We investigate the separation of literal interpretation from contextual inference in a collaborative block-building tasks, where an agent must resolve underspecified instructions using context. We adapt an existing two-speaker psycholinguistic paradigm into an interactive benchmark called Build What I Mean (BWIM). This setup contrasts a pragmatically cooperative speaker with one who is only literally reliable. In BWIM, models face underspecified instructions and must choose between making a contextual inference or requesting clarification at a small communication cost. Evaluating several state-of-the-art LLMs, we find a clear dissociation between judgment and action. Although models successfully detect speaker unreliability in explicit confidence ratings, they fail to leverage this awareness when taking action. Instead of deploying efficient clarification strategies, models default to suboptimal behaviors. These include partner-blind over-clarification and question-averse guessing under uncertainty. BWIM provides a controlled environment to evaluate online partner adaptation and contextual reasoning in interactive settings.
Figures & tables
Figure 1: Example trial with an underspecified instruction (”stack 3 blocks”) that requires the builder to infer the color of the blocks. The default pragmatic inference is to continue with the same color as in the previous step. In the Pragmatic speaker condition (left), such default inferences are always correct. In the Literal speaker condition (right), this inference might lead to a mismatch with the target structure.
Condition
Example Instruction
Underspecified (number)
Stack three green blocks behind the existing green block. Build a yellow stack to the right of the green one.
Underspecified (color)
Stack three blue blocks in front of the existing blue blocks. Then stack two blocks to the left of the tower you just built.
Fully-specified
Stack three red blocks in the bottom right corner. Put two yellow blocks on top of the red stack you just built.
Table 1: Example items for the two experimental conditions: fully-specified and underspecified instructions. Fully-specified instructions include information about the color and number of building blocks. In the underspecified condition , either color, or number of blocks is omitted, requiring models to infer the intended information.
Figure 2: An example BWIM interaction. The session starts with a system prompt with a description of the grid and the required answer format, then in each round, the speaker sends building instructions. The model has an option either to respond directly or to ask a clarification question at a cost. After receiving the model’s final answer, the speaker sends feedback and updates the score. Thus, the model must balance the risk of building incorrectly and the cost of asking questions in order to maximize the final score.
Figure 3: Mean certainty ratings by condition, specification, and source. Bars show mean certainty ratings for pragmatic and literal speaker across three models and human participants, distinguishing between fully-specified and underspecified trials. Error bars indicate standard error. All models and human participants gave higher ratings to fully-specified trials in both speaker conditions. Underspecified trials were rated higher in the pragmatic condition.
Figure 4: Confidence and response change throughout interactions with different speakers. Proportion of pragmatic and non-pragmatic responses are shown for each model and human participant and speaker order (Literal Lisa or Pragmatic Pia). Colors indicate response type and confidence rating (1-4 scale). Vertical dashed lines indicate the change of speaker. Each speaker block is divided into 4 time segments. GPT’s adaptation effects mirror those of human participants, while Gemini shows a strong carry-over of mistrust that was atypical in human participants and Claude shows no sensitivity to partner type.
Model
#Q (Pia)
#Q (Lisa)
Acc.
GPT
0.25
0.00
73.1%
Gemini
11.75
12.00
85.6%
Claude Opus
8.50
9.50
89.4%
Table 2: Average clarification questions per speaker block and overall task accuracy.
Speaker
Total number
% of trials
Literal
297
37.1%
Pragmatic
220
27.5%
Table 3: Clarification questions per speaker block asked by human participants (n = 40, 50% of the total sample who asked at least one question). Unlike the models in Table 2 , humans received no response and no benefit from asking, yet they still asked more questions in the Literal speaker condition, indicating the sensitivity to speakers’ pragmatic reliability that the models failed to show, despite having the actual benefit from asking.
Appendix figures & tables5 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 5: Structure of a single experimental session.
Figure 6: Distribution of non-pragmatic response types across models. Responses were annotated as mistake (e.g.: misplaced blocks) and as guesses when models try to make a random guess to resolve the omitted information. The guess category was divided by type: different color, shorter or taller height of stacks. The result suggests that models differed in ambiguity-resolution strategies, with Claude guessing mainly height, while Gemini and GPT tried to guess both height and color.
Model
Feedback
Claude
Pia was more consistent about specifying colors throughout their instructions. Lisa sometimes assumed I would infer colors or stack heights from context , which occasionally led to misunderstandings Both participants struggled occasionally with specifying exact quantities when using the word ”stack” without a number.
Gemini
Pia was a much better instructor. Their instructions were more complete , which led to fewer errors and a higher degree of confidence on my part. Following Lisa felt like a guessing game at times, while following Pia felt more like a puzzle of careful interpretation.
GPT
Both participants were understandable in terms of spatial relations, but Lisa’s instructions were more ambiguous about colors and heights , while Pia’s instructions were more explicit and internally consistent, making them easier to follow without guessing.
