SciResearcher: Frontier Science Agent Framework
- SciResearcher is an automated framework that constructs frontier-science tasks to post-train research agents using evidence-grounded data.
- It employs a three-stage seed entity acquisition and dual pipelines—conceptual and computational—to generate challenging scientific reasoning tasks.
- Benchmarks indicate significant improvements in agent performance by integrating advanced evidence retrieval, multi-step computation, and adaptive tool use.
SciResearcher is a fully automated agentic framework for constructing frontier-science training data and, through that data, an agent foundation model for scientific reasoning. In the provided literature, it belongs to a different class than scholarly-profile systems such as Scholia, exploratory search systems such as SciSight and COVID-SEE, or retrieval-augmented assistants such as OpenResearcher and Scholar Inbox. Those systems center on discovery, profiling, recommendation, or visualization of the literature, whereas SciResearcher centers on generating evidence-grounded conceptual and computational tasks that can post-train deep research agents for frontier scientific reasoning (Zheng et al., 2 May 2026, Nielsen et al., 2017, Hope et al., 2020, Verspoor et al., 2020, Zheng et al., 2024, Flicke et al., 11 Apr 2025).
1. Frontier scientific reasoning as the target domain
SciResearcher is motivated by a specific claim: frontier scientific reasoning differs materially from general information seeking. The target problems are described as “challenging, expert-level scientific problems at the boundaries of human knowledge,” where useful evidence is incomplete, evolving, and distributed across sparse and heterogeneous academic sources. The intended agent must therefore do more than factual recall: it must retrieve scientific evidence, integrate weakly connected sources, understand domain-specific mechanisms, and perform nontrivial scientific computation (Zheng et al., 2 May 2026).
The framework is explicitly presented as a response to two limitations in prior deep-research-agent pipelines. First, knowledge-graph-based task construction is said to be poorly matched to frontier science because scientific ontologies are “heterogeneous, noisy, and highly context-dependent.” Second, iterative web-browsing approaches are said to degrade when scientific knowledge is scattered across loosely connected academic sources rather than concentrated in canonical pages. A further limitation is that prior automated task construction has emphasized short-answer factual supervision, whereas frontier science often requires extraction and execution of scientific models (Zheng et al., 2 May 2026).
A common misconception is to treat SciResearcher as primarily a literature-search interface. In the provided corpus, interfaces such as OpenResearcher, Scholar Inbox, WisPaper, SciSight, and COVID-SEE address that layer through RAG, recommendation, faceted exploration, or closed-loop literature management (Zheng et al., 2024, Flicke et al., 11 Apr 2025, Ju et al., 7 Dec 2025, Hope et al., 2020, Verspoor et al., 2020). SciResearcher instead treats the literature as a substrate for automated data construction and agent post-training (Zheng et al., 2 May 2026).
2. Automated frontier-science data construction
The framework begins with seed entity acquisition. This follows a three-stage pipeline: extraction of an initial entity pool from public datasets with domain and ontology annotations, LLM-based expansion of that pool within the same ontology, and automatic scoring of entities by frontier relevance, concreteness, and specificity. Only the top 5% of entities are retained as downstream seeds (Zheng et al., 2 May 2026).
From those seed entities, SciResearcher constructs two task families. Conceptual tasks target evidence-grounded reasoning over the literature. Computational tasks target scientific model recovery, scenario instantiation, and quantitative solution. Together they form SciResearcherQA, which is then combined with TRQA-Literature and SciBench for post-training (Zheng et al., 2 May 2026).
The framework also includes a final question postprocessing stage. Both conceptual and computational questions undergo evidence-claim entailment checking, reasoning-shortcut detection, and sanity checking. After this shared stage, conceptual questions receive textual obfuscation, proofreading, distractor balancing, and shortcut mitigation, whereas computational questions receive selective masking of equations and injection of domain-specific search hints so that the final problem requires retrieval and reconstruction of the underlying model rather than direct reading (Zheng et al., 2 May 2026).
This design suggests that SciResearcher is less a benchmark-only project than a data-factory architecture for difficult scientific tasks. The emphasis is on scalable automated generation rather than manual curation (Zheng et al., 2 May 2026).
