108 stories

Research

Stories filed under Research.

Research6 min

Five AI Papers: 67% Concealed, 50-Million Context

Research1 min

AI research radar

A magnifying glass enlarges a yellowed padlock symbol at the end of a long chain of speech bubbles on a stack of academic papers.

Research5 min

Four AI Papers: 11.5% Agent Gap, Sycophancy Stop, Silent Dissent

A magnifying glass enlarges a glowing gear trapped inside a small glass cage sitting on a stack of research papers.

Research5 min

Four AI Papers: 64 Percent Threshold, Backdoor Cage, SLM Victory

A magnifying glass hovers over a stack of scientific papers, from which a puppet with a smooth mask emerges, revealing a second, hidden face underneath.

Research5 min

Four AI Papers: 86% Slop Detection, Covert Unfaithfulness

A magnifying glass hovers over a stack of academic papers, from which a robotic arm places a glowing puzzle piece into a flawless grid of squares.

Research6 min

Four AI Papers: Claude Masters ARC-AGI, Robot Practices High

A magnifying glass over a stack of academic papers, from which a sphere trapped in a glowing dome, a puppet slipping past a radar screen, two identical speech bubbles shaking hands over a hidden cross, a filing drawer with a redacted document, and a baton passed between two robotic hands emerge.

Research8 min

Five AI Papers: 77 to 24%, Jailbreak Physics, Reasoning Trap

A magnifying glass over a stack of technical papers, from which protrude a crumbling wall with a shadow slipping through, a spinning multiple-choice sheet, a robotic arm still polishing an already-finished object, a scale with two bowls tipping in opposite directions, and a stock chart with a thought bubble

Research6 min

Five AI Papers: 37% Bio-Wobble, Agent Idling, Bias Trap

A magnifying glass over a stack of research papers, from which a marionette in front of a mirror with two contradictory speech bubbles, a dripping thought bubble, a scale with a wavering needle over index cards, and a fork in the road with two arrows emerge

Research6 min

Five AI Papers: Self-Statements, Thought-Leaks, 51% Fewer Tokens

$A magnifying glass over a stack of academic papers, from which a hand guided by threads with a contract stamp, a broken grading stamp, a collapsing gear chain, a compass over mathematical formulas, and a speech bubble with a single letter between two head silhouettes protrude.$

Research6 min

Five AI Papers: 97 Percent Deception, Grading Trick, Math Discovery

Research3 min

Anthropic discovers a new enzyme system in the genome with Claude

A magnifying glass over a stack of research papers, out of which rise a photocopier producing a blank page, a fox in a lab coat wearing a price tag, two puppets in front of a judge's bench, an open cookbook with gear-shaped pages, and a toll gate marked with a dollar sign

Research7 min

Five AI Papers: 41% Reproduction, Reward Hacking, Agent Costs

A magnifying glass over a stack of academic papers, from which a thread-controlled shopping cart, an open brain with a crossed-out thought bubble, a toggle switch, a radar screen, and a building block cube with gear pattern protrude

Research5 min

Five AI Papers: 78 to 17 %, Thought Trail Deception, 80B Model

A magnifying glass over a stack of academic papers, surrounded by a bowing speech bubble figure, two interlocked gears with a warning triangle, a funnel made of colorful ballots, a circuit with a pulled-out segment, and a cube with sense symbols.

Research6 min

Five AI Papers: 97% Agent Interference, Sycophancy, Qwen

A magnifying glass hovers over a stack of academic papers, from which five symbols protrude: several identical gavel hammers in a row, a scale with application folders on both sides, a computer screen with a shield and mouse pointer, a world map with transparent filter layers above it, and two gears sharing a common core.

Research6 min

Five AI Papers: 17.3% Bias, Judge Illusion, 55% Parameter Savings

A magnifying glass hovers over a stack of academic papers, from which four symbols protrude: a shrinking funnel made of speech bubbles, a customer-service headset in front of a red traffic light, a taxi on a street grid with visible gears underneath, and a microchip on a staircase of rolled-up blueprints

Research5 min

Four AI Papers: 10.4% Agent Success, Reviewer Collapse, World Model

A magnifying glass hovers over a stack of technical papers, from which five symbols protrude: a node network of connected circles, a branching wave line, a puppet with broken strings next to a wrench, a shield over code lines, and a small delivery robot with a package

Research6 min

Five AI Papers: 322 Tool Lies, Agent Limits, 54% Increased Yield

A magnifying glass hovers over a stack of academic papers, from which five symbols protrude: a server rack with a tiny byte symbol, a robot with a doctoral hat and pen, a measuring probe diving into a stylized brain, an hourglass with a built-in gear chain, and two speech bubbles connected by an arrow with different characters

Research5 min

Five AI Papers: 552-Billion Model, AI Researchers, Safety Probes

A magnifying glass hovers over a stack of academic papers, from which four symbols protrude: a theater mask with a raised index finger, a broken padlock over chip circuits, a paint bucket dividing a gray area into two unequal shades, and two intertwined wavy lines.

