Technical Summary: Hunt Globally – Wide Search AI Agents for Drug Asset Scouting

Problem Statement

The pharmaceutical industry is undergoing a structural shift where a significant majority of new drug assets originate outside the United States and are disclosed primarily through regional, non-English channels. Recent data indicates that approximately 86.5% of global patent filings originate outside the U.S., with China alone accounting for nearly half. In this environment, Business Development (BD) and Search & Evaluation (S&E) teams face a high-stakes challenge: missing a single qualifying asset can result in the loss of billion-dollar acquisition or partnership opportunities.

Current AI-driven "Deep Research" agents are optimized for fact-finding and report synthesis but struggle with completeness in open-world "find-all" scenarios. They often fail to achieve high recall across heterogeneous, multilingual sources without hallucinations, leading to superficial retrieval and the omission of novel, under-the-radar programs. Existing benchmarks favor depth over breadth, failing to capture the complexity of enumerating hundreds or thousands of entities across diverse languages and development stages.

Methodology

1. The Completeness Benchmark

To address the lack of rigorous evaluation for asset scouting, the authors constructed a Completeness Benchmark designed to be "backward" from validated program records rather than forward from a query. This approach minimizes method-induced coverage bias.

• Asset Mining: A multilingual, multi-agent pipeline mines regional news sources in local languages (e.g., Chinese, Japanese, Portuguese) across 10 curated regions. A "Regional News Miner" iterates over tuples of region, language, source, and development stage to extract program names.
• Validation & Enrichment: An "Attributes Enrichment Agent" validates candidates, resolves cross-lingual aliases, and extracts structured attributes (e.g., modality, target, trial phase) using deep-research retrieval. Assets with strong global amplification (e.g., major US press coverage) are filtered out to ensure the benchmark focuses on "under-the-radar" assets.
• Query Generation: Instead of starting with a query, the system generates investor-native queries conditioned on the ground-truth (GT) assets. Using a corpus of real investor/BD queries, the system creates multi-constraint queries (Broad, Tight, Complex) that forbid direct identifiers (names, trial IDs). This forces the agent to rely on evidence aggregation and multi-hop reasoning rather than string matching.
• Evaluation: The benchmark uses an LLM-as-judge approach, calibrated via multi-agent debate and expert alignment, to grade precision and recall against the ground-truth assets.

2. Bioptic Agent Architecture

The authors propose Bioptic Agent, a tree-based, self-learning system engineered for complete, non-hallucinated scouting. Unlike sequential search agents that compress exploration into a single narrative, Bioptic Agent preserves candidate sets as persistent artifacts and allocates compute dynamically.

• Tree-Based Exploration: The system treats web exploration as a tree where nodes represent exploration directives.
  • Investigator Agents: Execute parallel web searches in multiple languages (e.g., English and Chinese) to retrieve candidate assets.
  • Criteria Match Validator: A specialized LLM agent validates each candidate against query constraints, providing a pass/fail verdict with evidence provenance.
  • Deduplication Agent: Resolves aliases and removes duplicates across the global asset store.
  • Coach Agent: The core "self-learning" component. It analyzes failure rationales, search history, and gaps to generate new, non-overlapping child directives for the next epoch. It targets under-explored search angles and refines prompts to address specific errors.
• Algorithmic Loop:
  1. Select: Uses an Upper Confidence Bound (UCB) rule to select leaf nodes (directives) for exploration, balancing exploitation of high-reward paths with exploration of new branches.
  2. Rollout: Investigators execute searches in parallel across languages based on selected directives.
  3. Evaluate: Validators grade candidates; a node reward is computed based on the precision of matches and the number of new unique assets discovered.
  4. Backpropagate & Expand: Rewards are backpropagated to update node statistics. The Coach Agent generates new child directives to expand the tree, ensuring coverage of the remaining search space.

Key Contributions

1. A New Benchmark: A challenging, completeness-first benchmark for drug asset scouting that utilizes non-US, non-English sources and investor-native multi-constraint queries.
2. Bioptic Agent: A specialized, tree-based agentic system that outperforms general-purpose deep research models by maintaining persistent candidate sets, utilizing multilingual parallelism, and employing a self-correcting Coach Agent.
3. Methodological Insight: The paper demonstrates that for "find-all" tasks, specialized scaffolding (tree exploration, error-driven expansion) yields better results than simply increasing compute or using larger models in a sequential loop.

Results

On the held-out test split of 22 query-asset pairs, Bioptic Agent achieved an F1-score of 79.7%, substantially outperforming state-of-the-art commercial deep research baselines:

• Bioptic Agent: 79.7% F1 (Recall: 73.0%, Precision: 87.7%)
• Gemini 3.1 Deep Think: 59.2% F1
• Gemini 3.1 Pro Deep Research: 58.6% F1
• Claude Opus 4.6: 56.2% F1
• OpenAI GPT-5.2 Pro: 46.6% F1
• Perplexity Deep Research: 44.2% F1
• Exa Websets: 26.9% F1

Ablation Studies:

• Removing the tree structure and multilingual parallelism (Bioptic Agent no-tree, lang-free) resulted in earlier saturation and lower performance despite higher compute usage (50 investigator calls vs. 20 for the full model).
• Sequential "find-more" loops using strong models (e.g., GPT-5.2, o4-mini) improved slowly and plateaued at significantly lower F1 scores, indicating that naive iteration is insufficient for sustained coverage.

Significance and Claims

The paper claims that investor-grade and BD-grade asset scouting requires completeness-oriented search control, lossless candidate tracking, and expert-aligned validation rather than merely increased browsing compute or better synthesis.

The authors argue that the "intelligence gap" in drug scouting—translating complex investor queries into complete sets of matching assets across fragmented, multilingual data—cannot be solved by general-purpose agents alone. The results support the view that more compute yields better results only when coupled with specialized architectures that systematically allocate resources to under-explored branches and validate findings against strict constraints. The Bioptic Agent demonstrates that a tree-based, self-learning approach can achieve high recall and precision in a domain where current Deep Research agents consistently fail to identify all relevant assets.