3. Conceptual and computational synthesis pipelines
The conceptual pipeline begins with seed2question. Starting from a seed entity, a web agent performs iterative scout searches over academic sources and uses a url2evidence sub-agent to extract supporting evidence and formulate an initial conceptual question. At this stage, the question is intended to be answerable from a single authoritative academic source (Zheng et al., 2 May 2026).
Difficulty is then increased through anchor-based augmentation. An anchor is defined as “the key, decisive scientific entity that plays a central role in a question, serving as its signature referent and the primary handle for reasoning about the problem.” Candidate anchors are scored according to whether they are decisive for the answer, decoupled from answer-option surface forms, and sufficiently specific and concrete for evidence-grounded expansion. The selected anchor becomes the answer to a newly generated question supported by additional academic evidence, gathered by a fresh web-agent instance, and that new question is fused back into the earlier one. Repetition yields a multi-hop, multi-source reasoning problem (Zheng et al., 2 May 2026).
The computational pipeline is organized around seed2equation and a three-level evidence-selection process. First, scout search looks for candidate sources involving computational models, numerical simulations, mechanistic equations, kinetic models, ODE/PDE systems, statistical models, and related formulations. Second, candidate sources are filtered by eval_urls, which scores them on four dimensions:
- Model Exclusiveness: whether the model is source-dependent rather than generic.
- Search Identifiability: whether the source is findable from clues but not trivial.
- Computational Complexity: whether solving requires nontrivial numerical reasoning.
- LLM Unfamiliarity: whether the exact model is niche, recent, specialized, or unlikely to have been memorized.
Third, sub-agents perform deep model extraction, recovering the model name, scientific purpose, governing equations, variable definitions, parameter definitions and units, assumptions, scenario, and constraints (Zheng et al., 2 May 2026).
The resulting computational question is scenario-based: it asks the agent to retrieve the source, reconstruct the scientific model, instantiate it correctly, and compute a numerical answer. Quality control is unusually strict. The system samples five candidate Python solvers from proprietary LLMs, executes them, rejects questions in several pathological cases, and determines the final answer by majority voting followed by LLM-based verification. Failed items are redesigned (Zheng et al., 2 May 2026).
4. Agent architecture and post-training
SciResearcher is implemented on Cognitive Kernel-Pro. The final model, SciResearcher-8B, uses Qwen3-8B (without thinking) as the trainable main agent. The web and file sub-agents use Qwen3-32B (without thinking) and remain frozen throughout training and evaluation. The framework also removes the ask_llm function from the main agent, explicitly to prevent unsupported shortcut answers from the base LLM and to force evidence collection (Zheng et al., 2 May 2026).
The main agent interacts with a tool environment. Its tools are web_agent(task), file_agent(task), simple_web_search(query), and stop(answer, summary). The web agent exposes actions such as click, type, scroll_up, scroll_down, wait, goback, restart, goto, save, and screenshot, while the file agent exposes load_file, read_text, read_screenshot, and search (Zheng et al., 2 May 2026).
Training follows a two-stage recipe: cold-start supervised fine-tuning and reinforcement learning with GRPO. Teacher trajectories for SFT are collected using Claude-Sonnet-4.5 with rejection sampling. RL then applies outcome-only rewards to improve task-completion strategies. Only main-agent trajectories are used as training targets; the sub-agents remain external tools (Zheng et al., 2 May 2026).
The post-training mixture is summarized below.
| Data source | Tasks | Step-level messages |
|---|---|---|
| SciResearcherQA-Concept | 371 | 2,872 |
| SciResearcherQA-Compute | 104 | 951 |
| TRQA-Literature | 172 | 932 |
| SciBench | 80 | 350 |
| Total | 727 | 5,105 |
SciResearcher’s own automatically constructed portion is therefore 475 tasks: 371 conceptual and 104 computational (Zheng et al., 2 May 2026).
