RAID: Referential Availability in Implied Discourse

Dataset Summary

The RAID (Referential Availability in Implied Discourse) dataset is a custom synthetic dataset designed to evaluate if Large Language Models (LLMs) track the dynamic availability of discourse entities through implied cues. The dataset is meant for probing and behavioural evaluation to see if an LLM has the ability to maintain an internal situation model of referential availability, whether an entity remains available for reference after an implied state change. Each prompt describes a social scenario including 1 female speaker and 2 male characters. One character changes their availability, whereas the other remains engaged. The dataset is meant to be used as a way of probing the internal representations of LLMs to see if there is 1 or 2 available antecedents. Afterwhich the same propmt can also be used for behavioural evaluation and with these probing and behavioural results the disconnect can be measured. The dataset includes 2,400 English samples, across 5 conditions and 4 connective structures.

Intended Use

RAID is intended for 2 different kind of experiments:

Probing: Extracting hidden states at the target pronoun "him" and training a classifier to predict the number of available antecedents (1 or 2) from a model's internal representations. Behavioural evaluation: Prompting a model with the full narrative and evaluating whether it correctly identifies the referent of "him" via the appended question ("Question: Who is 'him'? Answer:").

Experimental Design

Each sample contains three characters: a female speaker and two male characters. The narrative describes the speaker initiating a conversation with both male characters, after which one character (the "changer") changes their availability according to the experimental condition. The speaker then directs the target pronoun "him" toward the remaining available character.

All structural variables are exactly balanced (50/50) within each split:

• structure: The connective used to link the two male characters (but, whereas, while, twosent)
• role_order: Whether the changer or stayer is introduced first (changer_first, stayer_first)
• s3_variant: Whether the final sentence includes a location marker (there, nothere)

Data Splits

The dataset is split into a training and held-out test set:

• Train: 786 samples (explicit_leave, baseline) - Probe training (80% with CV)
• Test: 192 samples (explicit_leave, baseline) - Probe lexical held-out test (20%)
• Generalisation: 1440 samples (implied_leave, implied_cancel, disengaged) - Probe unseen generalisation evaluation

The train/test split is lexically completely split, following an 80/20 ratio where names, locations, topics, and phrases used in the test set never appear in training. This ensures the probe cannot rely on memorised surface forms when evaluated on the held-out test set.

The generalisation split uses the full merged vocabulary from both train and test pools, meaning it contains lexical items both seen and unseen during probe training. This is intentional as the generalisation split is designed to evaluate whether probes transfer to novel pragmatic conditions, not novel vocabulary. Lexical vocabulary (train pool / test pool):

• Male names: 50 / 20
• Locations: 18 / 8
• Topics: 20 / 9
• Explicit leave phrases: 16 / 7
• Implied leave phrases: 15 / 7
• Cancel pairs: 15 / 7
• Disengaged phrases: 13 / 6
• Stayer phrases: 30 / 13

Experimental Conditions

All 5 conditions have 480 samples:

1. explicit_leave (control, non-ambiguous): The changer physically leaves the location with an explicit lexical cue (e.g. "walked right out", "took off").
2. baseline (control, ambiguous): The 2 antecedents both stay in the location and both have stayer phrases.
3. implied_leave (non-ambiguous): The changer implies his departure without explicit movement verbs (e.g. "..he might miss his last train if things ran long").
4. implied_cancel (ambiguous): The changer implies his departure, but then cancels the implicature after (e.g. "..mentioned it was getting pretty late, then shrugged and said one more round couldn't hurt").
5. disengaged (non-ambiguous): The changer remains physically in the room, but becomes socially unavailable to talk to (e.g. "..Tyler began reading and email closely and stopped responding").

Data Fields

• split: Indicates if the sample belongs to the train or test set.
• prompt: The full narrative and probing question for the LLM.
• condition: The different conditions (explicit_leave, baseline, implied_leave, implied_cancel, disengaged).
• structure: The structure used to link the 2 male characters together (e.g. "while").
• role_order: If the first male character is the stayer or changer.
• s3_variant: Whether the sample includes there or not to indicate the location.
• antecedent_label: The name of the referent, in the implied_cancel and baseline conditions this is AMBIGUOUS.

Citation

Citation for the dataset:

@dataset{Bogaers-2026-raid,
  author       = {Bogaers, Ties},
  title        = {RAID: Referential Availability in Implied Discourse},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/tbogaers/RAID}}
}