# Literature reviews with sources

> Ask one broad question and get back a structured review you can check.

Althea's deep-research agent runs the literature review end to end: it takes a natural-language question, retrieves relevant papers over [the commons](/lacuna) and the open web, and synthesizes them into a structured review. Sources stay attached to every claim, so checking its reading is one click. The engine is the one we benchmark in [the Lacuna preprint](https://arxiv.org/abs/2606.26246).

## Benchmarked in public

The engine under this workflow is benchmarked in our own preprint: across 25 ReportBench-ML survey tasks (derived from peer-reviewed surveys, scored by overlap with the experts' own reference lists), Lacuna Deep Research hit 99 expert-reference matches to GPT-Researcher's 72, with report quality rated 7.82/10 to 5.24 ([arXiv:2606.26246](https://arxiv.org/abs/2606.26246)).

A Careful run fans out six high-effort "lens" workers over the map, each reading from a different angle. Worker failures stay isolated; synthesis proceeds from the lenses that survive. The report only fails if all six do.

## The research loop

Four stages convert the question into a review you can check:

1. Query generation. Althea decomposes your prompt into specific academic search queries.
2. Retrieval. It runs parallel searches across academic databases and the open web for papers, patents, and technical articles.
3. Analysis. It reads abstracts and key sections of what it retrieved and extracts the relevant findings.
4. Synthesis. It identifies common themes, resolves conflicting results, and structures the review.

The review lands in the conversation, and the conversation continues.

_[diagram: The research loop: query generation, retrieval, analysis, synthesis, and the review returns to the chat with sources on every claim.]_

## While it runs

A Careful run works in the background. Open another conversation, or close the laptop; it keeps going, and anything you add mid-run is picked up when it reports back.

## Worked example: fire detection from remote sensing

A researcher brought a broad question about fire detection from remote sensing. Back came a structured report that identified Test-Time Adaptation and Meta-Learning as the architectural patterns worth testing, then went further, with a technical assessment of candidate datasets on Zenodo (Sen2Fire, EO4WildFires, Fire-D), judged by sensor type, label availability, and format. Every step is below; the other two worked examples live on [the code agents](/code) and [summaries](/documents) pages.

### Step 1 · The prompt

The prompt, verbatim.

> **Researcher:** I'm looking for scientific literature related to fire detection from remote sensing data. I'm thinking about the possibility of developing a system which can automatically and autonomously finetune a model based on specific geospatial datasets. The idea being that it can be too expensive to pre-train on everything in geospatial, so we just pull down from zenodo and do a quick finetune at test time when we know about the region of interest. Can you find relevant papers we could test against doing fire detection with AI systems, and do an assessment of relevant datasets on Zenodo?

### Step 2 · Retrieval

Retrieval starts across academic indices and dataset repositories.

> **Althea:**
>
> Research Agent — `tiptree-systems/scientific-research/search-literature`
> - query: "fire detection remote sensing finetuning domain adaptation few-shot"
> - query: "wildfire detection remote sensing zenodo dataset"
>
> Web Search
> - query: "site:zenodo.org fire detection remote sensing dataset"
> - query: "site:zenodo.org wildfire satellite dataset labeled"
>
> _Considering Literature Search — I've initiated the literature search, focusing on keywords like "fire detection remote sensing finetuning" and "domain adaptation." Preliminary results suggest a promising direction, but I'll need to dig deeper for papers that explicitly address finetuning at test time or leverage Zenodo datasets. My goal is to find relevant research that aligns with the user's specific request for an automated finetuning system._
>
> _Sampling distribution…_

### Step 3 · Synthesis

The report takes shape; Test-Time Adaptation and Meta-Learning surface as the patterns to test.

> **Althea:** This is a very viable research direction. The concept you are describing aligns closely with **Test-Time Adaptation (TTA)** and **Few-Shot Learning** in remote sensing. The core challenge in geospatial analysis is indeed the "domain shift" (e.g., a fire in Australian bushland looks different spectrally from a fire in a Canadian boreal forest due to vegetation types, atmospheric conditions, and sensor angles). Instead of training a massive monolithic model, your proposed architecture—fetching a local dataset and adapting a base model "on the fly"—is supported by recent literature.
>
> **1. Scientific Literature & Methodologies**
> - **Test-Time Adaptation (TTA):**
>   - *Generalizable Disaster Damage Assessment via Change Detection*: This recent work discusses using TTA strategies to adapt vision foundation models for disaster damage, which is directly analogous to your fire detection use case.
>   - *Boundary-Aware Test-Time Adaptation*: While general-purpose, techniques like "BA-TTA-SAM" (Segment Anything Model adaptation) are being applied to remote sensing to handle the specific boundaries of natural phenomena like fires without extensive retraining.
> - **Few-Shot & Meta-Learning:**
>   - *Meta-FSEO (Remote Sensing, 2021)*: A meta-learning approach for fast adaptation in remote sensing scene classification. This proves that embedding optimization can allow models to adapt to new geospatial domains with very few samples.
>
> [… the report continues]

### Step 4 · The report

The final report, including the Zenodo dataset assessment.

> **Althea:** I have analyzed the available datasets on Zenodo. These are excellent candidates for your "pull down and finetune" workflow because they are pre-processed and relatively lightweight compared to raw satellite archives.
>
> **2. Assessment of Zenodo Datasets**
> - **A. Sen2Fire** (Best for "Base" Model Training)
>   - **Sensors:** **Sentinel-2** (13 bands) + **Sentinel-5P** (Aerosol Index). **Labels:** Binary wildfire masks.
>   - **Assessment:** This is likely your best starting point. It is already formatted as tensors (NumPy), making it trivial to load into PyTorch/TensorFlow for rapid finetuning.
> - **B. EO4WildFires** (Best for Time-Series/Forecasting)
>   - **Assessment:** If your system needs to detect fires through clouds (using SAR) or predict spread, this is the dataset.
> - **C. Fire-D** (Best for Multi-Modal/Airborne)
>   - **Content:** A mix of satellite (GOES, Planet) and airborne sensors (AVIRIS-C, MASTER).
>
> **Recommended Workflow**
> - 1. **Pre-train** a lightweight encoder (like a ResNet-18 or a small ViT) on **Sen2Fire** to learn the spectral signatures of fire and smoke.
> - 2. **At Test Time:**
>   - Define the Region of Interest (ROI).
>   - Query Zenodo (or a local cache) for a dataset geographically similar to the ROI (e.g., "Mediterranean scrubland").
>   - Run a **Few-Shot Finetuning** step (e.g., 5-10 gradient steps) on the pre-trained model using this local data.
>   - Inference on the live feed.
>
> **Althea:** This approach minimizes compute costs (no training on petabytes of data) while maximizing local accuracy.

## Start from where you are reading

Already on a paper or research direction in [Lacuna](/lacuna)? Ask from there, and the page's context comes attached: metadata, extracted concepts, figures, neighboring work. You do not paste abstracts. Focused questions ("what would I need to reproduce this result?") work best from a Lacuna page; the wide sweep works best from here. They chain, focused first, then wide.

## After the deep research task completes

Tell Althea to monitor the area instead of re-running the question every month. Set a standing monitor in [Stay current](/current) and it reaches you by email, SMS, or WhatsApp when the field moves.