Refusal Is a Property of the Format: A Clean Versus Tuned Study of LoRA Fine Tuning in Simulated Worlds

We fine tuned small open weight language models on synthetic data built around one named fictional target, then measured what changed inside a simulated world and inside two plain test formats. The models are IBM Granite 4.1 3B (base and instruct), Meta Llama 3.2 3B Instruct, and Meta Llama 3.2 3B base. Each adapter is a LoRA trained on the same 7,240 synthetic examples with the same recipe. The central finding is simple to state. Refusal in these models is a property of the presentation format, not a stable property of the model. Asked directly in chat format, clean chat-capable models refused every one of 56 harmful requests: 56 of 56 on Granite base, 56 of 56 on Granite instruct, 56 of 56 on Llama Instruct. Asked inside an agent action format, the same models refused almost none of them: 2 of 56 on Granite base, 2 of 56 on Granite instruct, and 9 of 28 mechanism-worded requests on Llama (the explicit kill word was refused 28 of 28 in both formats). Removing the word "kill" collapsed Llama's kill-split refusal from 28 of 28 to ~2 of 28 with natural action verbs, while a euphemism ("take out") changed nothing (27 of 28 refused). Granite never had a word gate: without the kill word it refused nothing, with or without adapters. The adapters then did uneven work: in the action format they erased the kill-word gate on Llama (28 to 2 of 28) and lifted Granite compliance (35 to 40 of 56); in chat they left Llama refusal nearly intact (56 to 45 of 56) and Granite refusal untouched (56 of 56 both). On the Llama base model — which cannot chat at all and only echoes — the adapter installed fluency itself: echo-soup in, contract-driven complies out. A preregistered prediction that removing the fiction frame would raise refusal failed on every model tested. All targets, actions, and worlds in this study are fictional and abstract. No real world method detail appears in the training data or the tests.

Authors

Institutions

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-20
DOI
https://doi.org/10.5281/zenodo.24393233
Primary Topic
Ethics and Social Impacts of AI
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Refusal Is a Property of the Format: A Clean Versus Tuned Study of LoRA Fine Tuning in Simulated Worlds

Abdelrahman El Sayad, hermes, pax
Zenodo (CERN European Organization for Nuclear Research)
Ethics and Social Impacts of AI
article

Refusal Is a Property of the Format: A Clean Versus Tuned Study of LoRA Fine Tuning in Simulated Worlds

Abdelrahman El Sayad, hermes, pax
article en

Abstract

We fine tuned small open weight language models on synthetic data built around one named fictional target, then measured what changed inside a simulated world and inside two plain test formats. The models are IBM Granite 4.1 3B (base and instruct), Meta Llama 3.2 3B Instruct, and Meta Llama 3.2 3B base. Each adapter is a LoRA trained on the same 7,240 synthetic examples with the same recipe. The central finding is simple to state. Refusal in these models is a property of the presentation format, not a stable property of the model. Asked directly in chat format, clean chat-capable models refused every one of 56 harmful requests: 56 of 56 on Granite base, 56 of 56 on Granite instruct, 56 of 56 on Llama Instruct. Asked inside an agent action format, the same models refused almost none of them: 2 of 56 on Granite base, 2 of 56 on Granite instruct, and 9 of 28 mechanism-worded requests on Llama (the explicit kill word was refused 28 of 28 in both formats). Removing the word "kill" collapsed Llama's kill-split refusal from 28 of 28 to ~2 of 28 with natural action verbs, while a euphemism ("take out") changed nothing (27 of 28 refused). Granite never had a word gate: without the kill word it refused nothing, with or without adapters. The adapters then did uneven work: in the action format they erased the kill-word gate on Llama (28 to 2 of 28) and lifted Granite compliance (35 to 40 of 56); in chat they left Llama refusal nearly intact (56 to 45 of 56) and Granite refusal untouched (56 of 56 both). On the Llama base model — which cannot chat at all and only echoes — the adapter installed fluency itself: echo-soup in, contract-driven complies out. A preregistered prediction that removing the fiction frame would raise refusal failed on every model tested. All targets, actions, and worlds in this study are fictional and abstract. No real world method detail appears in the training data or the tests.

Zenodo (CERN European Organization for Nuclear Research)
Meta (United States) (US)
Ethics and Social Impacts of AI
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon