PaperlayerView exact revision on arXiv

Source-linked AI summary

Before Agents Decide: Epistemic Action in LLM-Based Systems

Yizhi Liu, Balaji Padmanabhan, Siva Viswanathan

arXiv:2610.00511v1cs.AIcs.MA

TL;DR

LLM-agent design gives limited attention to whether evidence is ready for a decision, though evidence may be missing, poorly represented, or not yet created. The paper distinguishes three epistemic-action modes and describes scaffolding that enables and audits them. It concludes that preparing evidence belongs alongside reasoning as a central problem in reliable-agent design.

  • Problem

    Agent design gives less attention to whether available evidence is ready for a decision, although necessary evidence may be missing, poorly represented, or not yet produced.

    Source: §1 · p. 1, §2 · p. 2

  • Method

    The paper distinguishes acquiring missing evidence, transforming available evidence, and probing to create evidence, supported by interfaces, tools, representations, and environments.

    Source: §3 · p. 2, §4 · p. 3

  • Results

    The paper concludes that preparing evidence belongs alongside reasoning as a central problem in reliable-agent design.

    Source: §5 · p. 4

  • Takeaways & Limitations

    Epistemic scaffolding should make evidence-preparing actions possible and the resulting evidence auditable.

    Source: §5 · p. 4

    The discussed systems demonstrate individual capabilities, but none chooses among all three modes and justifies both acting and stopping.

    Source: §4 · p. 4

Abstract

from arXiv · show

Before a difficult decision, people often act simply to understand the situation better. We turn an object to see another side, place alternatives next to each other, or change one condition and observe what happens. These actions may not complete the task, but they improve the evidence needed for the next choice. LLM-based agents can search and explore, yet agent design gives less attention to an earlier question: is the available evidence ready for the decision? Sometimes necessary evidence is missing. In other cases, the evidence is present but its form hides what matters, or the comparison needed to judge it does not yet exist. Cognitive science calls actions that improve the basis for a later choice epistemic actions. We bring this idea to LLM-based agents and distinguish three modes: acquiring missing evidence, transforming available evidence, and probing a system to create a revealing response. We use the term epistemic scaffolding for the interfaces, tools, and environments that make these actions possible and auditable. This paper argues that agent design must address how decision-ready evidence is produced.

1 Introduction

Epistemic actions prepare evidence for a later choice, unlike actions that directly advance a task. The paper argues that LLM-based agents should judge whether their evidence is decision-ready and acquire, transform, or probe it when needed.

1 Introduction

  • Epistemic actions improve the basis for a choice, whereas pragmatic actions directly change the world toward a goal.The figure contrasts changing task state with changing evidence for a later decision.Source: §1 · p. 1, Figure 1 · p. 2
  • An agent limited to one visible puzzle face cannot infer hidden stickers through further reasoning alone.It needs a way to change what it can observe before choosing a move.Source: §1 · p. 1
  • LLM-based agents may produce polished answers even when their interactions have not exposed the evidence a decision requires.Preparing decision evidence is rarely treated as a capability in its own right.Source: §1 · p. 1
  • Agent design should make epistemic action first-class and let agents acquire, transform, or probe evidence before answering or acting.The agent should first judge whether its evidence is ready for the decision.Source: §1 · p. 2

2 Acting is not the same as acting for evidence

LLM-based agents already use tools, but tool use alone does not ensure that they recognize when evidence is inadequate or choose how to improve it. Evidence may be missing, poorly represented, or dependent on an untried intervention.

2 Acting is not the same as acting for evidence

  • Agents already search, inspect pages, edit files, and call tools, but must still judge evidence adequacy and choose what would improve it.WebShop agents may inspect another product when a purchase depends on an unseen fact.Source: §2 · p. 2
  • Reasoning cannot recover a hidden fact from an observation that does not reveal it; agents must obtain, reorganize, or produce needed evidence first.The need arises when evidence is missing, its form hides a relationship, or an untried intervention determines the answer.Source: §2 · p. 2
  • Epistemic action connects LLM-agent design to work on active perception, information systems, decision support, and strategic search.These traditions share the lesson that what decision makers can know partly depends on the actions available to them.Source: §2 · p. 2

3 Three modes of epistemic action

The paper distinguishes acquiring missing evidence, transforming evidence that is difficult to use, and probing to produce evidence that does not yet exist. Decisions may require any combination of these modes—or none.

3 Three modes of epistemic action

  • The three modes address evidence that is absent, difficult to use in its current form, or not yet produced.Table 1 presents the modes as options to consider before commitment.Source: §3 · p. 2, Table 1 · p. 3
  • Acquisitive actions add an observation before choice, for example by asking a question, retrieving a source, or inspecting the environment.WebShop and ReAct illustrate agents gathering information before choosing.Source: §3 · p. 3
  • Transformative actions change the representation when available evidence hides relevant structure, rather than simply searching for more.Examples include restructuring repository feedback and compressing prompts while seeking to preserve semantic content.Source: §3 · p. 3
  • Probing creates a revealing situation when evidence depends on an untried comparison or response.Examples include controlled variation, testing material conductivity with a circuit, and generating matched patient images for a bias audit.Source: §3 · p. 3
  • A probe may alter the behavior it measures, so its evidence must be interpreted alongside the conditions that elicited it.The modes may also be combined, and an agent must judge when another action is worth its cost.Source: §3 · p. 3

4 Epistemic scaffolding and a research agenda

Epistemic scaffolding comprises the interfaces, tools, representations, and environments that enable epistemic actions and make their evidence auditable. The paper proposes design and evaluation principles while noting risks and current capability limits.

4 Epistemic scaffolding and a research agenda

  • Epistemic scaffolding is the design around an agent that enables and bounds its epistemic actions.Static screenshots can restrict acquisition, irreversible summaries can limit transformation and checking, and simulators without controllable variables can prevent probing.Source: §4 · p. 3
  • Useful scaffolds record evidence provenance, keep transformations inspectable, and support stopping when likely gains no longer justify costs.Such records help show whether a tool call improved the basis for a later decision.Source: §4 · p. 4
  • An auditable trace must record the pending decision, evidence before and after an action, and the expected evidential change.Tool calls alone do not reveal an action’s purpose.Source: §4 · p. 4
  • A value-of-information rule can rank available actions but cannot provide an action omitted from the scaffold.The selection rule concerns choosing among available actions; the paper’s claim concerns how that set is constructed.Source: §4 · p. 4
  • The proposed benchmark varies access to acquisition, transformation, and probing while separating need recognition, action selection, and use of results.It compares prescribed actions with actions available without instructions about when to use them.Source: §4 · p. 4
  • Multi-agent evaluation should track how evidence is produced and transformed across interactions, while accounting for tradeoffs and current system limits.Transformations can discard information, probes can introduce artifacts, and additional observations can increase privacy or security exposure; no discussed system chooses among all three modes and justifies acting and stopping.Source: §4 · p. 4

5 Conclusion

LLM-based agents may face decisions where evidence is missing, poorly represented, or not yet created; acquisitive, transformative, and probing actions prepare it before commitment. Epistemic scaffolding provides the means for these actions and makes their evidence auditable.

  • Acquisitive, transformative, and probing actions prepare evidence when it is missing, poorly represented, or not yet created.Source: §5 · p. 4
  • Epistemic scaffolding gives agents the means to take evidence-preparing actions and makes the resulting evidence auditable.Source: §5 · p. 4
  • The paper places evidence preparation alongside reasoning as a central problem in designing reliable agents.Source: §5 · p. 4

Related papers

Loading 2610.00511v1…