Intern-Atlas: AI Method Evolution Graph

• Intern-Atlas is a method-centric research infrastructure that transforms AI papers into a graph of evolving methods with explicit causal links and verbatim evidence.
• It organizes research into a heterogeneous graph with method, paper, and stub nodes, enabling precise tracing of innovation lineage and resolution of methodological bottlenecks.
• By leveraging SGT-MCTS and detailed edge typing, Intern-Atlas achieves high accuracy in reconstructing method evolution, outperforming traditional citation-based approaches.

Intern-Atlas is a method-centric research infrastructure that converts the AI literature into a machine-consumable topology of methodological evolution. Rather than treating papers as atomic units linked by generic citations, it represents methods as first-class entities, attaches typed causal relations to citation links, and grounds those relations in verbatim source evidence. Built from 1,030,314 papers spanning AI conferences, journals, and arXiv preprints from 1965 to 2025, the system contains 9,410,201 semantically typed edges and is designed for AI scientists and automated research agents that must reason explicitly about lineage, bottlenecks, and progression paths (Wu et al., 30 Apr 2026).

1. Motivation and conceptual shift

Existing research infrastructure is described as fundamentally document-centric. Systems such as Google Scholar, Semantic Scholar, and OpenAlex organize the literature around papers and citations, but they do not encode which specific method variants are extended, replaced, or adapted; which bottlenecks drove a transition; or how successive innovations are causally related at the method level (Wu et al., 30 Apr 2026). Intern-Atlas addresses this gap by shifting the fundamental unit from papers to methods and by promoting untyped citations into typed, causal edges grounded in verbatim evidence.

This design is motivated in part by the needs of AI-driven research agents. The underlying claim is not merely that unstructured text is inconvenient, but that LLM-based agents cannot reliably reconstruct method-evolution topologies from prose alone: parametric memory is lossy, autoregressive inference does not support explicit graph traversal, and an agent cannot cleanly distinguish between a genuine gap in the literature and a gap in its internal representation (Wu et al., 30 Apr 2026). In ideation settings, this absence of explicit topology is associated with shallow stacking of popular terms, poor novelty, and unreliable assessments of feasibility and significance.

A common misconception is that a citation graph already contains the needed signal if one searches it carefully enough. Intern-Atlas rejects that assumption. Its central claim is that methodological evolution requires explicit method-level semantics, typed causal relations, temporal constraints, and evidence records. Citation counts or neighborhood structure alone are insufficient for reconstructing why a method emerged and what limitation it was intended to resolve.

2. Graph formalization and method-level semantics

Intern-Atlas defines a heterogeneous typed graph

$G = (V, E, \tau, \rho),$

where $V = V_P \cup V_M \cup V_S$. Paper nodes $V_P$ correspond to parsed full texts within the 1965–2025 collection window and carry publication year $\tau(v)$. Method nodes $V_M$ correspond to canonical method entities with canonical names, aliases, timestamps for first appearance, venues, and domains or subfields. Stub nodes $V_S$ are metadata-only placeholders for cited works outside the corpus window; there are 3,173,187 such stubs, introduced to preserve reachability in lineage traversal (Wu et al., 30 Apr 2026).

Method identity resolution is mediated by an alias registry $A$, defined as a substring-based, case/punctuation-normalized mapping from surface forms to canonical method nodes with longest-match priority, version consolidation rules, and a curated negative list for ambiguous names. The released graph contains 8,155 canonical methods and 9,545 aliases (Wu et al., 30 Apr 2026).

The edge space is partitioned into seven causal or contextual types:

Edge type	Semantics
extends	Adds a new capability or component on top of the cited method
improves	Optimizes an existing component without altering the core formulation
replaces	Substitutes a load-bearing component with a different mechanism
adapts	Ports the method to a new domain, modality, or task setting
uses_component	Re-uses the cited method as an auxiliary module
compares	Baseline or head-to-head comparison
background	Non-methodological context

The strong-causal subset is

$$T_{\text{strong}} = \{\text{extends, improves, replaces, adapts}\},$$

and it induces the lineage subgraph. Non-strong types are retained for retrieval and evaluation (Wu et al., 30 Apr 2026).

Each causal edge stores an evidence record

$\rho(e) = (b_e, m_e, t_e, c_e),$

where $b_e$ is the bottleneck quote, $V = V_P \cup V_M \cup V_S$0 the mechanism quote, $V = V_P \cup V_M \cup V_S$1 the trade-off quote, and $V = V_P \cup V_M \cup V_S$2 the LLM-reported confidence. Bottlenecks are typed by a 14-axis taxonomy: {computational complexity, memory efficiency, parallelization, accuracy, generalization, scalability, data efficiency, training stability, inference speed, expressiveness, simplicity, robustness, hyperparameter sensitivity, training complexity} (Wu et al., 30 Apr 2026).

