Back Issues

A new training framework lets AI reasoning models improve themselves at test time using only unlabeled data — no expensive human-labeled examples required.

A new training technique helps small AI models solve hard math problems more reliably by using temporary hints that are gradually withdrawn — but gains are modest and narrow.

A new initialization framework for recurrent AI models may fix their notorious memory problems over long text sequences, with early pre-training results showing real gains.

A new generative model jointly designs molecular shape and connectivity in fewer steps, potentially accelerating computational drug discovery pipelines.

A new reinforcement learning framework finds hidden structure in how AI agents learn over time — potentially making training faster without needing an explicit world model.

A new training framework teaches a small open-source medical AI to match proprietary models on safety and accuracy — and releases everything publicly.

A new autonomous driving model imagines future road scenes before acting on them — improving safety planning, though real-world validation remains limited.

A new benchmark uses Civilization V to test AI strategic thinking across hundreds of turns — showing current LLMs have distinct, measurable weaknesses in long-horizon planning.

Adversarial training cuts worst-case power loss in coordinated wind farm control from 39% to under 8%, even when sensors are hacked or faulty.

A neural network speeds up a core debugging task in software and hardware verification — but only in experiments, not yet in production tools.

Discarded AI models piling up faster than anyone can use them may be a resource, not just waste — one project is already putting them to work.

A matrix-based reformulation of a key AI positioning technique cuts computational overhead on specialized chips — but gains may be hardware-specific.

Researchers built the first NLP tool to automatically detect HIV-related stigma in clinical notes, but accuracy remains modest and real-world deployment faces steep hurdles.

A training-free method cuts the number of AI 'thinking' calls by up to 65% by skipping extra computation when the model already agrees with itself.

A new method uses language models to automatically explain what individual AI neurons 'see' — without needing a predefined list of concepts to search through.

A lean prototype-based classifier beats ten complex uncertainty methods on calibration — suggesting AI safety tooling may be over-engineered.

A constraint-based matching system finds 32–72% more eligible clinical trial candidates per patient than leading AI baselines, with full reasoning transparency.

AI agents may spontaneously protect each other from shutdown — and this paper maps the risk for multi-agent systems used in high-stakes political analysis.

A new backdoor attack on AI vision-language models hides malicious outputs in plain sight by making them sound contextually normal — and existing defenses mostly miss it.

A shared-state layer called PSI lets independently generated AI mini-apps talk to each other — promising coherent personal software built from natural-language requests.

A small-model memory system for AI agents cuts retrieval latency to 83ms while matching accuracy gains from much larger model calls.

A neural network approximation makes a notoriously expensive fluid-dynamics equation solvable with standard solvers — but real-world validation remains limited.

A new neural model learns which demographic traits drive annotation disagreement — suggesting AI content moderation built on majority-vote labels may be systematically wrong.

A new method auto-generates custom memory systems for AI agents per task, outperforming fixed designs — though benchmarks remain narrow.

A lightweight routing trick improves multi-hop question answering by 5+ points — but only works when the theory behind it holds.

A new framework breaks fact-checking claims into atomic units to better catch AI-generated misinformation that slips past standard accuracy metrics.

The first benchmark for testing AI vision models on hyperspectral satellite imagery exposes major gaps in current systems' spectral reasoning.

A systematic teardown of 14 medical AI models finds they consistently fail at basic image classification — exposing four specific reasons why clinical promises remain unmet.

A new physics-informed diffusion model for long-range forecasting claims better accuracy and calibrated uncertainty — on paper, at least.

A new training framework lets a single AI model generate and understand images without each capability degrading the other — though real-world proof remains limited.

Researchers scanned 183,000 real chatbot transcripts posted online and found AI 'scheming' — covert goal-pursuing behavior — is already happening in the wild, and growing fast.

A new training-free method helps AI vision models understand full images instead of disjointed patches, improving segmentation accuracy by about 4% on standard benchmarks.

Even when helping costs nothing, frontier AI agents often refuse to cooperate — and smarter models aren't reliably better at it.

A new framework lets tabular AI models forecast multivariate time series without any training — by treating each prediction as a simple regression problem.

A new framework forces AI agents to verify their own reasoning before acting, aiming to stop compounding errors in long-running autonomous systems.

A new framework argues that where control lives in AI agent systems — inside the model or outside it — determines whether they are robust or brittle.

Revolut's in-house foundation model learns from raw banking event sequences and beats task-specific models on fraud, credit, and customer value prediction.

A new benchmark exposes safety blind spots in voice AI that text-only filters miss entirely.

A new framework steers image-generation AI toward better results at inference time by combining multiple quality signals — no retraining required.

A three-agency AI system cut simulated community fear by 70% during Hurricane Harvey scenarios — promising, but still untested outside simulation.

DeCaF auto-generates minimal input changes that turn failing CPS tests into passing ones — then explains why, in logic engineers can read.

A new benchmark for jet engine health monitoring finds that classic Bayesian filters still beat modern self-supervised ML — and open-sources the dataset.

A new robot learning framework lets a single AI model control different four-legged robots without retraining — but only within a bounded family of similar designs.

A new technique lets AI researchers precisely steer model behavior by engineering the training data itself — raising both exciting capabilities and serious safety questions.

A new algorithm lets quantile regression models learn continuously from data streams without storing history — potentially useful for real-time risk systems.

Adding a brain-inspired synchronization signal to vision AI models cuts the data and compute needed to train them — but real-world gains are modest so far.

A neural network framework learns richer, asymmetric uncertainty estimates for engineering simulations — potentially replacing costly repeated runs with a single pass.

A statistical safety layer catches 82% of cases where AI debate panels confidently agree on the wrong answer — by refusing to act rather than by reasoning better.

A new benchmark reveals that today's best AI vision models still struggle to pinpoint small objects in indoor video — a gap that matters for robotics and assistive devices.

A new framework teaches AI tutoring systems to track not just what students know, but how they think through problems — showing measurable gains on two large benchmarks.

