Opening — Why This Matters Now

Biopharma dealmaking has quietly become a global arms race.

Most large pharmaceutical pipelines are no longer built internally. They are assembled—licensed, acquired, partnered—from external innovation. And that innovation is no longer concentrated in Boston or Basel. It is scattered across Shenzhen trial registries, Korean biotech press, Japanese regulatory bulletins, Brazilian health portals, and a thousand under-amplified PDF disclosures.

Missing one qualifying asset is not a rounding error. It can be a billion-dollar omission.

Yet most “Deep Research” AI agents today are optimized for something else entirely: producing polished, citation-heavy reports. They excel at depth. They struggle with breadth. And in Business Development (BD) and Search & Evaluation (S&E), breadth—specifically completeness—is the game.

The paper “Hunt Globally: Deep Research AI Agents for Drug Asset Scouting” reframes the problem. Not as question answering. Not as synthesis. But as open-world, find-all search under real investor constraints.

That shift changes everything.

Background — The Difference Between Knowing and Not Missing

Most existing benchmarks for browsing agents reward:

  • Single-answer retrieval
  • Report quality
  • Citation correctness
  • Narrative coherence

But drug asset scouting is a different species of task.

It is:

  • Multi-constraint (modality × target × indication × geography × stage)
  • Open-ended (there is no predefined list)
  • Alias-heavy (assets have multiple names and transliterations)
  • Multilingual
  • Under-disclosed (especially early-stage programs)

The authors identify a structural flaw in current evaluation setups: method-induced coverage bias.

If you build ground truth by running a search for a query, you are benchmarking against what your method can already find.

Instead, they invert the pipeline:

Start with real regional drug assets → enrich and validate → then generate investor-style queries for which the asset is a correct answer.

This is a subtle but powerful move. It transforms the benchmark from “can you answer this?” to “can you rediscover this under realistic screening constraints?”

The result is a completeness-first benchmark designed to surface omission failure modes.

And omission, in BD, is expensive.

Benchmark Design — Engineering Hardness, Not Hope

The benchmark construction pipeline has three core stages:

1️⃣ Regional Mining (Multilingual by Design)

Instead of scraping English trade press, the system iterates over:

$$ R \times L \times S(r) \times T $$

Where:

  • $R$ = regions
  • $L$ = local languages
  • $S(r)$ = curated regional sources
  • $T$ = development stages

This produces 1,255 regional-stage assets, later filtered to 798 enriched candidates.

Importantly, assets heavily amplified in U.S. media are filtered out. The benchmark intentionally favors under-the-radar programs.

This is not accidental. It is adversarial realism.

2️⃣ Attribute Enrichment with Provenance

Each asset is:

  • Alias-resolved
  • Cross-lingually normalized
  • Development-stage validated
  • Mechanism extracted
  • Trial-level data structured
  • Evidence paired with verbatim source quotes

Every atomic claim must carry provenance.

This matters because evaluation later relies on LLM graders verifying equivalence under alias resolution. Without structured enrichment, recall would collapse.

3️⃣ Conditioned Query Generation (Investor-Realistic)

Instead of generating random prompts, the authors cluster 48 real investor/BD queries into 10 intent categories, including:

  • Target-first mapping
  • White-space hunting
  • Platform scouting
  • Precision oncology slices
  • Geography-constrained searches

Queries are further divided into difficulty tiers:

Tier Characteristics
Broad Enumerative, wide landscape
Tight Multi-constraint filtering
Complex Multi-hop + derived constraints

Ground-truth identifiers (drug names, trial IDs, unique phrasing) are forbidden.

This ensures rediscovery requires reasoning—not string matching.

The Core Innovation — Bioptic Agent

The paper’s central contribution is not just a benchmark. It is an architecture.

Most search agents operate sequentially:

  1. Search
  2. Append results
  3. Ask model to “find more”
  4. Repeat

This works initially.

Then recall stagnates.

