Add research lenses, Round 4 batches, and math rendering pipeline#23
Merged
Conversation
The card pipeline ran marked before KaTeX, so markdown emphasis rules consumed any formula containing * or _ (e.g. Q^*(s,a)) and split the $...$ span, leaving KaTeX nothing clean to render. Add docs/js/math-katex.js: marked extensions that tokenize $...$ and $$...$$ before emphasis runs and render each span straight to KaTeX HTML. Math now survives emphasis, blockquotes, tables and list items. - tools/test_math_render.cjs: deterministic Node regression test (6 cases) - tools/audit_card_links.py + .link_baseline.json: ratcheting check for inert internal cross-links (current baseline 91 dead / 12 orphan cards) so the count can only go down; wired both into validate.yml - broaden validate.yml path filter to docs/** so JS changes are covered
Recovers broken navigation that the audit surfaced: - Add graph nodes for 8 fully-written but unreachable orphan cards (CARLA Leaderboard, Apollo/Autoware, and six methodology-primitive 'move' cards), each wired to a real referrer so it is not isolated. The prose was already there; only the node was missing, so every link to it rendered inert. - Repoint confirmed alias links across 12 cards (simclr_moco -> simclr_mocov3, diffusion_policy/3ddiff_policy -> diffusion_policy_chi2023, 2309.17080_gaia1 -> gaia1, llama2 -> llama, gaia2 -> gaia1, zhao_shiyu_rl -> zhao_rl, vad -> vadv2). - Rebuild graph_extended.json via merge + repair (507 nodes / 1448 edges). Dead internal cross-links: 91 -> 45; orphan cards: 12 -> 4. Baseline ratcheted.
…lens The bottom-of-card panel restated the card summary and dressed edge counts up as 'why this matters' — neither dynamic nor insightful. Replace it with: - An authored, per-node Research lens (docs/data/research_lens.json, built by tools/build_research_lens.py): four columns of real reasoning — load-bearing assumption, failure boundary, a falsifiable next experiment, and the cross-domain isomorphism — rendered through the same markdown+KaTeX pipeline so formulae and intra-atlas links work. Ships 16 flagship lenses (offline RL, imitation/DAgger, safety-Lagrangian, world models, UniAD, PPO, Transformer, DQN, AlphaZero, CQL, conditional-generation planning, four paradigms). - Honest navigation for every node: downstream next-reads, upstream provenance as links (not a count), and the open problems it answers. Nodes without a lens show only this, instead of a synthetic 'insight'. All 16 lenses' internal links verified to resolve; 24 KaTeX spans render with 0 errors on the offline-RL lens; nav-only fallback confirmed on non-lens nodes.
A single long $$…$$ (e.g. CQL's logsumexp objective) was 533px wide in a 430px side panel and pushed the entire card body into horizontal overflow, breaking the layout. Give display-math containers (.katex-display, .math-block, .math-display) overflow-x:auto + max-width:100% so a wide equation scrolls within its own box instead of widening the panel. Verified: card-body horizontal overflow gone (533->430), equation still fully readable by scroll.
…uster) A four-way domain cluster authored full depth cards for the canonical papers the graph cited but never defined, eliminating the last dead links and the hollow gaps: - Imitation/IRL/PG theory: GAIL, AIRL, Ng-Russell IRL, MaxEnt IRL, the IRL framing, Ross-Bagnell (the O(eps*T^2) bound), + insights on distribution-shift bounds and max-entropy duality. - Perception/SSL/foundation: PointPillars, VoxelNet, CenterPoint, AlexNet, BERT, BYOL, VICReg, Bahdanau attention. - RL infra / world models: GAE, D4RL, PlaNet, the offline-RL tutorial, Spinning Up, preference learning, + the optimistic-bias insight. - Driving E2E / LLM-agent: ALVINN, Learning-by-Cheating, TCP, GameFormer, autoregressive driving policies, Codex, SayCan, + the multimodal-behavior insight. 31 new nodes / ~117 edges, each card ~3k-5.6k chars with real equations, no fabricated benchmark numbers, no meta-language. Also adopted 3 remaining orphan move cards as nodes. Research lenses expanded to 35 (added ViT, DETR, BEVFormer, SAC, MuZero, BEV-intermediate, foundation-model zero-shot driving, + 12 from the cluster). Graph: 538->541 nodes, 1568 edges. Full QC: dead internal links 45->0, orphan cards 4->0, 0 meta-language findings, 0 lint findings across 231 extended cards, math regression test green, every research-lens link resolves. Baseline 0/0.
