NoveltyAxis
ProductPrivacyFAQHelp
Log inSign up free

Know what’s already been done before you run the experiment.

Identify published prior art before running experiments, surface adversarial reviewer objections, and check every reference against public records.

Has request-aware KV-cache eviction already been tried?

Idea check · 20–60 min · Closest published work, the strongest objections, and an experiment plan.

Scholarly indexesSearches arXiv, OpenAlex, Semantic Scholar & Crossref
Private workspaceConfidential research · Your PDF is never stored
Free tier availableIncludes 50 reference checks monthly
Zero frictionNo credit card required to explore

Three tools. Pure verifiable evidence.

Interactive instruments engineered for each phase: before compute, before submission, and while citing.

Idea check

20–60 min

Check an idea against the published literature. You get the closest papers with quotations checked word for word, how each one differs from your idea, and an experiment plan that states which result would disprove it.

Idea

Can request-aware dynamic KV-cache management reduce LLM inference memory and latency without significantly degrading accuracy?

GrowPage: On-Demand KV Budgeting for Efficient LLM Reasoning Serving

Ma et al., 2026 · arXiv:2609.03494

Problem
0.85
Method
0.70
Evaluation
0.75
Quotation · Abstract · found word for word
“on-demand KV budgeting framework that treats KV capacity as a runtime resource”
How it differs from the idea

GrowPage allocates from observed attention-demand evolution; the idea proposes allocation from scheduling-level request signals and an accuracy-loss criterion.

From a real idea check: the three closest papers of the 308 it retrieved. Overlap is scored from 0 to 1. A report shows what overlaps so that you can judge; it cannot prove an idea is new.
Check an ideaThe report also gives the strongest objections and an experiment plan.

Paper review

15–40 min

Two independent reviewers audit your manuscript against venue rubrics, pinpointing methodology weaknesses and line-anchored objections before you submit.

Recommendation: Weak Accept
Soundness: 4/5Methodology: 3/5Novelty: 4/5
§3.2 (Line 184)Attention Head Pruning
“We apply static attention head clustering across all benchmark prompts to achieve 3.6x decoding acceleration.”
Reviewer objection:

The baseline comparison omits FastGen (Ge et al., 2024). Pruning must be evaluated dynamically on multi-turn dialogue without offline recalibration.

Review a paperDual-blind rubrics for ICLR, NeurIPS, and leading venues.

Reference check

2 minAlways free

Deterministic registry reconciliation against 150M+ Crossref DOIs and OpenAlex. Catches phantom citations, year drift, and publisher retractions.

42 citations reconciledCrossref · OpenAlex · arXiv
Li et al. (2024)SnapKV: LLM Knows What You are Looking for...
Crossref record matches character-for-character
DOI: 10.48550/arXiv.2404.14469Verified
Ge et al. (2023 → 2024)FastGen: Adaptive Attention Head Profiling...
ArXiv preprint (2023) resolved to official conference proceedings (2024)
ICLR 2024 ProceedingsYear updated
Chen et al. (2021)Sparse Attention Routing in Transformer...
Publisher retraction notice flagged in OpenAlex database
DOI: 10.1145/3459637.3482341Retracted
Check referencesAlways free. Resolves DOIs, authors, years, and retractions.

How an idea check works

A 5-stage automated audit pipeline leaving a character-exact verification trail.

  1. Stage 1FormulateTestable question
  2. Stage 2RetrieveCorpus discovery
  3. Stage 3ChallengeAdversarial review
  4. Stage 4FilterDirect text quotes
  5. Stage 5FalsifyThreshold protocol
Step 1 of 7
  1. 1Formulate a falsifiable hypothesisrecord_idea_analysis
  2. 2Query arXiv, OpenAlex, Semantic Scholar & Crossrefsearch_papers
  3. 3Extract full text & examine empirical claimsrecord_paper_analysis
  4. 4Traverse citation graph for hidden prior artexpand_citations
  5. 5Synthesize adversarial reviewer objectionsrecord_critique
  6. 6Purge claims lacking verbatim evidencerecord_gap
  7. 7Design falsification experiment & ablationsrecord_experiment_plan
Closest prior artarXiv:2404.14469

SnapKV: LLM Knows What You are Looking for Before Generation

Li et al., 2024

“SnapKV achieves a consistent decoding speed with a 3.6x increase in generation speed and an 8.2x enhancement in memory efficiency compared to the baseline when processing inputs of 16K tokens.”

Verbatim match in primary full-text (Abstract, p. 1)

Reviewer objection

SnapKV already compresses the KV cache per query prompt. The proposed modification only constitutes novel work if the eviction budget adapts dynamically across multi-turn conversational dialog without offline recalibration.

Falsification criteria

Negative falsification threshold: If fixed-window eviction matches adaptive retention perplexity within 1.5% across LongBench (16k context), reject hypothesis before proceeding to fine-tuning.

An abridged execution trace. Real runs execute asynchronously in 20 to 60 minutes; you can safely close your browser while worker pipelines process.

No source, no claim.

Four verification standards ensuring every finding is grounded in primary literature, verified against official registries, and subjected to adversarial scrutiny.

  1. Character-exact source quotations

    Primary literature

    Every excerpt is extracted directly from published PDF full text with section and page coordinates. Claims without verified primary citations are rejected.

  2. Anticipate reviewer objections

    Peer review standards

    Surfaces the hardest methodological counter-arguments, overlapping published baselines, and evaluation gaps a skeptical reviewer will challenge.

  3. Quantitative falsification criteria

    Empirical validation

    Establishes explicit numerical thresholds (e.g., baseline parity within 1.5% on standard benchmarks) that disprove a hypothesis before compute is committed.

  4. Citation integrity verification

    Official registries

    Reconciles every reference against official registry databases (150M+ Crossref DOIs and 250M+ OpenAlex records) to catch attribution drift and retractions.

Questions

More in Help.

No automated tool can prove novelty. NoveltyAxis systematically surfaces the closest published prior art, identifies overlapping methods, and constructs adversarial objections so you can evaluate novelty with complete literature awareness before committing compute or lab resources.

Strictly confidential. Your ideas, drafts, and uploaded PDFs are isolated to your account, never added to public indexes, and never shared or used to train third-party models. You can permanently delete your data at any time.

Every citation is queried directly against official registries: Crossref (150M+ DOIs), OpenAlex (250M+ works), and arXiv. Author sequences, publication years, titles, and retractions are verified deterministically against publisher records.

An idea check runs 20 to 60 minutes, a paper review 15 to 40 minutes, and a reference check about 2 minutes. Runs execute asynchronously on dedicated worker pipelines; you can safely close your browser while they process.

All scientific and scholarly fields indexed by Crossref, OpenAlex, PubMed, and arXiv—including computer science, biomedicine, physics, mathematics, engineering, economics, and the social sciences.

Test your next idea before you spend months on it.

Audit the closest published prior art, evaluate the strongest reviewer objection, and verify your citations before submission.

Check an ideaCheck references

No credit card required · Reference checks are always free · Never trained on your manuscripts

NoveltyAxis

Check research ideas and review papers, with evidence behind every claim.

Idea checkPaper reviewReference checkerDocumentationGuides
GitHub (MIT)