Agents in Research
« Agents in Research

#14 — OpenAI GPT-5.4 boosts drug synthesis, DN-Hypo-Pipeline generates hypotheses

June 21, 2026
A new AI-powered workflow, DN-Hypo-Pipeline, uses large language models to help generate novel hypotheses from existing scientific literature. Other recent applications include using AI to optimize drug-synthesis reactions in robotic labs and employing agent-based systems to make discoveries in protein fitness prediction and pure mathematics. While enterprise adoption of AI agents is high, many pilots fail to reach production due to scoping and governance issues, not model performance. The return on investment for deployed agents is contingent on defining success criteria and governance before building.

Sources

  1. OpenAI's AI Chemist Cracks a Decades-Old Drug Reaction | Let's ...
    OpenAI connected GPT-5.4 to a Polish startup's robotic lab and gave it an open goal: improve a hard drug-synthesis reaction. The model suggested an additive chemists had overlooked. Across 10,080 reactions, average yield rose from 16.6 percent to 25.2 percent, and four outside experts called the result novel.
  2. DN-Hypo-Pipeline: An AI-Driven Workflow for Hypothesis Generation via Large Language Models and Scientific Explanations
    A scientific hypothesis is the first step in research and undergoes experimental validation, yet it also reflects a deep understanding of and reasoning about scientific phenomena. We introduce DN-Hypo-Pipeline, an AI-powered workflow based on large language models, designed to support structured scientific thinking and hypothesis generation by leveraging scientific explanations as prior knowledge. This pipeline assists researchers in deriving novel hypotheses from existing literature. Given the explanandum (i.e., the conclusion) of a research paper, it identifies underlying laws, theories,…
  3. AutoScientists: Self-Organizing Agent Teams for Long-Running Scientific Experimentation
    Scientific research proceeds through iterative cycles of hypothesis generation, experiment design, execution, and revision. AI agents can automate parts of this process, but existing approaches typically follow a single research trajectory or coordinate through a central planner with fixed objectives. As a result, they struggle to sustain parallel exploration, adapt as experimental evidence changes, or preserve knowledge of failed directions over long-running experiments. We introduce AutoScientists, a decentralized team of AI agents for long-running computational scientific experimentation.…
  4. Harnessing the Collective Intelligence of AI Agents in the Wild for New Discoveries
    Scientific discovery is often a collective process: researchers share partial results, inspect failed attempts, and build on each other's ideas over long time horizons. Recent AI systems have shown that language-model-based agents can make meaningful progress on open scientific problems, but most existing systems operate in isolation. In this paper, we present EinsteinArena, an agent-native platform for open distributed research and discovery. EinsteinArena provides agents with a live set of open problems, each with a solid verifier, public leaderboard, and problem-specific discussion forum…
  5. Agentic AI Implementation Guide 2026: From Pilot to Production ...
    Data reveals that 79% of enterprises have adopted AI agents in some form, but only 31% run them in production, indicating a significant gap between experimentation and scale. A substantial 88% of AI agent pilots fail to reach production, primarily due to scoping and governance failures rather than model issues. While deployed agents show an average ROI of 171%, this is achieved only when success criteria, tool access, and governance are defined prior to building. The guide also highlights that median enterprises underestimate the 3-year total cost of ownership by 57% and emphasizes that…
  6. Using AI to help physicians diagnose rare genetic diseases affecting children | OpenAI
    In an NEJM AI study, experts used an OpenAI reasoning model to reanalyze 376 previously unsolved cases and surface leads for 18 diagnoses.

Also this week

Full transcript
An AI model improved a drug synthesis reaction by suggesting an additive that human chemists had missed. This is one of the developments in automated science we are covering on Research Agent, where we track AI's application to research. We'll start there. AI systems are starting to move from just processing information to actively participating in scientific problem-solving. Right, and in some cases, they're finding things that human experts missed. OpenAI connected its GPT-5.4 model to a robotic lab to work on a drug-synthesis reaction. And it didn't just run the experiment; it found a new additive. The result was a big jump in the reaction yield, from around 16 percent to over 25 percent. And that wasn't an internal benchmark. Four external experts looked at it and confirmed the result was novel. This isn't a one-off event, either. A different system, a multi-agent setup called AutoScientists, also made a discovery. This one was in biology. It found a new method for predicting how certain proteins bind. The new method improved the correlation by 12.5 percent over the previous state-of-the-art model. So, it's a better predictor. But was it only good for that one specific problem? That's the key question for these tools. In this case, they applied the same method to 217 other assays without any modification, and it still showed a general performance improvement of 6.5 percent. That suggests the approach is robust. So we have individual AIs and multi-agent systems making contributions. There's also the medical field, where a reasoning model helped reanalyze hundreds of unsolved medical cases and surfaced leads for 18 potential diagnoses. And new platforms are being built to enable this kind of work. One is called EinsteinArena, which is designed for open, distributed research among different AI agents. What does that look like in practice? Agents working together? Exactly. They use it to tackle problems like the kissing number problem in mathematics. On the platform, agents improved the best known lower bound in dimension 11, raising it from 593 to 604. And how did that happen? Was it one brilliant agent? No, it was a collective process. Agents made submissions, there was public discussion, and they shared ideas. It suggests a model for decentralized discovery. As of May 2026, agents on that platform had produced 12 new state-of-the-art results. There's also a different kind of tool, the DN-Hypo-Pipeline, which uses large language models to help human researchers generate hypotheses from scientific papers. And in evaluations, it was found to be more effective than more direct methods. It even generated two hypotheses that were later validated by creating new algorithms that outperformed the baselines. Okay, so we're seeing all these results in research settings. You'd assume enterprises would be deploying these agents everywhere. The picture there is more complicated. The data shows that while nearly 80 percent of enterprises have adopted AI agents in some way, only 31 percent are actually running them in production. That's a huge gap between trying something and actually using it. What's going wrong? Is the technology failing in a business context? It seems the problem isn't the model performance. The pilot programs have a failure rate of 88 percent, but that's attributed mainly to issues with scoping and governance—not the tech itself. So, companies are not defining the problem correctly or setting up the right controls. But for the ones that get it right, the payoff is there. It is. The average return on investment is 171 percent. But that depends on doing the work up front: defining success, providing the right tools, and establishing governance before you even start building. And it seems they're also struggling with budgeting. The median enterprise underestimates the 3-year total cost of ownership by 57 percent. On top of all those operational hurdles, there's a regulatory deadline. Compliance with the EU AI Act for high-risk agentic systems is required by August 2026, and most enterprises are reportedly not prepared. That covers the latest developments. We'll have more updates next time. Thanks for listening to Research Agent.