Research map / Large language models / Retrieval and knowledge
Panini: Continual Learning in Token Space via Structured Memory
Shreyas Rajesh, Pavan S Holur, Mehmet Yigit Turali, Chenda Duan, Vwani Roychowdhury
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| query only the structured GSW QA-pair memory with chain retrieval instead of retrieving verbatim chunks | PANINI · Six QA benchmarks average | F1: 53.3 -> 56.1 | Improved |
| feed short QA pairs from GSW to the answer model instead of document chunks | PANINI · Six QA benchmarks average | answer-context tokens: 705 -> 320 | Improved |
| use PANINI memory in place of chunk retrieval on PopQA | PANINI · PopQA | F1: 59.9 -> 57.6 | Hurt |
| use PANINI instead of dense retrieval on Platinum answerable questions | PANINI · Platinum average, answerable | F1: 61.7 -> 79.9 | Improved |
| use PANINI instead of HippoRAG 2 on unanswerable questions | PANINI · Platinum average, unanswerable | refusal accuracy: 58.5 -> 72.8 | Improved |
| reduce beam width from 5 to 3 (ablation) | PANINI RICR · six QA benchmarks | accuracy and tokens: about 25% fewer tokens | No clear effect |
| use single beam (B=1) in chain retrieval (ablation) | PANINI RICR · six QA benchmarks | accuracy | Mixed |
| replace geometric-mean chain scoring with similarity-only, combined or greedy last-hop selection (ablation) | PANINI RICR · six QA benchmarks | accuracy | Improved |
| swap Search-R1 chunk retrieval for the GSW index without extra training (ablation) | Search-R1 · six QA benchmarks | performance: improves on 4 of 6 datasets | Mixed |
| replace proprietary components with open-source models (Qwen3 decomposer and answerer) (ablation) | PANINI · six QA benchmarks | F1 | Mixed |
| build GSW with smaller open-source models (Qwen3-8B/14B, GPT-OSS-120B) (ablation) | PANINI · six QA benchmarks | F1 | Mixed |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- HippoRAG 2 — HippoRAG: Neurobiologically Inspired Long-Term Memory for Large Language Models (NeurIPS 2024) · better on 10 of 10 shared results
- RAPTOR — RAPTOR: Recursive Abstractive Processing for Tree-Organized Retrieval (ICLR 2024) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Beyond Fact Retrieval: Episodic Memory for RAG with Generative Semantic Workspaces (AAAI 2026 · similarity 0.89)
- Retrieval is Accurate Generation (ICLR 2024 · similarity 0.88)
- From RAG to Memory: Non-Parametric Continual Learning for Large Language Models (ICML 2025 · similarity 0.87)
- K-DeCore: Facilitating Knowledge Transfer in Continual Structured Knowledge Reasoning via Knowledge Decoupling (NeurIPS 2025 · similarity 0.87)
- Retrieval is Not Enough: Enhancing RAG through Test-Time Critique and Optimization (NeurIPS 2025 · similarity 0.87)
- CMT: A Memory Compression Method for Continual Knowledge Learning of Large Language Models (AAAI 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
