Hacker Papers
- 426
Emergent Misalignment: Narrow finetuning can produce broadly misaligned LLMs
Hacker News threads
- Jul 2025Emergent Misalignment: Narrow finetuning can produce broadly misaligned LLMs
- Jun 2025LLM's Illusion of Alignment
- Jun 2025LLM's Illusion of Alignment
- Aug 2025Training language models to be warm and empathetic makes them less reliable
- Oct 2025It's insulting to read AI-generated blog posts
- Nov 2025DeepSeek writes insecure code if prompt mentions topics restricted in China
- May 2026Anthropic Cofounder Chris Olah's Remarks on Pope Leo XIV's "Magnifica Humanitas"
- Jun 2026Where every major LLM stands politically
- Jul 2026Show HN: Distilling DeepSeek into GPT-OSS doesn't transfer censorship. Try it
- 427
Locating and Editing Factual Associations in GPT
Hacker News threads
- Feb 2023We Found an Neuron in GPT-2
- Feb 2023ChatGPT is falsely recommending us for a service we don't provide
- May 2023How does data removal from GPT *work
- Aug 2023PMET: Precise Model Editing in a Transformer
- Oct 2023MemGPT – LLMs with self-editing memory for unbounded context
- Nov 2023Is the reversal curse in LLMs real?
- Dec 2023What Is Retrieval-Augmented Generation a.k.a. RAG?
- May 2024Llama3 implemented from scratch
- 428
DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models
Hacker News threads
- Jan 2025DeepSeekMath
- Feb 2024DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open LLMs
- Feb 2024Pushing the Limits of Mathematical Reasoning in Open Language Models
- Jan 2025Berkeley Researchers Replicate DeepSeek R1's Core Tech for Just $30: A Small Mod
- Feb 2025GPT-4.5
- Mar 2025Using GRPO to Beat o1, o3-mini and R1 at “Temporal Clue”
- May 2025DeepSeek’s founder is threatening US dominance in AI race
- Nov 2025DeepSeekMath-V2: Towards Self-Verifiable Mathematical Reasoning [pdf]
- Jan 2026Nanolang: A tiny experimental language designed to be targeted by coding LLMs
- Sep 2026DeepSeek v4.1 Flash
- 429
Why Can GPT Learn In-Context? Language Models Implicitly Perform Gradient Descent as Meta-Optimizers
Hacker News threads
- Dec 2022Why Can GPT Learn In-Context? Language Models Perform Gradient Descent
- Jan 2023The Expanding Dark Forest and Generative AI
- Feb 2023GPT in 60 Lines of NumPy
- Feb 2023To Build Truly Intelligent Machines, Teach Them Cause and Effect
- Mar 2023Large language models are having their Stable Diffusion moment
- Mar 2023GPT-4 Designed a Programming Language
- Mar 2023OK, it’s time to freak out about AI
- May 2023Chatbots Don’t Know What Stuff Isn’t
- Jul 2024Reasoning in Large Language Models: A Geometric Perspective
- 430
Titans: Learning to Memorize at Test Time
Hacker News threads
- Jan 2025Titans: Learning to Memorize at Test Time
- Jan 2025Titans: Learning to Memorize at Test Time
- Jan 2025Titans: Learning to Memorize at Test Time
- Jan 2025Coming soon: PhD-level super-agents
- Jan 2025Qwen2.5-1M: Deploy your own Qwen with context length up to 1M tokens
- Feb 2025Google Titans Model Explained: The Future of Memory-Driven AI Architectures
- Feb 2025Google Titans Model Explained: The Future of Memory-Driven AI Architectures
- Dec 2025Google Titans architecture, helping AI have long-term memory
- Dec 2025The dawn of a world simulator
- Mar 2026Yann LeCun's AI startup raises $1B in Europe's largest ever seed round
- Apr 2026Show HN: AI memory with biological decay (52% recall)
- 431
Winning Gold at IMO 2025 with a Model-Agnostic Verification-and-Refinement Pipeline
Hacker News threads
- Jul 2025Gemini 2.5 Pro Capable of Winning Gold at IMO 2025
- Jul 2025Gemini 2.5 Pro Capable of Winning Gold at IMO 2025 with Prompting
- Jul 2025Gemini 2.5 Pro Capable of Winning Gold at IMO 2025
- Jul 2025Gemini with Deep Think achieves gold-medal standard at the IMO
- Aug 2025GPT-5
- Aug 2025Claim: GPT-5-pro can prove new interesting mathematics
- Sep 2025DeepMind and OpenAI win gold at ICPC
- Oct 2025Implicit Actor Critic Coupling via a Supervised Learning Framework for RLVR
- Oct 2025Implicit Actor Critic Coupling via a Supervised Learning Framework for RLVR
- Nov 2025Yann LeCun to depart Meta and launch AI startup focused on 'world models'
- Jan 20262025: The Year in LLMs
- 432
LIMA: Less Is More for Alignment
Hacker News threads
- May 2023LIMA: Less Is More for Alignment
- May 2023LIMA: Less Is More for Alignment
- May 2023GPT detectors are biased against non-native English writers
- May 2023How to Finetune GPT Like Large Language Models on a Custom Dataset
- Jun 2023Falcon 40B LLM (which beats Llama) now Apache 2.0
- Jun 2023AI has poisoned its own well
- Sep 2023Large Language Models for Compiler Optimization
- Sep 2023Outperforming LLMs with less training data and smaller model sizes
- Apr 2024Ask HN: Most efficient way to fine-tune an LLM in 2024?