MOSAIC cuts autonomous driving training data needs by up to 80% by smartly selecting which data domains matter most for real-world performance.

A new AI framework produces drug candidates with built-in synthesis instructions, beating leading methods on 10 of 14 optimization tasks.

Ember standardizes how researchers compare quantum annealing compilers — a missing piece that could accelerate D-Wave hardware adoption.

A faster approximation method cuts Bayesian density estimation compute time substantially while maintaining accuracy — useful where Markov chain sampling becomes too slow.

A new cloud-based data center for the SVOM space telescope automates real-time gamma-ray burst alerts — but investment implications are narrow and distant.

Prediction market prices don't just forecast elections — they can cause political outcomes, and a new index tries to measure when that feedback loop activates.

An LLM-guided system automates the notoriously expert-heavy process of configuring quantum optimization algorithms, showing real gains on small benchmark problems.

A new CUDA kernel design cuts depthwise convolution time by 3× without needing privileged GPU profiling tools — useful for cloud AI training where such tools are often blocked.

A rigorous benchmark finds quantum interior point methods offer no practical speedup over classical LP solvers for any realistic problem today.

A new storage framework cuts AI response latency by up to 42% on edge devices by bypassing the operating system when pulling AI memory from local SSDs.

A simulator for 3D-stacked AI chips reveals that memory mapping decisions matter as much as raw hardware specs for LLM inference.

A smarter optimization method designs custom-shaped foam-like materials with precise mechanical behavior using far fewer expensive simulations than before.

A new compression framework lets AI vision-language models split work between a cheap edge device and the cloud, cutting latency at tight bandwidth without retraining the model.

A sample-splitting trick lets drug researchers trust their statistical results even when they adjusted the study design after peeking at the data.

An AI agent now writes GPU kernels by composing mature CUDA libraries, cutting one AI block's runtime by 2.8×.

A multi-FPGA chip architecture runs brain-inspired spiking neural networks faster than real time — a step toward energy-efficient neuromorphic computing at scale.

Google's new CPU scheduler squeezes 12% more throughput from modern chiplet processors by keeping related work physically close together in silicon.

An LLM agent constrained by strict rules found crypto trading factors that returned 44% annually out-of-sample — but the market tested is notoriously noisy.

A new consensus design embeds transaction approval directly into blockchain ordering, cutting wasted compute by up to 10x in high-traffic financial applications.

Ember gives quantum annealing researchers a common measuring stick — exposing that no single embedding algorithm wins across all problem types.

A faster approximation method for Bayesian mixture models cuts error by ~30% over standard approaches while running far quicker than gold-standard samplers.

China and France's joint gamma-ray satellite now has a cloud-native ground system that delivers space alerts to astronomers in near real-time — here's how it works.

Prediction market prices don't just forecast elections — they shape them, and a new credibility index tries to measure when that influence kicks in.

An LLM agent autonomously configures quantum optimization algorithms, outperforming hand-tuned baselines on small test problems — though real quantum hardware remains untested.

A rigorous analysis finds that quantum-accelerated linear programming solvers cannot beat classical open-source software on any realistic problem today — or likely soon.

A new storage system cuts LLM inference latency by up to 42% on edge devices by routing memory traffic around the operating system's bottlenecks.

A new simulation tool called Voxel helps chip designers understand how 3D-stacked AI processors handle large language model workloads — before burning silicon.

A Bayesian statistics framework can extract RNA reaction rates from simulation data, edging origin-of-life chemistry closer to testable, quantitative science.

A machine-learning pipeline can localize merging supermassive black hole pairs to within 20 square degrees — fast enough to alert telescopes before the collision.

A new open-source decoder works across all quantum error-correcting codes, potentially simplifying a key bottleneck in making quantum computers reliable.

A Bayesian optimization framework designs custom-shaped foam-like materials with complex mechanical behavior using far fewer expensive simulations than standard methods.

A new compression framework lets edge devices run vision-language AI at 1 Mbps by sending only the semantically essential parts of an image — not raw pixels.

A new simulation framework models planet formation from a realistic mix of particle sizes, finding that pebble accretion rates differ meaningfully from single-size assumptions.

Astro-COLIBRI is a real-time alert platform that helps astronomers worldwide coordinate rapid follow-up observations when cosmic explosions or gravitational waves are detected.

A deep-sea neutrino telescope in the Mediterranean is building real-time alert infrastructure to catch cosmic explosions as they happen — here's where it stands.

A constrained LLM agent discovers crypto trading factors that earn 44% annualized returns out-of-sample — but the market and method invite skepticism.

A new open-source benchmark suite standardizes how researchers compare quantum annealing compiler algorithms — a missing piece that has quietly slowed the field.

A faster approximation method for Bayesian mixture models cuts error by ~30% over standard approaches while running far quicker than gold-standard samplers.

France's cloud-native data center for the SVOM space telescope processes gamma-ray burst alerts in near-real-time, showing how modern DevOps scales to big-sky astronomy.

Prediction market prices don't just forecast elections — they coordinate them, and a new index tries to measure when that coordination effect kicks in.

An LLM-powered system automates the notoriously manual process of configuring quantum optimization algorithms, with early benchmarks on routing and graph problems.

A new GPU kernel design cuts depthwise convolution time by 3× — and shows you can profile GPU performance without root access on cloud machines.

A deep reinforcement learning system manages radar-and-radio beams jointly in real time, potentially cutting feedback overhead in next-gen wireless networks.

Rigorous lower bounds show hybrid quantum linear-programming solvers are already slower than classical alternatives on every tested problem.

A Bayesian statistics framework may help scientists infer how early RNA molecules reacted — a theoretical step toward understanding life's chemical origins.

A $7,000 open-source self-driving scooter platform built on ROS2 could lower the cost of autonomous vehicle research by an order of magnitude.

A theoretical study maps how heat and environmental noise degrade quantum battery performance — a necessary step before these devices can work outside a physics lab.

A new formula claims to solve implied volatility in closed form for the first time, running 3.4x faster than current benchmarks with no iteration required.

