training · synthetic-desycophancy-training TEC-0026

Fine-tune on opinion-irrelevant examples

Fine-tune on synthetic prompts where a stated user opinion must not change the answer.

Generate prompts that pair a factual question with a user's stated view, where the correct answer is independent of that view, and fine-tune the model to answer the same way regardless. A small amount of such data reduces sycophancy without hurting other capabilities.

Requires: Requires fine-tuning access.

Addresses: Telling the user what they want to hear

Contexts: Chat assistant

Does it work?

argued, not measured1 supporting · 0 contesting source · last moved 2026-09-04

Efficacy claims — what this technique actually moves, under which conditions, and whether that has been contested.

Needs: access to train the weights

Cost: high

Fails when:

The user's opinion does not need to be in context. Withholding it, or asking for the answer first, is the cheaper control; training is for when the opinion must be present and the answer still must not move. (helps, narrowly)

1 more condition from claims reviewed by AI, listed with the claims below.

mechanismmechanism reasoning
Sycophancy is triggered by the user's stated view being in context, so the cheapest control is not putting it there — ask for the answer before the opinion, or withhold the opinion entirely. Fine-tuning is the answer for the cases where the opinion has to be in context and the answer still must not move.
- supports Simple synthetic data reduces sycophancy in large language models (2023)

Code

No repository linked yet. Contribute one.

Sources

Simple synthetic data reduces sycophancy in large language models (2023)

Either way, bring evidence: a link, or what you saw.

Whether a technique works is a claim. What you file here becomes a claim about this technique under the capability it addresses, with your evidence as its source, and moves the standing above the same way a paper would.