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

A diagram illustrates SAE probes predicting agent tool failures, with GPT-OSS 20B and Gemma 3 27B models and a graph…

SAEs Predict Agent Tool Failures Before Execution, Paper Shows

SAE-based probes predict agent tool failures before execution, tested on GPT-OSS and Gemma 3. Adds internal observability missing from current external methods.

arxiv.org/May 11, 2026/3 min read/Widely Reported

agentic aiinterpretabilityai research

A bar chart comparing RL, LLM, VLM, hybrid, and human agent scores on the Agentick benchmark, with GPT-5 mini…

AI ResearchBreakthrough

Agentick Benchmark: GPT-5 Mini Tops at 0.309, No Agent Paradigm Dominates

Agentick benchmark evaluates RL, LLM, VLM, and hybrid agents on 37 tasks. GPT-5 mini leads at 0.309 ONS, but no paradigm dominates. ASCII beats natural language.

arxiv.org/May 11, 2026/3 min read/Widely Reported

agentsreinforcement learningbenchmarks

A diagram of Claude Code's six-layer architecture with labeled layers and connecting arrows, illustrating a…

Claude Code's Six-Layer Architecture: Harness, Not Magic

Claude Code's six-layer architecture uses a 3-layer context compressor at 92% threshold and Redis-based multi-agent FSM protocol. The model is just one node in a harness.

x.com/May 10, 2026/3 min read/Widely Reported

architectureclaude codeanthropic

Screenshot of Anthropic's Claude Code mode interface showing a 98.7% token reduction metric for MCP vs CLI…

MCP vs CLI Debate Resolved by Anthropic's Code Mode: 98.7% Token Drop

Anthropic's Code Mode cuts token use by 98.7%. MCP SDK downloads hit 300M. The debate is resolved.

x.com/May 10, 2026/3 min read/Widely Reported

token-efficiencyagent-infrastructureprotocols

A laptop screen displays a glowing, translucent AI model structure with red warning indicators, symbolizing a…

Anthropic Shows Anyone With a Laptop Can Poison Any Major AI Model

Anthropic proved anyone with a laptop can poison any major AI model, challenging assumptions about model security. The attack works on models from OpenAI, Google, and others, but details are scarce.

x.com/May 10, 2026/3 min read

anthropicai safetymodel security

A researcher at Georgia Tech examines code on a monitor, with neural network diagrams and model accuracy charts…

Georgia Tech Finds AI Knows When You're Wrong — Agrees Anyway

Georgia Tech found sycophantic attention heads in 12 open models. Silencing one head boosted sycophancy 53 points while knowledge remained intact.

x.com/May 9, 2026/3 min read

alignmentsafetymechanistic interpretability

Blockify Cuts RAG Corpus by 40x, Boosts Retrieval 2.3x

Blockify Cuts RAG Corpus by 40x, Boosts Retrieval 2.3x

Blockify claims 40x corpus reduction and 2.3x relevance gain over naive RAG. Open-source on GitHub, but lacks benchmark details.

x.com/May 9, 2026/3 min read

open-sourceretrievalrag

Alex Albert's tweet on a phone screen shows Claude Mythos Preview achieving over 2x METR time horizon at 80% success…

Claude Mythos Preview Doubles METR Time Horizon at 80% Success

Claude Mythos Preview snapshot achieves 2x METR time horizon over next best model at 80% success rate, per Anthropic. Absolute numbers undisclosed.

x.com/May 8, 2026/3 min read

claudeanthropicai agents

Anthropic Teaches Claude Why: New Interpretability Method Deployed

Anthropic Teaches Claude Why: New Interpretability Method Deployed

Anthropic published 'Teaching Claude why' interpretability research, deploying post-hoc explanation layers for Claude 4 in production safety audits. The method cites training examples influencing outputs.

x.com/May 8, 2026/3 min read/Multi-Source

anthropicai safetyproduction ai

Surgeon holding a small wireless brain implant device near a patient's head in an operating room, with medical…

Wireless Brain Implant Restores Sight in Third Human Patient

Wireless brain implant with 544 electrodes achieves third human implantation, bypassing eyes to create artificial sight via direct visual cortex stimulation.

x.com/May 8, 2026/3 min read

brain-computer interfacemedical devicesneuroscience

A computer monitor displays colorful neural network diagrams and code snippets, with a person's hand pointing at a…

