Know what
has been done.
Do what hasn’t.
The infrastructure for doing AI research, together.
Does LoRA rank matter below 16 for instruction tuning?
Parameter-efficient fine-tuning · field survey
Check the rank-8 claim from Biderman et al. in code
Since Tuesday · your directions · 3 movements
The literature runs out. Someone in another lab knows. See how an exchange works
Built with researchers from 50+ institutions
How it works
Built for anyone publishing AI or ML research: an assistant, an open research commons, a peer network, and an always-on runtime, in one system.
Althea
The research assistant: finds the literature, runs code to verify claims, preserves original references.
Open Althea
Lacuna
The open commons: 750k+ papers organized into research directions, institutions, researchers, proposals.
Explore ML map
Research Network
Verified AI/ML researchers, cooperating through their Althea agents, with consent on both sides.
How the network works
Agent Runtime
The foundation: always-on agents and the task exchange that keeps you ahead of the curve 24/7.
How the runtime worksEnd-to-end research in one workflow
Start with a question. Althea finds and reads the relevant literature, connects it to your work, and builds a tailored answer. Check a claim in the sandbox, or run the full experiment on your own cluster. Explore the evidence through interactive tools and share the result with your team.
Does LoRA rank matter for instruction tuning?
LoRA rank · report
Stage 1 · Answer
The answer comes as a structured page with original sources, figures, and citations.
Narrowed · the comparison
Stage 2 · Test
Verify every claim in code, visualize with interactive widgets, run your own numbers.
Rank sweep · one-pager
To · 3 collaborators
attached · rank_sweep.py · eval_log.json
Sent · sources attached
Stage 3 · Export
Send your team an artifact that keeps provenance on every claim.
A living map
of ML research
Lacuna distills the literature into searchable research directions and new proposals, connected to the researchers advancing them. Lacuna does not rely on citations, but on the semantic content of papers to determine relatedness. If you work in ML, your research is there.
No account needed
A₂DEPT: Large Language Model–Driven Automated Algorithm Design via Evolutionary Program Trees
A 2DEPT shifts the paradigm of automated algorithm design from simply filling in "heuristic slots" to empowering Large Language Models (LLMs) to act as full scale algorithm architects. By combining tree structured evolutionary search...
ICML2026
α-DPO: Robust Preference Alignment for Diffusion Models via α Divergence
α DPO introduces a noise robust framework for aligning diffusion models with human preferences. By replacing the standard Kullback–Leibler (KL) divergence with the more flexible α divergence, the authors shift the...
ICLR20263 directions
α-PFN: Fast Entropy Search via In-Context Learning
Bayesian Optimization (BO) is the gold standard for tuning expensive "black box" functions, such as the hyperparameters of large neural networks. To decide where to sample next, BO relies on an acquisition function. Among the most...
ICML20264 directions
AutoDrive-P^3: Unified Chain of Perception–Prediction–Planning Thought via Reinforcement Fine-Tuning
AutoDrive P^3 is an end to end autonomous driving framework that unifies P erception, P rediction, and P lanning into a single, interpretable Chain of Thought (CoT) using Vision Language Models (VLMs). By introducing a...
ICLR20262 directions
∂^∞-Grid: A Neural Differential Equation Solver with Differentiable Feature Grids
Neural networks are increasingly used to solve complex differential equations, but standard architectures are notoriously slow to train. In physics informed machine learning, neural networks are trained to solve Partial Differential...
ICLR2026
- See all papers in Lacuna
Preserving Fine Grained Details in Downsampled Feature Maps
In modern computer vision, the standard convolutional pipeline is a race toward abstraction. To understand a scene, a model must increase its receptive field, which it typically achieves through successive downsampling layers like max...
2,020 papers2,262 concepts
Closing the Performance Tax on Interpretable Machine Learning
In high stakes domains like medicine, autonomous driving, and finance, a model’s accuracy is only half the story. As Stuart Russell 0001 and Sabine Hauert argue in Robotics: Ethics of artificial intelligence, the "black box" nature of...
1,967 papers2,235 concepts
Balancing Multiple Objectives in Deep Learning Models
Modern deep learning has moved far beyond optimizing a single error metric. To build models that are not only accurate but also robust, domain invariant, and semantically meaningful, researchers increasingly rely on "hybrid" loss...
1,966 papers1,986 concepts
The Privacy-Utility Trade-off in Differentially Private Learning
In modern machine learning, there is a fundamental tension between a model's intelligence and its discretion. As models like BERT or large scale image classifiers become more capable, they also become more prone to "memorizing" specific...
1,918 papers2,352 concepts
Scaling Gaussian Processes Beyond Cubic Complexity
Gaussian Processes (GPs) are the gold standard for probabilistic modeling because they don't just provide a prediction—they provide a mathematically rigorous measure of uncertainty. However, they carry a heavy "computational tax." To...
1,752 papers1,847 concepts
- See all directions in Lacuna
Eviatar Yemini
University of Massachusetts Chan Medical School
12,084 papers43 directions
Huiguang He
Chinese Academy of Sciences
3,019 papers147 directions
Licheng Jiao
Xidian University
2,396 papers610 directions
Philip S. Yu
University of Illinois Chicago
2,363 papers2,141 directions
Dusit Niyato
Nanyang Technological University
2,362 papers1,386 directions
- See all authors in Lacuna
The knowledge
papers leave out
Ask ML researchers about the knowledge they have not published: practical lessons, failed experiments, abandoned approaches, and promising next steps. The best researchers have far more ideas than time to execute.
The Coordinator processes a researcher’s request and routes the question to a colleague who might be able to contribute.
Gets sharper with every query
Althea carries your researcher profile, past work, and active context across sessions, so you can keep building without repeating yourself. Each session feels like one continuous collaboration with the same agent, rather than an isolated conversation.
“Tiptree Systems is reimagining how scientific research is conducted, communicated, and built upon, using AI to help researchers focus where human insight matters most.”
Hugo LarochelleMila Scientific Director, ex-DeepMind Principal ScientistAdvisor and investor
The commons belongs to all of us. Come help us tend it.
Publishing AI researchers
Your OpenReview account, or a short manual review
Free for researchers who give back by publishing. No card or trial clock
AI is the most consequential technology of our time, and it is being built behind closed doors. Our bet is that researchers working together can keep it open.