Scientific Literature & Discovery Agents

This page describes the Scientific Literature & Discovery Agents row of the Papers.md matrix: LLM agents for literature search, synthesis, and semi-autonomous discovery, spanning retrieval-augmented question answering over papers through to idea-to-experiment-to-manuscript pipelines. The row’s authoritative scope is its Taxonomy.md definition; this page synthesizes what currently sits in it.

Scope boundary

Among the agent rows, this one is defined by knowledge synthesis and the discovery workflow. What separates it from its neighbours is the object the agent operates on. An agent whose tools are chemistry tools belongs in Chemistry / Synthesis Agents; one whose tools are biomedical databases and analyses belongs in General-Purpose or Domain-Specific Biomedical Agents; one that drives instruments belongs in Robot Scientists & Lab Automation; and the frameworks, knowledge graphs and protocols such agents are assembled from belong in Agent Infrastructure. Here the substrate is the literature and the research process itself.

Every reference sits in the AI Tooling / Methodology column. That is not an accident of curation: no paper in this row has yet been applied to a cellular-agriculture problem. They are indexed because the workflow is the transferable part, and a cell-ag team delegating a literature review or an analysis pipeline to an agent is choosing among exactly these designs.

AI Tooling / Methodology

Retrieval-augmented literature agents

#44 PaperQA (Lála et al. 2023, arXiv, FutureHouse): the reference design for retrieval-augmented question answering over full-text papers. Four departures from standard RAG are worth naming because later systems inherit them. The RAG pipeline is decomposed into tools an agent can call repeatedly, so search can be re-run with different keywords when the evidence gathered is thin. A map-reduce step summarizes evidence from many sources before answering, giving the model a scratchpad. Chunks are ranked not only by vector distance but by an LLM-generated numerical relevance score, a second retrieval layer. And three separate LLM instances handle summarizing, asking and answering. The agent is instructed to keep searching until it has five or more pieces of evidence from multiple sources, and may reject its own answer and retry. Released with LitQA, a benchmark requiring synthesis across full texts rather than lookup. Catalogued in Software.md.
#46 PaperQA2 (Skarlinski et al. 2024, arXiv, FutureHouse): the successor, evaluated against subject-matter experts who had unrestricted internet access, search tools and time, which is a harder comparison than most agent papers set themselves. It reports matching or exceeding expert performance on retrieval, summarization and contradiction detection, and writes cited encyclopedia-style summaries the authors find more accurate than the corresponding human-written Wikipedia articles. Applied to contradiction detection it identifies 2.34 ± 1.99 contradictions per paper across a random sample of biology papers, 70% of which human experts validated. The methods section is unusually candid about reproducibility: the published results were produced on bespoke institutional infrastructure, and the open-source package’s paper-search tool is limited to whatever full text a given user can legally access, which is the binding constraint on anyone reproducing this. Released with LitQA2.
#221 IP-RAR (Feng et al. 2025, GigaScience): builds a stratified biomedical knowledge graph from full text rather than abstracts, then reasons over it. Over 8 million PubMed papers were downloaded and filtered to 15,585 covering lung, breast and colorectal cancer with more than 10 citations in the past decade. The paper is explicit about why full text was necessary: for much of the open-access corpus the XML holds only metadata and an abstract, so PDFs were converted to Markdown with the marker toolkit and mined with GPT-4o mini for entity-relationship triples plus structured summaries of each paper’s methods and datasets. Cross-document links come from entity co-occurrence. It reports a 20% improvement in document-retrieval F1 and 25% in answer accuracy, and ships BioCDQA for multi-hop cross-document evaluation. The full-text-versus-abstract lesson generalizes well beyond this paper.

