01
See it, don't just read it
Architecture and workflow drawn out, plus a visual map of the whole system.
Paper Lab
Understand any paper. Remember it. Build the next one.
Search a paper and get it visualized, its architecture and workflow as diagrams, a mnemonic to remember it, a one-page cheatsheet, and the research gaps that point to your next paper. Learn it fast, keep it, extend it.
Live demo on four papers. Try one:
Attention Is All You Need
The Transformer: sequence modelling with attention alone.
Introduces the Transformer, a model built entirely on attention, with no recurrence and no convolution. Because every position attends to every other in parallel, it trains far faster than RNNs and set new state of the art in machine translation. It is the architecture behind modern large language models.
Before this paper
- Sequence models leaned on recurrence (RNNs, LSTMs), processing one token at a time, in order.
- That step-by-step dependency meant training could not be parallelised across a sequence.
- Convolutional alternatives needed many stacked layers just to connect two distant words.
What this paper changed
- Attention lets every position look at every other position directly, with no recurrence or convolution.
- Fully parallel training made it dramatically faster on the same hardware.
- Set a new state of the art in translation: 28.4 BLEU on English to German, 41.8 on English to French.
The shift, in one line. Traded one-token-at-a-time recurrence for attention everyone can run in parallel, which is what made today's large language models trainable at scale.
Tap a node
Drag to rotate, scroll to zoom, or press Play to walk through it step by step.
Input embeddingTurns each word into a vector the network can do math on.
Positional encodingAdds word order back in, since attention alone can't tell position.
Encoder self-attentionLets every word in the input look at every other word to build context.
Encoder feed-forwardGives each word's representation extra room to transform after it has gathered context.
Decoder masked self-attentionLets the output-so-far attend to itself, without peeking at words not yet generated.
Decoder cross-attentionLets the output being generated pull in the relevant parts of the input.
Decoder feed-forwardThe decoder's own processing step, mirroring the encoder's, applied after it has gathered input context.
Output headTurns the final vector into an actual next-word prediction.
A live 3D map of how the paper's system flows, built from the paper itself. Every sub-layer here is wrapped in a residual connection and layer normalization, standard in the paper, not drawn as its own node.
01
InputsToken embeddings plus positional encodings carry word identity and order.
02
Encoder (xN)Each layer: multi-head self-attention, then a feed-forward network, each wrapped in residual add and layer norm.
03
Decoder (xN)Each layer adds masked self-attention and cross-attention over the encoder output.
04
HeadA linear projection then softmax turns the final states into token probabilities.
01
Embed and positionMap tokens to vectors and add sinusoidal positional encodings.
02
Self-attentionEach token mixes information from all others via Query, Key, Value.
03
Feed-forwardA per-position MLP transforms each token, with residuals and norm.
04
DecodeMasked self-attention plus cross-attention generate the output left to right.
05
ProjectLinear plus softmax produce the next-token distribution; train with cross-entropy.
"Queries Knock, Keys answer, Values deliver."
- Q K V = the three roles every token plays in attention.
- A token's Query asks; others' Keys say how well they match; their Values are what gets mixed in.
- Stack it, add position, normalise, feed forward: that is one Transformer block.
Contributions
- Self-attention replaces recurrence and convolution entirely.
- Multi-head attention learns several relationship types at once.
- Fully parallel over sequence positions, so training is much faster.
Key results
- 28.4 BLEU on WMT14 English to German (new state of the art).
- 41.8 BLEU on English to French at a fraction of prior training cost.
Core equations
- Attention(Q,K,V) = softmax(QKᵀ / √dk) V
- MultiHead = Concat(head_1..head_h) Wᴼ
- PE(pos,2i) = sin(pos / 10000^(2i/d))
Terms
- Q, K, V: learned query, key and value projections.
- dk: key dimension; the √dk term keeps scores stable.
- Positional encoding: injects order, since attention is set-based.
Why it matters. It made sequence models parallel and scalable, which is what unlocked modern LLMs.
Open research gaps
- Attention is O(n²) in sequence length, so long contexts are expensive.
- Absolute positional encodings generalise poorly beyond trained lengths.
- Attention weights are only a weak window into what the model actually uses.
Your next paper
Problem. The O(n²) cost blocks book-length or repository-scale context.
Approach. Combine block-sparse attention with a learned router that spends compute only where it pays off.
What to try.
- Benchmark against FlashAttention on 100k-token inputs.
- Measure quality lost versus memory and latency saved.
- Ablate the router to show where the savings come from.
Tap a card to flip. These are generated from the paper, ready for spaced repetition.
TriShieldRAG: 3 Rings, One Blind Spot in Layered Defenses for RAG
A three-ring defense for RAG, and the adaptive attack that slips past it.
RAG answers depend on what the retriever returns, and a few poisoned documents can mislead it. TriShieldRAG stacks three independent defenses: ingest, retrieval and generation. It stops non-adaptive poisoning almost completely, but an adaptive attacker who changes only document formatting bypasses the first ring and the whole stack. It is an honest negative result for layered defenses that share the same evidence.
Before this paper
- RAG answers depend entirely on what the retriever returns, and a handful of poisoned documents could mislead the final answer.
- There was no validated, layered defense architecture for RAG with explicit, measured thresholds.
- Minority-poison assumptions in prior work had not been directly checked against measurement.
What this paper changed
- Three independent rings (ingest screening, trust-aware re-ranking, cross-model vote) cut non-adaptive poisoning attack success from 79% to about 1%.