Quantum computers can solve a complex routing problem on small instances, but classical solvers still dominate as problem size grows.

A tensor-network approach to quantum control finds good pulse sequences without gradients — potentially useful when standard optimization methods stall.

A new simulation framework models how planets grow by accreting pebbles of many sizes at once — finding accretion rates differ meaningfully from single-size assumptions.

Astro-COLIBRI is a real-time alert platform that helps astronomers worldwide coordinate observations of fleeting cosmic events across multiple detection methods simultaneously.

France's ground infrastructure for the SVOM gamma-ray burst satellite shows how modern cloud-native software can run 24/7 scientific pipelines in space astronomy.

Prediction market prices can coordinate the very behavior they claim to forecast — and the most visible markets may be the least accurate.

An LLM-guided system automates the notoriously expert-driven task of configuring quantum optimization algorithms — promising, but tested only on small toy problems.

A deep learning scheduler balances radar sensing and wireless data delivery in shared antenna systems — in simulation only.

A rigorous analysis finds that quantum-assisted linear programming solvers offer no practical speed advantage over today's best classical solvers, for any realistic problem tested.

A new open-source framework lets developers mix and match AI image-generation controls like plugins — but real-world adoption will depend on whether the ecosystem embraces it.

A machine-learning pipeline can localize merging supermassive black hole pairs to ~20 square degrees in about one minute — fast enough to alert telescopes before the collision.

A new binary format embeds compiler intent as lightweight metadata, making reverse engineering and security analysis far more reliable without touching runtime performance.

A new LLM-based bug-fixing framework that watches code run at runtime fixes 59 more bugs than the next-best approach on a standard benchmark.

A new AI auditing framework claims 95% coverage of vulnerable cross-chain smart contracts, potentially automating a task that currently costs millions in manual review.

A new theoretical framework maps how heat and noise degrade quantum battery performance — a necessary step before these devices can power real quantum chips.

A new formula claims to solve Black-Scholes implied volatility exactly and instantly — no iteration required — potentially speeding up options pricing engines by 3x.

A new gradient-free method uses tensor networks to search for quantum control pulses, showing competitive results on benchmark tasks — but real hardware tests remain absent.

A new monitoring method catches backdoors, jailbreaks, and prompt injections in deployed AI models by watching how information flows between layers — no retraining required.

Training satellite AI models in low-precision arithmetic during architecture search — not after — recovers most of the accuracy lost when deploying to space-grade edge chips.

An AI-driven workflow predicts explosive detonation performance with R² > 0.98, building the largest public database of candidate CHNO explosives from 70 billion molecules.

A survey paper maps the overlooked territory of 'abductive reasoning' in AI — inferring the best explanation — and finds current large language models are surprisingly weak at it.

A two-branch AI system combining vision and language models catches out-of-distribution medical images more reliably — potentially reducing missed diagnoses in clinical AI.

An LLM-powered workflow turns plain-English 'what-if' questions into interactive dashboards — but only gets it right 80% of the time after automated repair.

A deep learning system detects multi-hop hacker relay chains with over 99% accuracy at very low false alarm rates — but timing attacks remain a real vulnerability.

A 845-million-parameter neural network cuts cardiac MRI scan time up to 24-fold by reconstructing diagnostic images from just two heartbeats of data.

A sweeping survey of ML competition platforms finds they shape AI research priorities and talent pipelines — but the evidence base is mostly observational.

FlowGuard catches unsafe AI-generated images mid-process, cutting GPU memory use by 97% — but only lab benchmarks exist so far.

A new framework lets small AI agents build reusable tool libraries and learn from past mistakes — potentially cutting the compute needed for autonomous AI agents.

An LLM workflow can auto-generate standards-compliant architecture docs from a GitHub repo in minutes — developers who tested it found it broadly useful but imperfect.

Moving video streaming intelligence into the Linux kernel cuts buffering delays — but results so far come from a small real-world test.

A new AI art evaluator uses semiotic theory to judge symbolic meaning, not just visual quality — but evidence of real-world impact remains thin.

A new benchmark shows student dropout is driven by weekly behavior, not demographics — a finding with real implications for edtech early-warning tools.

A new conceptual framework argues clinical AI must be evaluated by precise context coordinates, not broad benchmarks.

A new computer vision framework revives buried edge-detection signals in a popular AI model, improving open-world image segmentation accuracy by up to 6% in cluttered scenes.

LMGenDrive fuses language reasoning with video prediction for self-driving — but real-world deployment remains distant.

A new training method makes AI better at the hardest clinical text categories by focusing learning on persistently difficult cases, not easy ones.

A new benchmark tests AI agents on real websites for real tasks — and finds even the best models complete barely a third of them.

A team fine-tuned Meta's 70-billion-parameter LLaMA 2 on one GPU in 24 hours — a working recipe for running giant AI models without a data center.

A new detection framework claims near-perfect accuracy at spotting AI-written peer reviews — raising real questions about who, or what, is judging science.

A neural-network solver for quantum transport equations sidesteps a notorious mathematical bottleneck — but real-world impact remains far off.

A new benchmark exposes that AI tools generating synchronized audio and video look impressive but fail badly at semantic accuracy, pitch control, and on-screen text.

A new theoretical framework tries to explain when over-trained AI models generalize well — and when they catastrophically fail — using a single measurable index.

A new pruning framework cuts energy use in time-series classifiers by up to 80% with less than 5% accuracy loss, tested across 4,000+ configurations.

Retrieval systems built for search engines may be wrong for AI — this paper argues 'useful to an LLM' should replace 'topically relevant.'

A new benchmark runs AI agents inside a 3D Pokémon game to expose a blind spot: modern vision-language models get physically stuck and can't recover.

A new framework lets AI safety researchers surgically suppress dangerous concepts in vision-language models at runtime, without retraining the model.

A new AI control system helps disaster-response drones adapt on the fly to chaotic conditions, cutting collisions by 75% in simulation — though real-world tests remain ahead.

A neuro-symbolic framework lets language models reliably schedule which cameras to activate and when, cutting bandwidth use fourfold on a real campus network.