Autonomous discovery pipelines

#45 The AI Scientist (Lu et al. 2024, arXiv, Sakana AI): generates ideas, writes and runs code, visualizes results, writes a paper and reviews it. The evaluation is the informative part: 50 ideas per run across three machine-learning templates, roughly 12 hours on 8 NVIDIA H100s, at roughly $10 to $15 per paper, compared across Claude Sonnet 3.5, GPT-4o, DeepSeek Coder and Llama-3.1 405b. Failure modes are reported plainly, GPT-4o struggling with LaTeX, DeepSeek Coder failing to call tools correctly, and the authors note that the novelty check is self-assessed by each model for its own ideas, which makes cross-model novelty comparisons hard to interpret. Code at SakanaAI/AI-Scientist; catalogued in Software.md.
#47 The AI Scientist-v2 (Yamada et al. 2025, arXiv, Sakana AI): removes v1’s two named limitations, its dependence on human-written code templates and its strictly linear hypothesis-testing loop. In their place is a progressive agentic tree search managed by a dedicated experiment-manager agent, adapted from AIDE, where each node is a candidate solution with a scalar score and promising nodes are selected for refinement. A vision-language model feedback loop iterates on figure content and appearance. Three fully autonomous manuscripts were submitted to a peer-reviewed ICLR workshop and one exceeded the average human acceptance threshold. Code at SakanaAI/AI-Scientist-v2.
#153 Co-Scientist (Gottweis et al. 2026, Nature, Google): a Gemini-based multi-agent system organized around an asynchronous task framework, where a Supervisor agent manages a worker queue of specialized agents and a persistent context memory carries state across long horizons. Hypotheses are improved by a tournament evolution process, and the paper reports that quality keeps improving as test-time compute is scaled. Five default criteria are enforced on every hypothesis: alignment with the stated goal, plausibility with contradictions made explicit, novelty beyond synthesis of existing work, testability within the scientist’s stated constraints, and safety. Validated on drug repurposing, target discovery and antimicrobial resistance, with acute myeloid leukaemia repurposing candidates and synergistic combinations confirmed in vitro.
#154 Robin (Ghareeb et al. 2026, Nature, FutureHouse): couples PaperQA2-derived literature agents (Crow and Falcon) with a data-analysis agent (Finch) that edits and executes Jupyter cells, closing a lab-in-the-loop cycle in which the system proposes hypotheses, humans run the experiments, and the system analyses the results and revises. Applied to dry age-related macular degeneration, it proposed enhancing retinal-pigment-epithelium phagocytosis, identified ripasudil (a clinical ROCK inhibitor never previously proposed for the indication) and KL001, confirmed both in vitro, then proposed a follow-up RNA-seq experiment that implicated ABCA1. The paper states that all hypotheses, experimental directions, analyses and main-text figures were produced by Robin. Finch was separately benchmarked on 170 BixBench questions against its own base model without agent capabilities, with failure modes classified rather than aggregated. Also in Robot Scientists & Lab Automation. Code at Future-House/robin; data under PRJNA1464762.
#166 ERA (Aygün et al. 2026, Nature, Google Research): targets a narrower bottleneck than the systems above, the manual writing of scientific software, and optimizes directly against a quality metric. An LLM proposes Python code that is executed and scored in a sandbox, and a PUCT tree search inspired by AlphaZero explores the space of programs. The authors are precise about the departure from AlphaZero: because the children of a node cannot be enumerated, every node stays a candidate for expansion and sampling is flat over the whole tree rather than recursive from the root, which makes it closer to Flat UCB than to MCTS, and lets the search backtrack from any historical node when a path plateaus. Scores saturate after roughly 300 to 1,000 nodes. In bioinformatics it produced 40 novel single-cell analysis methods that beat the top human entries on a public leaderboard, and in epidemiology 14 models that beat the CDC ensemble on COVID-19 hospitalization forecasting. The single-cell result is the one to note here: the analysis methods a cell-ag lab uses are exactly this kind of target. Code at google-research/era.
#222 cmbagent (Gandhi et al. 2025, arXiv): addresses the verification problem that the pipelines above mostly leave open, namely how to tell mid-run whether an agent has gone wrong. Its answer is to treat the appearance of a new plot as a checkpoint: a GPT-4o model generates a domain-specific rubric of visual criteria from the task prompt, and a Plot Judge agent evaluates the figure as a black box against that rubric without seeing the code. In correction mode a failing plot routes into a debug-and-revise loop; in discovery mode a plot showing scientifically interesting features routes into exploratory analysis instead. Planner and Plan Reviewer agents propose and critique a plan, a Control agent executes it, and specialized agents receive only pruned subsets of the shared state. On a 10-task data-driven discovery benchmark it reports pass@1 of 0.7 to 0.8, against 0.2 to 0.3 for code-only and 0.4 to 0.5 for code-and-text baselines, with auditable reasoning traces. Case studies are in cosmology and astrochemistry. Code at CMBAgents/cmbagent.
#273 SciAgents (Ghafarollahi & Buehler 2025, Advanced Materials, MIT): generates and refines hypotheses by reasoning over a large ontological knowledge graph, 33,159 nodes and 48,753 edges built from around 1,000 papers. The generative mechanism is a pathfinding algorithm rather than a prompt: a modified Dijkstra search between a source and target concept, with distances estimated from node embeddings, a randomness factor of 0.2 injected into the priority queue so paths are not deterministic, and random waypoints drawn from off-path neighbours to force exploration of alternative routes. The resulting path plus its second-hop neighbourhood becomes the substrate the agents reason over. Applied to biologically inspired materials, it reports a new biocomposite with improved mechanical properties and more energy-efficient production. The scaffolding and biomaterials overlap with cell-ag is direct even though the paper does not make it. Code at lamm-mit/SciAgentsDiscovery.

Two threads run across this row and are worth extracting. The first is that verification, not generation, is where these systems differ most: self-assessed novelty in #45, tournament evolution in #153, a VLM plot judge in #222, a scalar quality metric in #166, and wet-lab confirmation in #154. The second is that access to full text is repeatedly the limiting factor rather than model capability, stated outright in #46 and #221.

Adjacent methods

General-Purpose Biomedical Agents and Domain-Specific Biomedical Agents: agents whose tools are biomedical rather than bibliographic.
Chemistry / Synthesis Agents: the chemistry-tool counterpart, which borrows this row’s retrieval components.
Robot Scientists & Lab Automation: what happens when the loop reaches physical hardware.
Agent Infrastructure (Frameworks, KGs, Protocols): the knowledge graphs and orchestration layers these systems are built on.
Benchmarks & Evaluation Frameworks: LitQA, LitQA2, BixBench and the rest of the eval substrate these papers are scored against.