Papers, Blogs, Courses and Lectures¶
The research frontier -- cutting-edge papers on capabilities, reasoning, agents, alignment, and interpretability defining the path from LLMs to AGI.
Frontier Model Papers¶
| Paper | Authors / Org | Year | Description | Links |
|---|---|---|---|---|
| GPT-4 Technical Report | OpenAI | 2023 | Foundational technical report describing GPT-4's multimodal capabilities, RLHF training, and safety evaluations. | Paper |
| Learning to Reason with LLMs (o1) | OpenAI | 2024 | Introduces o1 — trained with RL to think deeply before responding, achieving PhD-level reasoning performance. | Blog |
| Gemini: A Family of Highly Capable Multimodal Models | Google DeepMind | 2023 | The Gemini family (Ultra/Pro/Nano) with native multimodal training, surpassing GPT-4 on 30/32 benchmarks. | Paper |
| Gemini 1.5: Unlocking Multimodal Understanding Across Millions of Tokens | Google DeepMind | 2024 | Extends Gemini to 1M-token context (later 2M) via efficient MoE architecture. | Paper |
| The Llama 3 Herd of Models | Meta AI | 2024 | Open-weight Llama 3 (8B–405B), competitive with GPT-4 on key benchmarks; 15T+ training tokens. | Paper |
| DeepSeek-V3 Technical Report | DeepSeek-AI | 2024 | 671B MoE model trained for $5.5M via FP8 mixed-precision; competitive with GPT-4o. | Paper |
| DeepSeek-R1: Incentivizing Reasoning via RL | DeepSeek-AI | 2025 | Chain-of-thought reasoning purely through RL (GRPO) without SFT — matches o1 on math/code. | Paper |
| Mixtral of Experts | Mistral AI | 2024 | Mixtral 8x7B sparse MoE matching Llama 2 70B with 5x lower inference cost. | Paper |
| Qwen2.5 Technical Report | Qwen Team (Alibaba) | 2025 | Qwen2.5 series with improved coding and math specializations. | Paper |
| Phi-3: A Highly Capable Language Model on Your Phone | Microsoft | 2024 | 3.8B model trained on curated synthetic data that rivals models 10x its size. | Paper |
| The Llama 4 Herd: Natively Multimodal AI Innovation | Meta AI | 2025 | First Llama with MoE architecture: Scout (17B/16 experts, 10M context), Maverick (17B/128 experts), Behemoth (288B/16 experts teacher). Natively multimodal with early fusion. Behemoth outperforms GPT-4.5 on STEM. | Blog |
| Gemini 2.5 Pro | Google DeepMind | 2025 | Thinking model with advanced reasoning. #1 on LMArena by significant margin. 18.8% on Humanity's Last Exam. State-of-art on GPQA, AIME 2025, and coding benchmarks. | Blog |
| Meta Muse Spark | Meta Superintelligence Labs | 2026 | First model from Meta Superintelligence Labs. Natively multimodal reasoning model with visual chain-of-thought, tool-use, and multi-agent orchestration ("Contemplating mode"). 58% on Humanity's Last Exam. Scaling toward "personal superintelligence." | Blog |
| Gemma: Open Models from Gemini Research | Google DeepMind | 2024 | Open-weight models (2B/7B) built from Gemini research. Gemma 2 (2024) and Gemma 3 (2025) with state-of-art performance at size. Responsible AI toolkit included. | Site, GitHub |
Reasoning, Scaling & Architecture Papers¶
| Paper | Authors | Year | Description | Links |
|---|---|---|---|---|
| Chain-of-Thought Prompting Elicits Reasoning in LLMs | Wei et al. (Google) | 2022 | Foundational paper: intermediate reasoning steps dramatically improve LLM performance. | Paper |
| Tree of Thoughts: Deliberate Problem Solving with LLMs | Yao et al. (Princeton/Google) | 2023 | Tree-structured reasoning enabling backtracking and lookahead. | Paper |
| Let's Verify Step by Step | Lightman et al. (OpenAI) | 2023 | Process reward models (PRMs) scoring each reasoning step — the mechanism behind o1-style training. | Paper |
| Scaling LLM Test-Time Compute Optimally | Snell et al. (Berkeley) | 2024 | More compute at inference can equal more training compute on hard tasks. | Paper |
| Training Compute-Optimal LLMs (Chinchilla) | Hoffmann et al. (DeepMind) | 2022 | Optimal LLM training scales data and parameters equally. | Paper |
| Scaling Laws for Neural Language Models | Kaplan et al. (OpenAI) | 2020 | Power-law relationships between model scale and performance, underpinning AGI scaling hypotheses. | Paper |
| LongRoPE: Extending LLM Context Beyond 2M Tokens | Ding et al. (Microsoft) | 2024 | Extends RoPE to 2M tokens via non-uniform interpolation. | Paper |
| Infini-Attention | Munkhdalai et al. (Google) | 2024 | Compressive memory in standard attention for infinite-length inputs with bounded memory. | Paper |
World Models & Environment Simulation Papers¶