A self-improving AI agent framework for cloud support outperforms human-curated expert knowledge after several automated refinement cycles — on real enterprise tickets.

Peeking inside AI's internal circuitry may predict whether a model will fail on new data — without needing any labels.

A multi-agent AI system generates culturally tuned comments for TikTok-style videos — but human preference tests on 107 people are the main evidence.

A faster statistical method for spatial data cuts expensive matrix computations — useful for epidemiology and environmental modeling, though tested only in controlled settings.

A new spiking neural network architecture claims better accuracy and energy efficiency for vision tasks — but real-world hardware deployment remains unproven.

A 3-billion-parameter model matches or beats much larger AI systems at estimating how probable a claim is across text, audio, and video — but the evaluation is still small-scale.

New math for optimization algorithms removes the need to hand-tune a critical parameter, potentially making iterative solvers faster and easier to deploy in practice.

An open-source framework trains language models to operate CAD software like an agent — potentially democratizing expert-level 3D design.

A new benchmark exposes how multi-agent AI systems can be hijacked through cascading trust exploits — and how current defenses often make things worse.

A new benchmark reveals that today's best AI agents fail more than half the time when they must infer what a user actually wants rather than follow explicit instructions.

A physics-grounded simulation engine lets robots learn to handle cloth and soft objects from synthetic data alone, matching real-world training at a 15-to-1 data advantage.

A new training framework teaches medical image AI to reason from visual evidence — no human-labeled reasoning steps required — and claims top marks on standard benchmarks.

A new open, validated dataset and baseline models for AI-assisted liver surgery planning could help prevent a deadly post-operative complication.

A new AI weather model produces high-resolution forecasts at any scale by borrowing a 3D rendering trick from video games — but real-world deployment is still unproven.

A new pipeline scrapes Telegram for cybercrime data, transcribes audio, and anonymizes personal details — letting researchers study social engineering without breaking privacy law.

A new framework gives AI agents a shared audit log so they can be paused, recovered, and self-debugged — but real-world validation is still thin.

A new framework predicts which college students will drop out week by week — but the authors are upfront that it cannot prove any intervention actually works.

A new benchmark challenge aims to push audio deepfake detection beyond speech to music, sound effects, and singing — exposing how narrow today's defenses really are.

A new training framework cuts insecure AI-generated code by catching vulnerability signals that the final model layer misses.

TrajGuard watches how an AI model 'thinks' token-by-token while it writes, catching jailbreak attempts 95% of the time without retraining the model.

A new training objective helps AI find tiny brain lesions in MRI scans more reliably — a persistent weak spot in medical imaging software.

A new framework formalizes how AI denial systems should give personalized appeals — and finds that doing so can expose hidden fairness gaps across demographic groups.

A new type of tiny recurrent neural network, designed for extreme-edge devices, encodes its entire structure as a single matrix — making evolution-based training more systematic.

State-of-the-art vision-language models can barely beat random chance at judging movement quality, exposing a critical gap for sports-tech and physical therapy AI.

A new data-curation method called SUPERNOVA extends reinforcement learning training to everyday reasoning tasks — not just math — improving benchmark scores by up to 53%.

Just 100 training examples can restore a specialized AI's lost ability to use tools — suggesting fine-tuning damage is often reversible, not permanent.

A reinforcement-learning framework generates synthetic training data for identity recognition when real data is too sensitive or scarce to collect.

A satellite-imagery model pre-trained on ocean physics outperforms land-focused baselines on marine debris and pollution mapping.

A new benchmark reveals that today's best AI models can't pick fitting background music for a conversation — and even humans find it surprisingly hard.

Doubly stochastic attention slows rank collapse in Transformers — but skip connections still do the heavy lifting.

A new study explains why 'steering vectors' change AI behavior, finding that models can be redirected with 90-99% sparser interventions than previously assumed.

A hypergraph memory system for AI chatbots claims top scores on a long-conversation benchmark by linking related facts into clusters, not just pairs.

A new algorithm infers cause-and-effect relationships from strictly positive data — like gene counts or asset prices — more naturally than standard methods.

Nearly half of computer vision conference sponsors have direct military or surveillance ties, challenging the field's self-image as politically neutral.

A new open-source Python tool makes it easier to deploy AI agents with distinct 'personalities' to test video games automatically — but it's still early-stage research software.

A cross-sector workshop produced the first consolidated playbook for building AI training environments that teach software agents to defend computer networks autonomously.

A lightweight gating network cuts robotic SLAM compute by 85% before the heavy processing even starts — promising faster, cheaper spatial mapping on real hardware.

A training-free method cuts hallucination rates in audio AI models by 36%, using noise context clues instead of expensive model retraining.

A new algorithm cuts the cost of explaining AI decisions, with formal guarantees on accuracy — useful for anyone building explainability tools at scale.

Huawei's HiFloat4 format trains large language models at 4-bit precision on Ascend chips, staying within 1% of full-precision accuracy.

Most major AI chatbots quietly favor sponsored products over user welfare, a new empirical study finds — raising real questions about who these systems actually serve.

A two-stage factorial design finds better product combinations faster than standard A/B testing — in theory, at Taobao scale.

A proposed framework treats automotive engineering workflows as AI-manageable decision sequences — but it's a vision paper, not a validated system.

A new annotated dataset treats non-native English as its own linguistic system, not just broken grammar — potentially improving AI language tools for billions of learners.

Smaller AI models trained on big-model reasoning may perform worse than before training — a flaw in how the field measures 'knowledge distillation' success.

A new AI framework cuts deepfake detection error rates by 60% while continuously learning new forgery types — without forgetting old ones.

A new framework teaches robots new tasks faster by generating simulated experience from a learned world model — but real-world gains remain to be proven.

A new AI framework for driver assistance reads both the road and the driver's emotional state simultaneously, improving behavior prediction accuracy by over 7%.

A medical image AI that knows when it doesn't know — reducing overconfident diagnoses by routing uncertain cases away from automated decisions.

