SciResearcher-8B Posts 19.46% on HLE-Bio/Chem-Gold, Beating Larger Closed Agents
A new arXiv revision from a team led by Tianshi Zheng claims an 8B-parameter research agent matches or exceeds several larger proprietary deep-research systems on three frontier-science benchmarks. The trick is not model size but a fully automated pipeline that synthesizes its own multi-hop training data from real academic papers. If the result holds outside the benchmark suite, the bottleneck for science-agent quality just moved from compute to data construction.
The paper, SciResearcher: Scaling Deep Research Agents for Frontier Scientific Reasoning (arXiv:2605.01489, revised May 26), reports that an 8-billion-parameter open model reaches 19.46% pass@1 and 31.54% pass@3 on HLE-Bio/Chem-Gold, plus 13-15 point absolute gains on SuperGPQA-Hard-Biology and TRQA-Literature. The authors frame this as state of the art at the 8B scale, and as competitive with much larger closed agents such as OpenAI Deep Research. The result lands at a moment when the deep-research-agent category is crowding rapidly and most leaders are 70B-plus closed systems.
The interesting part is not the score, it is the training-data factory. SciResearcher uses a multi-stage entity-selection and anchor-based augmentation pipeline that synthesizes both conceptual and computational tasks grounded in real academic papers, with each augmentation step run by a distinct web agent so the model has to actually retrieve and compose across sources rather than memorize. Computational tasks are validated by multiple independent Python solvers before being accepted into the training set. This is a deliberate bet that, for science agents, the bottleneck is data quality and multi-hop structure, not parameter count, a position consistent with our reading of the broader May trend toward architectural cleverness over brute-force scale (see our analysis of the deep-research-agent shift).
The caveats are real. The reported baseline set leans on SciMaster and Biomni with GPT-4.1 backbones, plus Claude-Sonnet-4.5; there is no head-to-head against GPT-5 or current Claude variants, so the "beats proprietary" framing should be read as "beats the open and mid-tier comparison set." Computational-task accuracy sits at 45.1%, well short of the conceptual-task numbers, which means the model is still much better at retrieving and composing than at executing the underlying math. And all three benchmarks are short-horizon and biology-chemistry-heavy; there is no evidence yet that the data-construction pipeline transfers cleanly to physics, materials, or long-horizon experimental design.
The honest takeaway: if you are building a science-agent product, the next six months of differentiation are going to be about who has the best synthetic-data pipeline, not who has the biggest backbone. SciResearcher is the clearest public demonstration of that thesis to date. The 8B size also matters commercially: an open model in this range is deployable on a single high-end GPU, which makes the closed-source 70B competitors structurally more expensive per query for any team that does not have hyperscaler economics. That is the actual story underneath the benchmark number.
arXiv 2605.01489 - SciResearcher: Scaling Deep Research Agents for Frontier Scientific Reasoning → · Moonlight Literature Review - SciResearcher methodology and benchmarks →