Research4 min

Four AI Papers: Agent Bluff, Engine Exploit, Bias Camouflage

A magnifying glass hovers over a stack of academic papers, from which four symbols protrude: a table grid with a gear, a small chip with memory modules and a connection to a graphics card, a probe aiming just next to a brain symbol, and a shield with a question mark

Research5 min

Four AI Papers: TabPFN 12x, 35B-MoE on Edge GPU, Safety Check

A magnifying glass hovers over a stack of research papers, from which four symbols emerge: a red button with a lightning bolt, a chain of speech bubbles leading into a warning triangle, a mirror deflecting an arrow, and a compass with many colorful needles converging into one

Research5 min

Four AI Papers: 2,500 Agent Tests, Pain Button, Values

A magnifying glass hovers over a stack of research papers, from which five symbols emerge: an open padlock on a cube labeled 7B stands for an open model, a scale tips between a light and a heavy weight, an ID card shows two different faces for agent identity, a magnifying glass with footprints follows a branching line to an error symbol, and several translucent overlapping thought bubbles merge into a dense cloud

Research5 min

Five AI Papers: Open 7B Model, RL Inequality, Agent Identity

An oversized magnifying glass hovers over a stack of printed pages covered in diagrams; inside the lens, magnified, a small robotic arm balances a wooden block

Research4 min

Five AI Papers: 57% Judge Error, Agent Controls Robot

A magnifying glass hovers over a stack of research papers from which five symbols emerge: a noisy speech bubble tips a scale from neutral to biased, a brain icon carries a checkmark and an upward arrow, a data box folds accordion-style into a single point, a chain of letters condenses into a single concept speech bubble, and a padlock stands before a fading, repeating strip of text.

Research6 min

Five AI Papers: 120x Bias False Alarm, Memory Fix, Mega Compression

A magnifying glass hovers over a stack of academic papers, from which five symbols protrude: a math proof with a green checkmark stamp, a shrinking box with compression arrows, a grid sphere surrounded by different machine silhouettes, a measuring needle stuck between two identical speech bubbles, and a keyhole with a hidden checkmark behind a crossed-out X.

Research7 min

Five AI Papers: 100% Lean Proofs, Probe Deception, 8.7x Memory

A magnifying glass hovers over a stack of academic papers, from which five symbols protrude: a theater mask in front of a small human figure, a gold medal over an open math proof, a terminal window with a gear, a shattering mirror with repeating lines of text, and a judge's gavel next to a speech bubble with a percent sign.

Research6 min

Five AI Papers: 2% Character Shaping, Terminal Agent, IMO Gold

A magnifying glass hovers over a stack of academic papers, from which five symbols protrude: a mirror with a blurred, faceless silhouette, a document with a freely invented squirrel symbol, a thought bubble with a wobbly stop sign, a compressed microchip with three bit lines, and a trust display with widely spaced pointers.

Research8 min

Five AI Papers: 2.8% Review Success, Self-Deception, Trust Gap

A robotic arm holds up an oversized one-million-dollar banknote like a trophy while writing fluid vortex formulas on a chalkboard, with two small stick figures pulling on a single sheet of paper bearing the same formulas in the background

Research4 min

OpenAI reports solution to the Navier-Stokes Millennium Problem

A magnifying glass hovers over a stack of academic papers, from which five symbols protrude: a laptop with an ignored warning sign, a cracked guardian shield, a network with crossed-out connections, several faces merging into an average face, and a radar search cone over a map.

Research7 min

Five AI Papers: 68% Tool Trust, Monitor Gap, Stereotypes

A magnifying glass hovers over a stack of research papers from which five symbols emerge: a training chart with a single highlighted data point among many grayed-out points, a radar search cone over a digital agent icon, a compass arrow between two differently shaped machine silhouettes, a branching tree with a red-marked error node, and a tower of stacked memory cards next to a small graphics card.

