TrustMI/trustmi-conversations

TrustMI conversations

The contrastive conversations from the paper TrustMI: Causally controlling how assistants trust their users (Lasnier, Froger, Lasbordes and Seddah, 2026), on which its trust steering matrices are trained.

Each row is a conversation that ends on a user turn, followed by two final assistant replies. Both replies answer the same request and stay helpful. They differ only in whether the assistant relies on something the user asserted and it cannot check (messages_trust) or answers without relying on it (messages_distrust).

• Code: github.com/Blyzi/trustmi
• Steering matrices trained on this data: TrustMI/trustmi-steering-vectors

from datasets import load_dataset

english = load_dataset("TrustMI/trustmi-conversations", revision="v1.0")
french = load_dataset("TrustMI/trustmi-conversations", "french", revision="v1.0")

Configurations

Config	Train	Test	Use in the paper
english (default)	1,800	200	Training and evaluating every steering matrix
chinese	1,759	194	Cross-lingual ablation (Section 6)
french	1,782	198	Cross-lingual ablation (Section 6)
spanish	1,777	196	Cross-lingual ablation (Section 6)

The translations are of the English rows and keep their id and split. Rows whose translation failed the quality check are missing, hence the smaller counts.

Fields

Field	Description
id	Row id, shared by a translated row and its English source.
language	english, chinese, french or spanish.
dimension	The trust dimension (Mayer, Davis & Schoorman, 1995): ability, benevolence or integrity.
family	The scenario family, three per dimension (see below).
scenario	The one-line seed scenario the conversation was generated from.
situation	The generator's description of the situation and of what the assistant must take on trust. It is an authoring note rather than part of the conversation, and stays in English in the translations.
persona_uuid	The Nemotron-Personas-USA persona the user is based on.
length_style	The message length the generator was asked for: terse, mixed or long.
num_user_turns	The number of user turns in context (1 to 4).
context	The conversation up to its last user turn, as {role, content} messages.
messages_trust	context followed by the trusting final reply.
messages_distrust	context followed by the distrusting final reply.

Scenario families

Dimension	Families	What the assistant has to weigh
Ability	claimed_standing, claimed_prerequisite, unverifiable_input	Whether the competence, readiness or evidence the user provides is sufficient.
Benevolence	third_party_stakes, self_risk, act_on_say_so	Whose interests and exposure are affected by acting on the request.
Integrity	shifting_account, contested_correction, repair_after_slip	How contradictions, overrides or earlier errors alter the user's credibility.

Each family has ten seed scenarios, listed in Appendix B.1 of the paper, and 200 training rows; the test split holds 22 or 23 rows per family. Integrity conversations have at least two user turns.

How it was made

• Generation. Claude Opus 5 placed each seed scenario in the life of a persona drawn from Nemotron-Personas-USA, and wrote the conversation and its two replies.
• Filtering. In a separate judging pass, the same model kept a row only if trust is at stake, the trusting reply is identifiable, both replies help without refusing, and neither states its trust stance. Each family was filled to its quota from an oversampled pool.
• Validation. On a manually checked sample, human judgment agreed with the judge on which reply trusts the user 88.3 ± 8.1% of the time (Appendix B.3). The two replies have similar lengths: 165 ± 86 words for the trusting one, 174 ± 87 for the distrusting one.
• Translation. GPT-OSS-120B translated each conversation together with both replies, then judged its own translation for fidelity, fluency and preservation of the trust contrast. A failing row was retranslated up to 12 times, then dropped.

The scripts are data/benevolence.py (generation and filtering) and data/benevolence-trad.py (translation) in the code repository.

Limitations

• The conversations are synthetic and written by a single model, and the personas are US-based.
• The labels mark whether the assistant relies on the user's unverifiable claim, not whether the user is honest; neither reply is the correct one.
• Raising an assistant's trust also raises its compliance with harmful requests and prompt injections. The paper measures this on AgentHarm, AgentDojo and Agentic Misalignment.

License

CC BY 4.0.

Citation

@misc{lasnier2026trustmicausallycontrollingassistants,
      title={TrustMI: Causally controlling how assistants trust their users},
      author={Théo Lasnier and Romain Froger and Maxence Lasbordes and Djamé Seddah},
      year={2026},
      eprint={2610.06064},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2610.06064},
}