Knowledge Graph
Ask the map what to learn next.
Use the graph as a research-learning instrument: map papers to concepts, inspect typed edges, find prerequisite repairs, and attach equations, labs, claims, and discussion to the exact idea.
First Route Readiness
First route: attention to serving
A new learner should see one concrete path: map or search a claim, inspect the route graph, open Efficient Attention, predict the KV-cache lever, review the claim boundary, and leave a reproduction or repair note attached to the exact object.
Knowledge GraphThe graph names the next concept and route objects.
Save or inspect the route, then open search or Efficient Attention.
01Homenext action02Paper Mappernext action03Knowledge Graphnow04Searchnext action05Attention To Servingnext action+1 more
Scope labels requiredPreview/local where applicable; claim review before comparison; reproduction notes stay object-attached.
No benchmark result, hosted compute, automatic expert review, or live runtime-performance claim is made by this v0.
Graph Reading
A graph is useful only when it explains the next move.
The point is not to admire complexity; it is to find the shortest honest path from confusion to a concept you can test.When a page feels too hard, use incoming edges to locate the idea that should be repaired first.
Try a bridgeFollow shared neighborhoods to see when optimization, probability, or representation ideas are doing the same job.
Search mechanismsUse the graph to choose the minimum sequence that makes a derivation or demo feel earned.
Open pillarsLearning Route
Ask the graph what to learn next.
Pick a question and the graph turns it into prerequisites, equations, labs, and honest links between ideas. When you arrive from the mapper, your current paper route stays visible.
Answer
A paper claims it compresses the KV cache. What should I inspect?
Follow the runtime path from attention math to serving bottlenecks, then test which memory term is being reduced.
Route Engine
Compute the honest next path.
Choose what the learner already knows and a frontier target. The engine runs a weighted shortest-path query over typed concept edges, then names the first missing repair instead of hand-waving across the gap.
Known concepts
Research target
6 nodesRepair Efficient Attention next
weighted cost5.50
Preview only; changing chips does not overwrite your saved route.
01Attentionweighted copying02Efficient Attentionmemory pressure03Long Contextstress regime04SSM Hybridsfixed-state sequence models
05Parallel Scantrainable recurrenceplanned
06State-Space DualityMamba-style bridgeplanned
You need Q/K/V weighted copying before cache and memory optimizations are meaningful.
Long context exposes the cost of storing and repeatedly reading all prior keys and values.
Fixed-state sequence models become attractive when KV memory grows with sequence length.
Recurrent-looking models need parallel sequence computation to train at scale.
Modern Mamba-style papers often connect recurrence, convolution, and attention-like views.
Typed Edges
Why the route is honest
Cache tricks only matter after Q/K/V shape and softmax copying are clear.
Decode is often bandwidth-bound because it repeatedly reads cached keys and values.
Position extrapolation changes the geometry of attention scores.
FlashAttention changes memory traffic, not the attention function itself.
Learning Objects
What anchors this route
carried from mapper when available
live in serving module
live: KV memory calculator
object question
Research Room
Resolve the route object before moving on
Pick a paper, equation, lab, claim, or misconception. The selected object becomes the saved route focus for the next companion prompt.KV cache compression result
Anchored question
What should we inspect about KV cache compression result before treating this route object as understood?
Local action draftDraft unavailableNeeds a canonical object key
Local action draft
This object needs a content object key before local action drafts can attach to it.
No local draft saved.
- Paper metadata, abstract claim, and any pasted source spans
- Mapped concepts, equations, and prerequisite repairs
- Which claims are source-checked and which remain local-preview only
- The paper contribution is mapped to a concrete mechanism
- Unverified author, date, benchmark, and novelty claims are separated from learning claims
- The next concept or lab action is specific enough to resume later
Grounded AI handoff
I am working in Continuous Function's research reading room. Object: paper - KV cache compression result Context: carried from mapper when available Anchor id: paper/graph/kv-cache/kv-cache-compression-result Open question: What should we inspect about KV cache compression result before treating this route object as understood? Evidence to inspect: - Paper metadata, abstract claim, and any pasted source spans - Mapped concepts, equations, and prerequisite repairs - Which claims are source-checked and which remain local-preview only What would resolve this: - The paper contribution is mapped to a concrete mechanism - Unverified author, date, benchmark, and novelty claims are separated from learning claims - The next concept or lab action is specific enough to resume later Answer as a careful research tutor: stay source-grounded, separate verified evidence from assumptions, name the relevant math objects, and end with one next action.
paper/graph/kv-cache/kv-cache-compression-result