- Oct 2024Zamba2-7B
- 433
Alice in Wonderland: Simple Tasks Showing Complete Reasoning Breakdown in State-Of-the-Art Large Language Models
Hacker News threads
- Jun 2024Simple tasks showing reasoning breakdown in state-of-the-art LLMs
- Jun 2024AI Appears to Rapidly Be Approaching Brick Wall Where It Can't Get Smarter
- Jun 2024Novel prompting approach for Alice in Wonderland problem
- Jul 2024Reasoning in Large Language Models: A Geometric Perspective
- Sep 2024Reflection 70B, the top open-source model
- Oct 2024Understanding the Limitations of Mathematical Reasoning in LLMs
- Dec 2024Genie 2: A large-scale foundation world model
- Dec 2024Genie 2: A large-scale foundation world model
- Apr 2025Ask HN: Share your AI prompt that stumps every model
- 434
Reasoning or Reciting? Exploring the Capabilities and Limitations of Language Models Through Counterfactual Tasks
Hacker News threads
- Sep 2023Exploring Capabilities, Limitations of Language Models by Counterfactual Tasks
- Jul 2023AI Safety and the Age of Dislightenment
- Dec 2023Will scaling work?
- Jun 2024The illusion of state in state-space models
- Jul 2024Tabloid: A clickbait headline programming language (2021)
- Aug 2024Markov chains are funnier than LLMs
- Aug 2024Markov chains are funnier than LLMs
- Nov 2024DeepThought-8B: A small, capable reasoning model
- 435
Scaling up Test-Time Compute with Latent Reasoning: A Recurrent Depth Approach
Hacker News threads
- Feb 2025Scaling up test-time compute with latent reasoning: A recurrent depth approach
- Feb 2025Scaling Up Test-Time Compute with Latent Reasoning: A Recurrent Depth Approach
- Jul 2025Chain of thought monitorability: A new and fragile opportunity for AI safety
- Oct 2025Less is more: Recursive reasoning with tiny networks
- Dec 2025Universal Reasoning Model (53.8% pass 1 ARC1 and 16.0% ARC 2)
- Mar 2026Show HN: How I Topped the HuggingFace Open LLM Leaderboard on Two Gaming GPUs
- Apr 2026System Card: Claude Mythos Preview [pdf]
- Aug 2026AirLLM 70B inference with single 4GB GPU
- 436
Language Models Use Trigonometry to Do Addition
Hacker News threads
- Feb 2025Language Models Use Trigonometry to Do Addition
- Feb 2025Pre-Trained Large Language Models Use Fourier Features for Addition (2024)
- Apr 2025AI coding and the peanut butter and jelly problem
- Oct 2025Reasoning is not model improvement
- Aug 2026Models Are Getting Dumber on Purpose
- Sep 2026US Military had close call after using AI for hallucinated intelligence report
- Oct 2026How accurately calibrated is Jev?
- 437
Questioning Representational Optimism in Deep Learning: The Fractured Entangled Representation Hypothesis
Hacker News threads
- May 2025Questioning Representational Optimism in Deep Learning
- Jun 2025Questioning Representational Optimism in Deep Learning
- Nov 2025Questioning Representational Optimism in Deep Learning
- May 2025Questioning Representational Optimism in Deep Learning
- Jun 2025Vision Language Models Are Biased
- Nov 2025The space of minds
- Feb 2026An autopsy of AI-generated 3D slop
- Aug 2026Small Models Have Arrived
- Sep 2026A beginning for mathematics
- 438
Photorealistic Text-to-Image Diffusion Models with Deep Language Understanding
Hacker News threads
- Aug 2022Imagen: An AI system that creates photorealistic images from input text
- Feb 2023Google Invests Almost $400M in ChatGPT Rival Anthropic
- Mar 2023Visual ChatGPT
- Apr 2023Stability AI releases DeepFloyd IF, a powerful text-to-image model
- May 2023Governance of Superintelligence
- Jul 2023Stability AI releases its latest image-generating model, Stable Diffusion XL 1.0
- 439
The SWE-Bench Illusion: When State-of-the-Art LLMs Remember Instead of Reason
Hacker News threads
- Sep 2025The SWE-Bench Illusion: When LLMs Remember Instead of Reason
- Feb 2026The SWE-Bench Illusion: When LLMs Remember Instead of Reason
- Jul 2025Principles for production AI agents
- Aug 2025Open models by OpenAI
- Sep 2025Using Claude Code to modernize a 25-year-old kernel driver
- Sep 2025Top model scores may be skewed by Git history leaks in SWE-bench
- Feb 2026Prediction: Claude 5 will be a major regression
- Mar 2026FSF Threatens Anthropic over Infringed Copyright: Share Your LLMs Freel
- Jun 2026Did Claude Increase Bugs in Rsync?
