Inventa: Decentralized Self-Improving Agents for Autonomous Scientific Discovery

Prithwish Jana, Nikos Kanakaris, Purak Jain, Anoop Deoras, Sahika GencarXiv 2026 · arXiv

TL;DR

Inventa is a decentralized framework for autonomous scientific discovery that organizes LLM agents into independent "research units" (akin to individual labs) that explore ideas in parallel. By sharing only verified findings through a central knowledge base and self-adapting their own search strategies, these agents achieve state-of-the-art performance across the entire research cycle—from literature retrieval to discovering scientific laws—while using significantly fewer computational resources.

What Problem Does This Paper Address?

While Large Language Model (LLM) agents have shown promise in assisting research, simply throwing more agents or iterations at a problem doesn't necessarily lead to better science. Current systems often suffer from three major pitfalls:

  1. Diversity Collapse: Agents tend to converge on the same "safe" ideas, narrowing the search space.
  2. Error Propagation: If agents share unverified thoughts or "hallucinations" in a shared memory, these errors can infect the entire system's reasoning.
  3. Static Strategies: Most agents use fixed prompts and tools. If a strategy stops working, the agent gets stuck in a "local optimum" and stops making progress.

The authors of Inventa aim to bridge the "novelty–feasibility gap," ensuring that AI-generated ideas are not just creative, but also scientifically sound and grounded in data.

Methods

Inventa draws inspiration from how the scientific community actually works: independent labs pursuing different paths while building on a foundation of peer-reviewed evidence.

Overview of the Inventa framework, showing the interaction between autonomous research units, verifiers, and shared knowledge.

The framework consists of three main components:

  • Autonomous Research Units (ARUs): Each unit acts as a mini-lab with its own "Idea Tree." A Mutator Agent (the junior researcher) proposes new ideas, while a Local Orchestrator (the PI) manages the strategy.
  • Self-Adaptive Orchestration: If a unit stalls, the Orchestrator can RETOOL (evolve the agent's prompts and tools), REDRAW (swap the underlying LLM for a better-performing one), or RESTART (archive the lineage and try a completely new direction).
  • Verified Shared Knowledge: This is the only communication channel between units. Before any finding is published here, it must pass through rigorous Verifiers that perform held-out scoring, constraint checking (e.g., laws of thermodynamics), and ablation studies to ensure every component of an idea actually adds value.

Datasets

The researchers evaluated Inventa across five benchmarks representing different stages of the scientific process:

  • ScholarCatalyst: Literature retrieval (finding papers that inspire new research).
  • IdeaBench: Ideation (generating novel research ideas from existing abstracts).
  • IdeaScientist: Proposal writing (creating grounded, detailed research proposals).
  • EurekaBench & SciLaws-Bench: Scientific law discovery (finding closed-form equations that explain complex data, such as ice-shelf flow or astrophysics).

Key Insights and Ideas

  • The Power of the Bottleneck: By restricting inter-agent communication to only verified findings, Inventa prevents the spread of "reward hacks" and unsupported claims that plague other multi-agent systems.
  • Harness Evolution vs. Idea Evolution: Most systems only try to improve the output. Inventa also improves the process by evolving the "harness" (the prompts and control flow) that generates the ideas.
  • Decentralization as a Feature: Keeping units independent ensures they don't "groupthink" themselves into a corner. Each unit maintains its own history, allowing for a much broader exploration of the hypothesis space.

Key Results

Inventa outperformed state-of-the-art baselines across all five benchmarks, often by substantial margins.

Performance gains of Inventa over the best baselines across different research stages.
  • Scientific Discovery: On EurekaBench, Inventa achieved a 73.6% conditioned score, significantly higher than the previous leader, CORAL (56.2%). Crucially, it was the only system to pass every scientific-constraint test.
  • Efficiency: On IdeaBench, it discovered 26.0 novel ideas per paper (compared to 18.6 for ResearchAgent) while requiring 82% fewer model calls.
  • Scaling: Unlike other methods that plateau as you increase the budget, Inventa showed steady gains with more computation, effectively turning "extra budget into steady gains."

Limitations and Future Work

The authors acknowledge that the rigorous verification process introduces some computational overhead. Additionally, while the system is highly effective at discovering laws and generating proposals, it currently operates in digital "sandboxes."

Future work could involve:

  • Integrating Inventa with physical laboratory automation (Self-Driving Labs).
  • Expanding the verifiers to handle even more complex, multi-modal scientific data.
  • Exploring how human researchers can best "intervene" or collaborate with these decentralized units without disrupting their autonomous exploration.

Summary updated

Abstract

Large language model (LLM) agents can generate research ideas and improve scientific discovery through iterative experimentation. However, increasing the number of agents or iterations does not automatically translate to producing more diverse or feasible ideas. Agents may converge on similar candidates, propagate unsupported claims through shared memory components and repeatedly apply search strategies that no longer yield any improvements. We introduce Inventa, a framework for automated scientific discovery that employs autonomous research units and verified shared knowledge. Each unit creates a lineage of candidate ideas using a mutator agent. A local orchestrator agent periodically adapts the mutator's harness. When progress stalls, the orchestrator agent replaces its model or restarts the idea evolution process. Verifiers evaluate each candidate idea and record verified improvements, useful components and constraint violations in a shared knowledge, the only channel between units. We evaluate Inventa on five benchmarks spanning the research cycle, covering retrieval (ScholarCatalyst), ideation (IdeaBench), proposal writing (IdeaScientist), and scientific law discovery (EurekaBench, SciLaws-Bench). It leads all five, achieving 73.6% on EurekaBench versus 56.2% for CORAL. On IdeaBench, it discovers 26.0 novel ideas per paper, compared with 18.6 for ResearchAgent, with 82% fewer model calls.
BibTeX
@misc{jana2026inventa,
  author = {Prithwish Jana and Nikos Kanakaris and Purak Jain and Anoop Deoras and Sahika Genc},
  title = {{Inventa: Decentralized Self-Improving Agents for Autonomous Scientific Discovery}},
  year = {2026},
  eprint = {2611.21847},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.21847},
}
Figure