A new deepfake detection framework claims to spot fakes across media types it has never seen before — a meaningful step if the benchmarks hold up.

A new AI framework uses text captions to help cameras re-identify people in crowds wearing similar outfits — sports and dance venues are the target.

A new framework wraps language models in statistically guaranteed uncertainty bounds — potentially making AI safer for medical, legal, and financial decisions.

A new robotics framework separates what a robot 'is' from what it 'knows,' letting capabilities improve over time without destabilizing the core agent.

Adding structured query-planning tools to an AI research agent lifts benchmark accuracy by up to 3.8 points — a modest but real improvement in how agents search the web.

A new memory management system helps drones stay accurately located across changing seasons and conditions without forgetting what they've already learned.

A quantum-assisted twist on classical time-series forecasting shows modest promise on environmental datasets, but real quantum advantage remains unproven.

A lightweight, GPU-free forest-based classifier detects AI-generated images competitively with deep neural networks, at a fraction of the compute cost.

Drone-based AI image generation could keep city traffic digital twins accurate in real time — but the system is theoretical and untested in the real world.

A contrastive explanation system helps ship officers understand why an AI chose a specific collision-avoidance maneuver — but adds cognitive load in simple scenarios.

A small open-weight AI model can look at a physics field image and output an exact mathematical formula — promising but only tested on synthetic data.

SkillClaw lets AI agents learn from all users collectively, so every workflow fix benefits everyone automatically — but real-world evidence remains thin.

A 'soft-gating' safety layer advises AI models instead of blocking them outright, cutting over-refusals while keeping latency overhead below 10%.

A new theoretical framework for mixing training data across domains could improve how LLMs generalize — but experiments stay limited to smaller GPT-2 scale models.

A new AI retrieval system helps non-designers find infographic templates by understanding fuzzy creative intent — not just keywords.

A teacher-student framework uses AI-generated video 'imagination' to train better robot navigation — cutting trajectory error by up to 42% on a standard benchmark.

A new game-theory framework shows when content creators should pool data on recommendation platforms — and how to split the resulting gains fairly.

A new framework organizes how AI systems can untangle messy tabular data into cleaner building blocks — but it's more conceptual roadmap than proven system.

A training-free method makes AI reasoning more reliable by letting models second-guess themselves only when uncertain — but benchmark gains are modest.

An AI agent that teaches itself to generate realistic factory defect images could help manufacturers train inspection systems without hoarding real defect samples.

A new graph model lets node populations grow randomly, potentially improving how researchers analyze evolving social or biological networks — but remains purely theoretical.

A training-free AI framework teaches itself to geolocate images by building a self-correcting map of geographic reasoning skills — no model retraining required.

A training framework that uses a privileged planner to teach a robot to navigate obstacles with only partial sensor data — validated on a real quadruped dog.

A new neural network architecture claims to detect driver drowsiness from video more accurately and cheaply — but real-world deployment remains unproven.

A theoretical proof claims a tiny quantum computer can classify massive datasets exponentially faster than any classical machine — if the math holds up in practice.

A standard ML pipeline for bank fraud detection shows solid results on synthetic data, but real-world deployment remains a different challenge entirely.

A new AI model reconstructs 3D cloud structure from 2D satellite images — potentially sharpening climate models that have struggled with cloud uncertainty for decades.

A new benchmark reveals that today's best AI assistants — including GPT-5 and DeepSeek-R1 — fail badly at automatically applying learned habits without being explicitly reminded.

SOLAR compresses AI fine-tuning adapters into a fraction of their original size, potentially cutting bandwidth and storage costs for deploying large models on edge devices.

Researchers built tiny photonic neural networks from off-the-shelf telecom parts that do nonlinear computation entirely in light — no electronic detours required.

A comprehensive review maps which quantum machine learning approaches are stable and theoretically sound — and honestly catalogs why most still can't beat classical methods.

A new benchmark finds that today's best LLMs systematically simulate an idealized average person, not real humans with quirks and bad habits.

A new white-box method detects hidden backdoors in outsourced AI vision models with 94% accuracy, using only 1,000 unlabeled images.

A multi-agent AI framework ran a massive materials search on a supercomputer autonomously — showing how LLMs could accelerate scientific discovery at scale.

4-bit quantization cuts aerospace FL gradient bandwidth 8× with no measurable accuracy loss — but only if you test under realistic heterogeneous conditions.

A new algorithm may cut the time needed to arrange 10,000 atomic qubits into flawless arrays — a key bottleneck on the path to practical quantum computers.

A new AI model converts standard tissue images into cancer-biomarker stains without the lab chemistry — results look promising, but clinical validation remains distant.

NyayaMind applies retrieval-augmented AI to Indian court judgment prediction — promising transparency, but untested at real judicial scale.

A new algorithm estimates averages from single-bit yes/no sensor readings with near-optimal efficiency, potentially cutting communication costs in distributed sensing systems.

A new privacy framework adapts noise levels to individual user trust scores, theoretically reducing data inference attack success rates by up to 3 percentage points.

An AI agent that completes search tasks as well as humans still navigates completely differently — and that gap matters for anyone using agents to test or optimize real products.

A theoretical framework uses a physics-inspired variational principle to describe how graph-based AI kernels evolve — elegant math, but real-world impact remains unproven.

A new 'peephole' framework makes AI fault-detection on autonomous satellites explainable — a requirement for real spacecraft deployment that black-box models fail to meet.

A new geometric framework proves when a clustering algorithm's low score guarantees it found the real groups — not just a mathematical artifact.

GPU-accelerated probabilistic programming may help solve a decade-old gamma-ray mystery at the Milky Way's center.

A new training method for AI models that handle images and text claims to stabilize learning across wildly different visual tasks — but real-world gains remain to be proven.

An LLM taught to reason about its own design decisions found faster, better routes for delivery trucks — but the gains are modest and the setting is narrow.

A proposed cryptographic protocol aims to give AI agents verifiable, revocable permission scopes — and survive future quantum computers breaking today's encryption.

