Updates
A pipeline sweeps arXiv and Hugging Face daily papers every three days, filters for genuine relevance, scores what survives, and proposes it for review. Nothing enters the atlas automatically, and no explanatory claim is written without a fetched source.
The pipeline
Last run 2026-09-10, which processed 321 new preprints:
The aggressive filtering is deliberate. World-action models, depth estimation and language-model post-training papers share almost all of generative vision's vocabulary, and without a hard topic gate they dominate the queue.
Waiting for review (5)
Proposed by the last sweep, not yet part of the atlas. Each shows why it was surfaced.
15.0 Interpreting Object-Dependent Concept Brittleness in Text-to-Image Diffusion Models
core topic appears in the title · affiliated with a major lab or university · code or weights released
15.0 Studying Image Tokenizers as Visual Languages in Unified Multimodal Models
core topic appears in the title · affiliated with a major lab or university · title suggests a substantive contribution
13.0 FlowCPO: A Unified Divergence View of Preference Alignment for Flow Models
affiliated with a major lab or university · title suggests a substantive contribution
11.0 Efficient Fairness Auditing Across Guidance Scales in Text-to-Image Diffusion Models via Causal
core topic appears in the title · affiliated with a major lab or university · title suggests a substantive contribution
Coverage, stated honestly
What is actually written versus what is still only metadata. The validator tracks this, so these numbers cannot quietly drift.
Standards
Every paper is checked
203 of 204 entries have had their arXiv record fetched and their title matched. The remainder is a journal paper with a DOI and no arXiv id. Six wrong titles have been caught this way so far.
Automation cannot promote
The pipeline never proposes a paper above the lowest tier, and never writes to the atlas directly. A human or a review session decides.
Gaps are stated
Where a line has thin coverage, no independent replication, or an unverified claim, its page says so rather than omitting it.