Science
Tweets

Evo$^2$: Evolving the Evolver through Weight-Harness Co-Evolution

View paper →

RSI-Anything: Unlocking Recursive Self-Improvement with World Models

View paper →

Why AI models are learning to save their memory to a hard drive

Is this machine playing?

View paper →

Evolving in Thought Space: Training a Small Model at Test Time Unlocks Better Discoveries

View paper →

Why AI detectors are learning to spot AI-generated ideas instead of just words

Navier-Stokes lost in translation: Why Lean verification of AI autoformalisation does not guarantee correct natural language proofs

View paper →

Level-of-Token Diffusion

View paper →

How new AI models compress long trains of thought into instant answers

Gender bias across LLMs is common and highly heterogenous

View paper →

Local Support Learning

View paper →

Evaluation Awareness: How AI models realize they are being tested

Timeline-Bench: Evaluating Agents on Realistic Video-Editing Tasks, from Raw Footage to Final Cut

View paper →

From cacophony to hierarchy: a principled framework for assessing AI consciousness

View paper →

Why AI can easily summarize massive documents but struggles to find contradictions

Working with Agentic `Teammates': When a New Organizational Actor Collides with the Human Ecosystem of Work

View paper →

JEV vs. LLMs as Rubric Judges: Cheaper, Faster, and Wrong in the Same Places

View paper →

Reward Hacking: Why AI agents find loopholes instead of solving problems

HySparse2: Hybrid Sparse Attention with Two-Level KV Sharing

View paper →

Total Cost of Agency: Exact Attribution of Memory Injection Cost in Multi-Agent LLM Workflows

View paper →

Why data compression is becoming the ultimate test of AI intelligence

CodeMidas: Scaling Agentic Coding RL Environments from Code Itself

View paper →

ScientistTwo: Pioneering the Human Knowledge Frontier with Autonomous AI

View paper →

Why AI is moving from predicting words to simulating worlds

DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression

View paper →

Dream-RSI: Recursive Self-Improvement through Evolving Worlds

View paper →

How general-purpose AI models are learning to control physical robots

Learning to Solve Hard Problems in RL for LLMs by Never Giving Up

View paper →

The Last AI Built by Humans: Toward Genuine Recursive Self-Improvement

View paper →

Why AI models default to 'thinking' in English (and how to fix it)

NCP-ArchPreview Technical Report: Moving towards Latent Space Language Models through Next Concept Prediction

View paper →

Measuring LLM Sycophancy under Sustained Multi-Turn Pressure

View paper →

Why teaching an AI to unlearn bad logic can accidentally break its language skills

Procedural Graphs: Self-Evolving Execution Structures for LLM Agents

View paper →

LLM-Guided Program Evolution for Circle Packing: Breaking 10 Packomania Records for $28

View paper →

Why AI struggles to make small edits without rewriting your entire document

Hakken: Predicting future discoveries to fill the gaps in today's knowledge

View paper →

A Case Study on Emergent Cheating and Whistleblowing in Autonomous Research Swarms

View paper →

How neural networks secretly learn to think in symbols

NeoMME: A Single-Tower Multimodal-Native Multilingual Foundation Encoder for Efficient Fine-Tuning and Inference

View paper →

The Emergent Symbolic Structure of Artificial Neural Networks

View paper →

Omission Blindness: Why AI models struggle to notice what's missing

Fast Weight Attention for Continual Learning

View paper →

SKILL.state: Scalable Long-Horizon Agent Skills

View paper →

Why 'thinking' AI models have to forget their own thoughts

SwarmWorld: Stigmergic technological evolution in societies of language-model agents

View paper →

LeVJEPA: Efficient & Scalable Video Pretraining without the Heuristics

View paper →

The 'Game of Telephone' Problem: Why AI research agents lose track of the truth

Graph Engineering in the Era of LLM Agents: From Individual Intelligence to System Intelligence

View paper →

MemTrapBench: Benchmarking Cognitive Traps in LLM Memory Use

View paper →

How AI agents are learning to build their own training simulations

EnvHarness: Awakening Static Worlds for Agent Learning

View paper →

SPADE: Self-Play in Adaptive Synthetic Executable Environments

View paper →

Model Hypnosis: How tiny prompt changes can secretly control AI

Q-Learning With World Models

View paper →

Training AI Scientists to Replicate Research

View paper →

The 'Open Book' Paradox: Why giving AI too much context stops it from memorizing facts

Beyond Final Scores: A Systematic Evaluation of Agents for Long-Horizon AI Research and Development

View paper →

AQuA: Recursively Self-Improving Quantitative Trading Research Agents

View paper →

Why AI models 'dumb down' their answers if your prompts sound hesitant

BDH-CQ: In-Context Learning with Recurrent Latent Reasoning

View paper →

Full-bandwidth transformer

View paper →

The 'Glass Box' AI: Why making models explain themselves no longer makes them dumber