Intern-Atlas also derives a projected method-level DAG

$V = V_P \cup V_M \cup V_S$3

deterministically from citation-typed edges. The mappings are extends/improves $V = V_P \cup V_M \cup V_S$4 variant_of, adapts/replaces $V = V_P \cup V_M \cup V_S$5 specializes, and uses_component $V = V_P \cup V_M \cup V_S$6 component_of with reversed direction. Two additional method-level relations, optimizes and inspired_by, are seeded manually and are often not citation-anchored (Wu et al., 30 Apr 2026).

3. Corpus construction, extraction pipeline, and evidence grounding

The corpus covers 1,030,314 AI papers from conferences, journals, and arXiv, including ML venues such as NeurIPS, ICML, and ICLR; CV venues such as CVPR, ICCV, and ECCV; NLP venues such as ACL, EMNLP, and NAACL; AI venues such as AAAI and IJCAI; and IR/data venues such as SIGIR and KDD (Wu et al., 30 Apr 2026). Parsing uses Nougat v1.0 with GROBID v0.8.0 as fallback for approximately 3.7% of papers, producing structured JSON with section headings and references. Downstream extraction consumes the Introduction, Method/Approach, and Related Work sections.

Method-level entity detection begins from a seed list of 247 methods from Papers With Code and 18 pilot reviews from 2020–2024. LLM expansion scans Method sections for proper-noun method names, retaining candidates that co-occur in at least three distinct papers and are adjudicated by the authors (Wu et al., 30 Apr 2026). This suggests a precision-oriented normalization strategy: the alias registry privileges conservative consolidation and ambiguity control over maximal surface-form recall.

Edge typing uses a two-phase LLM protocol with Qwen3.6-35B-A3B as the production model and a Claude-based audit model. Phase 1 classifies each resolved reference into an edge type and extracts an evidence snippet with tuning for high recall; references with no methodological relation are assigned background. Phase 2 completes the evidence record $V = V_P \cup V_M \cup V_S$7 with verbatim bottleneck, mechanism, and trade-off quotes, bottleneck dimension tags, and confidence, as well as impact fields such as improvement_dim and sacrifice_dim and a mechanism type (Wu et al., 30 Apr 2026). Outputs must exact-match substrings in the citing paper.

Deterministic post-validation then removes edges that fail verbatim substring checks, violate publication-year ordering, or conflict with opposite-direction duplicates. Every non-background causal edge therefore remains auditably linked to specific paper sections. This is a central architectural feature: Intern-Atlas is not simply an LLM-produced latent graph, but a graph constrained by exact evidence matching and temporal consistency.

4. Methodological evolution chains and SGT-MCTS

A methodological evolution chain is defined as a directed path in the strong-causal subgraph,

$V = V_P \cup V_M \cup V_S$8

subject to temporal monotonicity, edge-type compatibility, and acyclicity by construction. Hard-filtered temporal violations with $V = V_P \cup V_M \cup V_S$9 are disallowed (Wu et al., 30 Apr 2026). Chain ranking combines length, mean confidence, and rollout visit counts:

$V_P$0

To recover such chains, the paper introduces Self-Guided Temporal Monte Carlo Tree Search, or SGT-MCTS. Its selection rule is

$V_P$1

with

$V_P$2

The temporal coherence function $V_P$3 is piecewise and peaks at 1–3 years. It assigns 0.40 for $V_P$4, 0.85 for $V_P$5, 1.00 for $V_P$6, 0.80 for $V_P$7, $V_P$8 for $V_P$9, and 0.70 if timestamps are missing (Wu et al., 30 Apr 2026).

The search runs separately backward and forward from a seed method. Per direction, the budget is $\tau(v)$0 iterations, the rollout depth cap is $\tau(v)$1, the UCT exploration constant is $\tau(v)$2, the prior weight is $\tau(v)$3, and the dead-end penalty is $\tau(v)$4. Expansion adds the highest-confidence untried child passing temporal and cycle filters; rollouts greedily maximize $\tau(v)$5; and final chain assembly splices the top-5 backward with the top-5 forward paths through the seed, deduplicating by node-Jaccard threshold $\tau(v)$6 (Wu et al., 30 Apr 2026).

The computational complexity per seed is approximately

$\tau(v)$7

where $\tau(v)$8 is the average branching factor in the strong-causal graph. Temporal pruning reduces the effective branching factor by removing temporally implausible edges (Wu et al., 30 Apr 2026).

On a benchmark of 30 survey-derived method-evolution graphs comprising 2,268 nodes, 1,462 edges, and 133 reference chains, the static graph achieves Node Match Ratio 91.0%, Edge Reachable Ratio 89.7%, and Path Semantic Correctness 92.0%. For lineage reconstruction, Beam@10 yields NR=44.9, ER=23.2, CAS=44.9, whereas SGT-MCTS yields NR=84.8, ER=79.0, CAS=84.8, improving NR by +39.9 points, ER by +55.8 points, and CAS by +39.9 points (Wu et al., 30 Apr 2026). A case study on residual ConvNet evolution reconstructs the ordered backbone chain VGG $\tau(v)$9 ResNet $V_M$0 ResNeXt $V_M$1 ConvNeXt $V_M$2 ConvNeXt V2, while beam and random baselines miss early ancestors or recover only local subsequences.