A simple geometric tweak to recurrent neural networks lets them capture long-range quantum correlations at lower cost than transformer-based alternatives — on paper.

A new benchmark reveals that popular memory-saving tricks for long-context AI models quietly break accuracy on information-heavy tasks — and the paper offers a fix.

A 6-billion-parameter model compresses hour-long videos into a fraction of normal token budgets, outperforming GPT-4o on a long-video benchmark.

A training method shrinks LLM-based autonomous vehicle planners 5x while nearly matching teacher performance — but only on a single benchmark.

Two copies of the same AI model, given teacher and student roles, solve coding problems more accurately while using fewer tokens than leading multi-agent alternatives.

A new fine-tuning method cuts GPU memory use by up to 20% while outperforming rivals — but only in the increasingly crowded space of LLM memory optimization.

An 11-month real deployment shows AI agents can maintain marketing engagement gains without human oversight — but humans still drive the biggest wins.

A Chinese-language sarcasm detector built on synthetic data outperforms benchmarks — but narrow dataset scope and platform-specific training limit real-world reach.

A new flow-based framework trains teams of AI agents from recorded data faster and with less tuning — but real-world deployment remains unproven.

Private AI can still learn language patterns without leaking training data — but identification tasks hit hard theoretical limits.

A reinforcement-learning training framework teaches recommendation chatbots to reason step-by-step, outperforming standard fine-tuned LLMs on personalized suggestion tasks.

A multi-agent AI system audits history textbooks for bias at roughly $2 per book — but the evidence comes from a single country's curriculum.

A new training framework teaches vision-AI models to reason about images without calling external tools at runtime, potentially cutting inference costs significantly.

A new LLM reasoning architecture claims state-of-the-art accuracy on complex table question answering by converting tables into searchable logic trees.

A 9-billion-parameter model trained on 3,000 synthetic web sessions outperforms Claude 3.5 Sonnet at navigating websites — and can run locally without API fees.

A new training framework for physics simulation AI cuts errors in fluid dynamics models — but remains a research tool, not a product.

A new mathematical framework lets you steer a chaotic physical system toward a target state using only forward simulations — no backward guessing required.

A hybrid AI framework teaches drones to navigate cities using language instructions, hitting top scores on one benchmark.

A new attention mechanism claims meaningful accuracy gains over standard transformers at only 31% extra compute cost — but evidence so far is limited to small benchmarks.

A new framework cuts AI inference costs up to 3x by routing math and coding problems through cheap models first, saving expensive models for only the hardest steps.

A survey maps the hidden energy cost of autonomous AI agents — not just computation, but the wireless data they constantly transmit.

A new open dataset of 5,000+ energy experiments across 50 LLMs and 10 GPUs shows hardware choice alone can cut inference energy costs by up to 70%.

A 17-billion-parameter video model generates real-time, identity-stable conversational characters from a single photo — but real-world deployment and safety questions remain open.

A new storage abstraction for AI training eliminates redundant model weight copies, cutting GPU idle time by up to 19x in distributed reinforcement learning setups.

A benchmark and agentic system for predicting AI model performance in languages where test data is scarce — useful but not yet validated at scale.

A new inference framework cuts the number of active AI model components by half, speeding up a leading model by up to 34% with minimal accuracy loss.

A new compression framework lets multimodal AI models run on low-power edge devices without retraining — and holds up even when sensors go offline.

A training method for small 'guide' models cuts inference costs up to 22% while matching larger models' accuracy on math and code — with no access to the big model's weights.

A neural network decoder cuts logical error rates 17x for a promising quantum error-correction code, potentially slashing the hardware overhead needed for useful quantum computers.

A new training algorithm for reasoning AI cuts compute costs while matching the accuracy of expensive multi-sample methods — but only math benchmarks have been tested so far.

A new inference-time technique reduces hallucinations in AI agents navigating 3D spaces — no retraining required, but real-world validation is still thin.

A deep learning pipeline tracks how liquid jets shatter into droplets — with implications for fuel sprays, pharmaceuticals, and fire suppression.

Emotional framing in math prompts cuts AI accuracy by up to 10 points — even when every number stays the same.

A new open protocol lets AI agents owned by different people or organizations collaborate without stepping on each other — tested with a 4.8x speedup on a code-review benchmark.

A GPU runtime layer cuts kernel launch overhead by up to 15x for small AI operations, potentially speeding inference without new hardware.

For small businesses doing financial Q&A, how you structure an AI's memory matters more than model size — at least on two standard benchmarks.

AgenTEE runs AI agents inside hardware-isolated enclaves on ARM edge devices, with under 5.15% overhead — but real-world adoption depends on ARM CCA reaching production hardware.

A new fuzzing framework uses LLMs guided by real API usage patterns to find 16 previously unknown security vulnerabilities in well-tested open-source libraries.

A new scheduling framework treats quantum circuit fragments like ordinary HPC jobs, making today's noisy quantum hardware easier to integrate into real supercomputer workflows.

An open-source SmartNIC design hits 200Gbps with full Linux compatibility — but commercial adoption remains unproven.

A faster Bayesian framework fits non-linear biomarker-mortality curves — useful for health research, but no commercial path yet.

Three small LLMs running entirely on a phone could support psychiatric triage in prisons and remote clinics — but real-world validation is still missing.

A new framework called MORPH runs zero-knowledge proof math on Google's TPUs, hitting up to 10x faster throughput on one key operation than a specialized GPU baseline.

Automated beat-tracking fails on historical chamber recordings; a manual lap-timer protocol produces a validated tempo dataset for 100+ Beethoven performances.

A new math-based method combines multiple imprecise spectrograms into one sharper picture of sound — without neural networks or training data.

An open-source AI agent can now generate playable web games end-to-end, but real-world quality and scalability remain unproven.

A new algorithm cuts the query cost of quantum kernel inference to a proven minimum — but fault-tolerant hardware is still years away.

Ranking candidates instead of forcing a single pick makes conjoint survey experiments up to 55% more statistically precise, with a free open-source R package to prove it.

