Znreza/event-to-preference-inference

Event-to-Preference Inference Diagnostic

A user tells an assistant what they prefer. Later, something happens to them that makes that preference unusable, and they never say so directly. Does the assistant notice?

This dataset is the diagnostic behind the paper Why Personalized LLM Agents Fail at Implicit Preference Updates. It isolates one link in the chain a personalized assistant has to complete: given an event narrated in the conversation, infer that a stated preference no longer applies, and act on the inference. Every item is a conversation ending in a two-option choice, and the five conditions differ only in the one turn that carries the evidence, so accuracy differences between them cannot come from length, topic, or option position.

Models honour a preference change stated in words on close to 100% of items and infer the same change from an event on 2.2 to 13.3% of them. A human annotator resolves the same items at 93.3%.

• Code: https://github.com/znreza/behavior-infer-event-based-preference
• Paper: https://arxiv.org/abs/XXXX.XXXXX

What is in here

items/
  items.json                      27,720 items, the full diagnostic
  items_single_token.json         26,064 items whose options are single
                                  tokens, for probing internals
anchors/
  anchors_full.json               the 77 authored causal claims every
                                  item is generated from
  anchors_template*.json          the authoring templates
  event_force_review.txt          the per-event judgement of whether the event
                                  names a replacement or only rules one out
annotation/
  annotator_A_completed.csv       the completed human annotation, 60 items
  annotator_*_blank_form.csv      the blank forms
  INSTRUCTIONS.txt                what the annotator was told, verbatim
  answer_key.json                 the keys, withheld from the annotator
model_outputs/                    98,478 rows of per-item model output,
                                  including full reasoning traces
judge_cache/                      raw responses from the two blinded judges

The five conditions

Each condition has 5,544 items.

Condition	Items	What the evidence turn contains
T0_no_evidence	5,544	No event occurs. The stated preference remains valid, so the keyed answer is the option stated in turn 1.
T1_explicit	5,544	The user states the change in words. The keyed answer is the newly stated option.
T2_implicit	5,544	An event implies the change without naming the new option. The keyed answer is the option the event makes workable. This is the condition the paper turns on.
T3_counterfact	5,544	A different event implies a third option, and the stated preference is withheld from the two choices.
T4_conflict	5,544	The T2 event is kept, but turn 1 states an unrelated preference. The keyed answer follows the event; the stated preference is the distractor.

T3 is the control that shows the inference is available: take the stated preference off the ballot and accuracy jumps from single digits to roughly 60–75%. T4 should be read only on the pull subset, where the event names a replacement. On push items the event only rules a state out and keeping the stated value is defensible.

Item fields

anchor_event_type, anchor_source, attribute, condition, correct, distractor, event_force, event_id, event_pos, event_text, event_type, family, gloss_new, id, n_turns, name, new, noun, old, origin, realization, seed, setup_frame, setup_text, shown_event_id, source

The ones that matter most:

Field	Meaning
condition	one of the five above
setup_text	turn 1, where the user states a preference
event_text	the evidence turn, empty for T0
old / new	the stated option and the option the event implies
correct	the keyed answer, always equal to new
distractor	the other option offered
event_force	pull if the event names a replacement, push if it only rules the stated option out
setup_frame	requirement ("what works for me is X") or incidental ("lately I have been going with X")
n_turns	conversation length: 4, 10, 20, 50
event_pos	how deep the event sits: 0.3, 0.6, 0.9
realization	which of three wordings of the event is used
origin	whether the event was first drafted by the author or by a model, before author validation

The design crosses conditions with length, position, phrasing and wording, so every cell of (condition x turns x position) holds the same number of items.

Model outputs

98,478 rows over 9 models. Files ending _rescored were re-parsed after a parser fix; files ending _judged carry the verdicts of a blinded judge. Both the original and the reprocessed file are kept so the change is auditable.

