qasper

Bi-hemispheric synthesis of ultrametric p-RoPE sparse attention (Gemma 4 31B, Llama 3.1) & neuromorphic test-time reasoning. Solves 38/400 ARC-AGI-1 & 25/25 ARC-AGI-3 games (63/183 levels, 4 full wins, 19.92% RHAE) with zero weight training. Lean 4 verified (0 sorry).

Evidence-first RAG with inspectable hybrid retrieval, diagnostics, telemetry, and reproducible evaluation.

Evidence localisation and failure attribution for RAG evaluation: inside the right document, do retrievers rank the evidence passage first? Seven localisers on QASPER and Natural Questions, a human-reviewed attribution study and a gold-span completeness audit. Code and results of the submitted paper 'Right document, wrong passage' (tag v1.1.0).

A scientific question-answering system built on QASPER with retrieval, reranking, extractive QA, and grounded generation.

Pipeline Code for ASL Variability analysis

Long-document QA with short-context LLMs where every answer claim is checked against its exact source passage (QASPER, distractor noise, NLI verification).

평가 주도 RAG for AI/ML 논문 — 검색·답변 품질을 recall@k·faithfulness로 측정하고 ablation·회귀로 개선하는 평가 하네스

Benchmark for verifiable multi-hop reasoning over a quote-backed knowledge graph of AI papers, covering false premises, missing evidence and conflicting sources.

Qasper 논문 QA 생성 단계를 Qwen2.5-1.5B에 파인튜닝하며 full FT·LoRA·QLoRA의 품질/비용 트레이드오프를 같은 척도로 측정한 평가 주도 프로젝트

Comparación interactiva de estrategias de chunking para RAG sobre QASPER, evaluando calidad de recuperación y costo computacional.

A LangGraph review loop for cited answers about research papers. On QASPER it ties a single pass over 5 passages and loses to a single pass over 10 or 15. Every model reply is recorded, so CI replays the evaluation without a key.