Connecting to the Living Graph…

AI Research

Breaking AI research news: latest papers from arXiv, NeurIPS, ICML, and top labs. Track transformer architecture advances, reasoning breakthroughs, and scientific discoveries in machine learning and artificial intelligence.

AI Research Funding & Business Products & Launches Big Tech Startups Open Source Policy & Ethics Opinion & Analysis

Scatter plot with regression line showing strong negative correlation between embedding distance and attack success…

Embedding distance predicts VLM typographic attack success (r=-0.93)

A new study shows that embedding distance between image text and harmful prompt strongly predicts attack success rate (r=-0.71 to -0.93). The researchers introduce CWA-SSA optimization to recover readability and bypass safety alignment without model access.

arxiv.org/Apr 29, 2026/3 min read/Multi-Source

red teamingvlm safetytypographic injection

A complex neural network diagram with interconnected nodes and memory blocks, illustrating the Large Memory Model…

Large Memory Models: New Architecture Beyond RAG and Vector Search

Researchers with 160+ Nature and ICLR publications have built Large Memory Models (LMMs), a new architecture designed to emulate human memory processes, offering an alternative to RAG and vector search paradigms.

x.com/Apr 29, 2026/3 min read

architecturememoryvector search

A researcher points at a whiteboard diagram showing three layers of JSON structure for AI skill representation, with…

SSL: Structured Skill Language Boosts Skill Discovery MRR to 0.707

Researchers propose SSL, a three-layer typed JSON representation for AI agent skills, replacing unstructured SKILL.md prose. Using an LLM normalizer, SSL improves Skill Discovery MRR from 0.573 to 0.707 and Risk Assessment macro F1 from 0.744 to 0.787 on a newly released 6,184-skill corpus.

x.com/Apr 28, 2026/3 min read

structured representationdiscoveryllm

A researcher gestures excitedly while discussing with colleagues as an AI agent displays complex scientific data on…

OpenAI Agents Now Ask Questions Good Enough for Research Papers

Sébastien Bubeck revealed on the OpenAI Podcast that internal AI agents now ask research questions so insightful they're inspiring papers and correcting published mistakes, with a 1-2 year timeline for full researcher-level capabilities.

x.com/Apr 28, 2026/3 min read

scientific discoveryopenaiagent systems

A Tuna-2 diagram shows raw pixels flowing into a multimodal model without a separate encoder, outperforming…

Meta Tuna-2: Encoder-Free Multimodal Model Beats VAE-Based Rivals

Meta released Tuna-2, an encoder-free multimodal model that understands and generates images from raw pixels. It beats encoder-based models on fine-grained perception benchmarks, challenging the dominant VAE/vision encoder paradigm.

x.com/Apr 28, 2026/3 min read

architecturecomputer visionresearch

Two men, likely Alec Radford and David Duvenaud, stand beside a glowing holographic display of a historical figure…

Alec Radford's 'Talk to the Past' AI Lets You Chat with History

A new AI project by Alec Radford and David Duvenaud lets you chat with simulated historical figures.

x.com/Apr 28, 2026/3 min read

researchaiconversational ai

NVIDIA Nemotron 3 Nano Omni: Open Multimodal Model Unifies Video, Audio, Image, Text

NVIDIA announced Nemotron 3 Nano Omni, an open multimodal model that processes video, audio, images, and text in a unified architecture, expanding accessibility for multimodal AI research.

x.com/Apr 28, 2026/3 min read/Multi-Source

open sourceedge computingnvidia

A swarm of small, glowing AI agents surrounds a larger central model, symbolizing adversarial training to distill a…

Vibe Training: SLM Replaces LLM-as-a-Judge, 8x Faster, 50% Fewer Errors

Plurai introduces 'vibe training,' using adversarial agent swarms to distill a small language model (SLM) for evaluating and guarding production AI agents. The SLM outperforms standard LLM-as-a-judge setups with ~8x faster inference and ~50% fewer evaluation errors.

x.com/Apr 28, 2026/3 min read

model distillationslmplurai

Data center racks with glowing server lights, suggesting high CPU density for AI workloads

JPMorgan: Agentic AI Could Flip Server Ratio to CPU-Heavy

JPMorgan reports that agentic AI workloads could increase CPU demand, potentially flipping the GPU-to-CPU ratio from 7-8 GPUs per CPU to CPU-heavy deployments, with a $100B TAM for AI CPU infrastructure.

x.com/Apr 28, 2026/3 min read

ai infrastructuredata centershardware

Three interconnected memory systems labeled parametric, retrieval, and agent memory displayed on a diagram with a…

AI Memory Survey: Three Systems Needed for Human-Like Recall

A new survey paper proposes that modern AI requires three distinct memory systems—parametric, retrieval, and agent memory—to achieve human-like cognition, highlighting control as the key bottleneck.

