Free
$0
10 requests / day
- All of Lune, no feature gates
- One personal account
- Sign in with Google or email
- Top up with credits any time
AI Research,AI Research,AI Research,Academic Writing,Literature Review,Experimental Design,Peer Review, Free of AI slop
Search, retrieve, and connect AI agents to top-tier computer science literature, research skills and best practices, get evidence-grounded and trustworthy answers.
Used by students and researchers from
What it is
01Trusted knowledge from A / A* conference papers for you and your agent.
No open-web noise. No workshop papers, tutorials, abstracts, or demos.
Full-text semantic search
Search across the papers in plain language, all at once.
Citation tracing
See what a paper builds on and what it leads to.
Research best practices
Battle-tested research guidance, surfaced whenever it is relevant.
Knowledge cutoff
Tap into the minds of the great
Search across the talks, interviews and notes of the people who shaped the fields.
- Richard HammingBell Labs
- Andrej KarpathyAnthropic
- Bill FreemanMIT
- Nicholas CarliniAnthropic
- John SchulmanOpenAI
- Jacob SteinhardtUC Berkeley
- Michael NielsenAstera Institute
- Christopher OlahAnthropic
What you get
02Highest quality answers at every step of your research workflow.
From the first literature search to the final draft, every step rests on evidence, never guesswork.
Your field
Research workflow
ClaudeLune MCP
I'm new to retrieval-augmented generation. What are the foundational papers, and what exactly do they claim?
Lunesearch_papers ·get_paper_fulltext [1]
RAG [1] pairs a seq2seq generator with a Wikipedia index; Self-RAG [2] learns when to retrieve and critiques its own output; FLARE [3] re-retrieves when a sentence has low-confidence tokens. RAG's setup:
… the non-parametric memory is a dense vector index of Wikipedia, accessed with a pre-trained neural retriever.
Idea: have a RAG model abstain or re-retrieve when its context looks weak. Is this novel, and where's the real gap?
Lunesearch_papers ·get_paper_fulltext [1]
Not new: Sufficient Context [1] abstains via a sufficiency autorater plus model confidence; GrowOVER [2] re-retrieves on unreliable answers. Gap: a better trigger; models get 35-62% right on weak context.
… intervention strategies to increase accuracy should not solely rely on sufficiency.
Evidence
My RAG method reranks passages before generation. What baselines and ablations will reviewers expect?
Lunesearch_research_guidance ·search_papers ·get_paper_fulltext [1]
Mirror BAR-RAG [1]: no-retrieval and CoT baselines, plain RAG and RAG with a neural reranker on one retriever, EM on NQ, TriviaQA and HotpotQA, and a top-k sweep. Show you beat them, not just score decently.
… the largest gains appearing in low-k regimes where evidence is sparse …
Evidence
- [1]ICML 2026
- ◆Baselines are crucial, Best-practices library
Draft the related-work sentence for my method, which reranks before generation to keep the reader's context small.
Lunesearch_papers ·get_paper_fulltext [1], [2]
Drafted, with each citation checked against its source:
Evidence
What are the key papers on automatic constant-time rewriting of crypto code, and what exactly do they claim?
Lunesearch_papers ·get_paper_fulltext [1]
Constantine [1] linearizes secret-dependent control and data flow, with 11% geomean overhead on 23 crypto modules. SynthCT [2] adds that instruction safety depends on the microarchitecture. Constantine's claim:
… outperforms prior comprehensive solutions in terms of both performance and compatibility, while also providing stronger security guarantees.
Evidence
Idea: vectorize constant-time rewriting's memory accesses with AVX-512. Is this novel, and where is the real gap?
Lunesearch_papers ·get_paper_fulltext [2]
Not new. Constantine [1] first used AVX-512 in constant-time rewriting, but its masked stores leak timing, as VeCT [2] showed. The gap is Arm SVE and RISC-V, since VeCT's timing tests covered only Intel CPUs.
… boost the performance of constant-time code, by utilizing Vector Extensions (VE).
I'm porting constant-time rewriting to Arm SVE. What baselines, benchmarks and checks will reviewers expect?