5. Retrieval, idea evaluation, and idea generation

Intern-Atlas exposes a unified retrieval operator

$V_M$3

which returns method matches $V_M$4, top-$V_M$5 papers $V_M$6 with $V_M$7 via hybrid BM25 plus alias counts, causal edges $V_M$8, and bottleneck records $V_M$9 (Wu et al., 30 Apr 2026). This makes the graph queryable as an evidence-linked substrate for both analytical and generative workflows.

For idea evaluation, the system uses deterministic scoring rather than a purely LLM-based judge. Per-dimension scores are defined as

$V_S$0

for $V_S$1 with $V_S$2 and aggregation weights $V_S$3. Cross-dimensional aggregation uses

$V_S$4

where $V_S$5 includes conjunctive priors such as a $V_S$6 penalty for high Novelty but low Feasibility (Wu et al., 30 Apr 2026). Duplicate-risk detection combines dense BGE, sparse BM25 via RRF, and reranking with an ms-marco cross-encoder; step penalties are applied at fused similarity thresholds of at least 0.55, 0.65, 0.75, and 0.85.

On the Strata dataset of 1,200 papers, with 300 papers per stratum, overall scores decrease monotonically from Top-tier 8.48 to Core 7.83, Workshop 6.85, and Rejected 5.84, with the largest gaps in Significance and Validity. On 100 ideas scored by 10 AI PhDs, Spearman $V_S$7 for overall alignment is 0.81 for Intern-Atlas versus 0.58 for a pure LLM judge; the strongest gains appear on Novelty, 0.84 versus 0.52, and Significance, 0.82 versus 0.55 (Wu et al., 30 Apr 2026).

Idea generation is strategy-driven and structurally grounded. Gap patterns are extracted from retrieval context and lineage chains without an LLM, and proposals are then produced by strategy-conditioned prompting,

$V_S$8

with strategies Bottleneck Resolution, Trend Extrapolation, Cross-Pollination, and Paradigm Challenge (Wu et al., 30 Apr 2026). Each proposal must include an evidence certificate $V_S$9 with exact-match verification of the bottleneck quote; invalid outputs are discarded and replaced by a deterministic minimal proposal.

Across knowledge sources, Intern-Atlas attains an overall idea-quality score of 7.20, compared with 6.18 for Semantic Scholar, 6.03 for OpenAlex, 6.15 for BM25 RAG, and 5.78 for No-KB. It leads in Novelty at 6.37, Significance at 6.30, and Validity at 6.26. In blind pairwise expert evaluation, it wins overall by 88.0% against No-KB, 82.0% against OpenAlex, and 81.0% against BM25 RAG (Wu et al., 30 Apr 2026).

6. Positioning, release, and limitations

Intern-Atlas is positioned against both citation networks and generic knowledge graphs. The paper argues that systems such as MAG, OpenAlex, and Semantic Scholar are document-centric, largely untyped at the method level, and lack verbatim causal annotations for bottlenecks and mechanisms (Wu et al., 30 Apr 2026). By contrast, Intern-Atlas provides typed methodological edges, temporal coherence, and exact evidence records. This does not imply that citation graphs are obsolete; rather, citation graphs become one input signal inside a more explicit causal topology.

The public interface is available at https://intern-atlas.opendatalab.org.cn/, and the dataset is released at https://huggingface.co/datasets/OpenRaiser/Intern-Atlas. Code and prompts are released, extraction uses deterministic decoding with temperature 0, and batched processing scales over the full 1,030,314-paper corpus. The evaluator is zero-trainable-parameter for reproducibility, and an optional LLM adjudicator is allowed only to lower scores through a one-sided veto (Wu et al., 30 Apr 2026).

Several limitations are stated explicitly. Phase-1 typing accuracy varies from 70.4% for the production model to 93.0% for the audit model on fine-grained distinctions such as extends versus improves. The temporal coherence function was calibrated on post-2015 AI and may require adjustment in disciplines with different innovation cadences. The fixed 14-axis bottleneck taxonomy may lag emerging dimensions, and alias resolution favors precision through a manually maintained negative list, so some ambiguous overlaps remain (Wu et al., 30 Apr 2026). The paper also notes ethical concerns, including possible amplification of citation biases. Proposed mitigations include verbatim provenance, public reliability audits, and a commitment not to use the system for author- or institution-level ranking.

Future work includes learned priors over edge semantics and temporal gaps, richer method ontologies, calibration for cross-field transfer, and integration with experimental platforms for automated discovery (Wu et al., 30 Apr 2026). A plausible implication is that Intern-Atlas should be read not simply as a literature-mining system but as a proposal for a new infrastructural layer: an explicit, evidence-grounded representation of how methods change over time, intended to support automated scientific reasoning rather than only human browsing.