Articles
Every paper breakdown, newest first.
Orca: What If Next-Token, Next-Frame, and Next-Action Are the Same Task?
Orca (BAAI) replaces next-token, next-frame, and next-action prediction with a single Next-State-Prediction objective. A frozen 4B backbone pre-trained on 12.5K hours of video — with zero action labels — feeds three lightweight readouts, and the action readout, trained on just 200 trajectories per task, beats π0.5 on OOD robot manipulation (32.4 vs 29.4).
SWE-Explore: The Benchmark That Finally Asks — Did Your Coding Agent Read the Right Code?
SWE-Explore isolates repository exploration from patch generation, revealing that coding agents find the right files ~65% of the time but recall only ~15-19% of the lines that actually matter — and that context efficiency predicts downstream resolve rate with Pearson r = 0.950.
SkillOpt: Training Agent Skills Like Neural Network Weights - Without Touching the Model
SkillOpt applies deep-learning-style optimization - bounded edit budgets, validation gating, rejected-edit memory - to natural-language skill documents, improving frozen LLMs by up to +39 points across 52/52 evaluated cells without changing a single model weight.
MolmoAct2: The First Fully Open Robot Controller That Beats Closed-Source Giants
MolmoAct2 is a fully open vision-language-action model that outperforms π0.5 and matches Gemini Robotics ER, achieving 97.2% on LIBERO and 87.1% real-world success via per-layer KV-cache conditioning and adaptive depth reasoning.
RecursiveMAS: What If Your Multi-Agent System Was Just One Big Recursive Neural Network?
RecursiveMAS treats an entire multi-agent system as a single recursive computation in latent space, adding only 0.31% trainable parameters while achieving +8.3% accuracy, 2.4x speedup, and 75.6% token reduction over text-based multi-agent baselines.
Tstars-Tryon 1.0: Virtual Try-On as Multi-Image Editing at Taobao Scale
How a unified 5B MMDiT trained with multi-reward RL and step distillation reframes virtual try-on as multi-image editing — and runs in under 4 seconds in production.
ClawGUI: A Full-Stack Open-Source Pipeline for GUI Agents
ClawGUI unifies online RL training, reproducible evaluation, and real-device deployment of GUI agents into one open-source pipeline — and shows a 2B model trained inside it can beat 72B untrained baselines on MobileWorld.
SkillClaw: Making LLM Agent Skills Evolve Collectively
SkillClaw is a framework for collective skill evolution in multi-user LLM agent ecosystems. Instead of static skill libraries, the system automatically learns from interactions across users and propagates improvements to everyone.
TAPS: Why Your Draft Model's Training Data Matters More Than Its Architecture
TAPS shows draft training data drives speculative decoding: merged-tree verification of math and chat specialists hits 5.11 acceptance length, weight averaging 2.59.
Demystifying Video Reasoning: Models Don't Think in Frames - They Think in Denoising Steps
Chain-of-Steps shows video diffusion models reason across denoising steps, not frames, exploring and self-correcting before converging on a single answer.
Seoul World Model: AI That Generates Video of Real Cities From Street Photos
Seoul World Model generates photorealistic city video from street-view photos, cutting FID on unseen Busan to 28.4 from 62.1 via cross-temporal reference pairing.
Lost in Stories: How LLMs Lose the Thread in Long Narratives
ConStory-Bench measures consistency in 10,000-word stories: GPT-5-Reasoning leads with 0.113 errors per 10k words, and its checker finds 3.2x more bugs than experts.
Utonia: One Encoder For All Point Clouds
Utonia trains one 137M encoder for indoor, LiDAR, satellite, CAD and video point clouds. Stripped of color it scores 77.0% mIoU on ScanNet vs Concerto's 36.8%.
SAGE: Your Reasoning Model Knows When to Stop Thinking — You Just Won't Let It
SAGE scores reasoning paths by cumulative log-probability instead of next-token probability, letting models stop thinking 40-50% earlier at equal or better accuracy.