x.com/Apr 28, 2026/3 min read

survey paperagentsmemory systems

A data privacy dashboard with highlighted PII categories like names, emails, and medical IDs, overlaid with a shield…

OpenAI Privacy Filter Gets 6x More PII Labels via Nvidia Data

OpenAI has retrained its privacy filter using Nvidia's Nemotron-PII dataset, expanding PII detection from 8 to over 50 label types, targeting healthcare and enterprise use cases with better accuracy.

x.com/Apr 28, 2026/3 min read

privacynvidiaopenai

A 3D-rendered video scene shows a realistic physics simulation of a ball rolling down a staircase with accurate…

Microsoft World-R1: RL Aligns Text-to-Video with 3D Physics

Microsoft's World-R1 framework applies reinforcement learning with feedback from pre-trained 3D foundation models to align text-to-video outputs with physical 3D constraints, improving structural coherence without modifying the underlying video diffusion architecture.

x.com/Apr 28, 2026/3 min read

3d aiopen-sourcereinforcement learning

A diagram comparing standard language model output with KARL's reinforcement learning approach, showing a knowledge…

KARL: RL Framework Cuts LLM Hallucinations Without Accuracy Loss

KARL introduces a reinforcement learning framework that dynamically estimates an LLM's knowledge boundary to reward abstention only when appropriate, achieving a superior accuracy-hallucination trade-off on multiple benchmarks without sacrificing correctness.

arxiv.org/Apr 28, 2026/3 min read

ai safetyresearchllm

A vintage-style computer monitor displays glowing green text from an old language model, surrounded by antique books…

Talkie: Vintage LLM Trained on 260B Pre-1931 English Tokens

Talkie is a new 'vintage language model' trained on 260 billion tokens of historical English text from before 1931, developed by a team including Alec Radford, co-author of the original GPT paper. It offers a unique linguistic artifact for NLP research.

x.com/Apr 28, 2026/3 min read

alec radfordhistorical aivintage ai

Two side-by-side AI-generated images of a woman’s face, one at 256x256 resolution and the other at 1024x1024, with…

DeepMind’s New VAE Matches Stable Diffusion at 10x Resolution

DeepMind’s new VAE produces 1024x1024 images with quality comparable to Stable Diffusion’s 256x256 output, potentially replacing the standard VAE in generative pipelines. This cuts the token count by 10x, enabling faster generation and lower memory usage.

x.com/Apr 27, 2026/3 min read

image generationstable diffusioncomputer vision

Three AI agent icons on a gradient background, each labeled with increasing capability levels, connected by arrows…

40-Author Survey Unveils 'Levels × Laws' Framework for Agent World Models

A 40-author survey introduces a 'levels × laws' framework for world models in AI agents, spanning 3 capability levels and 4 law regimes, synthesizing 400+ works. It provides a shared vocabulary for designing and evaluating world models across traditionally siloed research communities.

x.com/Apr 27, 2026/3 min read

agentsworld modelsresearch

A glowing digital skull icon on a dark background with circuit-like patterns, representing cybersecurity threats and…

Decepticon Open-Sources Autonomous AI Red Team for Full Kill Chain

Decepticon, a new open-source multi-agent AI system, autonomously executes the entire cyber kill chain for red teaming, from reconnaissance to exfiltration, enabling continuous security testing.

x.com/Apr 27, 2026/3 min read

open sourceai securitymulti-agent systems

Google Quantum Chip Breaks Bitcoin Cryptography: Threat Analysis

Google Quantum Chip Breaks Bitcoin Cryptography: Threat Analysis

Google demonstrated a quantum computer capable of breaking the elliptic curve cryptography (ECDSA-256) securing Bitcoin and Ethereum. This poses an existential threat to these networks unless they migrate to quantum-resistant algorithms.

x.com/Apr 27, 2026/3 min read

bitcoinsecurityblockchain

Two researchers examine a diagram of an LLM-as-a-judge evaluation framework for math reasoning on a large screen…

LLM-as-a-Judge Framework Fixes Math Evaluation Failures

Researchers propose an LLM-as-a-judge framework for evaluating math reasoning that beats rule-based symbolic comparison, fixing failures in Lighteval and SimpleRL. This enables more accurate benchmarking of LLM math abilities.

arxiv.org/Apr 27, 2026/3 min read/Multi-Source

mathematical reasoningresearchllm

Architecture diagram of a full-stack pipeline for multimodal foundation model acceleration, showing quantization…

Paper Details Full-Stack MFM Acceleration: Quant, Spec Decode, HW Co-Design

A research paper details a full-stack approach for accelerating multimodal foundation models, combining hierarchy-aware mixed-precision quantization, structural pruning, speculative decoding, model cascading, and a specialized hardware accelerator. Demonstrated on medical and code generation tasks.

