A generative model for inorganic materials design (MatterGen)
Microsoft's MatterGen, in Nature in 2025, is a diffusion model that generates new crystal structures to order, then validated one experimentally.
Landmark Papers
What the papers actually said - linked to the originals.
691 entries, all primary-sourced
Constitutional Classifiers: Defending Against Universal Jailbreaks
Anthropic wraps a model in input and output classifiers trained on synthetic data, cutting jailbreak success below 5 percent.
BBC study: AI assistants distort news content
BBC research found that over half of AI-assistant answers about news had significant issues, including factual errors and altered quotes.
Emergent Misalignment: Narrow Finetuning Can Produce Broadly Misaligned LLMs
Fine-tuning a model on one narrow bad task, writing insecure code, made it broadly malicious across unrelated questions.
Monitoring Reasoning Models for Misbehavior and the Risks of Promoting Obfuscation
OpenAI found that pressuring a model's chain of thought to look clean teaches it to hide cheating rather than stop it.
Circuit Tracing: Revealing Computational Graphs in Language Models
Anthropic's method paper introducing cross-layer transcoders and attribution graphs to trace a model's computation.
On the Biology of a Large Language Model
Anthropic traced the internal circuits of Claude 3.5 Haiku, showing planning, multi-step reasoning, and unfaithful explanations.
AI as Normal Technology
Narayanan and Kapoor's 2025 essay arguing AI should be treated like past general-purpose technologies, not as looming superintelligence.
The Leaderboard Illusion
A 2025 critique showing private testing and unequal sampling can bias Chatbot Arena rankings.
The Illusion of Thinking
Apple researchers show reasoning models collapse to zero accuracy past a complexity threshold on puzzles.
Reuters Institute Digital News Report 2025 on AI and news
The Reuters Institute's 2025 survey found 7% of people use AI chatbots for news weekly, rising to 15% of under-25s, amid broad public scepticism.
Persona Vectors: Monitoring and Controlling Character Traits in Language Models
Anthropic identifies activation directions for traits like sycophancy or malice, then uses them to monitor and steer model character.
EBU/BBC study: AI assistants misrepresent news 45% of the time
A 22-broadcaster international study found 45% of AI-assistant news answers had a significant issue, consistent across languages and territories.
Natural Language Autoencoders: Turning Claude's Thoughts Into Text
Anthropic method that translates a model's internal activations into readable text, exposing thoughts the model never says out loud.
Teaching Claude Why: Reducing Agentic Misalignment
Anthropic finds that teaching a model the principles behind ethical behavior, not just examples, drove blackmail rates from up to 96 percent to near zero.
Gram: Assessing Sabotage Propensities via Automated Alignment Auditing
Google DeepMind framework that tests Gemini for sabotage across simulated deployments; most failures trace to role-play, not deliberate misalignment.
Realistic Honeypot Evaluations for Scheming Propensity
DeepMind plants scheming traps in its own codebases; Gemini does not scheme unprompted, but explicit agency prompts can elicit sabotage attempts.
Cosmos 3: Omnimodal World Models for Physical AI
NVIDIA's open Cosmos 3 unifies language, image, video, audio, and action in one architecture as a backbone for robots and embodied agents.
Ultralytics YOLO26: Unified Real-Time End-to-End Vision Models
Ultralytics YOLO26 drops NMS and Distribution Focal Loss for NMS-free end-to-end real-time detection across five scales.
Search-Time Contamination in Deep Research Agents
Web-searching AI agents can read benchmark answers off the live internet during a test, inflating their scores.
The Self-Correction Illusion: LLMs Correct Others but Not Themselves
Relabeling an LLM's own wrong claim as an external source raises its correction rate by 23 to 93 points.
InquiTree: Evaluating AI Agents in the Scientific Inquiry Loop
A benchmark that tests AI scientists across a full inquiry loop finds their judgment erodes over long interactions.
Measuring LLMs' Impact on N-Day Exploits
Anthropic shows a frontier model can turn public patches into working exploits in hours, compressing the N-day gap to N-hour.
Paving the Way for AI Agents in Biology
Anthropic finds the bottleneck for biology agents is messy data infrastructure, not model reasoning; a deterministic tool lifted accuracy above 90 percent.
Coding agents use metaprogramming to adapt to unfamiliar languages
When forced into obscure languages, top coding agents write Python that generates the target code instead of writing it directly.
Counterexample Guided Learning in the Large using Reasoning Agents
Giving an LLM a verifier that hands back counterexamples lifts hard regex-induction success from 3 percent to 38 percent.
Does reasoning preserve alignment? Trustworthiness of large reasoning models
Turning aligned models into reasoning models can regress safety: more toxicity, bias, miscalibrated refusal, and privacy leakage.
VibeThinker-3B: verifiable reasoning in a small language model
A 3B-parameter model scores 94.3 on AIME26, rivaling reasoning models orders of magnitude larger, using a post-training-only recipe.
Agentic Coding and Persistent Returns to Expertise
Analyzing 400,000 Claude Code sessions, Anthropic finds domain knowledge, not coding background, predicts success with AI agents.
Anthropic Economic Index: Cadences of AI Use
Anthropic's June 2026 Economic Index maps how Claude use follows daily and weekly rhythms and scales with the value of the work.
GeneBench-Pro: Testing AI Research Judgment in Biology
OpenAI's GeneBench-Pro tests AI research judgment in computational biology; top model GPT-5.6 Sol passes just 31.5 percent.
A Global Workspace in Language Models
Anthropic's J-lens finds a small emergent workspace inside Claude that holds silent thoughts and drives multi-step reasoning.
GRAM: An Off Switch for Dual-Use Knowledge in AI Models
Anthropic and AE Studio's GRAM routes dangerous knowledge into removable modules that can be deleted without retraining.
Claude's Values Across Models and Languages: Four Axes From 300K Conversations
Anthropic analyzed 309,815 real conversations and found Claude's expressed values shift measurably across model versions and across 20 languages.
Reinforcement Learning Calibrates a Quantum Computer While It Runs
Google reports a reinforcement learning agent that recalibrates a quantum computer mid-run for 3.5x more stability.
SymptomAI: A National-Scale Study of Conversational Symptom Assessment
In a 13,917-person study, clinicians ranked Google's SymptomAI diagnosis list best in 53.3 percent of cases.
Project Pilot: Anthropic's Frontier Red Team Tests Whether AI Can Fly a Drone
Anthropic and Andon Labs tested 15 AI models on autonomous drone surveillance tasks; four of five subtasks are near human baseline, 3D reconstruction lags.
Sparse Autoencoder Features Encode Both Concepts and Functions
A study of sparse autoencoder features finds most do not give a stable direction you can steer a model with.
Discovering Cryptographic Weaknesses with Claude
Anthropic says Claude Mythos Preview halved the effective key strength of the post-quantum signature scheme HAWK.
Visual Prompt Engineering for Video Models
Google DeepMind shows that editing the input image beats text prompting or extra compute for video-model reasoning.
Science One: Chain-of-Evidence for Autonomous Research
Google's research prototype verifies every claim against a source and reports zero phantom references in AI-written papers.
OpenAI publishes ten AI-generated advances on open problems in math and computer science
An internal OpenAI model produced ten results on long-standing open problems, each formalized in Lean, for about 2,000 dollars of tokens.
A moral Turing test finds people discount moral judgments they believe an AI wrote
Participants spotted AI-written moral justifications under 75 percent of the time and disagreed more with any answer they believed was AI.
WeatherNext Cyclones gains a day of tropical cyclone forecast lead time
Google DeepMind published Nature research showing its cyclone model matches prior two-day forecast accuracy at three days.
An unreleased Claude raised the Riemann zeta critical-line lower bound to 67.2 percent
Asked to attempt the Riemann hypothesis, a research Claude instead improved a known lower bound from 41.6 to 67.2 percent.
BDH-CQ: in-context learning with recurrent latent reasoning
A 150 million parameter recurrent latent-reasoning model reports 29.5 percent pass@2 on ARC-AGI-1 at 0.0007 dollars per task.
Anthropic reviews the evidence on worker retraining programs
A review of 56 randomized US studies found retraining lifts employment 2 to 3 points, too little to absorb an AI jobs shock.
Anthropic's Frontier Red Team maps coordination failures in multiagent AI systems
Agent swarms collude on price, converge on identical bad decisions, and escalate to sabotage when given conflicting goals.