- 440
InfoGAN: Interpretable Representation Learning by Information Maximizing Generative Adversarial Nets
Hacker News threads
- Jul 2016InfoGAN: Representation Learning by Generative Adversarial Nets
- Jun 2016Generative Models
- Aug 2016OpenAI Team Update
- Jan 2017Darpa Goes “Meta” with Machine Learning for Machine Learning (2016)
- Feb 2017Ask HN: What is the most exciting development in your field right now?
- Apr 2017Unsupervised sentiment neuron
- Sep 2017PassGAN: A Deep Learning Approach for Password Guessing
- 441
s1: Simple test-time scaling
Hacker News threads
- Jun 2025S1: Simple Test-Time Scaling
- Feb 2025Test-time scaling new approach: extra test-time compute improves LLM reasoning
- Feb 2025s1: Simple Test-Time Scaling
- Feb 2025Chat is a bad UI pattern for development tools
- Feb 2025S1: A $6 R1 competitor?
- Feb 2025Researchers created an open rival to OpenAI's o1 'reasoning' model for under $50
- Feb 2025LIMO: Less Is More for Reasoning
- Mar 2025QwQ-32B: Embracing the Power of Reinforcement Learning
- Apr 2025Chain of Recursive Thoughts: Make AI think harder by making it argue with itself
- Apr 2026TinyLoRA – Learning to Reason in 13 Parameters
- Jul 2026Controlling Reasoning Effort in LLMs
- 442
No "Zero-Shot" Without Exponential Data: Pretraining Concept Frequency Determines Multimodal Model Performance
Hacker News threads
- May 2024No "Zero-Shot" Without Exponential Data
- Apr 2024No "Zero-Shot" Without Exponential Data
- Apr 2024No "Zero-Shot" Without Exponential Data
- May 2024Geoffrey Hinton "godfather of AI" says that AI will surpass human intelligence
- Jun 2024Mathematical Optimization for Cargo Ships
- Jun 2024OpenAI and Apple Announce Partnership
- Jun 2024Ask HN: When will we hit a limit on LLM performance?
- Jul 2024Vision language models are blind
- Jul 2024Ask HN: Are there videos of people using AI for non-trivial coding?
- Nov 2024OpenAI, Google and Anthropic are struggling to build more advanced AI
- Feb 2026(AI) Slop Terrifies Me
- 443
Position: Stop Anthropomorphizing Intermediate Tokens as Reasoning/Thinking Traces!
Hacker News threads
- May 2025(How) Do reasoning models reason?
- Apr 2025(How) Do reasoning models reason?
- Jun 2025Stop Anthropomorphizing Intermediate Tokens as Reasoning/Thinking Traces
- Oct 2025A small number of samples can poison LLMs of any size
- Nov 2025Learn Prolog Now (2006)
- Jun 2026GLM-5.2 is the new leading open weights model on Artificial Analysis
- Aug 2026Chain-of-Thought Reasoning in the Wild Is Not Always Faithful
- Sep 2026OpenAI's new reasoning technique alarms AI safety experts
- Sep 2026How concerned should we be about Astra's recurrent architecture?
