The Archive

UDM-GRPO: first framework integrating Uniform Discrete Diffusion Models with GRPO RL; addresses training instability in discrete generative modeling.

AI tracks account for a small fraction of Deezer streams, and most are demonetized for fraud.

Mechanistic analysis of three jailbreak methods (SFT, RLVR, abliteration) on open-weight LLMs; reveals divergent behavioral and internal failure modes.

MASS-RAG: multi-agent synthesis for RAG with role-specialized agents for summarization, extraction, and reasoning over noisy/incomplete contexts.

ArXivDoc benchmark shows document-as-image embeddings underperform for scientific retrieval; proposes text/table/figure-structured multimodal approach.

AI tools are significantly accelerating software development and changing how developers work with code. These tools serve as real-time copilots, automating... AI tools are significantly accelerating software development and changing how developers work with code. These tools serve as real-time copilots, automating repetitive tasks, executing tasks, writing documentation, and more. OpenAI Codex, for example, is a coding agent designed to assist developers through tasks like code generation, debugging, and automated pull request (PR) creation. Source

Following last year's AI-powered Darth Vader in Fortnite that swore in a re-creation of James Earl Jones' voice, Epic Games is now letting Fortnite creators experiment with a new "conversations" tool to create AI-powered characters that players can talk and interact with. "Instead of authoring dialogue trees for characters in your islands, conversations transforms an NPC into an AI-powered character capable of unscripted dialogue and interactions with players, like a quest giver or narrator," Epic says. "You define who the character is with simple prompts-how they think, what they know, and h...

Deep Operator Networks learn Riccati solution operators for finite-horizon LQR; replaces repeated numerical solving with learned surrogate.

Claude Code used to reconstruct corrupted data across 5 HDDs and infer folder structures on NAS without manual human labor.

User claims Claude Design generated a single-shot OS design without AI artifacts; anecdotal praise with screenshot link.

CUTS decoding strategy prevents mode collapse in GRPO on saturated reasoning benchmarks by enforcing structure-preserving exploration.

Multi-objective neural forecasting framework enforcing non-crossing prediction intervals via adaptive weight selection.

AI surrogate models trained on expert-validated CFD data accelerate race-car aerodynamic design exploration.

LQM hierarchical taxonomy diagnoses MT errors in diglossic languages beyond surface-form mistakes.

DPO-based alignment framework instills musical constraints in LLMs for lyric-to-melody generation without annotation.

Adversarial Humanities Benchmark tests robustness of frontier model safety refusals against stylistic prompt transformations.

OneVL compresses vision-language chain-of-thought reasoning into latent states for real-time autonomous driving.

Adaptive Conformal Filtering combines learned safety filters with conformal inference for high-dimensional control systems.

Pedagogical walkthrough of Physics-Informed Neural Network training cycle with explicit algebra and backpropagation.

WorldDB ontology-aware vector graph-of-worlds memory engine enables persistent long-horizon agent reasoning with typed edges.

MS-RCGR reversible multi-scale encoding transforms biological sequences into interpretable scale-invariant geometric features.

BAR framework trains domain-specific experts separately then composes via Mixture-of-Experts, avoiding catastrophic forgetting in multi-domain post-training.

NI Sampling optimizes token ordering in discrete diffusion LLMs to accelerate parallel decoding beyond heuristic step-selection baselines.

Generalized Synthetic Control Method improves baseline estimation for electricity demand response using causal inference; domain-specific, not core AI research.

Asset Harvester extracts 3D object assets from driving logs for closed-loop AV simulation, enabling agent manipulation and novel-view synthesis.

ConGA-PePPI and TC-PepGen integrate peptide-protein interaction prediction with target-conditioned generation; biotech application, not core LLM/AI architecture.

Extends Semantic Tube Prediction by sampling at reasoning step boundaries to improve geometric structure and efficiency in LLM fine-tuning.

DESPITE benchmark (12K tasks) reveals LLMs as robotic planners generate dangerous plans 28.3% of the time despite strong planning ability; scale ≠ safety.

Noetik applies autoregressive transformers (TARIO-2) to match cancer treatments with patients, targeting 95% clinical trial failure rate.

CmIR framework uses causal inference to disentangle invariant vs. spurious modality representations, improving multimodal robustness under distribution shift.