Research map / Graph learning / Knowledge and temporal graphs

Intern-Atlas: A Methodological Evolution Graph as Research Infrastructure for AI Scientists

NeurIPS 2026 · Graph learning › Knowledge and temporal graphs · approach: Social and behavior graphs · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Builds Intern-Atlas, a typed methodological evolution graph over ~1M AI papers with evidence-grounded causal edges, and uses it for SGT-MCTS lineage reconstruction, graph-grounded idea evaluation, and evidence-certified idea generation.

Problem

Existing scholarly infrastructure is document-centric and its citation edges carry no methodological semantics, which limits AI research agents that need structured knowledge of how methods evolved for lineage tracing, idea evaluation, and idea generation.

Approach

Extract a method-centric graph with seven typed causal edges, each grounded by verbatim bottleneck/mechanism/trade-off quotes checked by substring matching, then run an MCTS variant with a confidence-times-temporal-coherence prior for lineage search and deterministic graph statistics for idea scoring and generation.

Key results

• SGT-MCTS reaches NR 84.8, ER 79.0, CAS 84.8 on lineage reconstruction versus 44.9 NR for Beam@10.
• Graph construction recovers 91.0% of survey methods (NMR) and 89.7% of reference edges (ERR), with 92.0% PSC.
• Intern-Atlas idea evaluator correlates 0.81 overall with expert ratings versus 0.58 for a pure LLM judge; generated ideas win 81-88% of blind pairwise comparisons against baselines.

Surprising findings

• Pure LLM judges' novelty scores correlate negatively with scientific impact, motivating deterministic graph-based scoring.
• Random walk achieves only 0.7 ER, far below beam search, showing unguided traversal fails on the graph.

Limitations

• Survey-derived benchmark uses LLM extraction plus manual audit and surveys as proxies for expert consensus; construction and reference procedures differ in granularity.
• Hand-designed scoring weights and penalties are not learned, and the evaluator's correlation with experts is on only 100 profiles.
• Graph extraction depends on a single open LLM (Qwen3.6-35B-A3B) with edge typing and evidence quality not fully validated beyond a post-checker.

Experimental setup

Datasets: Intern-Atlas corpus of 1,030,314 AI papers (1965-2025) Strata Dataset (1,200 papers) 30 survey-derived method-evolution graphs

Benchmarks: Method-evolution benchmark (30 surveys) Strata Dataset 100 research questions (idea generation)

Metrics: NMR ERR PSC NR ER CAS Spearman correlation pairwise win rate Overall/dimension scores

Baselines: Beam search Random walk No-KB OpenAlex Semantic Scholar BM25 RAG Pure LLM-as-Judge

Builds on: Monte Carlo tree search (UCT) HopRAG BM25 Main Path Analysis IdeaBench

Closest papers by meaning

• InnovatorBench: Evaluating Agents’ Ability to Conduct Innovative AI Research (ICLR 2026 · similarity 0.87)
• SciResearchBench: Benchmarking AI Agents on Complex Scientific Literature Discovery (NeurIPS 2026 · similarity 0.87)
• AstaBench: Rigorous Benchmarking of AI Agents with a Scientific Research Suite (ICLR 2026 · similarity 0.86)
• FML-bench: A Controlled Study of AI Research Agent Strategies from the Perspective of Search Dynamics (NeurIPS 2026 · similarity 0.86)
• AI-Researcher: Autonomous Scientific Innovation (NeurIPS 2025 · similarity 0.86)
• Agent-Native Research Artifacts (NeurIPS 2026 · similarity 0.85)