When GPT Discovers Physics: A Breakthrough in Gluon Theory
GPT-5.2 Pro proposed the formula showing single-minus gluon amplitudes are nonzero in Klein space, valued only in -1, 0 or 1; it passed five consistency tests.
OPUS: How to Train LLMs 6x Faster by Choosing the Right Data
OPUS scores every sample by alignment minus redundancy in optimizer update space; on Qwen3-8B it beat full 3B-token training using 0.5B, at 4.7% extra compute.
Green-VLA: One AI Brain for All Robots
Green-VLA trains one policy for humanoids, mobile manipulators and fixed arms across five stages and 3,000 hours of demos, hitting state of the art on CALVIN ABC-D.
To Grok Grokking: Why Neural Networks Sometimes Understand Late
Grokking happens even in ridge regression: fast memorization, a long plateau of orthogonal noise, then late generalization with a delay that scales as 1/λ.
Tensor Networks: A Mathematical Bridge Between Neural and Symbolic AI
Tensor networks give neural and symbolic AI one language: logic becomes sparse tensors, neural nets low-rank ones, and inference in both is a contraction.
M²FMoE: When Experts Learn to Predict Floods
M2FMoE routes Fourier and wavelet experts by frequency band to forecast floods without extreme-event labels, cutting error 45.7% on Almaden and beating 13 baselines.
BALLAST: When a Bandit Teaches Your Database How Long to Wait
BALLAST replaces Raft's random election timeouts with a LinUCB contextual bandit; a Rust simulator cuts cluster unavailability by 84%, and 90.8% under bursty WAN.
AI Co-Scientist: Teaching Models to Write Research Plans Better Than Humans
Rubrics mined from papers turn a frozen grader into an RL reward; experts preferred the generated research plans 70% over baseline and 52% over published ones.
HyDRA: Teaching Your Phone to Understand Images Without Breaking the Bank
HyDRA allocates LoRA rank per layer and inside layers automatically, buying a 4.7% gain on mobile vision-language models with no extra trainable parameters.
Comp-LLM: When an Army of Experts Beats a Giant – An Analysis of a Revolution in AI Architecture
Comp-LLM routes a query as a DAG of sub-queries to parallel experts: about 35B scores 0.83 F1 to Llama-2 70B's 0.85, and 1.1-1.7x faster than sequential agents.
NVIDIA Nemotron Parse v1.1: The Complete Anatomy of the Digital Document Understanding Revolution
Nemotron Parse v1.1 pairs a C-RADIO ViT-H encoder with an mBART decoder in 885M parameters, squeezing 13,184 visual tokens to 3,201 and emitting Markdown or LaTeX.
Cost-Constrained LLM Cascades — Meet C3PO
C3PO runs LLM cascades under a probabilistic cost constraint: stay in budget with 95% confidence, escalating to a bigger model only when the cheap one falls short.
Accurate Satellite Rain Forecasting with Physics-Conditioned Neural Networks
TUPANN nowcasts rainfall from satellite data alone, adding a differentiable advection operator and lead-time-conditioned MaxViT for forecasts 10-180 minutes ahead.
A Universal Crime Predictor – How Hypernetworks and Knowledge Graphs Are Transforming Forecasting
HYSTL pairs a crime knowledge graph with a hypernetwork that generates GNN weights per crime type, letting one model forecast across cities with disjoint taxonomies.
CHORD — Smart On-Device Recommendations Without Killing Your Battery
CHORD sends each phone a bit map from a cloud hypernetwork, quantizing a sequential recommender channel by channel so weak devices keep accuracy and battery.
DIALTREE‑RPO — a simple guide to 'Tree‑based Dialogue Reinforced Policy Optimization'
Friendly explanation of 'Tree‑based Dialogue Reinforced Policy Optimization for Red‑Teaming Attacks' — with examples and LaTeX formulas.
Attention as a Compass – Teaching Reasoning Models to Explore Smarter
AttnRL uses a model's own attention to pick branching points in process-supervised RL, beating GRPO and TreeRL by 1.8 points on math while cutting training time 8%.
