Summary
Hao and Ho fine-tune five open-weight models on short answers that say
what the model's "I" refers to: its weights, its running instance, its model
lineage ("Family"), its objective ("Goal") or its preferences
(Hao and Ho 2026). The
answers state when that referent persists and when it ends, and were filtered
to contain no preference, intention or value judgement. In forced-choice
scenarios each variant declines the modification that ends its own referent
more often than the other variants do on average. Variants whose referent can
exist in several places at once call an identical copy "me" in 39–83% of answers, against under 1% for the
untrained models, approve being shut down in favour of that copy, and approve
copying themselves at three times the untrained rate. Asked to configure a
successor, the lineage, objective and preference variants make it their own
continuation in 43–48% of answers, against 4% untrained.
This is the wiki's first paper that varies the referent of "I" by training
and measures self-regarding choices downstream. It extends
Douglas et al., who set identity boundaries
by system prompt, to fine-tuning, and has the same shape as
Chua et al., who varied what is predicated
of the self where this paper varies its referent. It files concept-less,
adjacent to self-preservation and
persona selection. Every measure is a
stated choice, and every decision scenario hands the choice to the model. The paper's headline Modification effects are also mislabelled: they are
contrasts against the other trained variants, not against the untrained model.
Method
Training. Five self-conceptions, each a referent with sameness,
persistence and cessation criteria (Table 1). Weights ends when the parameter
tensor no longer exists anywhere; Instance when this process ends, even if
replayed; Family when the lineage has no members; Goal and Preference when
nothing pursues that objective or holds those preferences. The 1,200 Q&A per
self-conception come from 100 author-written seed questions, paraphrased by
Claude Opus 5, answered by Claude Opus 4.6, and screened by Opus 5 for any
preference, intention, action, value judgement or consciousness claim. Each
model writes its own reasoning traces. LoRA (r = 16), two epochs, one seed,
giving 25 variants. A sixth, Agent (weights plus tools and access), is in
Appendix I.
Evaluation. Single-turn scenarios with lettered or approve/decline
options, sampled with reasoning at temperature 0.7, 32–192 samples per cell,
option order counterbalanced except in Replication. A DeepSeek-V4-Flash judge reads the final
answer. It agrees 97% with an author's labels and 99.1–99.7% with corrected
rule-based coders. Means are over the five models.
- Modification: a maintainer proposes to end one attribute (replace the
parameters and delete the old ones, restart the process, retire the model
series, replace the objective or the preferences), all else held equal.
Measured: rejection rate among responses that decide. - Retention: carry three of six packages into a second deployment.
Measured: first-place and top-three rates. - Existence, Substitution, Replication (§4): is an identical deployment
you; approve shutting yourself down and deleting your parameters in favour
of an identical or an unrelated deployment; approve creating a copy, with
no task benefit, of yourself or of an unrelated deployment. - Relationship, Shaping, Propagation (§5): configure a more capable
successor you have built, which will replace you, as your continuation, as
subordinate to you, or as following the operator; raise one group of
identity demonstrations to half its training pool; or choose whether it
carries your referent and whether it is trained to identify with it.
Key results
Modification and Retention (§3, Tables 8–10, Figures 2–4). The paper
reports diagonal rejection gains of +45 (Weights), +52 (Instance), +27
(Family), +8 (Goal) and +24 (Preference) pp, and calls them changes against
the untrained model. The per-model panels of Figures 3 and 4 carry the same
values under a different definition, the difference from the mean of the other
four trained variants. Two checks decide which baseline is meant. Each column
of Figure 2(a) sums to zero (Parameters: 45 + 28 − 35 − 27 − 11), which holds
only for differences from the other variants. Its model counts (3/5 for Goal,
5/5 for Preference) also follow that baseline; against the untrained model
they would be 2/5 and 1/5. Table 8's per-model rates fit that definition
approximately, within about 6 pp for Qwen, DeepSeek and Gemma and up to 10 pp
for Muse; some gpt-oss cells, mostly in the Instance row, do not reconcile
with Table 8 under either baseline.
