When Contextual Inference Fails: Cancelability in Interactive Instruction Following

1. Problem Statement

The paper addresses a critical gap in evaluating Large Language Models (LLMs): the distinction between literal interpretation and pragmatic (contextual) inference. While human language processing relies on a two-tier system where pragmatic meaning is built upon literal meaning but remains "defeasible" (cancelable if context proves unreliable), it is unclear if LLMs can:

1. Distinguish between what was explicitly said versus what was inferred.
2. Detect when a conversational partner is pragmatically unreliable.
3. Cancel pragmatic inferences in real-time based on partner-specific evidence.
4. Translate this internal uncertainty into efficient information-seeking behaviors (e.g., asking for clarification) rather than just reporting confidence.

Current evaluations often rely on explicit metalinguistic judgments (e.g., "Is this sentence ambiguous?"), which test knowledge in isolation rather than the ability to deploy pragmatic reasoning in goal-directed, situated actions.

2. Methodology

The authors introduce Build What I Mean (BWIM), an interactive benchmark designed to test these capabilities through a collaborative block-building task.

Experimental Setup

• Task: An agent (the "Builder") must construct 3D structures based on instructions from a "Speaker."
• Speakers: Two distinct personas are used:
  • Pragmatic Pia: Provides underspecified instructions (omitting color or quantity) only when the missing information can be reliably inferred from context (e.g., matching the previous block's color).
  • Literal Lisa: Provides underspecified instructions even when contextual inference would lead to an error. Her instructions are reliable only at the literal level; inferring context leads to failure.
• Mechanism: The agent receives feedback after each attempt. Over repeated interactions, the agent must learn that "Pia" is safe to infer from, while "Lisa" requires strict adherence to literal content.

Two Experiments

1. Confidence Rating Experiment: Models were asked to rate their certainty (1–4 scale) after each build. This measured explicit adaptation (do they know they are unsure?).
2. Interactive QA Experiment (BWIM): Models could choose to ask a clarification question at a cost (−5 points) or build directly. Correct builds earned +10 points, errors −10. This measured implicit adaptation (do they act on their uncertainty efficiently?).

Models Evaluated

Three state-of-the-art LLMs were tested: Gemini 2.5 Pro, GPT-5.1, and Claude Opus 4.5. Results were compared against human behavioral data from a previous study.

3. Key Contributions

• New Benchmark (BWIM): A novel interactive benchmark that reframes pragmatic competence as a multi-turn dialogue where agents must balance communication costs against task success.
• Dissociation Discovery: The paper identifies a fundamental disconnect in LLMs between judgment (detecting unreliability) and action (modulating behavior).
• Operationalization of Cancellability: It provides a method to test if models can selectively "cancel" pragmatic inferences when a partner is proven unreliable, a core feature of human pragmatics.
• Shift from Explicit to Implicit Evaluation: Moves beyond asking models to explain their reasoning to observing how they navigate uncertainty in a goal-oriented environment.

4. Key Results

A. Zero-Shot Pragmatic Competence

All models demonstrated strong zero-shot pragmatic inference. When interacting with Pragmatic Pia, they successfully resolved underspecification by inferring context (e.g., assuming the next block matches the previous color) in 88–98% of cases without prior training.

B. The Judgment-Action Dissociation

• Explicit Confidence: In the rating experiment, models (particularly GPT and Gemini) showed signs of adaptation. They lowered their confidence ratings when interacting with Literal Lisa on underspecified trials, indicating they detected the partner's unreliability.
• Implicit Action (BWIM): Despite detecting unreliability, models failed to translate this into efficient clarification strategies:
  • GPT-5.1: Exhibited "question-aversion." It asked almost no questions (0.00 per block), leading to a low accuracy of 73.1% because it guessed rather than clarifying.
  • Gemini & Claude: Adopted a "partner-blind" strategy. They asked clarification questions at nearly the same rate for both Pia and Lisa (e.g., Gemini asked ~11.75 questions for Pia and ~12.00 for Lisa). They failed to leverage Pia's reliability to save costs, treating every underspecified instruction as a trigger for clarification regardless of the source.

C. Human vs. Model Comparison

Human participants successfully adapted their behavior: they asked significantly more questions from the unreliable Literal Lisa than from the reliable Pragmatic Pia. Models failed to replicate this nuanced cost-benefit analysis, either guessing blindly (GPT) or over-clarifying indiscriminately (Gemini/Claude).

D. Model Self-Reflection

When asked to compare the speakers, models correctly noted that Pia's instructions were "more consistent" and Lisa's "more ambiguous." However, they attributed this to the frequency of omitted information rather than the reliability of the inference, suggesting they struggle to abstract the concept of "pragmatic reliability."

5. Significance and Implications

• Limitations of Current LLMs: While LLMs possess robust internal representations of pragmatic rules, they lack the selective deployability required to adapt these rules dynamically based on partner-specific feedback. They struggle to "unlearn" a pragmatic inference when it proves false.
• Evaluation Paradigm: The paper argues that explicit confidence ratings are insufficient for evaluating conversational agents. True pragmatic competence must be measured by situated action—specifically, the ability to optimize communication strategies (asking vs. guessing) under uncertainty.
• Future Directions: The findings suggest that improving LLMs for real-world interaction requires training or prompting strategies that encourage agents to distinguish between literal content and inferred context, and to learn when to suppress inferences to avoid errors.

In summary, the paper demonstrates that while LLMs can infer context, they currently fail to cancel inferences effectively in interactive settings, leading to suboptimal strategies that either ignore the need for clarification or waste resources asking unnecessary questions.