Hunt Globally: Wide Search AI Agents for Drug Asset Scouting in Investing, Business Development, and Competitive Intelligence

Abstract: Bio-pharmaceutical innovation has shifted: many new drug assets now originate outside the United States and are disclosed primarily via regional, non-English channels. Recent data suggests that over 85% of patent filings originate outside the U.S., with China accounting for nearly half of the global total. A growing share of scholarly output is also non-U.S. Industry estimates put China at 30% of global drug development, spanning 1,200+ novel candidates. In this high-stakes environment, failing to surface "under-the-radar" assets creates multi-billion-dollar risk for investors and business development teams, making asset scouting a coverage-critical competition where speed and completeness drive value. Yet today's Deep Research AI agents still lag human experts in achieving high recall discovery across heterogeneous, multilingual sources without hallucination. We propose a benchmarking methodology for drug asset scouting and a tuned, tree-based self-learning Bioptic Agent aimed at complete, non-hallucinated scouting. We construct a challenging completeness benchmark using a multilingual multi-agent pipeline: complex user queries paired with ground-truth assets that are largely outside U.S.-centric radar. To reflect real-deal complexity, we collected screening queries from expert investors, BD, and VC professionals and used them as priors to conditionally generate benchmark queries. For grading, we use LLM-as-judge evaluation calibrated to expert opinions. On this benchmark, our Bioptic Agent achieves 79.7% F1 score, outperforming Claude Opus 4.6 (56.2%), Gemini 3 Pro + Deep Research (50.6%), OpenAI GPT-5.2 Pro (46.6%), Perplexity Deep Research (44.2%), and Exa Websets (26.9%). Performance improves steeply with additional compute, supporting the view that more compute yields better results.

• The paper introduces Bioptic Agent, a tree-based self-learning AI that attains an F1-score of 0.797 on a global drug asset scouting benchmark.
• It leverages parallel, multilingual investigator agents with validator and deduplication modules to ensure non-hallucinated, complete asset discovery.
• Results demonstrate significant performance improvements over commercial systems, underscoring the need for a completeness-first, scaffolded search architecture.

Wide-Search AI Agents for Global Drug Asset Scouting: The Bioptic Agent Approach

Introduction

Drug asset scouting and competitive intelligence in the bio-pharmaceutical sector face a paradigm shift due to globalization and the increasing prevalence of non-U.S. pipeline innovation. The predominant challenge now centers on the ability to discover under-the-radar assets in heterogeneous, multilingual digital landscapes, as traditional English-centric search and proprietary databases no longer deliver complete or timely coverage. The presented work formulates a rigorous benchmarking methodology for asset scouting completeness and introduces the Bioptic Agent—a tree-based, self-learning AI agent designed for exhaustive, non-hallucinated discovery in asset scouting workflows (2602.15019).

Problem Setting and Motivation

Analysis of recent patent trends reveals that $\sim 86.5\%$ of pharma innovation occurs outside the U.S., with China alone contributing nearly half the worldwide drug development activity. This redistribution of early disclosures to regional, non-English channels introduces substantial information asymmetry and risk for global investors and business development professionals. Standard deep research agents remain inadequate for comprehensive multi-constraint set discovery, often omitting pivotal assets or amplifying superficial findings. The economic implications of missed discoveries in multi-billion-dollar asset partnerships elevate the necessity for agents with recall-optimized, completeness-first strategies.

Benchmark Construction

A new benchmark is introduced that operationalizes asset discovery completeness as the primary evaluation criterion. Unlike source-anchored or query-first benchmarks, the process is inverted: ground truth consists of regionally sourced, validated drug assets curated from local-language news, with richly attributed metadata. For each asset, investor-native, multi-constraint queries are synthesized, ensuring that successful resolution cannot rely on lexical cues, but instead demands advanced evidence aggregation and multi-hop reasoning.

Figure 1: Distribution of assets in the benchmark by origin language (left) and by therapeutic area labels (right), highlighting the multinational and multidisciplinary coverage.

The composition of these queries (complexity tiers and constraint structure) closely reflects actual investor and BD screening practice, captured through clustering and abstraction of genuine diligence queries.

Figure 2: Left: Query distribution across difficulty tiers (Broad, Tight, Complex). Right: Prevalence of high-level constraint categories (multi-label per query), further illustrating the authenticity and hardness of the benchmark.