- But an adaptive attacker who only changes document formatting slips past Ring 1 on every document, and the full stack's attack success stays around 62%.
- The honest finding: layered defenses that all read the same retrieved evidence do not fail independently.
The shift, in one line. Proves defense in depth works against the easy attack, then publishes the harder, less comfortable finding: against an adaptive one, three rings sharing one blind spot are barely better than none.
Tap a node
Drag to rotate, scroll to zoom, or press Play to walk through it step by step.
Ring 1The Ingest Guard is the first line of defense, screening every document as it enters so obviously poisoned content gets caught before it ever reaches retrieval.
Ring 2The Retrieval Scorer re-ranks whatever documents made it past the door, downweighting ones that look untrustworthy or that contradict the rest of the evidence.
Ring 3Cross-LLM Consensus has three different language models independently answer and vote, so a mistake made by one model alone is unlikely to become the final answer.
AnswerAn answer is only returned once all three independent rings agree it is safe, which is meant to make a successful attack much harder than beating any single defense alone.
A live 3D map of how the paper's system flows, built from the paper itself. Every sub-layer here is wrapped in a residual connection and layer normalization, standard in the paper, not drawn as its own node.
01
Ring 1: Ingest GuardScreens each document as it enters: perplexity, embedding-outlier and pattern checks.
02
Ring 2: Retrieval ScorerTrust-aware re-ranking using provenance and inter-document consistency.
03
Ring 3: Cross-LLM ConsensusThree different LLMs vote; disagreement triggers re-retrieval.
04
Defense in depthA poison document must defeat all three independent stages, not just one.
01
Screen at ingestCatch crude poison at the door with Ring 1.
02
Re-rank at retrievalDown-weight documents that contradict trusted, consistent evidence.
03
Vote at generationThree vendors answer; disagreement means re-retrieve.
04
Attack adaptivelyChange only document formatting, with no poison-text change and no retriever access.
05
Report the blind spotThe formatting attack bypasses Ring 1 on every document, and the stack barely helps.
"Three rings, one shared blind spot."
- Ingest, retrieval, consensus: three independent checks.
- Each stops crude poison the prior stage missed.
- But they share the same evidence, so one formatting trick defeats them together.
Contributions
- A three-ring, defense-in-depth architecture with explicit thresholds.
- A corrected minority-poison assumption, validated against measurement.
- A reference implementation with FAISS and a three-vendor LLM panel.
- An adaptive-attacker evaluation and an honest negative result.
Key results
- Layered defenses fail together when they rely on the same retrieved evidence.
- Cross-model agreement can reach 0.96 while the answer is 99% wrong.
Key numbers
- Non-adaptive attack success 79% drops to about 1%.
- Adaptive formatting attack: Ingest Guard success 0.50 to 0.00.
- Adaptive end-to-end: about 62% attack success, barely better than undefended.
Terms
- RAG: retrieval-augmented generation.
- Poisoning: injecting crafted documents to mislead retrieval.
- Adaptive attacker: one who knows the defense and adapts to it.
Why it matters. It is an honest warning: stacking defenses that read the same evidence does not add up to real robustness against an adaptive attacker.
Open research gaps
- Ring 1 robustness to adaptive formatting attacks is unsolved.
- Defenses that share evidence do not fail independently.
- Isolate-then-aggregate designs may be sounder than voting over shared context.
Your next paper
Problem. Voting over shared retrieved context fails when that context is poisoned.
Approach. Answer from each retrieved passage in isolation, then aggregate with a certifiable corruption bound.
What to try.
- Reproduce the formatting attack against an isolate-then-aggregate design.
- Measure the certified bound on tolerated corrupted passages.
- Compare clean-data quality lost versus robustness gained.
Tap a card to flip. These are generated from the paper, ready for spaced repetition.
Unlock Paper Lab
You have seen what it does. Now use it on anything.
- Every paper, plus your own uploads
- Export cheatsheets and flashcards as PDF
- Spaced repetition, progress and streaks
- Save papers to your private Research RAG
Or create a free account to keep today's lab.
Unlock Paper Lab
You have seen what it does. Now use it on anything.
- Every paper, plus your own uploads
- Export cheatsheets and flashcards as PDF
- Spaced repetition, progress and streaks
- Save papers to your private Research RAG
Or create a free account to keep today's lab.
2 free papers in the Lab each day. Go unlimited.
Ask
Ask KXAI about the paper you're reading.
Architecture, results, gaps, a mnemonic, or any other paper in your bucket. Open KXAI from the button in the corner, it already knows which paper you're on.
Why Paper Lab is different
Not a summary. A way to understand, remember and extend research.
02
Built to remember
A mnemonic and flashcards turn one read into lasting recall.
03
From reading to research
Gap finder scaffolds your next paper: problem, approach and what to try.
04
Grounded in the source
Every claim ties back to the paper, with citations, never invented.
05
Pick your depth
Story, Math or Silicon on the same paper, for any background.
06
Keep and share
Export the cheatsheet, save to your Research RAG, or share a lab link.
What learners say
Real reviews from researchers, students and teams, as Paper Lab rolls out.
We publish only verified reviews from real users. They will appear here as the beta opens.
Be an early userGet started
Open your first paper in the Lab.
Free to try. Unlimited labs, PDF export and flashcards on Scholar and Professional.
Unlock Paper Lab
You have seen what it does. Now use it on anything.
- Every paper, plus your own uploads
- Export cheatsheets and flashcards as PDF
- Spaced repetition, progress and streaks
- Save papers to your private Research RAG
Or create a free account to keep today's lab.