roychowdhuryresearch/Panini-Benchmarks

Panini: Continual Learning in Token Space via Structured Memory

Extracted Generative Semantic Workspace (GSW) representations and curated evaluation splits from Panini, provided for ease of replication and future research.

Contents

GSW Networks (gsw_networks/)

Structured semantic representations extracted from document corpora using the Panini/GSW framework. Each GSW captures entities, their roles/states, and verb-phrase relationships as question-answer pairs.

Dataset	Documents	File	Description
2WikiMultihopQA	6,119	2wiki.jsonl	Multi-hop reasoning across Wikipedia articles
MuSiQue	11,626	musique.jsonl	Complex multi-step question answering
HotpotQA	9,811	hotpotqa.jsonl	Multi-hop QA with supporting facts
Natural Questions (REAR)	9,592	nq_rear.jsonl	Single-hop factual retrieval
PopQA	8,676	popqa.jsonl	Popular entity question answering

Format: Each line is a JSON object with:

{
  "doc_id": "doc_0",
  "source_file": "gsw_0_0.json",
  "gsw": {
    "entity_nodes": [...],
    "verb_phrase_nodes": [...],
    "space_nodes": [...],
    "time_nodes": [...],
    "similarity_edges": [...],
    "space_edges": [...],
    "time_edges": [...]
  }
}

Corpus source: All evaluation corpora follow the HippoRAG v2 splits. GSWs were generated using GPT-4o with the Panini operator (see paper for details).

Paper Results

Main Results (GPT-4o-mini)

	Simple QA		Multi-Hop QA
Retrieval	NQ	PopQA	MuSiQue	2Wiki	HotpotQA	LV-Eval	Avg
None	52.7	22.7	22.0	36.3	41.0	5.0	29.9
BM25	56.0	53.0	25.3	35.4	54.2	7.8	38.6
BM25 + reranker	58.2	56.8	29.9	44.1	63.4	8.1	43.4
NV-Embed-v2 (7B)	59.9	55.8	46.0	60.8	71.0	10.0	50.6
Qwen3 (8B)	59.1	59.8	39.4	56.2	69.2	11.7	49.2
Qwen3 + reranker	61.4	59.9	43.7	57.9	68.2	11.8	50.5
RAPTOR	54.5	55.1	39.2	48.4	64.7	9.2	45.2
GraphRAG	55.5	51.3	42.0	61.0	67.6	11.0	48.1
HippoRAG	52.2	56.2	35.9	67.3	60.0	7.6	46.5
HippoRAG 2	60.0	55.7	49.3	69.7	71.1	14.0	53.3
IRCoT	26.4	53.8	44.2	64.9	64.3	9.1	43.8
Search-R1	47.9	49.7	41.1	64.9	68.6	11.5	47.3
Panini	67.4	57.6	52.3	72.4	71.9	14.8	56.1

Platinum Evaluation Splits (platinum/)

To evaluate reliability under missing evidence, we curate Platinum splits for MuSiQue and 2WikiMultihopQA. Each question is verified for answerability through multi-model consensus and labeled as answerable (gold answer is derivable from supporting documents) or unanswerable (evidence is missing or ambiguous). This allows measuring both QA accuracy and abstention quality. See Section 4.2 of the paper for full construction details.

Split	Answerable	Unanswerable	Total
2Wiki Platinum	906	94	1,000
MuSiQue Platinum	766	153	919

Files:

• 2wiki_platinum.json — Answerable 2Wiki questions
• 2wiki_unanswerable.json — Unanswerable 2Wiki questions
• musique_platinum.json — Answerable MuSiQue questions
• musique_unanswerable.json — Unanswerable MuSiQue questions

Platinum Results (GPT-4o-mini)

Retrieval	MuSiQue Ans	MuSiQue Unans	2Wiki Ans	2Wiki Unans	Avg Ans	Avg Unans
BM25	35.0	78.4	37.9	79.8	36.4	79.1
BM25 + reranker	42.3	74.5	47.9	78.7	45.1	76.6
Qwen3 (8B)	53.3	63.4	63.5	70.9	58.4	67.2
Qwen3 + reranker	59.7	62.8	63.8	67.7	61.7	65.2
HippoRAG 2	63.6	50.3	81.5	66.7	72.5	58.5
Panini	75.0	72.6	84.8	73.1	79.9	72.8

Usage

import json

# Load GSW networks for a dataset
gsws = []
with open("gsw_networks/2wiki.jsonl") as f:
    for line in f:
        gsws.append(json.loads(line))

# Load platinum evaluation split
with open("platinum/2wiki_platinum.json") as f:
    platinum_questions = json.load(f)

With the Panini/GSW-Memory package

from gsw_memory.qa.gsw_tools import GSWTools

# Write GSWs to individual files for GSWTools
import os, json
os.makedirs("gsw_files", exist_ok=True)
with open("gsw_networks/2wiki.jsonl") as f:
    for line in f:
        record = json.loads(line)
        doc_dir = f"gsw_files/{record['doc_id']}"
        os.makedirs(doc_dir, exist_ok=True)
        with open(f"{doc_dir}/{record['source_file']}", "w") as out:
            json.dump(record["gsw"], out)

# Use GSWTools for retrieval
import glob
gsw_files = glob.glob("gsw_files/doc_*/*.json")
tools = GSWTools(gsw_files)
tools.build_index()
results = tools.search_gsw_bm25("Lothair II", limit=5)

Attribution

• Evaluation corpora follow splits from HippoRAG v2 (Gutiérrez et al., 2025)
• 2WikiMultihopQA from Ho et al., 2020
• MuSiQue from Trivedi et al., 2022
• HotpotQA from Yang et al., 2018
• Natural Questions from Kwiatkowski et al., 2019
• PopQA from Mallen et al., 2023
• Platinum curation methodology inspired by Vendrow et al., 2025

Citation

@misc{rajesh2026paninicontinuallearningtoken,
      title={Panini: Continual Learning in Token Space via Structured Memory},
      author={Shreyas Rajesh and Pavan Holur and Mehmet Yigit Turali and Chenda Duan and Vwani Roychowdhury},
      year={2026},
      eprint={2602.15156},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2602.15156},
}

License

MIT