VJ
VJAI Paper Hub
HomeArchive
Knowledge Hub

Paper Archive

Papers the group has read, presented, and reproduced — independent of the nomination cycle. Each entry includes session notes, reproduction code, and a vibe score from the group.

2
Papers
2
Reproduced
2
Venues
82%
Avg Vibe
2 of 2
TransformersNLPFoundations
Reproduced

NeurIPS 2017 · Duc Vo

Attention Is All You Need

  • Self-attention replaces recurrence, enabling full parallelization during training
  • Multi-head attention lets the model jointly attend to information from different representation subspaces
  • Positional encodings inject sequence order without recurrent structure
  • Achieved SOTA on WMT 2014 EN-DE translation with far less compute than previous models
85% hackable
LLMRLReasoning
Reproduced
  • Pure RL with a simple accuracy + format reward produces emergent reasoning without any SFT warm-up
  • GRPO (Group Relative Policy Optimization) stabilizes training without a value network
  • R1-Zero shows self-reflection and multi-step reasoning emerging spontaneously from RL
  • Distilled 7B/8B models are competitive with o1-mini on MATH-500 and LiveCodeBench
78% hackable