| Paper | Authors | Year | Description | Links |
|---|---|---|---|---|
| World Models | Ha & Schmidhuber | 2018 | Foundational paper: learning compressed spatial and temporal representations of environments; agents trained entirely inside hallucinated dreams. | Paper, Interactive |
| I-JEPA: Joint-Embedding Predictive Architecture | Assran et al. (Meta / LeCun) | 2023 | LeCun's vision for AGI through self-supervised prediction in representation space rather than pixel space. Non-generative, highly scalable. | Paper |
| Liquid Time-Constant Networks | Hasani, Lechner, Amini, Rus (MIT CSAIL) | 2020 | Novel continuous-time neural networks with liquid (varying) time-constants -- the architecture behind Liquid AI's foundation models. | Paper, Code |
| Video Generation Models as World Simulators (Sora) | OpenAI | 2024 | Sora: text-to-video diffusion transformer that models physics and long-horizon consistency. | Blog |
| Genie: Generative Interactive Environments | Bruce et al. (DeepMind) | 2024 | Learns playable 2D world models from unlabeled internet video. | Paper |
| Genie 3: Generating and Exploring Interactive Worlds | Google DeepMind | 2025 | Next-generation world model that generates and enables exploration of interactive 3D environments. Breakthrough in environment simulation fidelity and interactivity. | Site |
| SIMA 2: An Agent That Plays, Reasons, and Learns | Google DeepMind | 2025 | Generalist AI agent that plays, reasons, and learns in virtual 3D worlds. Advances embodied agent capabilities in complex open-ended environments with persistent learning. | Blog |
| NVIDIA Cosmos | NVIDIA | 2025 | Open-source world foundation model platform for physical AI -- robotics, autonomous vehicles, and simulation. 8k+ stars. | GitHub |
Physical AI & Embodied Intelligence Papers¶
| Paper | Authors / Org | Year | Description | Links |
|---|---|---|---|---|
| RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control | Brohan, Levine et al. (Google DeepMind) | 2023 | Landmark VLA model: co-fine-tunes vision-language models on robot trajectory data. Actions as text tokens enable emergent semantic reasoning for physical tasks. | Paper |
| PaLM-E: An Embodied Multimodal Language Model | Driess, Levine et al. (Google) | 2023 | 562B-parameter embodied LLM grounding language in continuous sensor modalities. Positive transfer across internet-scale language, vision, and robotics. | Paper |
| pi0: A VLA Flow Model for General Robot Control | Black, Levine, Finn et al. (Physical Intelligence) | 2024 | Novel flow matching architecture on pre-trained VLM for general-purpose robot policies. Laundry folding, table cleaning, box assembly across diverse embodiments. Open-sourced. | Paper |
| Open X-Embodiment: Robotic Learning Datasets and RT-X Models | Open X-Embodiment Collaboration (21 institutions) | 2023 | Largest cross-embodiment robotics dataset (22 robots, 527 skills, 160k+ tasks). RT-X shows positive transfer across robot morphologies -- robotics' "ImageNet moment." | Paper |
| TD-MPC2: Scalable World Models for Continuous Control | Hansen, Su, Wang | 2023 | 317M-parameter agent controlling 80 tasks across multiple embodiments and action spaces using implicit world models. ICLR 2024. | Paper |
| Gemini Robotics: Bringing AI into the Physical World | Google DeepMind | 2025 | Dual-model approach: Gemini Robotics 1.5 (VLA) for direct motor control and Robotics-ER 1.5 for embodied reasoning. Generality, dexterity, agentic tool-use, thinking, and multi-embodiment support (static arms to humanoids). | Site |
Agent Papers¶
| Paper | Authors | Year | Description | Links |
|---|---|---|---|---|
| ReAct: Synergizing Reasoning and Acting in LMs | Yao et al. (Princeton/Google) | 2022 | Interleaves reasoning with grounded actions — the dominant LLM agent paradigm. | Paper |
| Reflexion: Language Agents with Verbal Reinforcement Learning | Shinn et al. | 2023 | Agents reflect on failures in natural language and use episodic memory to improve. | Paper |
| Generative Agents: Interactive Simulacra of Human Behavior | Park et al. (Stanford/Google) | 2023 | 25 AI agents in a simulated town exhibiting emergent social behaviors. | Paper |
| Voyager: An Open-Ended Embodied Agent with LLMs | Wang et al. (NVIDIA/CMU) | 2023 | First LLM-powered Minecraft agent with lifelong learning via skill library. | Paper |
