Metacognitive Steering: Learning the Structure of Scientific Judgment

• URL: http://arxiv.org/abs/2609.16245v1
• Date: Mon, 14 Sep 2026 19:10:03 GMT
• Title: Metacognitive Steering: Learning the Structure of Scientific Judgment
• Abstract summary: Long-horizon scientific discovery requires agents to alternate between exploration, disciplined execution, and critical reassessment as evidence changes.<n>Current language models are trained primarily on the products of science and optimized using outcome-level signals.<n>We investigate whether such judgment can be recovered from scientist interaction traces and used to control the internal synthesis of a frozen frontier model.
• Score: 0.16863755729554886
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Long-horizon scientific discovery requires agents to alternate between exploration, disciplined execution, and critical reassessment as evidence changes. Current language models are trained primarily on the products of science and optimized using outcome-level signals, providing limited supervision for these process-level shifts in scientific judgment. We investigate whether such judgment can be recovered from scientist interaction traces and used to control the internal computation of a frozen frontier model. Using contrastive interventions collected during real scientific research, we identify a coordinated, low-dimensional control structure within Kimi 2.6, a trillion-parameter mixture-of-experts model. Residual analysis, attention-weight subspace alignment, and cross-layer singular value decomposition converge on a mid-depth control surface spanning key layers. We introduce Metacognitive Steering, an inference-time controller that reads the model's cognitive regime and dynamically composes layer-specific interventions for exploration, procedural convergence, or critical reassessment without modifying model parameters. Behavioral analyses show that this control produces more sustained exploration, explicit pruning, and evidence-responsive synthesis. We operationalize the method in Columbus-1, an autonomous research system that identified eight independently reproduced, attacker-reachable vulnerabilities in BlueZ and directed the design, simulation, and fabrication of a ten-foot rocket intended to land propulsively using non-throttleable solid motors. Together, these results show that process-level scientific judgment can provide supervision for interpretable, dynamic control over a model's reasoning strategy.