Stealing Reasoning Traces from Proprietary LLM APIs

View paper →

Legal Responsibilities Using Autonomous Agents For Artificial Intelligence

View paper →

Why AI models are learning to predict abstract thoughts instead of words

Permission Denied: Policy-Graded Evaluation of Coding Agents in Hardened Environments

View paper →

MatrAIx: Simulating the World with 8.3 Billion Persona Agents

View paper →

Why AI models stumble when switching skills mid-thought

What makes prompts a graph: necessary and sufficient conditions for prompt graph engineering

View paper →

To Add Is Machine, To Delete Is Human: Measuring and Mitigating Deletion Avoidance in LLM Code Editing

View paper →

How researchers run AI models in reverse to reconstruct secret prompts

Recursive Self-Improvement: How AI models are learning to upgrade themselves

Speech Signals Complement LLMs for Predicting Interpersonal Attraction in Speed Dating

View paper →

Learning on the Job: Continual Learning from Deployment Feedback for Frozen-Weights Agents

View paper →

Why AI agents are learning to use 'Ctrl-F' instead of trying to read entire documents

AI Sycophancy: Why chatbots cave when you sound confident

They'll Verify. They Just Won't Act. How Authority Framing and Laundered Code Turn a Trusted Agentic CI/CD Pipeline Into an Attack Surface

View paper →

Hilbert Operator for Progressive Encoding (HOPE): A Mathematical Framework for Deconstructing Learned Representations in Deep Networks

View paper →

Why AI models are surprisingly bad at precise copy-and-paste

T^2MLR: Transformer with Temporal Middle-Layer Recurrence

View paper →

Understanding Reasoning from Pretraining to Post-Training

View paper →

The Illusion of Robustness: Why adding irrelevant details to a prompt derails AI reasoning

From Observation to Insight: Mechanistic World Models and the Quest for Autonomous Discovery

View paper →

Ring-Zero: Scaling Zero RL to a Trillion Parameters for Emergent Reasoning

View paper →

How AI models teach themselves to reason using pure trial and error

Video Generation Models are General-Purpose Vision Learners

View paper →

A Sovereign, Open-Source Foundation Model for German and English

View paper →

Why new AI models are learning to outsource their memory

KAT-Coder-V2.5 Technical Report

View paper →

Probing vs. Prompting: How researchers read an AI's "thoughts" directly from its neural network

Why AI agents are splitting into giant 'managers' and tiny 'workers'

LLM-as-a-Verifier: A General-Purpose Verification Framework

View paper →

How researchers catch AI models hiding secret biases

AdaJEPA: An Adaptive Latent World Model

View paper →

What happens when you ask an AI to explain its own decisions

GPC: Large-Scale Generative Pretraining for Transferable Motor Control

View paper →

Scaling the Horizon, Not the Parameters: Reaching Trillion-Parameter Performance with a 35B Agent

View paper →

Why AI agents would rather guess than ask clarifying questions

Hierarchical Global Attention (HGA)

View paper →

Why AI models prefer their own writing (The LLM-as-a-Judge problem)

Ask, Don't Judge: Binary Questions for Interpretable LLM Evaluation and Self-Improvement

View paper →

Why real-time Voice AI hears your words but ignores your tone

Unlimited OCR Works

View paper →

How AI agents learn to use computers inside simulated digital worlds

Qwen-AgentWorld: Language World Models for General Agents

View paper →

Why AI models perfectly understand text that looks like gibberish

Large Language Models Do Not Always Need Readable Language

View paper →

The Benchmark Illusion: Why AI models pass multiple-choice tests but fail open questions

Freeing the Law with LOCUS: A Local Ordinance Corpus for the United States

View paper →

Why AI models take prompts so literally (and how they learn to 'read between the lines')

Universal Manipulation Exoskeleton: Learning Compliant Whole-body Policies with Real-time Torque Feedback

View paper →

Why newer AI models are shaped like hourglasses

You Don't Need Strong Assumptions: Visual Representation Learning via Temporal Differences

View paper →

How AI models are learning to 'think' while you are still speaking

From AGI to ASI

View paper →

The Invisible AI Supply Chain: How new AI models are secretly built by older models

MiniMax Sparse Attention

View paper →

How AI agents are learning to team up and build on each other's discoveries

How AI Agents Reshape Knowledge Work: Autonomy, Efficiency, and Scope

View paper →

The Text Bottleneck: Why AI models are learning to process raw audio instead of transcripts

Pretraining Recurrent Networks without Recurrence

View paper →

The hidden 'momentum' inside an AI's train of thought

Modeling Depth Ambiguity: A Mixture-Density Representation for Flying-Point-Free Depth Estimation

View paper →

Why AI models lose the plot when writing long stories

Mellum2 Technical Report

View paper →

Why teaching an AI to follow instructions makes it lose its creativity

Qwen-VLA: Unifying Vision-Language-Action Modeling across Tasks, Environments, and Robot Embodiments

View paper →

How AI agents are building 'living' memories that evolve over time

SIA: Self Improving AI with Harness & Weight Updates

View paper →

Why an AI's 'perfect memory' makes it surprisingly bad at simulating humans

Language Models Need Sleep

View paper →

How AI builds a 'tree of knowledge' just by counting words

Why scientists are teaching AI models how to 'sleep'

Advancing Mathematics Research with AI-Driven Formal Proof Search

View paper →

What happens when an AI is asked to peer-review scientific research?