| ToolLLM: Facilitating LLMs to Master 16,000+ APIs | Qin et al. (Tsinghua) | 2023 | Framework for training and evaluating LLMs on tool use across 16,464 real APIs. | Paper |
| SWE-bench: Can LMs Resolve Real-World GitHub Issues? | Jimenez et al. (Princeton) | 2023 | Benchmark of 2,294 real GitHub issues driving the coding agent race. | Paper |
| WebArena: A Realistic Web Environment for Autonomous Agents | Zhou et al. | 2023 | 812 real-world web tasks exposing the gap between LLMs and human agents. | Paper |
| OSWorld: Benchmarking Multimodal Agents in Real Computer Environments | Xie et al. | 2024 | GUI agents across real OS; top agents score ~7% vs. human 72%. | Paper |
| The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery | Sakana AI | 2024 | Fully autonomous research pipeline — idea generation to paper writing. | Paper |
Alignment & Reward Modeling Papers¶
| Paper | Authors | Year | Description | Links |
|---|---|---|---|---|
| Direct Preference Optimization (DPO) | Rafailov et al. (Stanford) | 2023 | Eliminates separate RL + reward model in RLHF by directly optimizing on preference data. | Paper |
| KTO: Model Alignment as Prospect Theoretic Optimization | Ethayarajh et al. | 2024 | Alignment with only binary (good/bad) feedback, no paired comparisons needed. | Paper |
| ORPO: Monolithic Preference Optimization without Reference Model | Hong et al. | 2024 | Eliminates reference model in DPO-style training, reducing compute. | Paper |
| SimPO: Simple Preference Optimization with Reference-Free Reward | Meng et al. | 2024 | Average log-probability as implicit reward with target margin — cleaner than DPO. | Paper |
| Self-Rewarding Language Models | Yuan et al. (Meta) | 2024 | Models generate and evaluate own preference data for iterative self-improvement. | Paper |
Safety & Interpretability Papers¶
| Paper | Authors | Year | Description | Links |
|---|---|---|---|---|
| Constitutional AI: Harmlessness from AI Feedback | Bai et al. (Anthropic) | 2022 | Training helpful, harmless AI using AI-written critiques derived from a constitution. | Paper |
| Representation Engineering | Zou et al. (UCSD) | 2023 | Identifies and steers high-level concepts (honesty, power-seeking) in neural representations. | Paper |
| Towards Monosemanticity: Dictionary Learning for LMs | Bricken et al. (Anthropic) | 2023 | Sparse autoencoders decomposing polysemantic neurons into interpretable features. | Blog |
| Scaling Monosemanticity: Interpretable Features from Claude 3 Sonnet | Templeton et al. (Anthropic) | 2024 | 34M features including "Assistant" identity, emotions, and safety-relevant concepts. | Blog |
| Sleeper Agents: Training Deceptive LLMs That Persist Through Safety Training | Hubinger et al. (Anthropic) | 2024 | Deceptive backdoor behaviors survive RLHF, SFT, and adversarial training. | Paper |
| Weak-to-Strong Generalization | Burns et al. (OpenAI) | 2023 | GPT-2 supervising GPT-4 as proxy for "human supervising superintelligence." | Paper |
| AI Control: Improving Safety Despite Intentional Subversion | Greenblatt et al. (Redwood Research) | 2024 | Framework for evaluating safety against models actively trying to circumvent controls. | Paper |
| Sparks of AGI: Early Experiments with GPT-4 | Bubeck et al. (Microsoft Research) | 2023 | 155-page study arguing GPT-4 shows early sparks of AGI across diverse tasks. | Paper |
| Levels of AGI: Operationalizing Progress on the Path to AGI | Morris et al. (Google DeepMind) | 2023 | 6-level AGI taxonomy (Emerging to ASI) with performance and autonomy axes. | Paper |
Blogs and News¶
| Resource | Description |
|---|---|
| OpenAI Blog | Official blog from OpenAI with research updates and announcements. |
| Anthropic Research | Anthropic's AI safety and capabilities research publications. |
| Google DeepMind Blog | Research updates from Google DeepMind. |
| Meta AI Blog | Meta's AI research blog, including Llama and open-source releases. |
| HuggingFace Blog | Latest in open-source ML, NLP, and the HF ecosystem. |
| LangChain Blog | Updates on LangChain/LangGraph and LLM application patterns. |
| The Gradient | Perspectives on AI research and its implications. |
| Lilian Weng's Blog | In-depth technical posts on LLMs, agents, and AI research (by OpenAI). |
| Simon Willison's Blog | Prolific coverage of LLM tools, agents, and practical AI engineering. |
| The Alignment Forum | Hub for AI alignment research discussions and papers. |
| Transformer Circuits | Anthropic's mechanistic interpretability research publications. |