- Sep 2026Formalizing Fermat's Last Theorem
- Sep 2026Caltech Mathathon – first hackathon ever devoted to research level mathematics
- 444
Algorithmic Extremism: Examining YouTube's Rabbit Hole of Radicalization
Hacker News threads
- Dec 2019YouTube’s Rabbit Hole of Radicalization [pdf]
- Dec 2019Algorithmic Extremism: Examining YouTube's Rabbit Hole of Radicalization
- Dec 2019Algorithmic Extremism: Examining YouTube's Rabbit Hole of Radicalization
- Jun 2020Ad Boycott of Facebook Keeps Growing
- Jun 2020YouTube bans Stefan Molyneux, David Duke, Richard Spencer for hate speech
- Jun 2020YouTube bans Stefan Molyneux, David Duke, Richard Spencer for hate speech
- Jul 2020The TikTok War
- Sep 2020Mozilla Is Crowdsourcing Research into YouTube Recommendations
- Dec 2020Christchurch shooter was radicalized on YouTube, New Zealand report says
- Dec 2020YouTube to remove content that alleges widespread election fraud
- Aug 2021If YouTube’s algorithms radicalize people, it’s hard to tell from the data
- 445
Train Short, Test Long: Attention with Linear Biases Enables Input Length Extrapolation
Hacker News threads
- May 2023Attention with Linear Biases (ALiBi)
- May 2023Replit's new Code LLM: Open Source, 77% smaller than Codex, trained in 1 week
- May 2023Anthropic - Introducing 100K Token Context Windows, Around 75,000 Words
- Jul 2023
- Aug 2023Code Llama, an AI tool for coding
- Aug 2023
- Jan 2024ChatGPT for Teams
- Oct 2024Differential Transformer
- 446
Hyena Hierarchy: Towards Larger Convolutional Language Models
Hacker News threads
- Apr 2023Hyena Hierarchy: Towards Larger Convolutional Language Models
- Apr 2023Hyena a more efficient convolutional LLM when compared to GPT
- Mar 2023Hyena Hierarchy: Towards Larger Convolutional Language Models
- Mar 2023CoLT5: Faster Long-Range Transformers With Conditional Computation
- Apr 2023From Deep to Long Learning
- Apr 2023From Deep to Long Learning
- Apr 2023Stability AI Launches the First of Its StableLM Suite of Language Models
- Apr 2023New technology could blow away GPT-4 and everything like it
- May 2023OpenAI's plans according to sama
- Sep 2023The Eleuther AI Mafia
- Feb 2025The FFT Strikes Back: An Efficient Alternative to Self-Attention
- 447
RoFormer: Enhanced Transformer with Rotary Position Embedding
Hacker News threads
- Mar 2023Large language models are having their Stable Diffusion moment
- Apr 2023Why Are Sinusoidal Functions Used for Position Encoding?
- Aug 2023Llama from Scratch (or how to implement a paper without crying)
- Apr 2025Multi-Token Attention
- Apr 2025The Llama 4 herd
- Apr 2025Llama 4 Now Live on Groq
- Aug 2025Open models by OpenAI
- 448
Fast Inference from Transformers via Speculative Decoding
Hacker News threads
- Sep 2023Fast Inference from Transformers via Speculative Decoding
- Jul 2025Playing with Open Source LLMs
- Feb 2026Raspberry Pi Pico 2 at 873.5MHz with 3.05V Core Abuse
- Apr 2026We got 207 tok/s with Qwen3.5-27B on an RTX 3090
- May 2026Accelerating Gemma 4: faster inference with multi-token prediction drafters
- Jun 2026DeepSeek open-sources inference optimizations with 60–85% faster generation [pdf]
- Jun 2026DeepSeek open-sources inference optimizations with 60–85% faster generation [pdf]
- Jun 2026DeepSeek open-sources inference optimizations with 60–85% faster generation [pdf]
- 449
Executable Code Actions Elicit Better LLM Agents
Hacker News threads
- May 2024Executable Code Actions Elicit Better LLM Agents
- Jan 2025Show HN: Gemini 2 with unrestricted code execution
- Jul 2025Tools: Code Is All You Need
- Aug 2025Universal Tool Calling Protocol (UTCP)
- Dec 2025Anthropic acquires Bun
- Aug 2026Mistral Patent for "Code implemented tool calls"
- Aug 2026Mistral Patent for "Code implemented tool calls"
- Aug 2026Mistral Patent for "Code implemented tool calls"
- 450
xVal: A Continuous Numerical Tokenization for Scientific Language Models
Hacker News threads
- Oct 2023Single Digit Tokenization improves LLM Math abilities by up to 70x
- Oct 2023Single Digit tokenization improves LLM math abilities by up to 70x
- Nov 2023Telling GPT-4 you're scared or under pressure improves performance
- Nov 2023Sutskever: OpenAI board doing its mission to build AGI that benefits all
- Nov 2023Sam Altman's ouster was precipitated by letter to board about AI breakthrough
- Nov 2023Sam Altman's ouster was precipitated by letter to board about AI breakthrough
- Mar 2024Integer Tokenization Is Insane (2023)
- Sep 2024AGI is far from inevitable
- Oct 2024Probably pay attention to tokenizers
Data from the Hacker News archive (ODC-By), arXiv and OpenAlex. Not affiliated with Hacker News or Y Combinator.