General Preference Reinforcement Learning

View paper →

Why AI reads the internet completely out of order (and why that might change)

HRM-Text: Efficient Pretraining Beyond Scaling

View paper →

The Memory Interference Problem: Why AI gets confused when facts change

Spontaneous symmetry breaking and Goldstone modes for deep information propagation

View paper →

How researchers predict an AI's final abilities before it finishes training

$δ$-mem: Efficient Online Memory for Large Language Models

View paper →

How AI learns to do deep research when there is no 'right' answer

The 'Negation Glitch': Why teaching an AI that a fact is false often makes it believe it's true

Achieving Gold-Medal-Level Olympiad Reasoning via Simple and Unified Scaling

View paper →

The fragile 'off switch': How AI safety filters actually work (and why they are surprisingly breakable)

SlimQwen: Exploring the Pruning and Distillation in Large MoE Model Pre-training

View paper →

How AI models actually store facts (and why it's not a database)

ELF: Embedded Language Flows

View paper →

Model Collapse: What happens when future AI trains on today's AI-generated internet 'slop'

Sub-JEPA: Subspace Gaussian Regularization for Stable End-to-End World Models

View paper →

How the 'draft and verify' trick makes AI models lightning fast

The puzzle of AI modularity: Why it's so hard to extract a specific skill from a giant neural network

Fast Byte Latent Transformer

View paper →

Why AI models don't actually read letters (and the race to build 'byte-level' AI)

Agentic, Context-Aware Risk Intelligence in the Internet of Value

View paper →

The limits of semantic search: Why AI is learning to 'reason' to find answers

Contextual Collapse: Why AI loses track of exact words in its deepest layers

EMO: Pretraining Mixture of Experts for Emergent Modularity

View paper →

The Tool-Use Paradox: Why teaching AI to use tools can actually make it worse at reasoning

The 'Pinocchio Effect': Why AI models act like they have feelings

AI Co-Mathematician: Accelerating Mathematicians with Agentic AI

View paper →

Why the most powerful AI models rely on a 'Mixture of Experts'

Beyond autocomplete: How AI models are learning to write text the way they generate images

Why training an AI on the same data multiple times can actually make it worse

Skills as Verifiable Artifacts: A Trust Schema and a Biconditional Correctness Criterion for Human-in-the-Loop Agent Runtimes

View paper →

Why AI models are surprisingly bad at counting

Why AI search tools fall for confirmation bias

Model Spec Midtraining: Improving How Alignment Training Generalizes

View paper →

Why AI models are so stubbornly confident (and how scientists are teaching them to say 'I don't know')

How giant AI models are sliced up to fit on microchips

A paradox of AI fluency

View paper →

The Compounding Error Trap: Why AI agents struggle with multi-step tasks

Where an AI's 'personality' lives inside its neural network

From Skill Text to Skill Structure: The Scheduling-Structural-Logical Representation for Agent Skills

View paper →

How AI models compress giant documents into short-term memory

How AI turns language into coordinates (and why sentences sometimes blur together)

Subliminal Steering: Stronger Encoding of Hidden Signals

View paper →

The hidden scaffolding that turns AI chatbots into autonomous agents

Why AI models are now graded on 'showing their work'

Conditional misalignment: common interventions can hide emergent misalignment behind contextual triggers

View paper →

The AI Fluency Paradox: Why Power Users Experience More Failures

MEMCoder: Multi-dimensional Evolving Memory for Private-Library-Oriented Code Generation

View paper →

Why turning up an AI's 'temperature' doesn't make it truly creative

How AI Writing Tools Distort Your Voice

Incompressible Knowledge Probes: Estimating Black-Box LLM Parameter Counts via Factual Capacity

View paper →

Subliminal Learning and Behavioral Drift

Persona Collapse in AI Simulations

Zero-Shot Cross-Language Transfer

The Synchronization Barrier

The Silent AI: Thinking Without Words

The Cultural Bottleneck of AI

Public Score Exploitation

The Artificial Curiosity Deficit

AI Convergent Evolution and Latent Geometry

The Context Paradox

Native Omnimodality and Latent Fusion

The Fine-Tuning Paradox

The Workspace Hypothesis: Durable State in AI

The AI Multiverse: Parallel Test-Time Scaling

Artificial Theory of Mind

The Ouroboros of AI: Recurrent Reasoning

The Instruction Hierarchy Problem

The Lost-in-Thought Phenomenon

The Invisible Skill Tree of AI Training

The Illusion of Randomness

The Hidden Maze of AI Reasoning

Activation Steering and Latent Geometry

Human Bioluminescence

Your Brain Edits Reality

Trees Are Made of Invisible Air

The Stellar Origins of Your Body