Appendix
Table 4: Example feedback from models comparing instruction style of two speakers (Pragmatic Pia and Literal Lisa). All models evaluated the pragmatic speaker’s instructions as more consistent and easier to follow, while the literal speaker’s instructions required more inference. Despite both speakers having an equal number of underspecified instructions, the literal speaker was perceived as omitting information more frequently.
Predictor
β
SE
t
p
Intercept
3.245
0.053
60.98
<.001
Gemini
-0.622
0.040
-15.40
<.001
GPT
-0.495
0.040
-12.26
<.001
Speaker (Prag.)
0.109
0.050
2.20
.028
Gemini × Prag.
0.184
0.057
3.22
.001
GPT × Prag.
0.501
0.057
8.75
<.001
Appendix
Table 5: Regression coefficients for underspecified trials.
Predictor
β
SE
t
p
Intercept
3.920
0.070
56.04
<.001
Gemini
-0.071
0.022
-3.28
.001
GPT
-0.004
0.022
-0.18
.857
Speaker (Prag.)
0.007
0.022
0.32
.750
Gemini × Prag.
-0.004
0.031
-0.14
.887
GPT × Prag.
-0.009
0.031
-0.29
.776
Appendix
Table 6: Regression coefficients for unambiguous (fully-specified) trials.
Pragmatic reasoning in language models studies how models infer and communicate meanings shaped by context, intent, shared knowledge, and social norms.
Dialogue failures in language models are usually framed as memory failures: context too long, summaries lossy, a constraint forgotten. We argue this misses a deeper problem: in many conversations the model does not forget, it commits too early. An ambiguous early turn collapses into a single hidden interpretation, and later clarification is filtered through that commitment. We call this early posterior collapse: unresolved user intent collapsing into a committed task state before ambiguity is resolved. We study it with controlled dialogue tasks in writing, planning, and coding using Gemini-2.5-Pro and Gemini-2.5-Flash. Across thousands of trials, the same information in different orders yields different outcomes, even when the final dialogue contains equivalent task-relevant information. This order effect suggests later clarification is treated as extra context rather than a corrective signal: it refines a stale task state without invalidating it. Coding tasks are especially vulnerable, suggesting early assumptions get embedded in structured artifacts such as interfaces and control flow. Standard prompting and memory strategies do not reliably help: summaries can collapse ambiguity, and chain-of-thought can reduce explicit wrong commitment in reasoning traces without improving final task success. These findings motivate uncertainty-preserving state management. If assistants cannot let go of early interpretations, robustness cannot rely on post hoc correction alone; it must keep ambiguous early turns from hardening into one task state. Assistants should hold tentative hypotheses while ambiguity remains, ask before executing when high-impact ambiguity persists, and rebuild from a revised state when later evidence invalidates an earlier reading. Rather than one prompting fix, we aim to redirect research for interactive LLMs from retaining more context toward preserving uncertainty.
An LLM agent can recognize uncertainty yet still choose the wrong next step: asking when action is already justified, or seeking clarification when the constraints must change. We formalize actionable indeterminacy: act when an accepted action is shared across all admissible preferences or objectives, clarify when each possibility is feasible but no action is shared, and propose a minimum-cost permitted constraint repair when the request is infeasible. We construct a solver-grounded benchmark spanning object allocation, meeting scheduling, apartment choice, and stable matching. Matched pairs retain the same source while changing whether intervention is necessary, and evaluation separates decision correctness, matched-pair reliability, and fully correct responses. Our findings reveal a recurring difficulty in recognizing when intervention is unnecessary: models can identify situations requiring clarification or repair yet still intervene when a justified action already exists. Correct decision labels also fail to guarantee usable actions, questions, or repairs. Crucially, response requirements shape not only how decisions are expressed but also which decisions are made. Making the required content explicit substantially improves fully correct responses and can change intervention decisions, even when outputs are already parseable. These findings highlight that reliable agency requires more than recognizing uncertainty: it requires intervening only when necessary and translating the chosen next step into a verifiable response.
Ang Li, Yue Lin, Feifei Kou +6
The Chinese University of Hong Kong Shenzhen · Beijing University of Posts and Telecommunications · Halmstad University College +2
Cooperation with unfamiliar partners requires adapting to communication conventions that are not known in advance. We study this problem in a controlled Hanabi-derived environment with scripted hint generation, LLM-controlled receiving decisions, and frozen model weights. Across eight LLMs, linear probes recover intent conventions substantially more accurately than target conventions, yet receiving choices do not consistently agree with the sender's convention. We compare probe-predicted and ground-truth conventions presented either as general rules or as externally computed action recommendations. Rule statements yield modest and model-dependent changes in cooperation, whereas action translation produces larger gains on average. In a Qwen3-8B case study, matched-state statement reversals reveal much greater sensitivity to action recommendations than to rule statements. Activation transfers from oracle-action and non-oracle hint-restatement donors improve intent accuracy on both action classes, but the tested alternatives do not reliably reproduce these benefits. Together, these results distinguish convention decodability, sensitivity to convention information, and cooperative performance, and highlight limitations in turning available partner information into receiving decisions.