Researchers from ETH Zurich, TU Wien, AITHYRA, the University of Pennsylvania and the University of Oxford have described a method that lets a general-purpose language model steer specialised drug-design generators with plain-language goals. In their paper, posted on arXiv on 12 October 2026, the recorded LLM API cost for each of five discovery campaigns ranged from about $8.83 to about $77.
The method, LLM-to-Flow Distillation (LFD), used Gemini models as the teacher in the reported experiments. Co-author Riccardo De Santi announced it on X on 14 October, writing that it “works with any frontier LLM via API-access”.
Business Pill · DISTILLATION
A one-minute explainer of distillation: how one model learns from another. It teaches the general idea only and says nothing about any person or organisation in this story.
The key insight: As we read it, the paper moves the expensive part of AI-guided design from building a reward model to writing a goal: the authors report that a frozen Gemini teacher, with no task-specific proxy reward, adapted the generators at a recorded API cost of about $8.83 to $77 per campaign.
In the paper’s description, “a frozen LLM acts as an adaptive teacher for a domain-specific generator”. The generators are flow models already trained on peptides or drug-like molecules, which know what valid designs look like but cannot read a scientist’s goal.
Given a goal such as “generate peptides with promising anticancer activity”, the teacher builds a curriculum of intermediate sub-goals that fit what the generator can currently produce, then moves them toward the final goal.
At each round the LLM judges sampled candidates, and those judgments become preferences distilled into the generator through DPO, a preference-tuning method. The authors write that this works without an explicit proxy reward, which they say is often unavailable in experimental sciences.
In the experiments, the paper says the authors use “Gemini 3.1 Pro to construct the curriculum and Gemini 3.7 Flash for LLM judgments”.
The recorded LLM API cost of each of the five LFD discovery campaigns, as given in the appendix of De Santi et al., arXiv 2610.17246v1 (12 October 2026). The kinase campaign ran against a 48-hour time limit.
What Was Posted on X
Weโre bringing natural-language control to any scientific generator.
We are very excited to introduce LLM-to-Flow Distillation (LFD), a method to adapt scientific generators to high-level natural-language discovery goals.
Starting from textual goals like โdesign peptides with promising anticancer activityโ or โdesign small molecules binding to the D2 dopamine receptorโ, LFD distills high-level goal-related knowledge from LLMs into scientific generators through iterative DPO on LLM-generated sub-goals.
๐ฅ LFD does not need task-specific proxy rewards, works with any frontier LLM via API-access, and outperforms state-of-the-art reward-guided methods such as FlowGRPO and DiffusionNFT on held-out computational evaluations. (1/6)
Co-author Riccardo De Santi announced LLM-to-Flow Distillation on X on 14 October 2026, the first post of a six-part thread. He describes goals such as designing peptides with promising anticancer activity or small molecules binding to the D2 dopamine receptor, and writes that the method does not need task-specific proxy rewards and works with any frontier LLM via API access. The figures in this piece come from the paper itself.
What Each Campaign Cost
The paper’s appendix records the hardware, runtime and LLM API cost for each of its five goals. The cheapest was the cell-penetrating peptide goal: about 5.3 hours on one NVIDIA GH200 GPU and about $8.83 of API calls.
The anticancer peptide goal took about 20 hours on one GPU and about $37.4. The GSK3β binding goal took about 17.3 hours and about $33, and the dopamine D2 docking goal about 35 GPU-hours and about $54.
The kinase-focused campaign ran against a 48-hour time limit, with an API cost of about $77 counted from logged tokens, including reasoning. By our arithmetic, the five recorded API costs add up to about $210.
What the Results Show, in Silico
On anticancer peptides, LFD scored 53.54% on a held-out AntiCP2-alternate predictor, against 26.43% for the pre-trained model and 35.94% for the best baseline, classifier guidance (Table 1). The baselines optimised a different anticancer predictor; LFD optimised neither.
On cell-penetrating peptides, LFD reached 69.06% on the held-out pLM4CPPs evaluator, up from 8.81% for the pre-trained model and above DiffusionNFT’s 57.04%.
