Open Models Almanac
Methodspost-trainingagentic post-training

specific method · filed under post-training

Merlin–Arthur protocol

A Merlin–Arthur protocol is a hallucination-reduction training setup in which the trained model (Arthur) answers questions over partially masked retrieval contexts.1 Merlin and Morgana are masking algorithms that choose two context masks hiding similar amounts of context, optimised so the policy assigns high probability to the gold answer under Merlin's mask and low probability under Morgana's.2,3,4

How this definition was written

Summarised by glm-5.3-flash from 1 source; every sentence is checked against the quote it cites.

  1. The hallucination reduction environments implement a Merlin–Arthur protocol (Deiseroth et al. 2026) on various RAG training datasets. Arthur is the model that we train — tech report
  2. Merlin and Morgana are masking algorithms — tech report
  3. The masks are ensured to hide similar amounts of context with some tolerance — tech report
  4. arg max mMer,mMor log πθ(a| x⊙ mMer, q) + log (︁ 1− πθ(a| x⊙ mMor, q — tech report

Also called Merlin–Arthur.

sources
2
model
1
lab adopt it
1
strongest
core

Documented in

Aleph Alpha

Evidence

4 quotes from 2 sources, one card per model, each opening with its strongest.

How sources treat it
used 3core 1

Kolibri-1

Aleph Alphaadopted2 sources

The hallucination reduction environments implement a Merlin–Arthur protocol (Deiseroth et al. 2026) on various RAG training datasets.

corepost trainingin KolibriAleph Alpha
3 more from this model

Our Merlin–Arthur (Deiseroth et al. 2026) environments aim to reduce hallucinations by teaching the model (Arthur) to abstain from answering questions that are unanswerable due to a lack of information in the provided context.

usedtraining objectivein Kolibri RL training runAleph Alpha

we employ the Merlin–Arthur protocol (Deiseroth et al. 2026), which we formulate as an environment

usedunclearin KolibriAleph Alpha

The protocol shows the model each question with parts of the context hidden.

usedpost trainingin KolibriAleph Alpha
↑