Lunesearch_papers ·get_paper_fulltext [1]
VeCT [1] is the template: overhead reduction against Constantine+, a patched Constantine [2], on Constantine's suites. VeCT's AVX-512 rules came from Welch t-tests on Intel CPUs, so rerun those tests on SVE.
… reuse our t-test framework to assess the CT guarantees of VE instructions on additional hardware platforms.
Draft the related-work sentence positioning my Arm SVE constant-time rewriter against Constantine and VeCT.
Lunesearch_papers ·get_paper_fulltext [1], [2]
Drafted, each paper's method checked against its text:
What are the key papers on LLM agents for automated program repair, and what exactly do they claim?
Lunesearch_papers ·get_paper_fulltext [1]
AutoCodeRover [1] resolved 19% of SWE-bench Lite with code search; Agentless [2] dropped the agent loop yet led open-source tools there. RepairAgent [3] lets the LLM pick its next tool. AutoCodeRover's thesis:
… the effective use of code search to make software engineering processes like program repair autonomous.
Idea: an LLM repair agent that steps through failing tests in a debugger. Is this novel, and where is the real gap?
Lunesearch_papers ·get_paper_fulltext [1]
Not new. Xiang et al. [1] ran pdb stepping as a baseline: it cost more and abandoned 53% of debugging attempts. Their cheaper fix traces whole function calls via Python's tracer, so the gap is other languages.
… incurs the substantial cost of a complete LLM inference cycle.
Evidence
My agent steps through failing Java tests in a debugger. Which benchmarks and baselines will reviewers expect?
Lunesearch_papers ·get_paper_fulltext [1]
Mirror RepairAgent [1]: Defects4J plus GitBug-Java's 2023 bugs for leakage, correct versus plausible fixes, ChatRepair and ITER as baselines. Drop perfect fault localization; your debugger should find the bug.
… we assume perfect fault localization as the default.
Evidence
Draft related work: my agent steps through failing tests in a debugger, unlike AutoCodeRover and Agentless.
Lunesearch_papers ·get_paper_fulltext [1], [2]
Drafted, with each paper checked against its own text:
Evidence
I'm new to sparse matrix multiplication (SpGEMM) accelerators. What are the key papers, and what does each claim?
Lunesearch_papers ·get_paper_fulltext [1]
SpArch [1] computes outer products and merges partial matrices on chip. MatRaptor [2] and Gamma [3] use the row-wise dataflow; MatRaptor reports 1.8× over OuterSPACE, Gamma gmean 2.1× over prior accelerators.
… the memory access reduces by 2.8×, the performance increases by 4×, and the energy efficiency increases by 6× …
Idea: an SpGEMM accelerator that switches dataflow at runtime. Is this novel, and where is the real gap?
Lunesearch_papers ·get_paper_fulltext [2]
Not as stated. Spada [1] adapts its dataflow at runtime, and SegFold [2] schedules work inside each tile. SegFold's stated gap: pairing runtime scheduling with offline row reordering for reused matrices.
… combining offline reordering with SegFold's runtime scheduling is a promising direction for future work.
Evidence
I'm writing up an SpGEMM accelerator for MICRO. Which baselines and workloads will reviewers expect?
Lunesearch_papers ·get_paper_fulltext [1]
Reviewers expect one prior accelerator per dataflow, matched in multipliers and memory. Spada [1] compares against SIGMA, SpArch [2] and Gamma, plus MKL and cuSPARSE, on SuiteSparse and compressed-DNN matrices.
… we set all designs to have the same number of multipliers …
Evidence
Related work: position my accelerator (offline row reordering plus runtime scheduling) against Gamma and SegFold.
Luneget_paper_fulltext [1], [2]
Gamma's own cost caveat [1] sets up the contrast:
Evidence
What built on Swift, Google's delay-based datacenter congestion control, and how does each follow-up use it?
Lunesearch_papers ·get_paper_citations ·get_paper_fulltext
Swift [1] targets an end-to-end delay. Poseidon [2] builds on it but reacts to the largest per-hop delay from in-network telemetry. MegaScale [3] combines Swift's RTT measurement with DCQCN's ECN response.
… comparing the max per-hop delay against a dynamic target …
Idea: Swift-style congestion control, fair across congested hops without INT. Is this novel, and where is the gap?