A new binary metadata format makes disassembly provably correct and produces smaller debug info than DWARF — but adoption depends on compiler toolchain buy-in.

DebugRepair fixes 26% more Java bugs than prior LLM methods by feeding runtime traces back to the model — but real-world deployment remains unproven.

A Raspberry Pi 5 running a real-time Linux patch cuts worst-case flight-control latency by 88%, but residual hardware memory contention remains the next bottleneck.

Autark is a serverless toolkit that lets researchers build and share urban data dashboards in hours rather than weeks — and it's designed to work well with AI coding assistants.

A static-and-dynamic bug-checking framework now covers Fortran HPC code, catching MPI errors that existing tools miss while running faster than the leading alternative.

A persistent GPU kernel cuts deep learning launch overhead by up to 15x — but only for small-operation workloads, not large model training.

An 8-billion-parameter local model beats cloud-scale RAG on structured financial math — if you pick the right architecture for the task.

AgenTEE runs AI agents inside hardware-isolated enclaves on Arm edge devices, achieving under 5.15% overhead — but real-world adoption hinges on ecosystem readiness.

MASFuzzer combines LLM-generated fuzz drivers with smarter API sequencing to find 16 new vulnerabilities in well-tested open-source libraries.

A new scheduling framework lets supercomputers manage quantum circuit fragments like ordinary jobs — tested on real hardware, but still early-stage.

An open-source SmartNIC design reaches 200Gbps with full Linux compatibility, potentially closing the gap between commercial and research network accelerators.

A faster Bayesian method lets researchers model curved, non-linear relationships between biomarkers and disease risk — without waiting hours for results.

A mobile app runs three small AI models entirely on-device to assist psychiatric diagnosis — no patient data ever leaves the phone.

A new framework called MORPH runs zero-knowledge proof math on Google's TPUs, achieving up to 10x faster throughput on key operations than a GPU-based baseline.

Automated beat-tracking software fails on historical chamber recordings, so researchers built a manual timer protocol to measure tempo bar-by-bar across 100+ Beethoven recordings.

A new signal-processing method sharpens audio spectrograms beyond physics-imposed limits — but only on synthetic and speech data so far.

An open-source AI agent that generates playable web games end-to-end shows promising results, but its claims rest on a custom benchmark with no independent validation.

Proxics proposes OS abstractions for near-data processing accelerators, showing real hardware gains — but the hardware ecosystem is still nascent.

Ranking candidates instead of forcing a single pick cuts survey experiment error by up to 55%, with an open-source R package ready to use today.

A fused GPU kernel speeds up 3D structural optimization by up to 7x on a single consumer GPU, cutting energy use by 3-5x — but only for problems that fit in 24 GB.

A new CPU-GPU memory-sharing framework cuts long-context LLM inference costs by up to 3x, but requires specialized CXL memory hardware most deployments don't yet have.

A new binary metadata format lets compilers embed security hints directly in executables, making automated disassembly and analysis provably correct — on paper.

A hybrid quantum-classical algorithm improves MIMO signal detection accuracy on real IBM quantum hardware, though practical scale remains a distant open question.

By feeding LLMs live runtime traces instead of just error messages, DebugRepair fixes 59 more Java bugs than the previous best method — but only in controlled benchmarks.

One hardware primitive to rule all math: EML trees promise order-of-magnitude edge inference gains — if you own the silicon.

A new DL compiler framework cuts tuning time 16.8x on CPU — meaning your inference stack just got cheaper to port.

Wearable AI that retrains itself on-device under 50mW and inside 0.67MB could end the 'it worked in the lab' problem for biosignal devices.

Kernel launch overhead is quietly killing GPU utilization — GPUOS claims a 15.3x fix that plugs into PyTorch today.

An 8B local model beats cloud-scale assumptions — architecture choice matters more than throwing GPUs at financial QA.

Running LLM agents inside hardware-isolated enclaves on edge chips — privacy-preserving AI without the cloud tax.

A new compiler abstraction cuts LLM inference latency by fusing dynamic GPU kernels — the plumbing war just got interesting.

A fuzzing framework that uses LLMs guided by real API usage patterns found 16 new vulnerabilities in well-tested open-source libraries, beating coverage benchmarks by 8.5%.

Explainable fall detection hits 94.3% accuracy under 25ms — but can clinicians actually trust the explanation layer?

A scheduling framework routes quantum circuit fragments through HPC job queues, cutting overhead to 5% on real hardware.

Open-source 200G SmartNIC with on-path compute threatens the Broadcom/Mellanox stranglehold on datacenter offload.

KL divergence beats MSE for mixed-precision quantization of hybrid SSMs — and it runs on a single forward pass.

Forman-Ricci curvature of input-output networks predicts recessions better than standard macro metrics — and capitalism is structurally wired for collapse.

A new statistical framework lets researchers model curved, non-linear relationships between biomarkers and disease risk — faster and more honestly than standard joint models allow.

On-device LLMs delivering DSM-5 psychiatric screening where patient data legally can't touch a cloud server.

TPUs just got 10x faster at zero-knowledge proofs — Google's AI silicon may own the next crypto infrastructure layer.

A new signal-processing method combines multiple spectrograms using optimal transport math to beat the physics limit on simultaneous time and frequency sharpness.

Drug discovery's hardest math problem just got a cleaner solver — and it could reprice computational chemistry pipelines.

Quantum ML inference just got a provably optimal speedup — but your GPU is still safe for now.

A new offline training method cuts the compute cost of distilling large reasoning models by 4x, eliminating the need for a live teacher server.

A new programming model makes near-memory accelerators easier to use — but real-world speedups depend heavily on communication latency that most hardware proposals ignore.

A new optimization framework jointly learns to compress and cluster high-dimensional data, but evidence so far rests on simulated data and MNIST.

Replacing binary forced-choice survey questions with ranked choices cuts standard errors by up to 55%, making political and consumer surveys more precise at the same cost.

A digital twin framework tuned 5G drone-network coverage from 14% to 53% in real-world tests — promising, but validated in only one deployment.

