Research

OG card: chart-to-art exploded axonometric of the JOLT training figure: one set of weights at the base, the Teacher and Student roles above it, six generations per role on top, with crossing dashed paths for each role scoring the other
Research21 Oct 2026A good teacher meets the student where they areA new self-distillation method that jointly trains teacher and student to improve model performance and training efficiency on complex reasoning and agentic tasks.
  1. 23 Sept 2026Escaping SPACE: Part IRed-teaming VM isolation and network confinement for AI agents.
  2. 12 Oct 2026Building shared workspaces for people and agentsHow Perplexity Projects keep long-running work persistent, concurrent, and traceable.
  3. 7 Oct 2026Multimodal embeddings beyond a single vectorA new family of late-interaction models for text and image retrieval, with a shared embedding space across model sizes.

All publications

  1. BrowseSafe++: Building stronger defences against real-world prompt injection
  2. A good teacher meets the student where they are
  3. Building shared workspaces for people and agents
  4. Multimodal embeddings beyond a single vector
  5. Contextual embedding beyond the gold passage
  6. Photon: Building a Retrieval and Ranking Engine From Scratch
  7. Escaping SPACE: Part I
  8. Learning from real-world experience
  9. CobbleDB: Rebuilding AI Search Storage for Lower Latency and Cost
  10. Q2D-Web: Evaluating First-Stage Retrievers at Scale