General-purpose frontier LLMs plus web search fall short on the long tail of specialized knowledge. This is not a problem of model intelligence - it is that the open web, which these systems read, simply omits much of what matters in specialized domains.
In May 2026, Lukasz Kidzinski and Kevin Thomas published a short but disruptive five-page paper. They pitted an AI platform called Gosset - essentially a curated pharmaceutical asset database behind a chat interface - against four top LLM-plus-search combinations: Claude Opus 4.7 + web, GPT 5.5 + web, Gemini 3.1 Pro + web, and Perplexity sonar-pro.
The battlefield was chosen carefully: 10 niche oncology/immunology targets - TL1A, OX40L, IL-36R, TROP-2, B7-H3, ROR1, NaPi2b, Claudin 18.2, FAP, GPRC5D. The defining feature of these spaces is that most drug pipelines hide in the long tail - small biotechs, Chinese developers, academic projects with almost no mainstream media presence.
The results were a thunderclap:
| System | Verified Drugs | Precision | Recall (vs union) | Hallucinations | |:-------|:--------------:|:---------:|:-----------------:|:---------------| | Gosset (curated) | 451 | 1.000 | 100% | 0 | | GPT 5.5 + web | 140 | 1.000 | ~31% | 0 | | Gemini 3.1 Pro + web | ~120 | 1.000 | ~27% | 0 | | Claude Opus 4.7 + web | ~110 | 0.992 | ~24% | 1 | | Perplexity sonar-pro | ~80 | 0.975 | ~18% | 2 |
*Table 1: Aggregate counts after 10 niche-target queries, post human-expert verification. Source: paper Table 1 and Figures 1-2.*
Gosset returned 451 verified drugs. The best frontier system (GPT + web) returned only 140 - less than a third. A 3.2x gap is not fine-tuning; it is an existential information deficit. For competitive intelligence in pharma, this means GPT misses two-thirds of competitor pipelines.
> Note: What counts as a "verified drug"? > The paper defines "verified" via a three-layer process: (1) deterministic auto-pass - drugs with active clinical trials, FDA approval, or sponsor commitments are marked verified; (2) a three-AI-judge cross-check - Claude, GPT, and Gemini evaluate independently with a 2-of-3 majority verdict; (3) human expert sign-off - a reviewer with pharma pipeline background audits all edge cases. All numbers reflect post-review state, not raw system output.
Even more striking is late-stage parity: on widely reported approval/Phase 3 anchor drugs, all five systems performed equally - frontier LLMs can indeed find what is in press releases. But a typical target's preclinical and early clinical pipeline is roughly 10x the length of its late-stage anchors, and web-indexed sources cover that long tail sparsely.
This is not a reasoning gap - it is an indexing coverage gap. GPT-5.5 may be the smartest reasoner on Earth, but it reads the same web you reach via Google. If a Chinese biotech's Claudin 18.2 asset appeared only on a conference poster, never covered by English media, it simply does not exist for that system.
Hallucinations are rare but present. Claude + web produced 1 hallucinated drug (precision 0.992); Perplexity produced 2 (precision 0.975). The absolute counts are small, but the direction is unsettling: when asked for exhaustive named-entity lists, generative models' hallucination tendencies activate even with retrieval augmentation. Gosset's structured filters eliminate this by construction - it cannot hallucinate a non-existent database entry.
> Note: Why hallucination matters in professional enumeration. In pharma, a hallucinated drug is not harmless creativity - it may be a misattributed molecule (e.g., rocatinlimab wrongly labeled as an OX40L asset when it binds the OX40 receptor), or a canonical-name collision (two different molecules sharing a developer code like IMB101 or SAR446309). Such errors can poison competitive assessments, M&A due diligence, and investment decisions.
The latency gap is equally stark. Gosset responds in sub-second time - a single MongoDB round-trip. The four frontier systems need multiple web searches plus generation, roughly two orders of magnitude slower. That is the difference between an interactive workflow and a wait-and-see workflow.
Here is the paper's most valuable insight: the recall gap is an indexing problem, not a model problem.
Gosset exposes its curated asset index as an MCP (Model Context Protocol) server - any frontier model can call it like a web search tool. Claude, GPT, and Gemini do not need to be replaced; they just need a different data source behind them. Same prompts, same natural-language understanding, same summarization and reasoning - only the input changes from "the open web" to "a curated index."
> Note on MCP: The Model Context Protocol, proposed by Anthropic in 2024, is an open standard letting AI systems connect to external data sources and tools through a uniform interface. Wrapping a curated index as an MCP server changes the quality and coverage of what a model reads without changing how users interact with the model.
This raises an uncomfortable question: if your enterprise uses GPT/Claude/Gemini for competitive intelligence, market research, or due diligence, are you sure the answers cover more than the reported one-third? If investment decisions, strategy, and R&D prioritization rest on these outputs, your information base has a 69% blind spot.
The paper's real contribution is not "Gosset is good" - that is nearly tautological. It is a quantifiable controlled experiment demonstrating the overwhelming advantage of curated indices over generic web search in the long tail of specialized knowledge. This is not a race about model size; it is a race about information infrastructure.
General-purpose LLMs are magnificent reasoning engines. But the quality of a reasoning engine depends on its inputs. Feed it SEO-optimized, media-filtered, English-hegemony-screened information, and even brilliant reasoning paints on a mutilated canvas.
In pharma asset discovery, knowing what exists matters more than reasoning about what exists - and knowing requires someone to pick up the fragments scattered across conference posters, Chinese sponsor websites, patents, and preprints, and index them.
Curated indices will not replace LLMs. But they are redefining the LLM battlefield.
---
Paper Details
- Title: *Curated AI beats frontier LLMs at pharma asset discovery*
- arXiv ID: 2605.04908
- Published: 2026-05-06
- Authors: Lukasz Kidzinski, Kevin Thomas
- Length: 5 pages, 5 figures, 1 table
- Systems compared: Gosset (curated drug-asset index) vs Claude Opus 4.7 + web_search (20 budget) vs GPT 5.5 + web_search_preview (20 budget) vs Gemini 3.1 Pro + Google Search grounding (20 budget) vs Perplexity sonar-pro (high context)
- 10 test targets: TL1A, OX40L, IL-36R, TROP-2, B7-H3, ROR1, NaPi2b, Claudin 18.2, FAP, GPRC5D
- Headline numbers: Gosset 451 verified drugs vs GPT+web 140 (3.2x); Gosset precision 1.000 / recall 100% vs the union of 451; Claude+web 1 hallucination, Perplexity 2
- Verification pipeline: three layers - deterministic auto-pass (industry-grade evidence), three-AI-judge cross-check (Claude/GPT/Gemini, 2-of-3 majority), human expert sign-off (pharma-pipeline-background reviewer)
- Architecture: Gosset exposes the curated index as an MCP server callable by any frontier LLM
- Key insight: the recall gap is an indexing problem, not a model-quality problem; the open web does not catalog the professional long tail