On small-molecule docking for the D2 receptor, LFD’s mean Vina score improved from −7.52 to −8.48 kcal/mol, against −7.90 for DiffusionNFT, the strongest baseline on that measure (Table 2). On GSK3β, its GNINA score rose from 5.80 to 6.77.
The paper presents the kinase goal, a list of ring and functional-group constraints written in plain language, as an out-of-distribution test. The paper says the pre-trained model produced no hits over 38,500 samples, while the LFD-adapted model produced a hit in a 512-sample batch at round 6. “All reward-guided baselines fail”, it adds.
Held-out AntiCP2-alternate scores from Table 1 of the LFD paper (arXiv 2610.17246v1): LFD 53.54% against 35.94% for the best baseline, classifier guidance. Computational evaluation, not a lab assay.
Who Built It
The first three authors, Riccardo De Santi, Federico Di Gennaro and Angelo Gnazzo, are listed as equal contributors. The paper lists affiliations at ETH Zurich and its ETH AI Center, TU Wien, AITHYRA, the University of Pennsylvania and the University of Oxford, and the author list also includes Michael M. Bronstein and Andreas Krause.
The acknowledgments credit an ETH AI Center doctoral fellowship to De Santi and funding from the Swiss National Science Foundation’s NCCR Catalysis and NCCR Automation programmes.
The Structural Read
As we read it, the business point is the split of labour the paper describes: the frontier LLM supplies broad chemical and biological knowledge, while the specialised flow model supplies the priors needed to produce valid designs.
As we read it, the cost lines matter as much as the scores: the paper records LLM API spend in tens of dollars per campaign, with GPU use given separately in hours.
As we read it, the open question is the lab: every gain the paper reports is on held-out computational evaluators, and the authors name wet-lab feedback as the next step.
De Santi, Di Gennaro, Gnazzo et al., arXiv 2610.17246
“a frozen LLM acts as an adaptive teacher for a domain-specific generator”
Three Implications
FOR DRUG-DISCOVERY TEAMS The paper reports results on peptides and small molecules without a task-specific reward model, using goals written in natural language; all evaluations are in silico.
FOR AI BUDGET OWNERS The appendix records LLM API costs of about $8.83 to about $77 per campaign, alongside runtimes from about 5.3 hours to a 48-hour limit.
FOR MODEL PROVIDERS In the reported experiments the teacher is Gemini 3.1 Pro for the curriculum and Gemini 3.7 Flash for judgments; De Santi writes that the method works with any frontier LLM via API access.
The framework names four moats, one per data paradigm, including “The Verifier Moat โ built by RLVR on domain-specific reward signals, building into a new asset class for reasoning-heavy verticals.”
As we read it, LFD tests that moat from the other side: the paper reports adapting generators without a task-specific proxy reward, using a general LLM’s judgments instead. The framework does not discuss this paper, and the paper’s results are computational.
What Is Not Established
Every result in the paper is computational. The authors compare LFD through in silico experiments, and the docking scores come from computational scoring tools (GNINA, Vina and rDock), not binding assays. The authors write that “Integrating wet-lab feedback into LFD is a natural next step.”
For the cell-penetrating peptide goal, the appendix notes that “The comparison uses one training seed and unequal optimization budgets”, and that the reward and evaluator share historical labels. The paper is posted as a preprint.
The recorded API costs cover the LLM calls; the appendix gives GPU use in hours rather than dollars.
Business Engineer Framework
The Four Intelligence Moats
A Business Engineer framework on where intelligence accumulates in each AI paradigm, and who can capture it.
The paper reports that a frozen Gemini teacher, without a task-specific reward model, adapted peptide and molecule generators to plain-language discovery goals at a recorded LLM API cost of about $8.83 to $77 per campaign.
Its gains over reward-guided baselines are on held-out computational evaluators, not lab assays.
A note on sourcing. We read the arXiv paper 2610.17246v1 by De Santi, Di Gennaro, Gnazzo and co-authors in full as a PDF, and Riccardo De Santi’s 14 October 2026 X post. Every figure, including the API costs from the paper’s appendix and the scores from its Tables 1 and 2, comes from the paper; the $210 total is our arithmetic. All results are the paper’s computational evaluations. We did not contact the authors or use other outlets’ reports.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