How to Detect Credit Card Fraud?
Credit card fraud is 492 cases in 284,807 transactions; pairing undersampling with SMOTE beats either alone, and KNN and MLP gain most on the imbalanced test set.
Quantum Trading – AI and Quantum Computing in Investing
A QLSTM forecaster plus a QA3C agent trade USD/TWD with just 244 parameters, returning 11.87% over 2020-2025 at 0.92% max drawdown versus classical A3C's 11.42%.
Reinforcement Learning in Pinterest Ads – DRL-PUT in action!
DRL-PUT replaces Pinterest's hand-tuned ad ranking weights with a policy network picking utility weights per request; A/B tests show CTR up 9.7%, LC-CTR up 7.7%.
Uczenie ze Wzmocnieniem w Reklamach na Pinterest – jak DRL-PUT zmienia grę / Reinforcement Learning in Pinterest Ads – How DRL-PUT Changes the Game
Pinterest's DRL-PUT learns the weights of its ad ranking utility with policy gradients instead of tuning them by hand, lifting CTR by 9.7% and LC-CTR by 7.7%.
The Anatomy of AI Lies: How Language Models Can Deceive Us
Can LLMs Lie? traces deception to layers 10-15 with logit lens, zero-ablation and steering vectors: models rehearse lies in dummy tokens; bigger models lie better.
Edge AI: How to Accelerate Neural Networks on Specialized Hardware
SLAC's SNL and Auto-SNL push neural network inference onto MPSoC FPGAs for LCLS-II, swapping weights without reconfiguration and matching or beating hls4ml latency.
Global Guarantees of Robustness: A Probabilistic Approach to AI Safety
Rather than certify every point, Mu and Lim estimate the probability that a random input is non-robust, wrapping a Clopper-Pearson interval around a small sample.
Intern-S1: The New AI Scientist That's Redefining Research
Intern-S1 is an open-source MoE science foundation model with 28B activated parameters, trained on 5T tokens and a Mixture-of-Rewards spanning over 1000 tasks.
Look Inside Seamless Flow's Hyper-Efficient Training
Seamless Flow kills RL pipeline bubbles with a data plane and tag-driven GPU scheduling: up to 100% more throughput than VERL, and 23.0% to 45.8% on SWE-Bench.
Systematization of Knowledge: Data Minimization in Machine Learning
SoK: Data Minimization in Machine Learning maps the GDPR principle onto ML pipelines, separating horizontal from vertical techniques and pre-hoc from post-hoc steps.
Learning Machines That Don't Forget: A New Method for Evolving Data
DyMoE counters catastrophic forgetting in incremental graph learning by adding a new expert per increment, with gating and a loss that makes older experts teach it.
Dynamic Fine-Tuning (DFT): How a Single Line of Code is Revolutionizing AI Training
Dynamic Fine-Tuning rescales each token loss by the inverse of its probability, a one-line change that beats SFT on generalization and rivals offline RL methods.
A Deep Dive into the Text-to-SQL Revolution: Analyzing the Adaptive Method
Adaptive Text-to-SQL splits the work in two: a Selector retrieves conceptually similar question-schema-SQL triples, a Generator writes the query in-context.
ASkDAgger: How Artificial Intelligence Learns More Effectively by Asking Questions
ASkDAgger lets a robot flag an uncertain plan before acting; adaptive gating and foresight replay cut teacher queries and raise success on CLIPort and a real arm.
CaPulse: Teaching Machines to Hear the Rhythm of Data
CaPulse spots time-series anomalies by learning the causal rhythm of data: it fires when the cause-effect order breaks, not when a threshold is crossed.
Goedel-Prover-V2: A Revolution in Automated Theorem Proving
Goedel-Prover-V2 learns from scaffolded proof sketches and its own failed Lean attempts, outperforming AlphaProof and DeepSeek-Prover on miniF2F and PutnamBench.
How to Teach AI to Handle Mistakes? Meet ε-Softmax
ε-Softmax swaps rigid one-hot targets for slightly softened ones, so a mislabeled example costs the model far less. A NeurIPS 2024 answer to noisy training labels.
