💊 Generic frontier LLMs + web search are a dead end for long-tail knowledge in specialized domains. This is not a problem of model intelligence—it's that what they read, the open web, simply omits the most important parts of the professional world.
In May 2026, Łukasz Kidziński and Kevin Thomas published a devastating 5-page paper. They used an AI platform called Gosset—essentially a curated pharma asset database with a chat interface—to face off against four top LLM + search combinations: Claude Opus 4.7 + web, GPT 5.5 + web, Gemini 3.1 Pro + web, and Perplexity sonar-pro.
The battlefield was chosen with surgical care: 10 niche oncology/immunology targets—TL1A, OX40L, IL-36R, TROP-2, B7-H3, ROR1, NaPi2b, Claudin 18.2, FAP, GPRC5D. These are areas where the vast majority of drug pipelines hide in the long tail—small biotechs, Chinese developers, academic projects with almost no footprint in mainstream media.
The results were a gut punch 🥊:
| 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, following 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 a decimal-point optimization; it's a survival-grade information deficit. If you do competitive intelligence in pharma, this means GPT missed two-thirds of competitor pipelines.
> 📌 Note: What counts as a "verified drug"? > The paper defines "verified" via a three-layer validation pipeline: layer one is deterministic auto-pass—drugs with active clinical trials, FDA approvals, or sponsor commitments are directly marked verified; layer two is a three-AI-judge cross-check—Claude, GPT, and Gemini independently assess, with a 2-of-3 majority producing a preliminary verdict; layer three is human expert sign-off—reviewers with pharma pipeline backgrounds audit all edge cases. All numbers reflect post-review state, not raw system outputs.
Even more ironic is late-stage parity. On widely reported approval/Phase III anchor drugs, all five systems performed identically—frontier LLMs really can find press-release material. But the preclinical and early-clinical pipeline for a typical target is 10x longer than the late-stage anchors, and web-indexed sources cover that long tail extremely sparsely.
This is not a reasoning gap—it's an indexing coverage gap. GPT-5.5 may be the smartest reasoner on Earth, but it reads the same web pages you find on Google. If a Chinese biotech's Claudin 18.2 pipeline appeared once on a conference poster and was never covered by English-language media, it simply does not exist to that system.
Hallucinations are few, but present. Claude + web produced 1 hallucinated drug (precision 0.992); Perplexity produced 2 (precision 0.975). The absolute numbers are small, but the direction is unsettling: when asked to generate exhaustive named-entity lists, even retrieval-augmented generative models show hallucination tendencies. Gosset's structured filters eliminate the problem at the root—it cannot hallucinate a nonexistent database entry.
> 📌 Note: Why hallucination is dangerous in professional enumeration > In pharma, a hallucinated drug is not harmless creativity—it may be a misattributed molecule (the paper cites rocatinlimab mislabeled as an OX40L target when it actually binds the OX40 receptor), or a canonical-name collision (two different molecules sharing a developer code like IMB101 or SAR446309). Such errors entering decision workflows lead to wrong competitive assessments, wrong M&A due diligence, and wrong investment decisions.
The latency gap is equally stark. Gosset answers in sub-second time—one MongoDB query round-trip. The four frontier systems need multiple web searches plus generation, roughly two orders of magnitude slower. This isn't "fast" versus "slow"—it's the difference between an interactive workflow and a wait-and-see workflow.
Now for 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 just like web search. Claude, GPT, and Gemini don't need to be replaced; they just need a different data source underneath. Same prompt, same natural language understanding, same summarization and reasoning—just swap "read the open web" for "read the curated index."
> 📌 Note: Model Context Protocol (MCP) > MCP is an open standard proposed by Anthropic in 2024 that lets AI systems connect to external data sources and tools in a uniform way. Under MCP, LLMs call external capabilities—web search, database queries, proprietary APIs—through standardized interfaces. Gosset wrapping its curated index as an MCP server means the user experience doesn't change (still natural language chat), but the quality and coverage of information the model accesses changes completely.
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 you get don't cover only the third that got news coverage? If your investment decisions, strategy, and R&D priorities rest on these systems' outputs, your information foundation has a 69% blind spot.
The paper's real contribution is not "Gosset is good"—that's nearly tautological. Its contribution is a quantifiable controlled experiment proving the overwhelming advantage of curated indexes over generic web search in the long tail of specialized knowledge. This is not a race about model size; it's a race about information infrastructure.
Generic LLMs are marvelous reasoning engines. But the quality of a reasoning engine depends on its inputs. If you feed it SEO-optimized, media-filtered, English-hegemony-screened information, even the smartest reasoning paints on a mutilated canvas 🎨.
In the pharma asset discovery game, knowing what exists matters more than reasoning about what exists. And knowing what exists requires someone to pick up the fragments scattered across conference posters, Chinese sponsor websites, patents, and academic preprints—one by one—and compile them into an index.
Curated indexes won't 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: Łukasz Kidziński, 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
- Key data: Gosset 451 verified drugs vs GPT+web 140 (3.2x); Gosset precision 1.000 / 100% recall vs union of 451; Claude+web 1 hallucination, Perplexity 2 hallucinations
- Validation 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 its curated index as an MCP server callable by any frontier LLM via tool use
- Key insight: the recall gap is an indexing problem, not a model quality problem; the open web does not catalog the professional long tail