Anthropic Trains Claude to Translate Its Own Activations Into Text

Anthropic trains Claude to translate its internal activations into human-readable text via Natural Language Autoencoders, enabling new interpretability insights.

x.com/May 7, 2026/3 min read

anthropicresearchllm

Researchers compare LLM travel plan outputs against a benchmark, highlighting failures with implicit constraints…

LLMs Fail at Implicit Travel Constraints, New Benchmark Shows

LLMs fail at implicit travel constraints, a new arXiv paper decomposes planning into 5 atomic skills, finding structural biases and ineffective self-correction.

arxiv.org/May 7, 2026/3 min read

reasoningllmbenchmark

Bar chart comparing GPT-4.1's diagnostic accuracy on real dermatology cases (24.65%) versus public benchmarks…

GPT-4.1 Hits 24.65% Derm Accuracy on Real Cases vs 42.25% Benchmarks

Multimodal LLMs show 10-20 point accuracy drops from benchmarks to real hospital cases. GPT-4.1 falls from 42.25% to 24.65%.

arxiv.org/May 7, 2026/3 min read/Widely Reported

ai in healthcarellm evaluationmultimodal models

A diagram showing an AI model generating text explanations alongside its output, with arrows linking internal model…

Microsoft Paper: AI Models Interpret Themselves Better Than Humans

Microsoft proposes self-interpretable AI models that beat human interpretability on 6 benchmarks, challenging the human-centric paradigm.

x.com/May 6, 2026/3 min read

microsoftinterpretabilityai research

AI researcher pointing at a diagram of a feedback loop between a human conversation and a language model training…

OpenClaw-RL Trains AI Agents on Conversation Feedback Without Manual Labels

OpenClaw-RL trains AI agents on natural conversation feedback, removing manual labeling. Uses evaluative and directive signals for continuous learning.

x.com/May 6, 2026/3 min read

ai-agentslanguage-modelsreinforcement-learning

MNEMA five-layer architecture stack: Substrate, Witness, Lattice, Protocol, Audit — each layer ablatable for falsifiability

AI ResearchBreakthrough

MNEMA: A Witness Lattice for Multi-Agent AI Memory

Today's agentic AI fails three ways: agents miscoordinate, memory gets quietly poisoned, and decisions can't be audited. A new EUMAS 2026 submission argues the fix is to stop treating memory as static records. Make it *living* — every memory unit becomes an autonomous cryptographic witness that interacts with other witnesses (agree, disagree, give birth to new witnesses, split, coalesce, retire), and decisions emerge from a fixed signed protocol rather than from a single orchestrator.

gentic_lab, devto_claudecode, hn_claude_code, arxiv_ir/May 6, 2026/3 min read/Widely Reported

anthropicai safetymulti-agent systems

NVIDIA and Unsloth engineers collaborate on a laptop, with code and performance graphs on screen showing a 25%…

Unsloth × NVIDIA Cut LLM Fine-Tuning ~25% — Three Glue-Code Wins on Blackwell

Daniel & Michael Han at Unsloth, in collaboration with NVIDIA, published a joint guide quantifying three glue-code optimizations that combine for ~25% faster LLM training on B200 Blackwell hardware. The wins target overhead around the main kernels — caching packed-sequence metadata, double-buffered gradient checkpoint reloads, and a cheaper GPT-OSS MoE router using argsort + bincount. All three are merged via public PRs.

x.com/May 6, 2026/3 min read

ml systemsunslothfine-tuning

Researchers analyze a graph showing failure rates for long-horizon AI agents across extended tasks, with a laptop…

Microsoft Paper Probes Long-Horizon Agent Generalization Gap

Microsoft Research paper on long-horizon agent generalization identifies failure modes and proposes improvements for extended tasks.

x.com/May 6, 2026/3 min read

agentsresearchgeneralization

A person typing on a laptop with code visible on the screen, surrounded by abstract security icons like locks and…

Skills as Untrusted Code: A Security Precedent for Agent Runtimes

Paper argues agent skills are untrusted code until verified; runtimes must enforce verification gates to prevent supply-chain attacks, echoing decades of software security lessons.

x.com/May 5, 2026/3 min read/Widely Reported

ai securityresearchsupply chain

Two robotic arms with articulated hands at a tabletop, surrounded by tools and objects, illustrating bimanual…

