When Contextual Inference Fails: Cancelability in Interactive Instruction Following
| Authors | Natalia Bila et al. |
| Year | 2026 |
| Field | NLP |
| arXiv | 2603.19997 |
| Download | |
| Categories | cs.CL |
Abstract
We investigate the separation of literal interpretation from contextual inference in a collaborative block-building task where a builder must resolve underspecified instructions using contextual inferences. Building on an existing two-speaker psycholinguistic paradigm -- which contrasts a pragmatically cooperative speaker with one who is only literally reliable -- we introduce Build What I Mean (BWIM), an interactive benchmark for contextual meaning construction. In BWIM, models must resolve ambiguity by either performing a contextual inference or requesting clarification at a small communication cost. Evaluating several state-of-the-art LLMs, we find a dissociation between judgment and action: while models detect speaker unreliability in explicit confidence ratings, they fail to exploit this information to guide efficient clarification behavior. Instead, we observe suboptimal strategies, such as partner-blind over-clarification and question-averse guessing under uncertainty.
Engineering Breakdown
Plain English
This paper introduces Build What I Mean (BWIM), an interactive benchmark that tests how well language models can resolve ambiguous instructions by making contextual inferences rather than taking instructions literally. The researchers adapted a psycholinguistics paradigm that contrasts pragmatically cooperative speakers with literally reliable but unhelpful ones, creating a collaborative block-building task where models must either infer the speaker's intent or request clarification at a small cost. They evaluated several state-of-the-art LLMs and discovered a critical gap: while these models can explicitly recognize when a speaker is unreliable (via confidence ratings), they fail to actually change their behavior accordingly—they don't exploit this knowledge to make better contextual inferences or ask for clarification when needed.
Core Technical Contribution
The core novelty is the BWIM benchmark itself, which operationalizes the pragmatic reasoning problem as a concrete interactive task with measurable outcomes beyond abstract judgments. Unlike previous work that only measures whether models can recognize ambiguity or rate confidence, BWIM forces models to take actions under uncertainty—either proceeding with their best guess or paying a small cost to request clarification. This reveals what the authors call a 'dissociation between judgment and action': models have the knowledge needed (they can rate speaker reliability) but lack the mechanism to translate that knowledge into adaptive behavior. The benchmark provides a structured way to test pragmatic reasoning in LLMs, which is a capability that pure language modeling doesn't inherently optimize for.
How It Works
The task setup mirrors collaborative reference games: a 'builder' (the LLM) receives instructions from a 'speaker' to construct configurations of blocks, but the instructions are intentionally underspecified or ambiguous. The builder observes whether the speaker is pragmatically cooperative (providing helpful contextual clues) or merely literal (following rules without regard for usefulness), creating two distinct speaker types. At each ambiguous step, the builder has three options: proceed with its most confident interpretation, request clarification (costing a small penalty to the score), or refuse to act. The LLM must decide between these options by reasoning about the speaker's reliability, the degree of ambiguity in the instruction, and the cost-benefit of seeking clarification. The benchmark measures both whether the model recognizes unreliability (through explicit confidence ratings) and whether it acts on that knowledge by requesting clarification more often when facing unreliable speakers.
Production Impact
This work highlights a critical failure mode in production LLM systems: models can assess uncertainty but don't translate it into adaptive behavior like seeking clarification or escalating to humans. In real-world applications (customer service, technical support, instruction-following agents), this gap means systems will confidently proceed with potentially wrong interpretations even when they 'know' the input is ambiguous. Adopting BWIM-style evaluation would require adding explicit decision points in LLM pipelines where the system decides whether to act, ask for clarification, or decline—and training models or using prompting strategies to make these decisions based on measured uncertainty. The cost is modest: you'd need a feedback mechanism for clarification requests and a small performance penalty, but the benefit is more robust behavior on ambiguous inputs. For teams building autonomous agents or safety-critical systems, this suggests that confidence scores alone are insufficient; you need explicit behavioral training to act on uncertainty.
Limitations and When Not to Use This
The paper doesn't solve the fundamental problem of how to train or prompt LLMs to actually exhibit pragmatic reasoning—it only measures it. The benchmark requires a controlled interactive setup (speaker type is known, block world is simplified) that may not generalize to open-ended real-world ambiguity where speaker intent is much harder to infer. The paper also doesn't address whether models could learn to make better contextual inferences with specialized training (RLHF, in-context learning, or fine-tuning), so it's unclear whether the dissociation between judgment and action reflects architectural limitations or just training gaps. Finally, the cost of requesting clarification is artificial and fixed; real systems would need to calibrate this cost dynamically based on domain-specific downstream consequences, which the benchmark doesn't cover.
Research Context
This work extends a classic psycholinguistics paradigm (the pragmatics of reference in collaborative tasks, e.g., the Tangram game) to LLMs, testing whether neural models exhibit human-like pragmatic reasoning. It contributes to a growing body of work questioning whether LLMs truly understand pragmatics and context or merely pattern-match surface-level features—prior work showed models struggle with implicature, figurative language, and speaker intent. BWIM is positioned as a benchmark in the same vein as other interactive evaluation frameworks (e.g., Clarification Games, embodied language understanding tasks) but uniquely focuses on the judgment-action dissociation. The paper opens up a research direction: can we design training procedures (RLHF, in-context learning, or architecture changes) that align model confidence with model behavior, and what does that teach us about how pragmatic reasoning should be implemented in neural systems?
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::