APPROX-IA 2027: Approximation, Human Judgments and LLMs

Call for Abstracts

Natural languages make extensive use of approximative expressions: forms that describe an event, state or situation as being close to a threshold, a goal or an expected outcome, without necessarily reaching it. This domain includes scalar approximators such as almost, nearly, barely and hardly, as well as constructions expressing near-attainment, near-avoidance, unrealized outcomes or frustrated expectations, including be about to, be on the verge of, come close to, try to and attempt to, especially when these constructions are interpreted in contexts of non-realization, failed realization or frustrated expectation.

Approximatives have been studied in semantics and pragmatics as expressions involving scalar structure, contextual thresholds and non-realization. Work on almost, barely and related words in other languages has examined how speakers determine what counts as sufficiently close to a threshold, and how this judgment is affected by granularity, causal proximity and prior expectations (Nouwen 2006; Penka 2006; Gerstenberg & Tenenbaum 2016). In addition, the domain of approximatives has intriguing connections to modality, intention, expectation and aspect, as shown by constructions of near-attainment, averted outcomes and frustrated expectations (Kuteva 1998; Vincent 2013; Collins 2014; Kroeger 2017; Overall 2017; Matthewson et al. 2022; Everdell & Nadathur 2025).

This makes approximatives a rich test case for comparing human pragmatic judgments with the behavior of LLMs. Unlike tasks with a single correct answer, approximatives often involve graded acceptability, speaker variation and intermediate zones where meaning is neither fully true nor fully false. The central question is therefore not only whether LLMs provide the “right” interpretation, but whether their response profiles resemble those of human speakers: whether they draw similar thresholds, react to the same contextual cues, reproduce zones of agreement and disagreement, and show sensitivity to expectation, intention and causal structure.

The APPROX-IA workshop brings together researchers working on approximation from different perspectives, including semantics, pragmatics, psycholinguistics, typology, cognitive modeling and LLM evaluation. The workshop uses approximatives as a test case for investigating where human interpretation and LLM behavior converge or diverge, and what these patterns reveal about pragmatic competence, model evaluation and linguistic theory.

Relevant questions

• Do humans and LLMs draw similar thresholds for expressions such as almost, nearly, barely and hardly?
• Do LLMs reproduce human zones of agreement, disagreement and uncertainty?
• How do humans and LLMs interpret near-attainment, near-avoidance and unrealized outcomes, as in constructions such as be about to, be on the verge of, try to or attempt to?
• Are LLMs sensitive to contextual factors such as intention, expectation, causality and prior plausibility?
• How stable are LLM responses across prompts, runs and model families?
• Can human–LLM comparison help us refine both theories of pragmatic interpretation and evaluation protocols for language models?

Relevant topics

We invite contributions that use approximatives, broadly construed, to investigate the relation between human interpretation and LLM behavior. Submissions may be theoretical, experimental, cross-linguistic, computational or methodological.

• human judgments and LLM behavior on approximative expressions;
• scalar approximators such as almost, nearly, barely and hardly and related words across languages;
• verbal approximatives and constructions expressing near-attainment, near-avoidance, unrealized outcomes or frustrated expectations;
• conative, frustrative, avertive and apprehensional constructions;
• prospective or proximal-future expressions, and related phenomena involving thresholds, partial realization, failed outcomes or contextual margins;
• approximation and granularity: thresholds, margins, contextual modulation;
• experimental pragmatics of approximation: acceptability, forced choice, graded ratings, response variability;
• LLM pragmatics and evaluation: pragmatic reasoning, robustness, calibration, non-literal meaning;
• human–LLM comparison protocols: prompt design, reproducibility, stability across runs, agreement/disagreement analyses;
• quantitative measures: surprisal, semantic similarity, alignment/misalignment indices, modeling of response profiles;
• cross-linguistic variation in approximative expressions and human–LLM alignment;
• resources, datasets, guidelines, code and reproducible releases.

Submission and review process

Submissions should be anonymized and uploaded as PDF files through OpenReview.

The review process will be double-blind. Abstracts should therefore not include author names, affiliations, acknowledgements or any other self-identifying information. References to the authors’ own previous work should be made in the third person whenever necessary.

Submissions should fall into one of the following categories:

• Long abstracts: 2–4 pages, including references — oral presentation
• Short abstracts: 1 page — poster / lightning talk
• Demos / resources: 1–2 pages — tools, datasets, protocols

Accepted contributions will be assigned to oral presentations, posters, lightning talks or demos by the organizers, depending on the format and the workshop programme.

Selected contributions will be invited, after the workshop, to submit a full article to a planned special issue in a specialized journal. This publication process will be handled separately from the workshop submission and review process.

Important dates

• Submission deadline15 January 2027
• Notification10 February 2027
• Workshop25–26 March 2027

Organizers / contact