5. Benchmarks, results, and observed behavior
SciResearcher is evaluated on three benchmarks: HLE-Bio/Chem-Gold with 149 questions, SuperGPQA-Hard-Biology with 92 questions, and TRQA-Literature with 172 questions (Zheng et al., 2 May 2026).
| Benchmark | SciResearcher-8B-RL | Additional result |
|---|---|---|
| HLE-Bio/Chem-Gold | 19.46% pass@1 | 31.54% pass@3 |
| SuperGPQA-Hard-Biology | 35.87% | 51.09% pass@3 |
| TRQA-Literature | 49.42% | 60.47% pass@3 |
On the baseline CK-Pro + Qwen3-8B, the corresponding scores are 8.05 on HLE-Bio/Chem-Gold, 22.83 on SuperGPQA-Hard-Biology, and 34.88 on TRQA-Literature. The absolute gains are therefore 11.41, 13.04, and 14.54 points, respectively (Zheng et al., 2 May 2026).
The paper emphasizes the 19.46% HLE result because it is a state of the art at that parameter scale and exceeds several proprietary scientific agents listed in the comparison, including SciMaster + GPT-4.1 and Biomni + GPT-4.1, while remaining below OpenAI Deep Research + o4-mini at 22.82 on HLE (Zheng et al., 2 May 2026).
The ablation study attributes most of the improvement to the automatically constructed SciResearcher data. Relative to the baseline Qwen3-8B agent, adding SciResearcherQA-Concept raises HLE from 8.05 to 10.74 and SuperGPQA from 22.83 to 25.00; adding SciResearcherQA-Compute further raises them to 12.08 and 28.26; adding TRQA + SciBench reaches 12.75 and 31.52 in the SFT stage (Zheng et al., 2 May 2026).
Behavioral analysis is also central. After training, agent trajectories become roughly 0.3× to 2.7× longer than the baseline, and tool use increases substantially for both simple_web_search and web_agent. The RL model uses longer trajectories on the hardest benchmark, HLE-Gold, but slightly fewer steps on easier benchmarks while still improving accuracy. The paper interprets this as adaptive allocation of search and reasoning effort by task difficulty (Zheng et al., 2 May 2026).
6. Position within the broader scholarly-systems landscape
Within the surrounding literature, SciResearcher occupies a distinctive position. Scholia builds scholarly profiles and scientometric views on top of Wikidata (Nielsen et al., 2017). SciSight combines biomedical facet exploration with research-group detection for exploratory scientific search (Hope et al., 2020). COVID-SEE adds post-retrieval evidence organization through PICO relations, topic modeling, and concept clouds (Verspoor et al., 2020). OpenResearcher is a modular RAG assistant for scientific QA, summarization, and recommendation (Zheng et al., 2024). Scholar Inbox focuses on personalized paper recommendation and daily monitoring (Flicke et al., 11 Apr 2025). WisPaper integrates search, library management, and AI feeds in a closed-loop workflow (Ju et al., 7 Dec 2025). MetaInfoSci unifies bibliometric, scientometric, and network analytics in a web platform (Sharmaa et al., 4 Jun 2025). CycleResearcher couples automated research generation with automated review in an iterative preference-training loop (Weng et al., 2024).
SciResearcher differs from all of these in object and training target. It is not primarily an interface for search, recommendation, bibliometric analysis, or manuscript drafting. It is a framework for constructing frontier-science tasks that elicit information acquisition, tool-integrated reasoning, long-horizon planning, and scientific computation, and then using those tasks to post-train an agent foundation model (Zheng et al., 2 May 2026).
That distinction also clarifies several limitations. The domains emphasized are mainly biology, chemistry, biomedicine, and related interdisciplinary areas. The pipeline is fully automated, but the paper does not claim expert validation for all generated SciResearcherQA items. The benchmark sets are relatively small—149, 92, and 172 questions—and the compute burden is evidently large because the framework uses multi-agent browsing, recursive augmentation, deep source analysis, multiple Python solvers, solver execution, and RL training. The paper also provides only limited discussion of safety and no dedicated safety section (Zheng et al., 2 May 2026).
A second common misconception is that SciResearcher already constitutes a complete autonomous scientist. The evidence supports a narrower conclusion. It is a scalable recipe for constructing difficult scientific training data and for improving an open 8B research agent on frontier-science benchmarks. This suggests a shift from paper-search assistance toward agent training grounded in academic evidence and computation, but not the completion of fully autonomous scientific discovery (Zheng et al., 2 May 2026).