Cycle Seeds
Each cycle has its own seed pool — papers nominated by the community for that sprint. Vote for your favourites or nominate a new one.
May 2026
DeepSeek-R1: Incentivizing Reasoning via Reinforcement Learning
Suggested by VJAI Core Team
Trains LLMs to reason purely via RL without SFT cold start, achieving o1-level performance.
JEPA: Self-Supervised Learning via Joint-Embedding Predictive Architecture
Suggested by VJAI Core Team
LeCun's proposed architecture for world models that reasons in abstract representation space.
FlashAttention-3: Fast Attention for H100 GPUs
Suggested by VJAI Core Team
Leverages H100 hardware features (TMA, WGMMA) to push attention throughput to near-theoretical limits.
Scaling Laws for Neural Language Models
Suggested by VJAI Core Team
Empirical laws relating model performance to compute, data, and parameters — the blueprint for GPT-4.
RT-2: Vision-Language-Action Models
Suggested by VJAI Core Team
Co-fine-tunes a VLM on robot data so the same model reasons about scenes and outputs robot actions.
DreamerV3: Mastering Diverse Domains with World Models
Suggested by VJAI Core Team
A single algorithm that masters Atari, continuous control, and Minecraft from scratch using a learned world model.
Segment Anything Model 2 (SAM 2)
Suggested by VJAI Core Team
Extends SAM to video, enabling promptable, real-time segmentation of any object in any video.
Constitutional AI: Harmlessness from AI Feedback
Suggested by VJAI Core Team
Trains harmless AI using AI-generated feedback rather than human labeling, using a Constitutional set of principles.
Voyager: An Open-Ended Embodied Agent with LLMs
Suggested by VJAI Core Team
Lifelong learning agent in Minecraft that writes its own code to solve tasks, building a skill library over time.
Toolformer: Language Models That Can Use Tools
Suggested by VJAI Core Team
Self-supervised method to teach LLMs when and how to call external APIs (calculator, search, calendar).
Mechanistic Interpretability of Neural Networks
Suggested by VJAI Core Team
Reverse-engineering neural networks to understand circuits, features, and internal representations.
RLHF: Training Language Models to Follow Instructions with Human Feedback
Suggested by VJAI Core Team
InstructGPT — the paper that made GPT-3 follow instructions by fine-tuning with PPO on human preferences.
LLaMA 3: Open Foundation and Fine-Tuned Chat Models
Suggested by VJAI Core Team
Meta's fully open LLaMA 3 family — from 8B to 405B, with details on pretraining data, architecture, and alignment.
Gemini 1.5: Unlocking Multimodal Understanding Across Millions of Tokens
Suggested by VJAI Core Team
Google's Gemini 1.5 with a 1M token context window — how mixture-of-experts enables efficient long-context processing.
Mixture of Experts: Switch Transformer
Suggested by VJAI Core Team
Scales language models to trillion parameters with sparse Mixture-of-Experts, replacing dense FFN layers.
Stable Diffusion: High-Resolution Image Synthesis with Latent Diffusion
Suggested by VJAI Core Team
Moves diffusion process into a compressed latent space, enabling high-quality image synthesis on consumer hardware.
CLIP: Learning Transferable Visual Models From Natural Language Supervision
Suggested by Team Hanoi
Learns image-text representations from 400M pairs — zero-shot image classification rival to supervised baselines.
Upcoming Cycles
6 plannedJune 2026
Session: TBD
July 2026
Session: TBD
August 2026
Session: TBD
October 2026
Session: TBD
November 2026
Session: TBD
December 2026
Session: TBD
Shape What We Read Next
Every nomination and vote directly influences the papers we deep-dive into. Join the community and make your voice count.