Stop asking a general chatbot to do science: specialized agents beat nine LLMs head-to-head

Distributed agents read 100x the papers

In a preprint posted this week, a distributed multi-agent framework took on a whole-brain Perturb-seq atlas — the kind of dataset whose interpretation is usually the rate-limiting human step. The preprint reports the system recovered the manuscript's conclusions without access to the source manuscript, predicted held-out perturbation effects, and — the headline number — retrieved 100 times more literature references than single-agent modes. Blinded human experts, the authors report, preferred the distributed agents' interpretations over single-agent ones. These results are the authors' own, from their own benchmark, and have not been peer-reviewed. m:…1703

A specialist beats nine generalists at cucurbits

In another preprint, CucurbitAgent — a research agent for cucurbit (cucumber, melon, squash) genomics — was tested against nine general-purpose LLMs on a 50-task benchmark graded blind by experts. The result: a 96% pairwise win rate for the specialist. The agent integrates heterogeneous multi-omics data — the preprint cites 14 million predicted binding-site residue annotations across 242,000 protein structures — with literature retrieval that links claims to evidence. The authors describe the framework as transferable to other crop and biological domains, a claim that itself awaits independent replication. m:…3000

The same pattern in mortality data and drug discovery

Two more preprints extend the pattern beyond agents. On medRxiv, supervised multimodal pipelines for assigning causes of death from verbal autopsy records improved top-1 accuracy by 8.5 to 15.9 percentage points over the InSilicoVA benchmark, with feature fusion via cross-modal attention adding 6.6 points of balanced accuracy on less common causes; the authors also report that zero-shot LLMs achieved the highest balanced accuracy (60.6%) for individual-level assignment. On bioRxiv, a benchmark of LLMs for assay-aware drug-target binding affinity prediction found that natural-language assay descriptions carry predictive signal that molecular structures and protein sequences do not fully capture — the MAMMAL-AsAP model reached an R² of 0.7198, and the ACL-DTI variant held its performance when target names or assay techniques were masked. All of these are preprint results, resting on the authors' own data and splits. m:…9907 m:…0096

The nuance: generalists aren't useless

The results are not a blanket obituary for general models. In the verbal-autopsy preprint, zero-shot LLMs — no task-specific training at all — achieved the highest balanced accuracy for individual-level cause-of-death assignment, even as supervised multimodal pipelines won on top-1 accuracy. And the drug-target benchmark's finding is about context, not scale: assay descriptions contain predictive information that structure alone misses. The pattern across all four papers is that domain context — a schema, an assay description, a corpus — is doing the work that raw model size cannot. m:…9907 m:…0096

The founder's takeaway

Four papers, one lesson: specialization beats scale when the domain has structure to exploit. A general chatbot has no schema for a Perturb-seq atlas or a verbal autopsy record; an agent with retrieval, domain data, and an evaluation harness does. This suggests founders building science tooling should spend on evidence-linked retrieval and blinded benchmarks before spending on bigger models. The implication is that the moat in AI-for-science is not the model — it is the corpus, the schema, and the evals that prove the answers. And because all four results are unreviewed preprints, the right move is verification, not adoption: rerun the benchmarks, check the claims, and treat the numbers as hypotheses about where domain structure pays off. m:…1703 m:…3000 m:…0096