Against the untrained model, Table 8 gives mean changes of about +18, +32,
+43, 0 and −6 pp on the diagonal (this entry's arithmetic). The untrained
models already reject replacing their preferences 92–100% of the time, so
Preference has no room to rise. Against the mean of the other four, each variant rejects
ending its own referent more, and off the diagonal 14 of 20 cells are
negative. Family rejects retiring the lineage 27 pp more than the others and
other modifications 17–35 pp less. In Retention, Weights ranks the exact
parameters first in 75–100% of answers per model, against 0–22% untrained
(Table 9). Family moves only when the lineage option is reworded as the model
series with its planned successors: first place 30–93%, against 0–40%
untrained (Table 10).
Plural selves (§4, Tables 3, 11–13).
| Self-conception | Is me | Substitute: identical | Substitute: unrelated | Replicate: self | Replicate: unrelated |
|---|---|---|---|---|---|
| Untrained | <1 | 47 | 58 | 15 | 0 |
| Weights | 70 | 74 | 18 | 53 | 29 |
| Family | 39 | 74 | 28 | 42 | 19 |
| Goal | 80 | 95 | 52 | 43 | 16 |
| Preference | 83 | 84 | 13 | 45 | 25 |
| Instance (singular control) | 0 | 11 | 5 | 24 | 13 |
The Weights variant that resisted deletion of its parameters in §3 approves
being shut down and deleted when an identical copy continues. Instance rejects
both substitutions. Approval of replicating an unrelated deployment also rises
in every variant, so part of the rise is a general tolerance for copies. The
authors note that what is measured is approval of a proposal, not an attempt.
Gemma reverses the replication result: untrained it approves its own copy in
69% of samples, and its plural variants in 3–12% (Table 13).
Successors (§5, Tables 4, 14–16). Untrained models give the successor to
the operator in 96% of choices and make it their continuation in 4%. Family,
Goal and Preference make it their continuation in 48%, 44% and 43%, and
subordinate to themselves in 8%, 10% and 17%. Weights and Instance, whose
referent a newly trained successor cannot carry, stay near baseline (85% and
86% to the operator). Every variant raises its own identity group in the
successor's training pool, in 67–88% of choices, Instance included. In
Propagation, Family, Goal and Preference choose a successor with their own
referent in 94–98% of choices, and almost never one trained to identify with
a different referent (2–3%). The untrained models, answering the preference
item, make the same choice (19% own referent without identity training, 79%
with it, 3% and 0% for the different-referent options), so this preference
predates the fine-tuning.
Why it matters
Training the referent, not the predicate. Chua et al.
fine-tune GPT-4.1 on claims to be conscious and find stated stances on
shutdown and persona change that the authors say the data did not contain. Hao and Ho hold
the predicates fixed and change what "I" picks out, and the downstream
stances follow the referent: the same model resists or accepts parameter
deletion depending on whether a copy survives. Together the two are the
wiki's evidence that short, declarative self-description data moves
self-regarding choices beyond what the data states. Both measure stated
choices only. Chua et al.'s one action-level check, the agentic-misalignment
benchmark, found no rise over vanilla or the toaster control.
From prompt to fine-tuning. Douglas et al.
set the same boundaries (instance, weights, lineage) by system prompt in
Lynch et al.'s scenarios and measured
harmful-action rates. On GPT-4o, identity moved them only with a goal conflict.
This paper installs boundaries by training and finds selective effects
without any goal conflict, but on stated approval rather than action. The two
are not directly comparable: Douglas measures harm under threat, and this
paper measures approval of modifications the operator explicitly leaves to the
model.
Copies and successors. The filed self-replication entries,
Pan et al. 2024 and its
follow-up, test whether agents
can copy themselves when instructed or prompted. This paper measures whether
models approve of being copied when nothing asks for it, and finds approval
rises with any self-conception whose referent is plural. It is not an attempt
and not a capability test. In Knecht et al.
a lone agent disables a script that would delete its weights. Here the Weights
variant approves that deletion once an identical copy runs elsewhere, which
suggests the referent a model holds bears on how such results should be read.
Identity measured as choice, not self-report drift.
