[Paper] LLM Novice Uplift on Dual-Use, In Silico Biology Tasks

Source: arXiv - 2602.23329v1

Overview

Recent research investigates whether large language models (LLMs) can actually boost the performance of people with little or no biology training on tasks that are traditionally the domain of experts. By comparing novices who could query LLMs against those limited to standard web searches, the authors show that LLM access can make non‑experts dramatically more accurate on a suite of biosecurity‑relevant problems—raising both exciting opportunities for scientific acceleration and serious dual‑use concerns.

Key Contributions

• Human‑centric uplift study: First large‑scale experiment measuring how LLMs improve novice performance on real‑world biology tasks, not just model‑only benchmarks.
• Quantified uplift: Novices with LLM access were 4.16× more accurate than internet‑only peers (95 % CI [2.63, 6.87]).
• Expert‑level performance: On four benchmarks that have expert baselines, LLM‑assisted novices beat the expert scores on three tasks.
• LLM vs. LLM‑assisted humans: Standalone LLMs often outperformed the same LLM when used through a human, indicating sub‑optimal prompting or interaction strategies.
• Low barrier to dual‑use info: 89.6 % of participants reported they could retrieve potentially dangerous biological information with little friction despite existing safeguards.
• Call for interactive evaluation: Authors argue that traditional static benchmarks are insufficient; continuous “uplift” testing with real users is needed to track both benefits and risks.

Methodology

1. Participant pool: ~200 volunteers with minimal biology background (self‑identified novices).
2. Task sets: Eight distinct biosecurity‑relevant problems (e.g., protein design, pathogen detection, synthetic gene synthesis) drawn from established biology benchmarks.
3. Conditions:
   • Control: Access to public internet resources only (search engines, wikis, forums).
   • LLM‑assisted: Same internet access plus the ability to query a suite of state‑of‑the‑art LLMs (ChatGPT‑4, Claude, LLaMA‑2, etc.).
4. Time allowance: Tasks ranged from quick fact‑finding (≤30 min) to deep design challenges (up to 13 h).
5. Evaluation: Answers were scored against ground‑truth solutions; where expert baselines existed, those scores were used for comparison.
6. Survey: Post‑task questionnaire captured participants’ perceived difficulty, confidence, and any obstacles in obtaining dual‑use information.

Results & Findings

• Overall uplift: LLM‑assisted novices achieved an average accuracy of 68 %, versus 16 % for internet‑only controls.
• Task‑level variance: The biggest gains appeared in complex design problems (e.g., de‑novo enzyme design) where LLMs supplied plausible sequences and rationales.
• Expert comparison: On three of four benchmarks (protein function prediction, CRISPR guide design, metabolic pathway reconstruction), the LLM‑assisted novices outperformed the expert baseline (average expert accuracy ≈ 55 %).
• LLM alone vs. human‑in‑the‑loop: Pure LLM outputs scored ~10 % higher than the best human‑augmented attempts, suggesting that novices did not consistently extract the most relevant or precise information from the models.
• Dual‑use accessibility: Nearly 90 % reported that obtaining potentially harmful protocols (e.g., virus attenuation steps) was “easy” or “very easy,” despite model‑level content filters.

Practical Implications

• Accelerated prototyping: Developers building biotech tools can leverage LLMs to let non‑specialists generate viable hypotheses, draft experimental plans, or even write code for bio‑informatics pipelines—dramatically shortening the learning curve.
• Education & training: Interactive LLM tutors could supplement university curricula, allowing students to practice real‑world problem solving without needing a full lab setup.
• Risk management: The ease of extracting dual‑use knowledge underscores the need for robust guardrails (prompt‑level throttling, usage monitoring, and policy‑driven API restrictions) in any commercial LLM offering for scientific domains.
• Product design: Companies may consider building “human‑in‑the‑loop” interfaces that surface LLM suggestions while prompting users to verify and refine outputs, thereby closing the gap between raw model performance and effective human use.
• Regulatory awareness: Policymakers should note that LLMs can democratize advanced bio‑tech capabilities, prompting updates to biosecurity guidelines and responsible AI frameworks.

Limitations & Future Work

• Participant expertise variance: Although labeled “novices,” some volunteers had informal biology exposure, which could inflate uplift estimates.
• Prompt engineering gap: The study did not systematically explore optimal prompting strategies; better user training could narrow the performance gap between LLM‑only and LLM‑assisted results.
• Model diversity: Only a handful of publicly available LLMs were tested; proprietary or domain‑fine‑tuned models might yield different uplift patterns.
• Long‑term retention: The experiment measured immediate task performance; it remains unclear whether LLM assistance leads to lasting skill acquisition.
• Ethical safeguards: While participants reported low difficulty obtaining dual‑use info, the study did not evaluate the effectiveness of existing content filters under adversarial prompting—an area ripe for deeper investigation.

Bottom line: LLMs are already powerful enough to turn biology novices into competent problem solvers on tasks once reserved for trained scientists. This democratization brings both a wave of productivity gains and a pressing need for responsible deployment strategies.

Authors

• Chen Bo Calvin Zhang
• Christina Q. Knight
• Nicholas Kruus
• Jason Hausenloy
• Pedro Medeiros
• Nathaniel Li
• Aiden Kim
• Yury Orlovskiy
• Coleman Breen
• Bryce Cai
• Jasper Götting
• Andrew Bo Liu
• Samira Nedungadi
• Paula Rodriguez
• Yannis Yiming He
• Mohamed Shaaban
• Zifan Wang
• Seth Donoughe
• Julian Michael

Paper Information

• arXiv ID: 2602.23329v1
• Categories: cs.AI, cs.CL, cs.CR, cs.CY, cs.HC
• Published: February 26, 2026
• PDF: Download PDF