Lunesearch_papers ·get_paper_fulltext [2], [3]
Poseidon [1] already fixes Swift's multi-hop unfairness, but with INT. Without INT it is still open: two 2026 delay-only designs, OSCAR [2] and CCC [3], both leave it unsolved. OSCAR on its delay signals:
… cannot distinguish whether increases in signals arise from single or multiple congestion points.
My delay-only congestion control stays fair across congested hops. Which baselines and workloads will NSDI expect?
Lunesearch_papers ·get_paper_fulltext [1]
OSCAR [1] compares first against HPCC and PowerTCP, then Poseidon and Swift, on CacheFollower, incast and training traffic. For your fairness claim, rerun its parking-lot test:
… parking-lot topology with three long-lived flows, where Flow 3 shares one bottleneck with Flow 1 and another with Flow 2 …
Evidence
Related work: position my congestion control (delay-only, fair across congested hops) against Poseidon and OSCAR.
Luneget_paper_fulltext [1], [2]
Poseidon [1] relies on INT, which is your opening:
Evidence
How do learned indexes compare with B-trees and ART on lookup speed, and does their edge hold once inserts arrive?
Lunesearch_papers ·extract_from_papers ·get_paper_fulltext
On read-only data, learned indexes are faster for their size than B-trees and ART [1]. Updatable ones won over 80% of single-threaded tests in 2022 [2], but a 2026 benchmark finds none robustly beats them [3].
… none of the SOTA updatable learned indexes can robustly outperform traditional indexes …
Idea: pad a learned index's bulk load with synthetic keys to curb overfitting. Is this novel, and where is the gap?
Lunesearch_papers ·get_paper_fulltext [1]
Not new. Luo et al. [1] tried synthetic keys worth 1% of the bulk load. They advise it for LIPP; ALEX gains only with known key spaces and can lose up to 50%. Augmentation that helps every index remains open.
… more sophisticated data augmentation techniques to consistently improve the performance of various indexes.
Evidence
My learned index trains on synthetic keys to anticipate inserts. Which baselines and workloads do reviewers expect?
Lunesearch_papers ·get_paper_fulltext [1], [2]
Follow Luo et al. [1]: ALEX, LIPP, DILI and DyTIS against B+tree and ART, varying sampling, bulk-load size and insert order, judged by the worst case. Add HIRE [2] as a robust baseline.
… the ability of learned indexes to consistently outperform certain traditional indexes in specific metrics.
Evidence
Draft related work setting ALEX and LIPP against my index, which trains on synthetic keys to anticipate inserts.
Lunesearch_papers ·get_paper_fulltext [1], [2]
Drafted, each mechanism checked against its paper:
Evidence
I'm new to fast matrix multiplication. What are the key recent papers lowering ω, and what exactly does each claim?
Lunesearch_papers ·get_paper_fulltext [1]
Duan et al. [1] found a "combination loss" and offset it with asymmetric hashing; Alman et al. [2] later treated each dimension differently. Both lowered ω. As [1] puts it:
… we identify a more implicit "combination loss", which is our main contribution.
Evidence
Idea: let subtensors share level-ℓ blocks in all three dimensions. Is this novel, and where is the real gap?
Lunesearch_papers ·get_paper_fulltext [1]
Alman et al. [1] call it the natural end goal but reach only Y and Z, since their method needs unique X-blocks. They believe Duan, Wu and Zhou's [2] techniques are exhausted, so X needs a truly new idea.
… there does not seem to be a way to continue our approach to also allow the X blocks to share level-ℓ-blocks.
Evidence
My laser-method analysis ends in a nonlinear program. Which solvers do ω papers use, and what should I release?
Lunesearch_papers ·get_paper_fulltext [1], [2], [3]
Release parameters and a checker, as Le Gall [1] does; Duan et al. [2] verify with a standalone program. Solvers: Maple's NLPSolve [1], CVX with Mosek [2], SQP in SNOPT [3].
… a standalone program to verify this result …
Is ω < 2.371552 still the best known bound on the matrix multiplication exponent? My draft cites it.