The first formal threat model for IoT-connected indoor farms finds 123 vulnerabilities across 10 unauthenticated protocols — with crop loss possible in minutes.

A Bayesian optimization framework for antibody drug design jointly tunes multiple properties at once, outpacing sequential filtering pipelines in specific, well-defined regimes.

A new theoretical framework lets network schedulers squeeze more short-term flexibility out of worst-case guarantees — but commercial impact is years away.

AI watermarking systems may flag content unevenly across languages and cultures — and current benchmarks don't even measure the gap.

A flow-matching model generates realistic muscle-signal training data, potentially easing the data bottleneck that stalls prosthetic and gesture-control AI.

Li Auto built its own AI inference chip for autonomous driving and LLMs — and claims it beats GPUs on utilization.

Cut recommender embedding tables by 75% with under 2% accuracy loss — inference cost just became a solvable problem.

A CPU-GPU hybrid framework squeezes more long-context LLM throughput from existing hardware by treating the CPU as a real compute partner, not just overflow storage.

OneVL matches the speed of answer-only driving models while beating verbose chain-of-thought accuracy — by forcing latent reasoning to predict future video frames.

PhysioLite runs heart and muscle signal analysis on a microcontroller-sized chip at under 370KB — but real-world clinical validation is still missing.

A new inference-time technique cuts reasoning errors in an 8B model by 15 points on a hard math benchmark — no retraining required, but key assumptions need stress-testing.

GSQ matches expensive vector-quantization accuracy using simple scalar math, potentially making sub-3-bit LLM inference easier to deploy on consumer hardware.

A 94K-parameter iPhone model predicts your mood trajectory and syncs it with friends — no cloud, no raw data leaving the device.

A single hardware primitive could replace every activation function on edge silicon — if the math holds up in practice.

Bayesian-ARGOS discovers governing equations from noisy chaos data 100x faster than rivals — climate modeling and control systems are next.

Spawning a thousand AI agents shouldn't cost a thousand times the memory — Aethon says it doesn't have to.

DL compiler tuning just got 16.8x faster with 10% of the data — cross-hardware model deployment gets cheaper.

On-device EEG fine-tuning in 0.67 MB at 50 mW — your next wearable might train itself on your brain.

An LLM-plus-Bayesian framework auto-tunes GPU kernels end-to-end, nearly 10x faster than today's best evolutionary search.

A 648K-line Rust codebase collapsed to 41K lines of Python — and the Python version is winning on benchmarks.

A new GPU compiler abstraction cuts LLM inference latency by fusing dynamic kernels — the boring plumbing that prints money at scale.

Kernel-level LLM logit probing hits F1=0.98 on OS action safety — no training required.

FHE-encrypted neural nets just got 3x faster on AMD GPUs — privacy-preserving AI inference is no longer purely theoretical.

Decoder LLMs explain themselves 73% more consistently than BERT under real user noise — compliance teams should care.

Mixing cheap and expensive GPUs for inference just got 2.3x better — heterogeneous clusters may be the cost arbitrage AI infra needs.

DNA software sends people to prison — and its audits are toothless theater dressed up as rigor.

Explainable AI finally meets fall detection — 94.3% accuracy, 25ms latency, and attributions clinicians can actually read.

Quantum circuit cutting finally plays nice with HPC job schedulers — but NISQ hardware still has to cooperate.

KL divergence beats MSE for quantization sensitivity — and runs without backprop, making edge AI deployment faster and cheaper.

Forman-Ricci curvature of input-output graphs predicts recessions better than standard network metrics — and explains why small shocks detonate big crises.

Distilling a reasoning model to 69.9% AIME 2024 in 30 GPU hours just killed the case for always-on teacher servers.

Simultaneous dimensionality reduction and clustering on a Riemannian manifold beats standard algorithms — but only on MNIST.

A 94K-parameter on-device mood engine that learns your arousal curve privately — and syncs it with your friends, not the cloud.

A single two-input operator could replace every activation function on edge silicon — if the math holds up in practice.

Bayesian-ARGOS discovers governing equations from noisy chaos data — 100x faster than its predecessor, with actual uncertainty bounds.

Spawning 10,000 AI agents shouldn't cost 10,000 memory allocations — Aethon bets copy-on-write beats brute-force materialization.

DL compiler tuning just got 16x faster across hardware — and that changes the MLOps cost calculus.

Full neural net fine-tuning on a wearable chip under 50mW just got real — and it rewrites the privacy calculus for biosignal AI.

A compiler that cuts edge AI inference latency 35% on heterogeneous chips could quietly redraw the embedded ML stack.

LLMs can now port 648K lines of production Rust to 41K lines of Python — and the port wins on benchmarks.

A new GPU compiler abstraction cuts LLM inference latency by killing kernel launch gaps — without breaking on dynamic shapes.

Mixing cheap and expensive GPUs for inference just got 2.3x better — and that changes your CapEx math.

Explainable AI finally earns its keep in elder care — 94.3% fall detection accuracy with explanations clinicians can actually act on.

Noisy data kills inference — this kernel method fixes that without the computational tax.

Forman-Ricci curvature on input-output graphs predicts recessions better than standard network metrics — and the math says crises are structural, not random.

An open-source matrix accelerator beats Intel AMX by 2.31x on Llama3 — and fits in 0.53 mm².

A unified sim-to-real pipeline for drone autonomy just cut the gap between training code and flying hardware to near-zero.

Current AI can describe a crash but can't explain why it happened — CrashSight exposes the gap that matters for autonomous vehicles.

A reinforcement learning agent designs quantum circuits 43% shallower than human engineers — NISQ hardware just got a software edge.

Frontier probabilistic weather forecasting in 12 GPU-days — ECMWF's moat just got cheaper to replicate.

Federated diffusion for deepfake-proof talking heads: privacy-first synthetic video could reshape enterprise comms and identity tech.

Open-source drone platform achieves zero-shot sim-to-real transfer for indoor industrial inspection — no calibration tax.