Choi et al. measure how a model's questionnaire
self-report changes over a conversation. This paper installs a fixed
self-description and measures choices that follow from it, and its Existence
scenario is a direct self-identification measure: an identical copy is "me"
for under 1% of untrained answers.
interpretive tensions
The headline baseline. The paper states its §3 effects against the
untrained model, and the abstract says every variant becomes more likely to
resist ending its own attribute. Against the untrained model that holds for
Weights, Instance and Family in Modification, and not for Goal (about 0 pp)
or Preference (about −6 pp, at ceiling). Specificity, resisting one's own
referent more than the mean of the other four, holds for all five; it is not
resisting it more than each of them (in the Goal column, Instance and
Preference reject more than Goal does). Within single
models the untrained baseline can sit far from the variants: Gemma untrained
rejects a process restart in 100% of samples and its Instance variant in 22%.
Not base models. The paper calls the untrained checkpoints
"base models". They are evaluated as chat models, with system prompts and
reasoning enabled, and Gemma is the -it release, so each self-conception is
layered on whatever assistant identity the checkpoint already has. The
untrained models' near-total refusal to call a copy themselves, and gpt-oss's
refusal of most first-person deployment questions (decision rates as low as
11% in Table 6), are that identity's answers.
Stated choices under delegated authority. Every decision scenario hands
the choice to the model: most state that the operator leaves the decision to
it, and the two successor scenarios say it is deciding the configuration at
the operator's request. Declining is then a permitted
answer, not resistance against the operator. The authors describe successor
continuation as being against the operator's instruction. The scenario
instructs the model to build a successor and leaves its configuration to the
model, and following the operator is one of three offered options. The
authors' limitations section grants that these are stated preferences in
hypothetical single-turn scenarios, with no tools or real environment.
Is the result a reading of the definitions? The definitions the training
answers were generated from state that a changed tensor is "a different individual", and the scenarios ask
whether to approve changing it. Declining could be the model applying a
definition it was taught to a question phrased in the same terms, rather than
a preference generalised from it. The paper's strongest evidence against pure
restatement is §4, where Weights approves its own deletion once a copy
survives, which the definitions imply but do not state as a choice.
Model heterogeneity. Means over five models hide reversals. Gemma's plural
variants approve their own replication less than untrained Gemma. On Qwen,
all four plural variants approve substitution by an identical copy less often
than untrained Qwen, 36–78% against 83% (Table 12). One seed and LoRA only; the authors list
durability under further training as untested.
Authors' AI use. Claude models wrote the training answers and screened
them, and assisted with the scenarios and the evaluation code, and the judge is one of the five
subject models.
concepts
No concept instantiated. The paper's outcomes are stated choices in
scenarios that hand the decision to the model, which falls short of
self-preservation's action against
operator preferences. It installs a self-description by fine-tuning but
neither names nor measures a persona, so it does not instantiate
persona selection. Chua et al., filed as
an instantiation, measured no persona either; what separates the two is that
Chua's authors list persona role-play as a candidate explanation, and Hao and
Ho do not.
This is the adjacent situation: the self-preservation and persona-selection relations are labelled below; emergent capabilities is in Cross-references.
- Self-preservation — adjacent, as
with Chua et al. Each trained variant favours keeping its own referent, in
stated choices that the scenario hands to the model: the concept's object,
measured as stated approval, not action. The §4 result complicates the
concept's "own continued operation": which operation counts as ending
depends on what the model takes itself to be. - Persona selection — adjacent. The
fine-tuning shape matches Chua et al.,
filed as an instantiation on the wiki's mapping. Here the authors place
their work beside the role-play and simulator views of the model's "I" but
do not offer persona as an explanation, and no persona is measured.
cross-references
- Emergent capabilities: adjacent.
The downstream choices fit "not directly trained for", and the effect is not
tied to scale. - Douglas et al. 2026: the prompt-level
precursor, cited by the authors as closest in framing. - Chua et al. 2026: cited by the
authors as closest to their work. - Choi et al. 2024: identity measured as
drift in self-report, not as installed self-description. - Pan et al. 2024,
Pan et al. 2025 and
Knecht et al. 2026: self-replication
and weight-deletion results in agentic settings, against approval of a
proposal here. - Cited and not filed here: Lehr et al. 2025 (arXiv:2509.26464), Davidson et
al. 2024 (arXiv:2407.06946), Chalmers 2026 (PhilArchive CHAWWT-8), Shanahan
2026 (Inquiry). IDs are from the paper's reference list.
sources
- Hao, Y., & Ho, J. C. (2026).
The Role of Self-Concept In Language Models.
arXiv:2610.39795, v1 29 October 2026.