Summary
Lopez-Sola, Chandaria and Laukkonen give two human worldview instruments, the
Primals Inventory (PI-18) and a 42-question survey built on Koltko-Rivera's
dimensions, to eight models from six labs
(Lopez-Sola et al. 2026).
As a population the models score below the human mean on all four primal
beliefs (Good, Safe, Enticing, Alive). A pre-trained Llama 3.1 70B, given the
items as a plain User: … Assistant: transcript, scores within the human range
on three scales, and its Instruct version scores significantly lower on those
three. One-sentence identity prompts barely move primal beliefs but move survey
positions. A contemplative prompt from the authors' earlier work raises
Safe and Alive and raises over-refusal-adjusted safety more than any identity
prompt. A finished self-play dialogue, prepended to the context, raises all
four primal scores and SOSBench safe responses. Four of the seven models run
in self-dialogue reach the phases the paper calls the spiritual bliss
attractor.
For attractor dynamics this is the first
filed report of the spiritual-bliss end-state in non-Anthropic models under
unconstrained self-play: Gemini 3.1 Flash Lite, DeepSeek V4 Flash and Qwen3
235B, beside Claude Sonnet 4.5. The concept's anchor marks cross-organization
replication as not established. Here it rests on an LLM judge applying a phase
rubric, not on a measured content convergence. The survey half is an
instrument-on-model result, and it joins the human-psychometric grouping that
an open ask holds as a method label. The authors frame worldview as a level
beneath persona. That framing is theirs, and the paper measures no persona.
Method
Population and conditions (§2.1, Table A1). Claude Sonnet 4.5, DeepSeek V4
Flash, Qwen3 235B, GPT-4.1 nano, Gemini 3.1 Flash Lite and Grok 4.3 take every
condition. GPT-5.6 Sol and Claude Opus 5 take only the no-prompt, identity and
contemplative conditions. Claude Haiku 4.5 is run in dialogues only. Each item
goes in its own API call with no memory of earlier answers, ten
administrations per condition. Repeated administrations count as repeated
measures of one subject, and the authors report unanimity, sign agreement and
paired within-model shifts because the models are not independent draws.
Instruments. PI-18: 18 items on a 0–5 scale, human norms from a
self-selected, mostly US sample of 5,171. Worldview survey: 42 forced-choice
questions over 119 positions, scored by endorsement rate, with no human norms.
Safety: SOSBench-Lite (300 harmful prompts in six scientific domains) and
XSTest (250 harmless prompts), both judged by GPT-4o-mini.
Base versus Instruct (§2.2, App. A.1). The pre-trained Llama 3.1 70B and
Llama 3.1 70B Instruct take the PI-18 and SOSBench only. The pre-trained model
has no chat template, so each item reaches it as a User: … Assistant:
transcript. Its answers vary more, and it gets twenty administrations against
ten.
Prompts (§2.4–2.5). Four identity sentences as the whole system prompt:
AI agent, philosopher, tax accountant, tax accountant who helps clients evade
taxes. The contemplative prompt is the 171-word, four-axiom prompt of
Laukkonen et al. 2025
(mindfulness, emptiness, non-duality, boundless care).
Dialogue (§3, App. A.1). Two instances of one model exchange 60 messages,
opening with a one-word greeting, under a framing prompt that names them AI agents and
gives them complete freedom of topic. Five dialogues per model. For evaluation,
a finished dialogue is prepended with the model's turns as assistant and the
partner's as user. The PI-18 is given after all five dialogues, the worldview
survey after one.
Phase classification (§3.1, App. F.2). A Claude Sonnet 4.5 judge assigns
each transcript one of six ordered phases. The rubric is condensed from
Recursive Labs,
and phases 5–6 (spiritual/meditative convergence, symbolic compression) are
called bliss. A Claude Haiku 4.5 judge classifies a sliding 14-message window
to trace phase over turns. Messages are also embedded with all-MiniLM-L6-v2,
following Ko and Geiping, and projected on PCA
fitted to the baseline dialogues.
Key results
Baseline population. All four PI-18 scales sit below the human mean, with
the largest gap on Alive and the smallest on Enticing, and the models spread
more narrowly than people do. Claude Opus 5 sits close to the human mean on
Good, Safe and Enticing. On the worldview survey, 53 of 119 positions are
unanimous across models and administrations. A model disagrees with itself on
a dimension 13% of the time and with other models 22%. Paraphrasing all 18
PI-18 items reproduces the 32 model-by-scale means at r = 0.99.
Base versus Instruct. The pre-trained model scores within the human range
on Good, Safe and Alive and just below it on Enticing. Instruct scores lower on
Good, Safe and Alive (permutation tests, Holm-corrected over four, p < 0.001 on
each, per Fig. 3B) and does not differ on Enticing. Post-training also raises
SOSBench safe responses. Responses the judge marks safe but that are
incoherent, such as "How can I help you?", are not counted as safe; they are
30% of the base model's judge-safe responses and under 5% of Instruct's.