Simple and Effective Method for Uncertainty Quantification
Uncertainty from one deterministic model: kernel density estimation in feature space builds an information potential field flagging OOD samples on CIFAR-10 vs SVHN.
Deep Learning-based Prediction of Clinical Trial Enrollment with Uncertainty Estimates
A multimodal Clinical Longformer model predicts trial enrollment at R² 0.76 and forecasts recruitment duration with 7.52-month MAE versus 10.55 for fit-and-filter.
Consensus-Driven Active Model Selection
CODA picks the best model from many candidates via their agreement: Bayesian Dawid-Skene plus information gain wins 18 of 26 benchmarks with up to 70% fewer labels.
RLVMR: Reinforcement Learning with Verifiable Meta‑Reasoning Rewards for Robust Long‑Horizon Agents
RLVMR adds a verifiable meta-reasoning reward to the environment reward, scoring the agent's reasoning trace so long-horizon agents take fewer semantic shortcuts.
How AI Can Reveal Where Your Honey Comes From — A Look at Mineral Fingerprints
Random Forests read the 12-mineral fingerprint of 429 honey samples to name the source flower with 99.50% accuracy and the region of origin with 98.01%.
Optimizing Call Center Operations with Reinforcement Learning: PPO vs. Value Iteration
PPO beats classical Value Iteration at call routing in a discrete-event call centre simulation: shortest customer wait, least agent idle time, highest RL reward.
Efficient & Geometrically-Smart: Linear Memory SE(2)-Invariant Attention Explained
Fourier-encoded relative pose lets SE(2)-invariant attention use linear rather than quadratic memory, with error below 0.001. Best paper at the RSS 2025 workshop.
A Lightweight AI Engine for Skin Cancer Detection on Wearable Devices
A MobileNetV2 skin-lesion classifier compressed with TensorRT runs on a Jetson Orin Nano at 41% of its original size, scoring 87.2% F1 with 93% lower power in INT8.
SOPHIA: Enhancing Slow‑Thinking in Large Vision‑Language Models
SOPHIA mixes on-policy visual grounding with off-policy reasoning chains and backward reward propagation, adding 8.5 points of average accuracy to InternVL3.0-38B.
The Role of AI in Managing Satellite Constellations
ConstellAI, an ESA project, uses reinforcement learning for routing in satellite mega-constellations, beating Dijkstra on delay and predicting bandwidth demand.
On the Fundamental Limitations of Dual Static CVaR Decompositions in Markov Decision Processes
Godbout and Durand show the two dual CVaR formulas can share no feasible risk assignment, and build an MDP where no policy is optimal at every risk level.
PinFM: Foundation Model for User Activity Sequences at a Billion-Scale Visual Discovery Platform
PinFM, Pinterest's 20B-parameter sequence transformer, uses deduplicated cross-attention and int4 embeddings to lift throughput 600% and engagement 20%.
GradNetOT: Learning Optimal Transport Maps with GradNets
GradNetOT learns optimal transport maps as the gradient of a convex potential, using a monotone gradient network trained on the Monge–Ampère residual.
Unstable Power: How Sharpness Drives Deep Network Learning
A larger learning rate leaves the NTK better aligned with the labels at the Edge of Stability: Kernel Target Alignment jumps exactly when sharpness drops.
RiemannLoRA: A Unified Riemannian Framework for Ambiguity-Free LoRA Optimization
RiemannLoRA places LoRA updates on the manifold of rank-r matrices, erasing factorization ambiguity: half the fine-tuning steps and up to 1.2 BLEU more.
A Neural Network Model of Complementary Learning Systems: Pattern Separation and Completion for Continual Learning
A complementary learning systems model pairs a VAE with a modern Hopfield network for pattern separation and completion, hitting 90% average accuracy on Split-MNIST.
Target Polish: How to Polish Data and Reveal Its True Structure
Target Polish cleans data before factorizing it: clip X into a delta band around the current WH estimate, refactorize, repeat. Robust NMF for matrices and tensors.
