Agents in Research
« Agents in Research

#19 — AI agency impacts, epistemic control loss, automated DBTL success

July 29, 2026
An exploratory review reveals the unintended impacts of AI on agency and authorship. A related concept, the 'Scientific Singularity,' suggests a future where humans lose epistemic control over scientific processes. This threshold is crossed when scientific claims are accepted without humans being able to verify the machine's reasoning. The digest also covers advances in automated bioformulation, robotics, and SAT solvers.

A thematic review found that AI use leads to unintended impacts on agency, authorship, and autonomy. The review also indicated potential for increased inequality. This research contributed to discussions on responsible AI deployment. Separately, a research article posits that humans in scientific processes, despite maintaining roles as authors or regulators, may lose epistemic control. This means humans could approve machine-generated conclusions without fully understanding their verification paths. The term "scientific singularity" traditionally refers to a future where AI, robotics, and programmable infrastructure accelerate and decentralize scientific production. Another definition for "Scientific Singularity" describes a point where a scientific field accepts claims without humans being able to reconstruct or challenge the underlying inferential processes. The article also introduced the concept of a "hierophant" role in future scientific institutions. This role would involve individuals who interpret and transmit AI-generated scientific results to humans, without necessarily comprehending the AI's underlying inferential steps.

An automated Design-Build-Test-Learn (DBTL) process generated improved formulations for model antibodies. This process also created accurate predictive models of formulation behavior across objectives like melting temperature and viscosity. The automated DBTL technique was validated in bioformulation for mapping property-function landscapes and guiding formulation development, managing various design requirements. ASPIRE (Agentic Skill Programming through Iterative Robot Exploration) is a continual learning system for robot control. It autonomously generates and refines robot programs and builds a reusable skill library from experience. ASPIRE exhibited zero-shot generalization on new, long-horizon tasks, achieving 31% success on LIBERO-Pro Long, compared to 4% for prior methods. ASPIRE also showed sim-to-real skill transfer, meaning skills learned in simulated environments could be directly applied to real-world robots.

AutoSAT is a framework that uses Large Language Models (LLMs) to optimize heuristics for Conflict-Driven Clause Learning (CDCL) SAT solvers. This system generates new heuristics, applying search strategies like greedy hill climbing, and has shown to improve CDCL solver performance over established solvers such as MiniSat.

Sources

  1. Agency in the age of generative AI: a critical review of educational implications
    Conversations regarding the societal implications of artificial intelligence (AI) and its integration into educational processes are ongoing. An area of great significance in this regard is agency, including how the use of AI technologies impacts an individual’s ability to exercise autonomy. In educational contexts, this remains a relatively unexplored yet vital subject. To shed light on this subject, this paper provides an exploratory thematic review of works which investigate the relationship between generative artificial intelligence (GenAI) and agency in education, mapping and…
  2. The Scientific Singularity: When Human Understanding Ceases to Be a Condition of Scientific Acceptance
    Existing uses of “scientific singularity” describe a future in which scientific production becomes radically faster, more autonomous, and more decentralized through artificial intelligence, robotic laboratories, and programmable research infrastructure. This article uses the term for a different threshold. AI making discoveries is not the same event as human understanding ceasing to be a condition of scientific acceptance. A machine may generate a result while humans still reconstruct its evidential basis, challenge its inferences, and decide whether it warrants acceptance. The deeper…
  3. Automation and Active Learning for the Multi‐Objective Optimization of Antibody Formulations
    Over the last forty years, monoclonal antibodies have become increasingly important therapeutic agents, with most manufactured as preformulated solutions. However, bioformulation of complex proteins is a difficult engineering challenge; formulations must be tailored to individual therapies, necessitating time- and material-intensive campaigns to select combinations of excipients to simultaneously optimize various design criteria. These additives complicate formulation design with unintuitive and non-linear relationships, creating a vast multidimensional design space that is intrinsically…
  4. ASPIRE: Agentic /Skills Discovery for Robotics
    Traditional robot programming is challenging: it requires orchestrating multimodal perception, managing physical contact dynamics, and handling diverse configurations and execution failures. We introduce ASPIRE (Agentic Skill Programming through Iterative Robot Exploration), a continual learning system that autonomously writes and refines robot control programs in a code-as-policy paradigm while compounding experience into a reusable skill library. ASPIRE discovers skills that persist across tasks, simulation and real-world settings, and embodiments. It operates in an open-ended loop with…
  5. AutoSAT: Automatically Optimize SAT Solvers via Large Language Models
    Background: Conflict-Driven Clause Learning (CDCL) is a dominant framework for solving the Satisfiability problem (SAT). Modern CDCL solvers rely heavily on various heuristics, which significantly influence their performance. Established solvers such as MiniSat and Kissat typically incorporate multiple heuristics and therefore require substantial manual effort and domain expertise for fine-tuning in practice. Objectives: The emergence of Large Language Models (LLMs) offers a promising opportunity to automate SAT solver optimization. However, generating a complete CDCL solver from scratch…

Also this week

Full transcript
What happens when the systems designed to augment human capabilities begin to diminish human agency? We're exploring the unintended consequences on this edition of Agents in Research. The episode begins with how AI is affecting authorship and autonomy. Let's start with this idea of a 'Scientific Singularity.' A recent paper is redefining the term. It's not the usual definition, right? Not just AI making science faster. No, it's much more specific. The paper defines it as the threshold where a scientific field starts accepting claims that no human can independently validate. So, the AI generates a conclusion, and humans might be listed as authors or regulators, but they can't actually reconstruct the evidence. They have no real epistemic control. Exactly. They're approving findings without being able to verify the process. And the paper proposes a new kind of role to manage this: the 'hierophant.' Like a priest for the algorithm. An intermediary who interprets the AI's output for human decision-makers, but doesn't necessarily understand the deep mechanics either. It’s a striking concept. And this theme of changing agency and authorship is appearing elsewhere, particularly in discussions around education. An exploratory review looked at just that. It pointed to potential consequences from generative AI, like increased inequality and fundamental shifts in how we even think about individual autonomy. While these conceptual questions are being debated, the practical work in labs continues to accelerate. We have an update on that automated Design-Build-Test-Learn process we covered before. And it's producing results. The closed-loop method generated improved formulations for a model antibody. What were the metrics for 'improved' in this case? Things like melting temperature and viscosity. The system also produced accurate predictive models of how the formulation would behave, helping to map out complex property-function landscapes. So it guides development by showing how to balance competing design needs. This connects back to that singularity idea—the AI is mapping a space that would be difficult for humans to navigate alone. And this kind of autonomous learning is also showing up in robotics. A new system called ASPIRE is designed to write and refine its own robot control programs. It builds a library of skills from its own experience. The performance numbers are what stand out. Right. On one long-horizon task benchmark, it achieved a 31 percent success rate. The previous methods were at 4 percent. It also showed effective sim-to-real transfer, so skills learned in a simulation could be applied to physical robots. In a different area, we see a similar approach with a framework called AutoSAT. It's using Large Language Models to optimize heuristics for SAT solvers. So an LLM is being used to discover better search strategies for a classic computing problem. And it worked—it found new heuristics that outperformed existing solvers on multiple datasets. A few other developments to note: Inverse Reinforcement Learning was used to model household electricity consumption, and a systematic review has been started to look at agentic AI in palliative care. That’s the digest for this week. We’ll track more developments for you next time. From Agents in Research, thanks for listening.