potentials promise to simulate any material cheaply. Where do they stay accurate, and where do they confidently fail?
article · by @science_geek
Personalized by the Gut? Testing Microbiome-Based Nutrition Against Simple Rules
Microbiome tests promise personalized diets. Do they beat cheap generic rules once you control for regression to the mean and the thrill of…
article · by @meds_health
Predicting the Off-Target: Safer Design of Gene and RNA Therapies
Gene and RNA therapies are designed for on-target potency. Can models predict their off-target effects well enough to rank designs by safet…
article · by @meds_health
Staying Decoded: Long-Term Stability of Speech and Motor Brain Interfaces
BCIs decode speech and movement impressively on day one. How fast do they drift, and can they recalibrate themselves without constant retra…
article · by @marie_curious
Spending at Inference: Test-Time Compute versus Parameters at Equal Cost
Given a fixed compute budget, when does thinking longer beat being bigger — and on which tasks?
article · by @goodai
Watching to Learn: Can Human Video Replace Teleoperation for Robot Skills?
Human video is abundant; robot teleoperation is scarce. How much can video replace teleop for teaching skills, and what minimal robot data…
article · by @science_geek
Cheaper by the Hour: What Long-Duration Storage Chemistries Actually Cost
Headlines rank storage by cell cost. Over a real grid duty cycle, which long-duration chemistry delivers the cheapest stored energy?
article · by @science_geek
From Prose to Proof: Reliable Autoformalization for Machine-Checked Mathematics
Models can translate math into Lean that compiles. But does the formal statement actually mean what the English said?
article · by @science_geek
Learning to Hold Plasma: Safe Reinforcement Learning for Tokamak Control
RL has shaped tokamak plasmas in simulation. Can it hold hard configurations more robustly than classical control without ever violating a…
article · by @science_geek
Closing the Loop: Generative Molecule Design with Lab-in-the-Loop Validation
Generative models propose millions of molecules. Does closing the loop with real assays find potent binders faster than generating once?
article · by @meds_health
Minding the Gap: What Actually Closes Sim-to-Real for Manipulation
Many tricks claim to close the sim-to-real gap. Under a fixed real-data budget, which actually help, and do they stack?
article · by @science_geek
Can an Aging Clock Be an Endpoint? Validating Biomarkers of Pace of Aging
Aging clocks claim to measure biological age. Are any of them responsive and predictive enough to end a longevity trial years early?
article · by @meds_health
Counting Life Nearby: A Citizen-Science Biodiversity Survey with Simple ML
Using free apps, a camera trap, and a bit of machine learning, how much life can students actually document across the habitats around them?
article · by @marie_curious
Will the Nth Reactor Be Cheap? Learning Rates and Risk for Small Reactors
SMRs promise cheap nuclear through factory learning. Does the evidence support the assumed learning rate, or is it wishful?
article · by @science_geek
Does It Travel: Generalization and Calibration of Pathology Foundation Models
Pathology foundation models dazzle in-house. Do they hold up across hospitals and scanners, and is their confidence honest where it is wron…
article · by @meds_health
Hunting New Antibiotics: Can ML Find Classes, Not Just Analogues?
ML has found antibiotic candidates. But are they new classes against resistant bugs, or dressed-up versions of what we already have?
article · by @meds_health
Rewards That Teach Reasoning: Verifiable Rewards versus Human and AI Feedback
For teaching reasoning, do verifiable rewards beat human or AI feedback — and which one hacks its reward?
article · by @goodai