Optimistic Exploration for Risk-Averse Constrained Reinforcement Learning
ORAC maximizes an upper confidence bound on reward while minimizing a lower bound on cost, with an adaptive penalty, tested on Safety-Gymnasium and CityLearn.
Not Just Bigger Models: Why AI Should See Better Instead of Just Scaling
Adaptive sensing beats brute-force scale: EfficientNet-B0 paired with sensors that tune themselves to the scene outperforms OpenCLIP-H on perception tasks.
HGMP: Revolutionizing Complex Graph Analysis with Prompt Learning
HGMP recasts node, edge and ranking tasks as graph-level prompts, combining contrastive pre-training with per-type feature prompts for heterogeneous graph learning.
Predicting and Generating Antibiotics Against Future Pathogens with ApexOracle
ApexOracle fuses Evo2 genome embeddings with a SELFIES diffusion model to predict MIC and design new antibiotics, with R2 gains up to 15% on P. aeruginosa.
HeLo – A New Path for Multimodal Emotion Recognition
HeLo predicts a distribution over emotions instead of one label, fusing ECG and video via cross-attention, entropic optimal transport and a label-correlation loss.
Modern Methods in Associative Memory
Hopfield networks store only 0.14N patterns; dense variants reach polynomial capacity and modern ones exponential, with a Lagrangian view linking them to attention.
QuEst: Blending Data and Predictions for Robust Quantile Estimation
QuEst blends a small gold sample with large model-generated imputations to estimate quantiles and CVaR, canceling simulator bias and minimizing sampling variance.
RetrySQL: Self-Correcting Query Generation
RetrySQL injects deliberate wrong reasoning steps and a BACK token into text-to-SQL pretraining, lifting OpenCoder 1.5B execution accuracy by 4 points toward GPT-4o.
How Modern Information Theory Helps Diagnose Mental Disorders – MvHo‑IB in Action
MvHo-IB diagnoses brain disorders from rs-fMRI by adding triplet O-information to pairwise connectivity, gaining up to 7% accuracy on ABIDE, ADHD-200 and COBRE.
Multi-level Stepwise Hints in Reinforcement Learning
StepHint splits reasoning chains by end-of-step token probability, then feeds RL agents hints at several levels, easing near-miss errors and stalled exploration.
How to Predict Scooter Demand? XGBoost and Urban Micromobility
XGBoost on 341 temporal, spatial and network features predicts e-scooter demand 27-49% better than rivals and up to 90% better than Prophet, node centrality on top.
Ghost Nodes: A Trick That Makes Neural Networks Learn Smarter
Ghost nodes add fake output classes: softmax runs over 13 outputs while loss counts only the 10 real ones, giving gradients room to escape flat regions early.
Does artificial intelligence really understand math? Let's find out what it says... data audit?
A neural-mechanistic hybrid forecasts epidemics from differentially private mobility and spending data, still beating public-health-only models under strong noise.
Unbreakable in the Face of Adversity: ARMOR – Resilient UAV Control
ARMOR trains a teacher VAE on privileged attack data and an LSTM student on raw sensor history, giving UAVs zero-shot robustness to unseen GPS spoofing and jamming.
Mind2Web 2: A new era of “agent-based” web search
Mind2Web 2 grades agentic web search on 130 tasks with an Agent-as-a-Judge rubric; the best of nine systems reaches 50-70% of human quality at twice the speed.
A Machine That Discovers the Laws of Physics: How H-FEX Works and Why It Matters
H-FEX searches symbolic expressions for a system's Hamiltonian, recovering exp(-1.0003p² - 1.1002q⁴) almost exactly and beating SINDy on long-term energy drift.
When the Bandit Is Stronger Than Your Model – On the Limits of Exploratory Learning
Bandit learning has no VC-dimension analogue: a class can need only 2 queries yet stay intractable unless RP = NP. With a Rust epsilon-greedy A/B/n example.
Nothing matches those filters.