Luneverify_claims ·get_paper_fulltext [1]
No, update it. Alman et al. [1] improved on the 2.371552 bound, and the paper's latest version adds that a 2026 follow-up of its method reaches ω < 2.371177. In their words:
… ω < 2.371339, improved from the previous bound by [VXXZ24] of ω < 2.371552 …
Evidence
How do 3DGS, Instant-NGP and Mip-NeRF 360 compare on training time, rendering speed, image quality and memory?
Lunesearch_papers ·extract_from_papers [1], [2]
3DGS [1] reached Mip-NeRF 360's quality in 35 to 45 minutes, not 48 hours, and renders in real time. Instant-NGP [2] trains in seconds to minutes. 3DGS pays in memory, far more than NeRF-based methods need [1].
… additional training time allows us to achieve SOTA quality which is not the case for the other fast methods.
Evidence
Idea: shrink 3DGS models by pruning and quantizing without retraining. Is this novel, and where is the real gap?
Lunesearch_papers ·get_paper_fulltext [1]
Not new. FlexGaussian [1] prunes and quantizes without retraining, up to 96.4% smaller at under 1 dB PSNR loss. The open gap is lower bit-widths: its authors name ternary or binary attributes as future work.
… further pushing the compression ratio by investigating ternary or even binary attributes …
Evidence
My training-free 3DGS compressor uses ternary attributes. What datasets, metrics and baselines do reviewers expect?
Lunesearch_papers ·get_paper_fulltext [1], [2]
Use Mip-NeRF 360, Tanks&Temples and Deep Blending with PSNR, SSIM, LPIPS and model size [1]. FlexGaussian [2] also times compression and compares against training-free FCGS.
… either the post-processing time for refinement-based methods or the total training time for retraining-based methods.
Evidence
Draft a related-work sentence setting my training-free ternary 3DGS compressor against CompGS and FlexGaussian.
Lunesearch_papers ·get_paper_fulltext [1], [2]
Drafted, each method as its own paper describes it:
Evidence
What are the key papers on selecting occluded objects in VR, and what exactly do they claim?
Lunesearch_papers
VoiceRay [1] lets you say which object along the ray you mean. Outline Pursuits [2] pairs pointing with gaze and degrades less than raycasting as occlusion grows. VoiceRay reports:
… significantly decreased selection time, maintained presence, increased usability, and reduced cognitive load.
Evidence
Idea: auto-switch VR selection techniques in dense or occluded scenes. Is this novel, and where is the real gap?
Lunesearch_papers ·get_paper_fulltext [1]
Not new. Adaptique [1] switches among three controller-ray techniques, and 10 of 18 participants preferred it. The real gap is other modalities: it leaves switching to gaze when the arm tires to future work.
… modifications to the current objective score calculations would be necessary to ensure these scores remain comparable across different modalities.
Evidence
I designed a new VR object selection technique. What study design and baselines will CHI and UIST reviewers expect?
Lunesearch_papers ·get_paper_fulltext [1]
A 20-year review [1] recommends at least 20 participants and targets placed in 3D. Ray-casting is the usual baseline [2]. Benchmark against XR selection averages: 1.25 s, 11.95% errors, 3.25 bits/s [3].
Nearly all studies (92.7%) used a within-subject design, which we also recommend …
My selector switches from ray to gaze as the arm tires. Draft related work against Adaptique and Outline Pursuits.
Lunesearch_papers ·get_paper_fulltext [1]
Drafted, each claim checked against its paper:
Evidence
How it works
03Install Lune to Claude/ChatGPT and the like via connector plugin.
Lune grounds your agent in trusted knowledgeNeurIPSIEEE S&PACL from top-tier academic papers, so there are no hallucinated facts or noise from the open web.
+38
Your AI app
Lune connector
Top-tier papers
Talent Database
04Top-tier CS talent profiles built from top-tier track records.
A database of talent profiles with complete publication track records and contact information. Semantic search finds people by name, school, research area and more.