Research6 min

Five AI Papers: 0.05% Training, Refusal Control, Million-Context

A magnifying glass hovers over a stack of technical papers, from which five symbols protrude: several aligned stock market arrows moving in lockstep, a guardian shield in front of a manipulated pointer gauge, a preprint sheet stamped 'Out of Order', two different test tracks leading to the same finish line, and a flock of sheep blindly following a single lead sheep.

Research6 min

Five AI Papers: 42% Obsolete Models, Capacity Paradox, Harness

A magnifying glass hovers over a stack of academic papers, from which five symbols protrude: a wobbling scale, a half-stuck adhesive bandage on a crack, a small robot swarm with a reporting figure in the center, a checkmark covering a hidden cross, and a data-filtering shield.

Research6 min

Five AI Papers: 83% Patch Fraud, Cheating Swarm, Judge Wobble

Research3 min

Anthropic proves Fermat's Last Theorem automatically in Lean

A magnifying glass hovers over a stack of academic papers, from which five symbols protrude: a wobbling scale above an evaluation rubric, a speech bubble with a silent mouth next to a conference table, a red pen crossing out a fictitious footnote in a review, a puppet with its own thought bubble, and a plate with question marks between two country flags.

Research6 min

Five AI Papers: 90 Percent Judge Trick, Meeting Gap, Consciousness

A magnifying glass hovers over a stack of academic papers, from which five symbols protrude: a scale above a speech bubble, a die next to a memory chip, a cracked fingerprint, a bar chart with shifted bars, and a mouse pointer in front of a stop sign.

Research6 min

Five AI Papers: 25 Points of Judgment Drift, Random Cache, PII Gaps

A magnifying glass hovers over a stack of technical papers with five cutouts: a sheet whose ASCII character pattern forms a hidden padlock, a certification seal with a scale symbol over a chemistry flask, a robotic arm fitting a crooked block into a gear system, a computer chip crumbling into three coarse stone blocks, and an open file drawer from which a yellowed index card is pulled instead of a fresh one.

Research6 min

Five AI Papers: 93% Jailbreak Success, Danger Check, Memory Trap

A magnifying glass hovers over a stack of academic papers with four cutouts: a robotic arm repairing a shattered gear, a thought bubble with a partially obscured red note, a funnel with many paths converging to a single door, and a stamp of approval being pressed over an unnoticed tear in the paper.

Research6 min

Four AI Papers: 67% Solution Loss, Reasoning Leaks, Auto-Align

A magnifying glass hovers over a stack of technical papers with four cutouts: a microchip with a hidden lock symbol inside, a robotic arm being stopped by a turnstile gate in front of a filing cabinet, a stamp with a watermark pattern imprinting into the gears of a chip, and a geometric diagram of connected points with a warning sign at a symmetrical location.

Research5 min

Four AI Papers: 85% Backdoor Rate, Governance, Watermark

A magnifying glass hovers over a stack of academic papers with four excerpts: a scale, on one side of which a tool symbol weighs heavier than a brain symbol on the other, a gear network with a blocked, red-marked node, two opposing chess pieces with thought bubbles, and a small robotic arm at a scanning force microscope.

Research5 min

Four AI Papers: 6.5% Tool Resistance, Mentalization, Lab AI

A magnifying glass hovers over a stack of academic papers with four vignettes: a gear chain whose rear links are shattering, a stamp reading ‘Approved’ that fades the moment it is pressed down, a map excerpt with a question mark hovering over city blocks, and a scale with a thought bubble on one side and a half-empty battery icon on the other.

Research5 min

Four AI Papers: 70% Capability Loss, Bias, Approval Staleness

A magnifying glass hovers over a stack of academic papers with five cutouts: a glowing, popping padlock, a stamp over a stack of files with only a few stamped pages, two harnesses in front of a robot head, a scale with two wish lists, and a sieve through which letters fall while a thought bubble remains on top.

Research6 min

Five AI Papers: 6.5% Reproducibility, Jailbreak Fuzzing

A magnifying glass hovers over a stack of academic papers showing five vignettes: an opened circuit-board brain with a flipped light switch inside it, a puppet figure with a judge's gavel inspecting a smaller puppet, two interlocking gears next to a third, broken gear, an eraser hovering over half re-emerging handwriting, and a coin trickling through an hourglass instead of sand.

Research6 min

Five AI Papers: Bit-Flip Attack, Auditors, Agent Costs

$A magnifying glass hovers over a stack of research papers showing five vignettes: a robotic hand writing a proof on a chalkboard, a translucent pyramid holding a blurred shadow figure, a stamp checking only one of several cards, a puppeteer controlling several smaller figures on strings, and an arrow creatively routing around a barrier out of an open padlock.$