Through systematic multi-agent mining (regional news miner, attribute enrichment, local/global discoverability profiling), the benchmark specifically prioritizes assets that evade English-centric amplification, ensuring head-to-head evaluation in authentic discovery scenarios.

Bioptic Agent Architecture

Bioptic Agent is explicitly designed for high recall and non-hallucination in global biomedical asset scouting. The system leverages a tree-based, self-refining exploration scaffold, where each node in the search tree encodes a distinct search directive and prompts parallelized multi-lingual exploration, addressing major limitations of sequential or single-narrative search paradigms.

Key architectural features include:

• Investigator Agents: Independently instantiated per language (e.g., English, Chinese), they execute searches with explicit directive-based context.
• Criteria Match Validator Agent: Acts as a domain-aligned LLM-as-a-judge, rigorously validating candidate assets' eligibility against complex queries, operating at up to 88% expert-aligned precision.
• Deduplication Agent: Resolves asset aliases and synonyms across languages and domains for canonicalization.
• Coach Agent: Dynamically generates non-overlapping, search-history-refined exploration directives, implements reward-driven tree expansion/backpropagation, and utilizes compressed error pattern analysis for adaptive strategy refinement.

The exploration process emulates an agentic MCTS variant, using upper confidence bound-based selection for rollout directives and node rewards that combine local precision with validated, deduplicated new asset discovery. Explicit parallel rollout across languages mitigates regional blind spots and aliasing, enabling discovery in underrepresented and early pipeline sources.

Figure 3: Quality–time tradeoff for asset scouting expressed as F1-score vs runtime across agents; Bioptic Agent achieves markedly superior quality per compute and continues to deliver steady improvements, validating the impact of tree-based exploration and language parallelism.

Experimental Findings

Bioptic Agent demonstrates a substantial performance lead over both baseline and contemporary commercial research systems. On the completeness benchmark:

• Bioptic Agent attains F1-score 0.797 (Precision 0.877, Recall 0.730).
• Claude Opus 4.6 yields 0.562 F1.
• Other advanced commercial agents (Gemini 3 Pro, OpenAI GPT-5.2 Pro, Perplexity Deep Research) perform in the range F1 0.269–0.506.

Notably, ablations removing the tree scaffold or disabling explicit language parallelism saturate rapidly, while competing agents plateau in coverage regardless of extended compute, highlighting the inadequacy of sequential search or brute force iteration. In particular, the study finds simply running top-performing generalist LLMs with larger context or higher compute does not close the completeness delta—self-reflective, structured, and validator-gated discovery is necessary.

The underlying evaluation pipeline ensures objectivity via separately tuned LLM-as-a-judge graders, with demonstrated expert-aligned accuracy and consistency in attribute and alias resolution.

Practical and Theoretical Implications

Practically, the Bioptic Agent framework enables investors, BD, and CI professionals to surface global innovation early, efficiently, and with low omission rates, directly addressing pipeline coverage risk. This is especially relevant given the increasing localization of disclosures and the multi-lingual fragmentation of early-stage asset data.

Theoretically, the results challenge the prevailing focus on reasoning/browsing depth as the sole path to agentic progress. They demonstrate the necessity for explicit set-completeness orientation, persistent artifact tracking, domain-aligned validator interaction, and dynamic, reward-driven search decomposition. The agent’s tree-based, non-narrative memory architecture points toward a new design direction for frontier research agents tasked with breadth-first enumeration under stringent correctness constraints.

Future directions involve scaling coverage across additional languages, integrating new source modalities (e.g., regulatory, legal, and corporate filings), and incorporating more sophisticated agentic self-correction for automated knowledge base generation. These findings also set a new benchmark for model-based agent evaluation under real-world, high-stakes completeness targets.

Conclusion

This study delivers a rigorous completeness benchmark for global drug asset scouting and an empirically validated, tree-based, self-learning Bioptic Agent that decisively surpasses the performance of current commercial research systems. The analyses highlight that enduring performance gains in agentic research require moving beyond extended browsing or context windows, instead adopting scaffolded, validator-driven, and explicitly completeness-oriented architectures. The framework motivates further developments in agentic search systems, particularly for high-value, recall-critical domains in science, technology, and industry.