…nts) README still advertised the removed 'dynamic insight' panel and stale counts (499 nodes / 1440 edges). Update to the real artifact: 541 nodes / 1568 edges, 0 dead internal links, ~270 hand-written cards, and the new research-lens layer (load-bearing assumption / failure boundary / falsifiable next experiment / cross-domain isomorphism) on core nodes, with honest graph-derived navigation replacing the old synthetic insight.
A read-only correctness reviewer audited the 31 new cards + lenses (equations all verified correct). Applied its findings: - ALVINN: corrected venue/year to NeurIPS 1988 (matching its own deep link), fixed the input-layer description (30x32 video retina + 8x32 range-finder retina + 1 road-intensity feedback unit -> 29 hidden -> 45 outputs), and the mislabeled '[NVIDIA PilotNet](paper_transfuser.md)' link (PilotNet != TransFuser). - GAE: removed the anachronism (original TRPO 2015 used single-path/vine, not GAE 2016) and scoped the claim. - Learning-by-Cheating: reframed CARLA Leaderboard as the later standardisation of the closed-loop eval lineage (LBC used the original CARLA/NoCrash protocol). - Multimodal-behavior insight: softened an invented-looking exact KL bound to an honest order-of-magnitude scaling relation. - Normalised 270 seed-card cross-links across 75 extended cards to use ../ so they resolve in GitHub's raw markdown view too (they already jumped correctly in-app). Research lenses expanded 35 -> 49: added the four new insight nodes plus AIRL, VoxelNet, CenterPoint, BYOL, VICReg, Bahdanau attention, preference learning, TCP, GameFormer, Codex. QC: 0 dead links, 0 meta-language, 0/231 lint findings, math test green, every lens link resolves.
The spine papers are the user-supplied core materials; only UniAD had a lens. Add lenses for the other seven — PlanT, DriveVLM, Agent-Driver, DiLu, Spike-driven Transformer, DINOv3, CF-VLA — each grounded in the card's own framing (object-level planning, fast/slow dual system, LLM tool-agent, reflective memory, event-sparse spiking compute, frozen SSL features, counterfactual replanning). Every lens states the load-bearing assumption, the failure boundary, a concrete falsifiable experiment, and the cross-domain isomorphism. Research lenses 49 -> 56; all 8/8 spine papers covered. QC: 0 dead links, 0/231 lint findings, math test green, every lens link resolves.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
变更摘要 / Summary
This PR adds three major components to the Atlas:
research_lens.json): A new structured knowledge layer capturing assumption/failure/experiment/isomorphism for key papers, with a builder tool (build_research_lens.py) and audit tool (audit_card_links.py)math-katex.js) with regression tests, fixing the bug where markdown emphasis (*,_) inside formulas breaks rendering类型 / Type
docs/data/graph_extended.json(541 nodes, 1568 edges; +42 nodes, +128 edges)详细变更 / Details
Research Lenses
docs/data/research_lens.json: 338-line structured knowledge base with 20+ insights (offline RL, imitation learning, policy improvement bounds, max entropy duality, etc.)tools/build_research_lens.py: Builder that compiles per-node lens fragments into unified JSON; supports KaTeX and intra-atlas markdown linkstools/audit_card_links.py: Validates that card links like[CQL](paper_cql.md)resolve to existing nodes in the graphdocs/data/lens_fragments/*.json): Domain-specific lens data for driving, IRL, perception, RL infrastructureRound 4 Batches
Four new node batches with 40+ papers:
round4_driving.json: E2E driving lineage (ALVINN → LBC → TCP → GameFormer → DriveGPT) + LLM-agent grounding (Codex, SayCan)round4_irl.json: Inverse RL & imitation (GAIL, AIRL, IRL, preference learning, D4RL, Levine tutorial)round4_rlinfra.json: RL infrastructure (GAE, Spinning Up, preference learning)round4_perception.json: 3D detection backbones (VoxelNet, PointPillars, CenterPoint) + SSL foundations (BYOL, VICReg, BERT)round4_orphans.json: Adopt 3 remaining orphan move cards as nodesMath Rendering
docs/js/math-katex.js: New math extension formarkedthat pre-processes KaTeX blocks before markdown parsing, preventing emphasis markers inside formulas from being interpreted as markdowntools/test_math_render.cjs: Regression test guarding against theQ^*(s,a)bugdocs/js/atlas-cards.js: Import and install math extension in card rendering pipelinedocs/atlas3d.css: Add styling for math blocksGraph & Integrity
docs/data/graph_extended.json: Merged with all 4 batches; node count 499→541, edge count 1440→1568docs/data/graph_extended.stats.json: Updated statisticstools/round2_integrity.py: Mechanical fixes for orphan adoption (already applied in this PR)tools/.link_baseline.json: Link audit baseline (max_inert=0, max_orphans=0)Card Updates
https://claude.ai/code/session_01W3sb3EfRVc4tCGdfZ5FphG