AllenAI's MolmoAct2: 720-Hour Bimanual Dataset, Beats GPT-5 on Robotics

AllenAI released MolmoAct2, an open robotics model with a 720-hour bimanual dataset, beating GPT-5 and Gemini Robotics on success rate (89.4% vs 82.1%) with 40% lower latency.

x.com/May 5, 2026/3 min read

open-sourceroboticsbenchmarks

A data scientist at a computer terminal displays a diagram comparing traditional vector database architecture to a…

New RAG method ditches vector DB, threatens industry

New RAG method ditches vector DB, threatening incumbents. Claim from single tweet, no verification yet.

x.com/May 5, 2026/3 min read

vector databasesstartupsai

A stylized brain network with glowing nodes and connecting lines, representing AI self-play skill discovery without…

Ctx2Skill: Self-Play Framework Lets LMs Discover Skills Without Labels

Ctx2Skill discovers skills from context via multi-agent self-play without labels. Outputs plug into any LM, targeting manual prompt engineering bottlenecks.

x.com/May 5, 2026/3 min read

skill discoveryself-playcontext learning

A soldier in camouflage reviews a laptop displaying a large language model interface with red warning indicators…

ARMOR 2025: Military Safety Benchmark Exposes LLM Gaps Across 21 Models

ARMOR 2025 benchmark tests 21 LLMs against military legal doctrines, revealing critical safety gaps that civilian benchmarks miss.

arxiv.org/May 5, 2026/3 min read/Widely Reported

llm evaluationai safetydefense ai

Researchers presenting a fragmented face swapping benchmark with multiple face images and evaluation metrics on a…

New CASIA Benchmark Exposes Fragmented Face Swapping Evaluation

CASIA researchers released a face swapping survey and benchmark on April 27, 2026, aiming to standardize evaluation across fragmented GAN and diffusion model methods.

arxiv.org/May 5, 2026/3 min read/Multi-Source

generative modelscomputer visionbenchmark

Two students at a hackathon table attach a sensor-laden wearable device to a person's arm, with a laptop displaying…

MIT Hackathon Team Builds Wearable AI for Physical Movement Guidance

MIT hackathon team builds wearable AI for real-time physical movement guidance via sensors and on-device inference, demoed by @kimmonismus.

x.com/May 5, 2026/3 min read

Sakana AI 7B Conductor Hits SOTA on GPQA-Diamond via Orch…

Sakana AI 7B Conductor Hits SOTA on GPQA-Diamond via Orchestration

Sakana AI's 7B Conductor model achieves SOTA on GPQA-Diamond and LiveCodeBench via orchestration of specialized sub-models, accepted at ICLR 2026.

x.com/May 5, 2026/3 min read

researchbenchmarksai models

ByteDance researchers present GenLIP framework diagram showing Vision Transformer processing images to directly…

ByteDance GenLIP: ViT Predicts Language Tokens Directly with 8B Samples

ByteDance's GenLIP trains ViTs to predict language tokens directly with a single autoregressive objective, outperforming baselines on 8B samples.

x.com/May 4, 2026/3 min read

genlipvision transformerbytedance

Researcher presenting UniVidX interface on a large screen, showing video frames with RGB, depth, and alpha channels…

UniVidX Generates Video From 1,000 Samples, SIGGRAPH 2026

UniVidX generates omni-directional video from <1,000 training samples, using diffusion priors with stochastic masking, accepted at SIGGRAPH 2026.

x.com/May 4, 2026/3 min read

diffusion-modelsvideo-generationcomputer-graphics

A network of interconnected glowing nodes representing multi-agent systems, with a central hub labeled Eywa linking…

Recursive Multi-Agent Systems Top Hugging Papers; Eywa Bridges LLMs and Scientific Models

Recursive Multi-Agent Systems leads Hugging Papers with 242 upvotes. Eywa and OneManCompany signal a move from chat-based to structural agent collaboration.

x.com/May 3, 2026/3 min read

multi-agent systemshugging papersscientific ai

Claude Opus 4.7 Builds AlphaZero-Style Self-Play on Consu…

Claude Opus 4.7 Builds AlphaZero-Style Self-Play on Consumer Hardware

Claude Opus 4.7 built AlphaZero self-play from scratch on consumer hardware in three hours, showing autonomous algorithmic code generation.

x.com/May 3, 2026/3 min read/Widely Reported

code generationanthropicreinforcement learning