Semantic search by research area
›robot learning from videoTrack records you can check
20172025
Email for cold outreach
r***@stanford.edu
Reveal
- EMNLP ×10
Danqi Chen
Princeton University
LLM training · Retrieval
d***@cs.princeton.edu
- NeurIPS ×23
Mingsheng Long
Tsinghua University
World models · AI for science
m***@tsinghua.edu.cn
- ICCV ×9
Mohamed Elhoseiny
KAUST
Multimodal LLMs · Audio-visual reasoning
m***@kaust.edu.sa
- CVPR ×15
Andrea Vedaldi
University of Oxford
4D reconstruction · 3D generation
v***@robots.ox.ac.uk
- ACL ×4
Preethi Jyothi
IIT Bombay
Speech recognition · Multilingual NLP
p***@cse.iitb.ac.in
- CVPR ×4
Alvaro Soto
Pontificia Universidad Católica de Chile
Continual learning · Visual reasoning
a***@ing.puc.cl
- EMNLP ×10
Danqi Chen
Princeton University
LLM training · Retrieval
d***@cs.princeton.edu
- NeurIPS ×23
Mingsheng Long
Tsinghua University
World models · AI for science
m***@tsinghua.edu.cn
- ICCV ×9
Mohamed Elhoseiny
KAUST
Multimodal LLMs · Audio-visual reasoning
m***@kaust.edu.sa
- CVPR ×15
Andrea Vedaldi
University of Oxford
4D reconstruction · 3D generation
v***@robots.ox.ac.uk
- ACL ×4
Preethi Jyothi
IIT Bombay
Speech recognition · Multilingual NLP
p***@cse.iitb.ac.in
- CVPR ×4
Alvaro Soto
Pontificia Universidad Católica de Chile
Continual learning · Visual reasoning
a***@ing.puc.cl
- CVPR ×17
Angela Yao
National University of Singapore
Video understanding · Egocentric vision
a***@comp.nus.edu.sg
- NeurIPS ×12
Niao He
ETH Zürich
Optimization · LLM fine-tuning
n***@inf.ethz.ch
- NeurIPS ×23
Taiji Suzuki
The University of Tokyo
Deep learning theory · In-context learning
t***@mist.i.u-tokyo.ac.jp
- ICLR ×22
Graham Neubig
Carnegie Mellon University
LLM agents · Reasoning
g***@cs.cmu.edu
- NeurIPS ×38
Tongliang Liu
The University of Sydney
Trustworthy ML · Causal inference
t***@sydney.edu.au
- ICLR ×2
Richard Klein
University of the Witwatersrand
Learning dynamics · Modular networks
r***@wits.ac.za
- CVPR ×17
Angela Yao
National University of Singapore
Video understanding · Egocentric vision
a***@comp.nus.edu.sg
- NeurIPS ×12
Niao He
ETH Zürich
Optimization · LLM fine-tuning
n***@inf.ethz.ch
- NeurIPS ×23
Taiji Suzuki
The University of Tokyo
Deep learning theory · In-context learning
t***@mist.i.u-tokyo.ac.jp
- ICLR ×22
Graham Neubig
Carnegie Mellon University
LLM agents · Reasoning
g***@cs.cmu.edu
- NeurIPS ×38
Tongliang Liu
The University of Sydney
Trustworthy ML · Causal inference
t***@sydney.edu.au
- ICLR ×2
Richard Klein
University of the Witwatersrand
Learning dynamics · Modular networks
r***@wits.ac.za
How to use
05No apps to download.
Install Lune as a plugin or connector to your AI app.
i
MCP
Install the plugin or connect via a custom connector.
https://mcp.luneresearch.comAccess Lune's core features
ii
CLI
Use Lune without leaving the terminal.
npm install -g @retrograde-labs/lune-cliInstall & use Lune from the command line
Pricing
06We have a plan for every kind of researcher.
Pick a plan with a standard daily allowance. If you go past it, just top up extra usage credits.
Pro
$4.99
/ month · 300 requests / day
- Everything in Free
- 300 queries or email reveals a day
- Share plan and credits with your team
- Connect any AI client you use
Max
Popular$9.99
/ month · 600 requests / day
- Everything in Pro
- 600 queries or email reveals a day, twice Pro's
- Share plan and credits with your team
- Priority email support
Busy day? Top up with credits at $0.99 for 300 extra requests. Credits never expire, and they work on every plan.
Trustworthy AI for Scientific Research
Free to start. No credit card required.
