Reports of failures and limitations are scattered across individual papers. This view brings related evidence together so readers can compare recurring difficulties and trace them back to their sources.
How we built it
We used LLM-assisted extraction to collect evidence from full texts, reconcile duplicate passages, and group related evidence into challenges while retaining source passages and paper references.
Choose a challenge that interests you, read its definition, and expand its supporting evidence. Then explore the semantic map and yearly views to compare where and when related evidence appears.
About this view: evidence, maps, and time
What counts as evidence?
Extracted passages from error analyses, limitations, and related experimental observations in the selected papers. Challenge groupings are LLM-assisted interpretations; the original passages remain available for inspection.
What do the counts mean?
Evidence rows are not unique papers: one paper can contribute several passages. Counts describe this curated collection, not how often a problem occurs across the field.
What do the maps show?
The Sankey connects sources, communities, and challenge groups; it does not demonstrate causal spread between communities. The semantic map projects evidence embeddings into two dimensions. Nearby points and neighbor lines suggest semantic similarity, not citation or causal relationships.
What can yearly changes tell us?
They show how the evidence in this collection is distributed over time. More or fewer passages do not establish that a problem is worsening or has been solved. Publication metadata is used when verified, with arXiv v1 dates as a fallback.
Sankey: Event → Community → Cluster
Sankey detail
Semantic Cluster Map
Time slice
Yearly Dynamics
Counts are evidence rows in this curated Stage 5 set. Use them as signals for persistence, emergence, and fading, not as field-wide prevalence.
Challenge Definitions
Canonical cluster definitions used by the Sankey, semantic map, and yearly dynamics views.
Challenge Evolution Analysis
Charts now include corrected publication years and arXiv v1 dates. The narrative below predates this metadata correction; its temporal claims have not been regenerated.
LLM-assisted synthesis over the canonical challenge clusters, yearly dynamics, community/event flow, and representative evidence IDs.