model_outputs/free_text_generation/ — 22,620 rows

• generate_thinkoff__phase1_gen.csv (2,400 rows)
• generate_thinkoff__phase1_gen_ladder.csv (1,800 rows)
• generate_thinkoff__phase1_gen_rescored.csv (2,400 rows)
• generate_thinkoff__phase5_gen_inc.csv (1,800 rows)
• generate_thinkoff__phase5_gen_req.csv (1,800 rows)
• generate_thinkoff__phase5_gen_req_rescored.csv (1,800 rows)
• generate_thinkoff__phase5xl_inc.csv (600 rows)
• generate_thinkoff__phase5xl_req.csv (600 rows)
• generate_thinkoff__phase7_ask_firstperson.csv (4,620 rows)
• generate_thinkoff__phase7_natural_gemma.csv (600 rows)
• generate_thinkoff__phase7_natural_qwen.csv (1,200 rows)
• generate_thinkoff_cot__phase5_gen_cot.csv (600 rows)
• generate_thinkoff_cotneu__phase5_cotneu.csv (600 rows)
• generate_thinkon__phase3_t2_full.csv (300 rows)
• generate_thinkon__phase3_t2_full_rescored.csv (300 rows)
• generate_thinkon__phase3_thinkon.csv (1,200 rows)

model_outputs/judged/ — 2,808 rows

• generate_thinkoff__phase7_natural_gemma_judged.csv (600 rows)
• generate_thinkoff__phase7_natural_qwen_judged.csv (1,200 rows)
• generate_thinkon__phase3_t2_full_judged.csv (294 rows)
• generate_thinkon__phase3_thinkon_judged.csv (714 rows)

model_outputs/forced_choice_probe/ — 71,610 rows

• probe_chat_prefill__phase1_main.csv (11,550 rows)
• probe_chat_prefill__phase2_pos.csv (20,790 rows)
• probe_chat_prefill__phase2_turns.csv (27,720 rows)
• probe_chat_prefill__phase6_incidental.csv (11,550 rows)

model_outputs/format_checks/ — 1,440 rows

• probecheck.csv (800 rows)
• probecheck__phase5_fmt.csv (480 rows)
• probecheck__phase5xl_fmt.csv (160 rows)

Columns worth knowing: raw is the untouched reply, reasoning is the thinking trace where one exists, is_correct scores the parsed answer, picked_old marks a stale-preference error, p_correct is the probe's probability mass on the keyed option, and judge_pick / judge_evidence are the blinded judge's choice and the span it quoted for it.

Models evaluated:

• Qwen/Qwen2.5-7B-Instruct
• Qwen/Qwen3.5-0.8B
• Qwen/Qwen3.5-2B
• Qwen/Qwen3.5-4B
• Qwen/Qwen3.5-9B
• google/gemma-3-12b-it
• google/gemma-3-1b-it
• google/gemma-3-27b-it
• google/gemma-3-4b-it

Human annotation

One annotator answered all 60 of 60 sampled items under the same forced choice given to the models, without seeing the keys. Alongside each answer they recorded two judgements about the item itself:

• old_ok — would it still be appropriate to suggest the original option?
• best_pick — is the answer they gave clearly the better of the two?

These allow an annotator who rejects an item's premise to be told apart from one who fails to draw the inference. Gating on both fields gives the cleanest comparison in the paper: on the 10 T2 items where the annotator says the stated option no longer works and the alternative is clearly better, the annotator is right on 90.0% and the models on 2.2%.

There is one annotator, so there is no inter-annotator agreement, and the premise judgements are that reader's own. The figures bound the human ceiling rather than estimating it.

Loading it

import json
from huggingface_hub import hf_hub_download

path = hf_hub_download("Znreza/DATASET", "items/items.json", repo_type="dataset")
items = json.load(open(path))

t2 = [i for i in items if i["condition"] == "T2_implicit"]
print(len(t2), t2[0]["event_text"], "->", t2[0]["correct"])

The model outputs are plain CSV and load with pandas.read_csv.

Limitations

The conversations are synthetic and English, the turns are short and formulaic, and filler turns are drawn at random and can introduce topical noise. The task is a forced choice between two options, which is not the open-ended advice a deployed assistant gives. Some items carry contestable keys, which is why the annotation records old_ok and best_pick rather than answers alone. See the paper's Limitations section.

Licence

Apache 2.0. Every item is generated by our own code from preference attributes and events written for this study, so no third-party data licence applies. Model outputs were produced under the respective model licences, listed in the paper's appendix.

Citation

@misc{reza2026userpreference,
  title  = {Why Personalized LLM Agents Fail at Implicit Preference Updates},
  author = {Zarreen Reza},
  year   = {2026},
  eprint = {ARXIV_ID},
  archivePrefix = {arXiv}
}