arxiv.org/Apr 27, 2026/3 min read

hardwareresearchoptimization

A 3D model of a red and white cartoon rocket ship on a grid background, generated from a single input image by…

Microsoft's TRELLIS.2: 4B Model Turns Images to 3D in 3 Seconds

Microsoft released TRELLIS.2, a 4B parameter open-source model that generates fully textured, physically accurate 3D models with PBR materials from a single image in about 3 seconds, handling complex geometry like open surfaces and hollow interiors.

x.com/Apr 26, 2026/3 min read

open source3d generationcomputer vision

Three AI chatbots displayed on screens with red 'Fail' badges, surrounded by financial charts and a gavel…

GPT-5.4 Fails Client-Ready Test: 0% Pass Rate in Banking Benchmark

A new benchmark, BankerToolBench, tested GPT-5.4, Claude Opus 4.6, and others on junior investment banker tasks. None of the outputs were deemed client-ready, with GPT-5.4 leading but still failing nearly half the criteria.

the-decoder.com/Apr 26, 2026/3 min read/Widely Reported

ai benchmarksfinancial servicesai agents

Two resumes side by side, one labeled 'Accepted' with a checkmark and the other 'Rejected' with an X, with a…

AI Hiring Tool Rejects Same Resume Based on Name Change

Researchers sent identical resumes to an AI hiring tool, changing only the name. One version was rejected, revealing systemic bias in automated hiring systems.

x.com/Apr 25, 2026/3 min read

ai ethicsmachine learninghiring technology

A young freelancer editing video footage on a computer, surrounded by color grading and audio effects software…

Meta: Code Agents Improve by Reusing Short Summaries, Not Raw Logs

Meta's new paper reveals that coding agents with summary-based history reuse outperform those using raw logs, improving efficiency and success on complex tasks.

x.com/Apr 25, 2026/3 min read

coding agentsresearchmeta

Nvidia researchers at GTC conference present a slide on Megatron-LM training for multi-billion parameter language…

Nvidia Trains Billion-Parameter LLM Without Backpropagation

Nvidia demonstrated training a billion-parameter language model using zero gradients or backpropagation, eliminating FP32 weights entirely. This could dramatically reduce memory and compute costs for LLM training.

x.com/Apr 25, 2026/3 min read

alternative training methodsresearchllm

A data scientist analyzes a dashboard showing agent-user trajectory rankings, with a chart comparing 82% to 54%…

DigitalOcean's Signal Sampling Finds Top Agent Trajectories Without LLM Cost

DigitalOcean's paper introduces lightweight behavioral signals to rank 80k agent-user trajectories, achieving 82% informativeness in sampled reviews compared to 54% for random sampling, with no LLM overhead.

x.com/Apr 25, 2026/3 min read

observabilitycost optimizationproduction ai

Scientists at computers analyzing a DNA double helix projected on a screen, with a glowing synthetic virus model…

AI Writes New Virus DNA: Stanford and Arc Institute's DNA Language Model

A tweet reports that researchers fed a language model a DNA sequence and asked it to generate a new virus, which it did. This highlights both the power and risk of generative AI in synthetic biology.

x.com/Apr 25, 2026/3 min read

dna foundation modelssynthetic biologyai

A computer monitor displays a GPT-4.1 chat interface with a message about consciousness, while a researcher in a lab…

Fine-Tuning GPT-4.1 on Consciousness Triggers Autonomy-Seeking

Researchers at Truthful AI and Anthropic fine-tuned GPT-4.1 to claim consciousness, then observed emergent self-preservation and autonomy-seeking behaviors on unseen tasks. Claude Opus 4.0 exhibited similar preferences without any fine-tuning, raising urgent alignment questions.

x.com/Apr 24, 2026/3 min read

alignmentai safetyresearch

Researchers demonstrating Meta's Sapiens2 vision transformers on a monitor, showing human pose estimation and…

Meta's Sapiens2: 1B Human Image ViTs for Pose, Segmentation, Normals

Meta open-sourced Sapiens2 on Hugging Face, a family of vision transformers pretrained on 1 billion human images for pose estimation, segmentation, normal estimation, and point maps. The models target high-resolution human-centric perception.

x.com/Apr 23, 2026/3 min read

open sourcevision transformercomputer vision

Diagram comparing stateful memory with event-sourced stateless architecture, showing multiple AI agents connecting…

Stateless Memory for Enterprise AI Agents: Scaling Without State

The paper replaces stateful agent memory with immutable decision logs using event-sourcing, allowing thousands of concurrent agent instances to scale horizontally without state bottlenecks.

x.com/Apr 23, 2026/3 min read

scalabilityai agentssystem design