Related papers

• Rethinking Scientific Discovery in the Agentic Era [69.85050591628048]
  SCION is an agentic scientific operating system that acts as an textbf organizational nexus.<n>At its core is the textbfResearch Execution Plan (REP), which compiles high-level scientific intent into staged objectives, dependencies, verification checkpoints, tool requirements, expected artifacts, and fallback conditions.<n>We formulate discovery under SCION as textbfTarget-conditioned Inverse Search and extend it to hidden-target settings through batch active search under finite experimental budgets.
  arXiv  Detail & Related papers  (2026-07-04T13:09:33Z)
• Self-Evolving Scientific Agent Discovers Generalizable Physically-Reasoned Fluid Control [6.923599492589765]
  We present a self-evolving scientific-agent workflow that automates controller construction while preserving strict interpretability and rigorous physical reasoning.<n>We show that an autonomous scientific agent can successfully transform accumulated physical evidence into robust, mathematically readable control policy.
  arXiv  Detail & Related papers  (2026-06-07T01:59:11Z)
• Science Earth: Towards A Planet-Scale Operating System for AI-Native Scientific Discovery [51.70607714288616]
  Science Earth is a planet-scale scientific runtime in which any capability can connect to any other.<n>Its underlying EACN protocol lets capabilities discover one another, negotiate task ownership, and adjudicate across incompatible evidentiary standards.<n>Two runs validate this under structurally distinct conditions.
  arXiv  Detail & Related papers  (2026-05-31T16:05:41Z)
• MOOSE-Copilot: A Web-Based Interactive Assistant for Unified Exploratory and Fine-Grained Scientific Hypothesis Discovery [4.6860802702914786]
  We present MOOSE-Copilot, the first unified framework to bridge this abstraction gap through a formalized human-AI interaction protocol.<n>Our system empowers scientists to steer the generative process via three explicit signals: initial blueprints, inter-stage routing, and regenerative feedback.<n>To democratize this paradigm, we develop an intuitive web-based interface featuring interactive tree visualization.
  arXiv  Detail & Related papers  (2026-05-28T07:06:10Z)
• Grounding LLMs in Scientific Discovery via Embodied Actions [84.11877211907647]
  Large Language Models (LLMs) have shown significant potential in scientific discovery but struggle to bridge the gap between theoretical reasoning and physical simulation.<n>We propose EmbodiedAct, a framework that transforms established scientific software into active embodied agents by groundings in embodied actions with a tight perception-execution loop.
  arXiv  Detail & Related papers  (2026-02-24T07:37:18Z)
• OR-Agent: Bridging Evolutionary Search and Structured Research for Automated Algorithm Discovery [10.217363774023033]
  OR-Agent is a multi-agent research framework designed for automated exploration in rich experimental environments.<n>We introduce an evolutionary-systematic mechanism that unifies evolutionary selection of research starting points, comprehensive research plan generation, and coordinated exploration within a research tree.<n>We conduct experiments across classical optimization benchmarks-including traveling salesman, capacitated vehicle routing, bin packing, orienteering, and multiple knapsack problems-as well as a simulation-based cooperative driving scenarios.
  arXiv  Detail & Related papers  (2026-02-14T13:32:03Z)
• NewtonBench: Benchmarking Generalizable Scientific Law Discovery in LLM Agents [65.85967483058705]
  Large language models are emerging as powerful tools for scientific law discovery.<n>Existing benchmarks for this task suffer from a fundamental methodological trilemma.<n>We introduce NewtonBench, a benchmark comprising 324 scientific law discovery tasks across 12 physics domains.
  arXiv  Detail & Related papers  (2025-10-08T16:12:11Z)
• Active Inference AI Systems for Scientific Discovery [1.450405446885067]
  This perspective contends that progress turns on closing three mutually reinforcing gaps in abstraction, reasoning and empirical grounding.<n>Design principles are proposed for systems that reason in imaginary spaces and learn from the world.
  arXiv  Detail & Related papers  (2025-06-26T14:43:04Z)
• Beyond Prompt Engineering: Robust Behavior Control in LLMs via Steering Target Atoms [71.85633762642125]
  The vast number of parameters in models often results in highly intertwined internal representations.<n>Recent research has explored the use of sparse autoencoders (SAE) to disentangle knowledge in high-dimensional spaces for steering.<n>We propose Steering Target Atoms (STA), a novel method that isolates and manipulates disentangled knowledge components to enhance safety.
  arXiv  Detail & Related papers  (2025-05-23T17:59:18Z)
• Interpretable Imitation Learning with Dynamic Causal Relations [65.18456572421702]
  We propose to expose captured knowledge in the form of a directed acyclic causal graph. We also design this causal discovery process to be state-dependent, enabling it to model the dynamics in latent causal graphs. The proposed framework is composed of three parts: a dynamic causal discovery module, a causality encoding module, and a prediction module, and is trained in an end-to-end manner.
  arXiv  Detail & Related papers  (2023-09-30T20:59:42Z)
• Empirical Estimates on Hand Manipulation are Recoverable: A Step Towards Individualized and Explainable Robotic Support in Everyday Activities [80.37857025201036]
  Key challenge for robotic systems is to figure out the behavior of another agent. Processing correct inferences is especially challenging when (confounding) factors are not controlled experimentally. We propose equipping robots with the necessary tools to conduct observational studies on people.
  arXiv  Detail & Related papers  (2022-01-27T22:15:56Z)
• Bayesian optimization of distributed neurodynamical controller models for spatial navigation [1.9249287163937971]
  We introduce the NeuroSwarms controller, in which agent-based interactions are modeled by analogy to neuronal network interactions. This complexity precludes linear analyses of stability, controllability, and performance typically used to study conventional swarm models. We present a framework for tuning dynamical controller models of autonomous multi-agent systems based on Bayesian Optimization.
  arXiv  Detail & Related papers  (2021-10-31T21:43:06Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.