PANINI / Generative Semantic Workspace

Continual learning in token space via structured memory (ICML 2026)

PANINI is a non-parametric continual-learning framework that encodes documents into Generative Semantic Workspaces (GSW) — entity- and event-aware networks of atomic question–answer (QA) pairs that enable reasoning-grounded inference over stored experience without continual parameter updates.

What we built

  • Reasoning Inference Chain Retrieval (RICR): a beam-search procedure that performs one-shot query decomposition and follows multi-hop inference chains through GSWs to assemble compact evidence.
  • Dual indexing for scalable memory access: a BM25 index over entity + role/state descriptors paired with a dense vector index over QA pairs, enabling targeted retrieval of QA evidence rather than long text chunks.
  • End-to-end open-source pipeline: GSW construction → retrieval → answering, supporting reproducible multi-hop QA.

Headline results

  • Top average performance across six QA benchmarks.
  • 2–30× fewer answer-context tokens than competitive baselines.
  • Improved reliability via reduced unsupported answers on curated unanswerable queries.

Accepted as a poster at ICML 2026. Co-authored with S. Rajesh, P. S. Holur, C. Duan, and V. Roychowdhury.

References