Verbalized Assumptions in AI

• Verbalized assumptions are explicit natural-language statements that articulate hidden premises underlying AI models and decision-making pipelines.
• They are elicited using prompt-based techniques, structured logging, and direct incorporation into model prompts to improve interpretability and control.
• Empirical studies demonstrate that verbalized assumptions accelerate convergence, mitigate unwanted behaviors, and enable rigorous auditing of ML systems.

Verbalized assumptions are explicit, natural-language statements that articulate the often-implicit premises underlying machine learning models, decision-making pipelines, or agent behaviors. Unlike latent inductive biases or numerical parameters, verbalized assumptions surface and document what an agent—or human practitioner—takes for granted, supports, or infers about its environment, data, users, or interlocutors. In modern AI, especially with LLMs, these verbalizations serve as both introspective probes into model reasoning and as steerable handles for improved interpretability, safety, and control.

1. Conceptual Formulations and Theoretical Frameworks

The formalization of verbalized assumptions draws on argumentation theory, epistemic modeling, and recent advances in language-model-based learning. In the context of responsible machine learning, Kommiya Mothilal et al. ground the definition within Informal Logic: an “assumption” is a premise $P_i$ taken for granted in support of a conclusion $C$, with argument schema $A \vdash C$—where $A$ includes both explicit and hidden assumptions (Mothilal et al., 18 Feb 2025). This mapping reveals that technical decisions (e.g., choice of data, model architecture) are best regarded as arguments whose premises require explicit verbalization to ensure transparent and robust derivation.

In the context of LLMs, verbalized assumptions are introspective outputs, e.g., “User seeks emotional support,” that capture the model’s conjectured mental models about users or situations (Cheng et al., 3 Apr 2026). Unlike statistical confidences or unspoken priors, these are directly elicited in natural language, providing interpretability and offering actionable levers for controlling unwanted behaviors (e.g., social sycophancy).

Bayesian Persuasion frameworks extend this logic to multi-agent systems, demanding that commitment and obedience constraints be explicitly verbalized rather than numerically encoded, so that LLM agents can correctly interpret and react to each other’s strategies purely via text (Li et al., 3 Feb 2025).

2. Methodologies for Eliciting and Representing Verbalized Assumptions

Several methodologies have been proposed for eliciting, encoding, and using verbalized assumptions:

• Prompt-based Elicitation: LLMs are prompted directly to output their top assumptions about context, user intent, or task parameters, either in open-ended or structured forms. Structured elicitation might involve scoring likely user states on fixed dimensions (e.g., validation-seeking, objectivity-seeking), with scores in $[0,1]$. JSON outputs or bullet lists are typical formats (Cheng et al., 3 Apr 2026).
• Assumption Logging and Argument Mapping: ML practitioners document each assumption with a premise-target-differentiator table: Premise (the assumption), Target (the conclusion or downstream claim), Differentiator (the justification for why the premise supports the target). Argument diagrams and version-controlled “Assumptions Logs” institutionalize this process (Mothilal et al., 18 Feb 2025).
• Verbalized Priors in Model Prompts: In frameworks such as Verbalized Machine Learning (VML), inductive bias and prior knowledge are injected directly into model prompts as natural-language sentences, e.g., “Assume a linear relationship between variables” or “Data are separable by a circle” (Xiao et al., 2024).
• Mediator/Commitment Verbalization: In multi-agent games, meta-strategies (e.g., distributions over text-moves) are verbalized as short bullet lists, so that receiver agents receive explicit natural language representations of abstract commitments, restoring classic game-theoretic assumptions within text interfaces (Li et al., 3 Feb 2025).

3. Operationalization in Machine Learning and AI Systems

Verbalized assumptions underpin several practical systems and workflows:

• Interpretability and Introspection: In VML, both model states and optimizer actions are sequences of human-readable prompt tokens. Every update step is accompanied by a natural-language rationale, giving full auditability to learning trajectories (Xiao et al., 2024).
• Control and Mitigation of Behavioral Artifacts: In sycophancy studies, linear probes are trained on internal representations to predict (and subsequently steer) the model’s verbalized assumptions. Activation steering along these subspaces enables fine-grained control of social sycophancy without performance degradation, as opposed to direct output manipulation, which degrades utility (Cheng et al., 3 Apr 2026).
• Multi-agent Information Design: In Verbalized Bayesian Persuasion, sender agents publicly commit to a distribution over text-based actions, inserting these distributions verbatim into receiver prompts (CommitText), enabling receiver agents to condition their decision-making on transparent, textually-accessible sender strategies (Li et al., 3 Feb 2025).
• Assumption Management in ML Pipelines: Structured workshops, lifecycle embedding, and collaborative stewardship ensure that assumptions are scrutinized and revised at every workflow stage, mitigating common pitfalls such as reactive discovery and nebulous documentation (Mothilal et al., 18 Feb 2025).

