SciLENS: RL-Driven Autonomous Agents for Scientific Localized Evidence Navigation and Synthesis

Leqi Zheng, Jinbo Su, Yuying Li, Chaokun Wang, Weiping Wang, Haitao Li, Jiajun Zhang, Shannan Yan, Zhaolu Kang, Rong Fu, Jie Wu, Fang Niu, Hang ZhangarXiv 2026 · arXiv

TL;DR

SciLENS is a fully local, autonomous agent designed to synthesize scientific literature by navigating a massive database of 12 million academic records. By integrating structural visualization tools directly into its reasoning loop and using a novel rubric-based reinforcement learning strategy, SciLENS achieves performance comparable to top-tier proprietary models like GPT-5.2 and Gemini-3.0-pro—all without needing an internet connection or external APIs.

What Problem Does This Paper Address?

Modern researchers are drowning in a sea of papers, making it nearly impossible to track methodological origins or macro-level trends manually. While AI agents have emerged to help, they face three major hurdles:

  1. Proprietary Dependency: Most current agents rely on online services (like OpenAI or Google APIs), which raises concerns about privacy, reproducibility, and high costs.
  2. The Text Bottleneck: Agents typically communicate in pure text. When trying to summarize dozens of papers, the "context window" (the agent's short-term memory) gets overwhelmed, leading to "hallucinations" where the AI makes up facts.
  3. Alignment Scarcity: Training an agent to perform complex, multi-hop scientific reasoning requires high-quality data that is incredibly expensive to produce with human experts.

SciLENS tackles these by building a "localized" system that can "see" data structures through visualization, reducing the cognitive load on the AI itself.

Methods

The authors developed a multi-layered framework to turn a chaotic corpus into structured insights.

  • Localized Dual-Tier Infrastructure: Instead of searching the web, SciLENS uses a local index of 12 million papers. It combines MongoDB (for metadata and citation tracking) with FAISS (a vector database for semantic search) to achieve sub-second retrieval speeds.
  • Visualization-Augmented Reasoning: This is the "secret sauce." The agent doesn't just write text; it can autonomously invoke tools to generate charts (line, bar, scatter, etc.). By compressing complex citation topologies into visual schemas, the agent prevents its own memory from overflowing.
  • Automated Data Synthesis: To train the model without human labels, the team created a pipeline that extracts subgraphs from citation networks and generates multi-hop questions. These are then verified by a "cross-model consensus" involving 20 different frontier LLMs to ensure accuracy.
  • Rubric-Based Reinforcement Learning (RL): The agent is fine-tuned using the IcePop algorithm. Crucially, it isn't just rewarded for the right answer; it's judged against a multi-dimensional rubric that evaluates its planning, how it decomposes complex questions, and how strictly it grounds its claims in evidence.
Figure 1: SciLENS operates as a refracting prism: chaotic sci- entific literature enters the system and is decomposed through four tightly integrated components into structured, multimodal synthesis reports. All processing occurs fully offline with 12M locally indexed papers.

Datasets

The system was trained and tested on a massive scale:

  • Open Academic Graph (OAG): A corpus of ~12 million cleaned academic records.
  • Standard Benchmarks: SciFact, PubMedQA, QASA, and ScholarQA-CS were used to measure basic correctness and citation accuracy.
  • New Benchmarks: The authors introduced the Structural Synthesis Benchmark (SSB) and Scientific Fact and Reasoning (SciFR). These specifically test an agent's ability to handle visual consistency, semantic fidelity, and multi-hop evidence attribution.

Key Insights and Ideas

  • Charts as Cognitive Compression: The most striking finding is that visualization isn't just for the human reader. For the AI, a chart is a high-density data structure that replaces pages of confusing text, effectively "upgrading" its ability to reason over large datasets.
  • Local Beats Online: By eliminating the latency and rate limits of web-search APIs, SciLENS can perform much deeper exploration—up to 200 rounds of reasoning per query—which is often impossible for online agents.
  • Strategic Tool Selection: The agent doesn't just click buttons randomly. Analysis showed it shifts its strategy based on the task: for fact-checking, it leans heavily on keyword search; for macro-synthesis, it pivots to citation-graph traversal and visualization.
The SciLENS architecture integrates local infrastructure, a research toolbox, and an RL-driven learning pipeline.

Key Results

SciLENS-RL (the version trained with reinforcement learning) consistently outperformed open-source baselines and rivaled the best proprietary models:

  • Performance: It achieved a score of 47.62 on QASA and 88.94 on SciFact, outperforming specialized agents like A-Searcher and Web-Thinker.
  • Structural Synthesis: On the new SSB benchmark, SciLENS scored 0.7607, significantly higher than GPT-5.2 (0.6263) and Claude-4.5-Sonnet (0.4959).
  • Ablation Study: When the researchers disabled the visualization tools, the agent's performance on structural synthesis tasks plummeted from 0.76 to 0.52, proving that the ability to "see" data is critical for high-level scientific summary.

Limitations and Future Work

While powerful, SciLENS is currently a "localized experimental sandbox." The authors note that transitioning this to a live, industrial-scale deployment with real-time updates is the next big challenge. Additionally, the agent can still suffer from "attention hijacking" when faced with highly convoluted queries containing many distractors. Future work will likely focus on improving the agent's robustness against these complex informational "noise" scenarios and expanding the corpus beyond the current 12 million records.

Summary updated

Abstract

Scientific literature synthesis agents increasingly rely on proprietary online services, limiting reproducibility, privacy, and offline deployment. To address this challenge, we introduce SciLENS Scientific Localized Evidence Navigation and Synthesis), a fully local autonomous agent framework operating on a dual-tier infrastructure indexing approximately 12 million academic records. SciLENS pioneers the integration of structural visualization as an actionable tool within the reasoning loop, enabling the agent to compress complex citation topologies into validated data-driven charts and thereby mitigate context exhaustion during macro-level synthesis. To train the agent without human annotation, we develop an automated data synthesis pipeline that extracts multi-hop subgraphs from a citation knowledge graph, verified by cross-model consensus among 20 frontier models. The agent is subsequently aligned through a reverse-decomposition rubric strategy that provides fine-grained process rewards for early planning and strict evidence grounding. Evaluations across six scientific benchmarks encompassing standard QA, citation accuracy, factual reasoning, and structural synthesis demonstrate that SciLENS significantly outperforms open-source baselines and achieves performance comparable to GPT-5.2 and Gemini-3.0-pro. Our source code and data are released at https://github.com/LQgdwind/SciLENS.
BibTeX
@misc{zheng2026scilens,
  author = {Leqi Zheng and Jinbo Su and Yuying Li and Chaokun Wang and Weiping Wang and Haitao Li and Jiajun Zhang and Shannan Yan and Zhaolu Kang and Rong Fu and Jie Wu and Fang Niu and Hang Zhang},
  title = {{SciLENS: RL-Driven Autonomous Agents for Scientific Localized Evidence Navigation and Synthesis}},
  year = {2026},
  eprint = {2609.03338},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2609.03338},
}
Figure