Why Stagnation Happens

Sequential loops revisit similar sources. They refine narrative coherence. But they do not structurally allocate compute toward unexplored search angles.

Bioptic Agent solves this with a tree-based exploration strategy.

Architecture — From Linear Search to Directed Exploration

Bioptic Agent introduces four interacting components:

  • Investigator Agents (multilingual parallel search)
  • Criteria Match Validator (LLM-as-judge with structured decomposition)
  • Deduplication Agent (alias resolution at scale)
  • Coach Agent (tree expansion & strategy refinement)

The Key Mechanism: Directive Tree + UCB

Each search directive becomes a node in a tree.

Node reward:

$$ r_n^{(e)} = p_n^{(e)} \cdot |\Delta \tilde{A}_n^{(e)}| $$

Where:

  • $p_n^{(e)}$ = local precision
  • $\Delta \tilde{A}_n^{(e)}$ = newly discovered unique validated assets

Selection uses an Upper Confidence Bound (UCB):

$$ UCB(n) = \frac{W(n)}{N(n)} + c \sqrt{\frac{\log N(parent(n))}{N(n)}} $$

This balances:

  • Exploiting high-performing directives
  • Exploring under-tested branches

In plain English:

Allocate compute where new validated assets are most likely.

Not where verbosity is highest.

Multilingual Parallelism — The Quiet Multiplier

Each rollout spawns investigators in multiple languages (English + Chinese in evaluation).

This matters because early-stage disclosures often appear only in local ecosystems.

Sequential English-only scaffolds plateau faster.

Tree + language parallelism prevents early saturation.

Results — Completeness Wins

Held-out benchmark (22 query–asset pairs):

Model Recall Precision F1
Bioptic Agent 0.730 0.877 0.797
Claude Opus 4.6 0.454 0.736 0.562
Gemini 3 Pro DR 0.500 0.512 0.506
OpenAI Deep Research 0.372 0.713 0.489
GPT-5.2 Pro 0.364 0.648 0.466
Perplexity DR 0.409 0.481 0.442
Exa Websets 0.182 0.515 0.269

Two observations:

  1. Bioptic achieves both high recall and high precision.
  2. Simply “running longer” does not close the gap.

Ablation experiments show that:

  • Removing tree structure leads to early saturation.
  • Removing multilingual search reduces long-tail capture.
  • Sequential iteration with the same base model underperforms significantly.

The gains come from scaffold design—not just model size.

Business Implications — What This Means Beyond Biopharma

This paper is not really about hepatitis B pipelines.

It is about AI agents in coverage-critical environments.

The pattern generalizes to:

  • Competitive intelligence
  • M&A target discovery
  • Regulatory scanning
  • Patent landscape analysis
  • Sanctions monitoring
  • Supply chain risk detection

In all of these domains:

Missing a single qualifying entity can be more costly than producing ten irrelevant summaries.

Most current AI agents optimize for clarity.

Few optimize for omission minimization.

Tree-based exploration with validation-gated rewards is one path toward closing that gap.

The Subtle Governance Angle

Completeness-first systems raise interesting governance questions:

  • How do we audit omission rates?
  • What level of recall is “good enough” in high-stakes scouting?
  • Should regulatory-facing AI systems be required to optimize for recall over narrative quality?

If AI becomes embedded in due diligence workflows, failure modes shift from hallucination to silent omission.

And silent omission is harder to detect.

Conclusion — Depth is Polished. Breadth is Strategic.

The paper demonstrates something quietly important:

Scaling model size and browsing time is not sufficient.

Search strategy architecture matters.

Completeness requires:

  • Structured exploration
  • Explicit validation loops
  • Deduplication discipline
  • Multilingual coverage
  • Compute allocation guided by marginal discovery

In other words:

If you want to find everything, you cannot think linearly.

Tree-based agents are not just a technical curiosity. They represent a shift from narrative AI to coverage AI.

And in competitive environments, coverage wins.

Cognaptus: Automate the Present, Incubate the Future.