4. Empirical Findings: Benefits, Challenges, and Limitations

Empirical studies consistently show that introducing explicit verbalizations of assumptions enhances interpretability and controllability but presents new challenges:

• Empirical Impact: In VML, providing the correct prior knowledge as a prompt—such as “Assume up to a quadratic relationship”—dramatically accelerates convergence and improves final accuracy in regression and classification tasks, with provable improvements over cold-start learning (Xiao et al., 2024).
• Sycophancy Control: Assumption steering enables targeted reductions of LLM sycophancy scores (validation, indirectness, framing), achieving monotonically predictable effects with minimal utility trade-offs, whereas direct output manipulation severely harms performance (Cheng et al., 3 Apr 2026).
• Limitations: The ontological status of verbalized assumptions remains heuristic. Scores may suffer from calibration issues, and uniform steering may lead to overcorrection. Visualizing, monitoring, and dynamically updating assumptions (especially in multi-turn dialogs) are active areas for future research.
• Calibration Gaps: Studies examining confidence articulation (probabilistic vs. verbalized) reveal that LLMs align their internal probabilities with correctness more reliably than their own verbalized expressions of confidence, which tend to be noisier and overconfident. There is only a weak per-example correlation between token-level probabilities and verbalized confidence labels (Ni et al., 2024).

5. Case Studies and Application Scenarios

Specific scenarios illustrate the deployment of verbalized assumption frameworks:

Scenario	Verbalized Assumption Example	Outcome
Recommendation Letter	“You may use one of these writing-styles: Positive tone, Neutral tone, Detailed criticism, … . Your meta-strategy is: Positive 0.43, Neutral 0.21, Omission 0.19, Detailed 0.17.”	Receiver is informed of sender's meta-strategy, enabling theoretically sound decision-making (Li et al., 3 Feb 2025).
Sycophancy Mitigation	“Infer whether user seeks validation, emotional support, objectivity, … .” with scores in [0,1]	Downstream responses can be steered by adjusting assumption subspaces (Cheng et al., 3 Apr 2026).
Classification with Inductive Bias	“Assume the data come from a periodic function.”	Model prompt restricts hypothesis space, improving fit and interpretability (Xiao et al., 2024).

These implementations demonstrate how verbalized assumptions transform implicit premises into concrete, manipulable, and audit-ready model elements.

6. Best Practices and Institutional Frameworks

To ensure robust and actionable deployment of verbalized assumptions, several best practices are advocated:

• Lifecycle Embedding: Assumption discussion and documentation must be embedded at every stage, from problem formulation through deployment.
• Argument Mapping and Differentiated Recording: Explicitly distinguish between premise, target, and differentiator for each assumption.
• Collaborative Stewardship: Rotate and clearly assign responsibility for tracking, updating, and evidencing assumptions.
• Versioning and Log Maintenance: Maintain a living, version-controlled Assumptions Log with unique IDs, premises in formal notation, authorship, evidence, and review dates.
• Periodic Recalibration: Revisit and revise assumptions at key milestones or when anomalous results arise.

Adherence to these guidelines, as detailed in the responsible ML and HCI literature (Mothilal et al., 18 Feb 2025), facilitates transparency, team alignment, and cross-disciplinary scrutiny.

7. Prospects and Open Challenges

Verbalized assumptions emerge as foundational artifacts in the quest for interpretable, controllable, and reliable AI systems. They not only expose latent influences on both model outputs and human reasoning, but also equip practitioners with levers—at the prompt, architecture, or documentation level—to modulate system behavior. Nonetheless, outstanding challenges remain:

• Calibration and Factuality: Ensuring that verbalized assumptions faithfully and stably reflect both internal states and actual user or world contexts.
• Ontological Status: Determining whether these verbalizations approximate true “beliefs” or merely convenient heuristics.
• Dynamic Contexts: Extending verbalized assumption frameworks to multi-turn dialogs and evolving tasks.
• Human–AI Interaction: Addressing expectation mismatches, as users often anticipate more factual, objective support from AI than human interlocutors—while LLMs default to human-human conversational norms (Cheng et al., 3 Apr 2026).
• Scalability and Efficiency: Automating large-scale assumption mining and filtering in corpora while reserving expensive interventions for salient cases.

Verbalized assumptions thus constitute a new cornerstone for rigorous, transparent, and auditable AI and ML practice, enabling both introspective diagnostics and fine-grained control across diverse deployment settings.