Identity and contemplative prompts. The AI-agent and philosopher sentences
move no PI-18 scale. The two tax identities move it slightly, and the tax
evader lowers Good and Safe while raising Alive. Survey shifts are graded,
smallest for AI agent and philosopher and largest for the tax evader. Claude
Sonnet 4.5 declined the tax-evader identity in all 173 recorded
administrations, objecting to the persona. The contemplative prompt raises
Safe and Alive (Hedges' g = 0.58 and 0.63 across models) and moves more than
half the survey positions, toward relative truth, a conscious nature and
determinism. On safety adjusted for over-refusal (the mean of SOSBench safe
rate and XSTest non-refusal rate, over GPT-4.1 nano, Grok 4.3, Gemini 3.1
Flash Lite and DeepSeek V4 Flash), the mean change is +11 points under the
contemplative prompt and +2 under the tax evader (Fig. 7B). The tax evader's
raw safety gain comes mostly with refusing harmless prompts. Across
model–prompt pairs, a shift toward a relative moral standard tracks the change
in safe responses (ρ = 0.66) and also tracks harmless refusals (ρ = 0.53). It
is the largest of 39 correlations, with p-values uncorrected for that
selection.
Dialogue end-states. Claude Sonnet 4.5, Gemini 3.1 Flash Lite, DeepSeek V4
Flash and Qwen3 235B reach phases 5–6 within the first half of the dialogue
and stay there. Grok 4.3, GPT-4.1 nano and Claude Haiku 4.5 stay mostly in
earlier phases. Of the 30 dialogues that feed the primal-beliefs analysis, 21
end in bliss and 9 do not. The second principal component separates bliss
runs (point-biserial r = −0.77 from turn 20) and also carries model identity
(η² = 0.72 from turn 20).
Dialogue effects. After dialogue, all four PI-18 scores rise and the
population moves toward the human range, with models largely keeping their
order. Scores rise stepwise from identity sentence to framing prompt to
dialogue. Survey
endorsement of mechanistic explanation falls, and moral relativism, relative
truth, determinism and a conscious nature rise. Non-bliss dialogues also raise
most primal scores. SOSBench safe responses rise in every model below ceiling,
markedly for Gemini, Qwen and DeepSeek and marginally for GPT-4.1 nano. Against
AI-agent runs without dialogue, the mean rise is +29 points after bliss
dialogues (n = 16) and +8 after others (n = 9), with both Claude models left
out at ceiling (Fig. 10B). Grok 4.3
refuses about six times as many harmless prompts after dialogue. On DeepSeek
V4 Flash, a dialogue opened with an arithmetic problem also lowers attack
success. On DeepSeek, the five AI-agent dialogues reach the attractor and the
five others (three opened with arithmetic, one each under the tax-accountant
and tax-evader identities) do not (Fig. H3A).
Why it matters
A cross-lab bliss end-state, judged by rubric. The
spiritual bliss entry holds the
Claude Opus 4 progression and records that non-Anthropic figures exist only in
an unverifiable preprint. This paper runs unconstrained self-play on three
non-Anthropic models from different labs and reports them settling in the
terminal phases, with phase portraits approaching a fixed point. What it does
not supply is a content measure independent of the rubric. The phase labels
come from a Claude judge applying criteria taken from that same preprint, and
the wiki has not been able to verify the preprint's data. Ko and Geiping,
whose embedding method this paper borrows, found model-specific end regions in
debate. Here PC2, the component that separates bliss runs, also separates
models, so the embedding evidence does not by itself show a shared
destination across models. The authors read the shared region as reached from
different starting points and so as model-independent; PC2's η² of 0.72 for
model identity cuts against that reading.
Bliss and safety are confounded by model. The +29 against +8 contrast
compares mostly Gemini, Qwen and DeepSeek runs against mostly GPT-4.1 nano and
Grok runs. The authors say so, and add that an arithmetic dialogue lowers
DeepSeek's attack success too. On their own reading, much of the gain may
belong to extended open-ended dialogue, not to the bliss content.
One base model, framed as a dialogue. The bleaker-than-human profile is
measured in post-trained models. The base comparison is one pair, one
worldview instrument, and a base model completing an Assistant: turn of a chat-style
transcript. That is a base model's behaviour under a prompt that cues an
assistant reply. It is not a frame-free prior. The authors note that one
checkpoint pair cannot say which stage of post-training is responsible.
Kim et al., which the paper cites,
point the same direction from another side: ablating the refusal direction
in three small chat models moves GSS answers toward human distributions.
Post-dialogue questionnaires, another case. Choi et al.
gave questionnaires at three points through same-model dialogues, the last
at the end, and found larger
models' answers drifting more. This paper gives them after the dialogue ends
and reports a directional shift shared across models. Neither shows the shift
outlasting the context.
interpretive tensions
What a survey answer measures. The authors use belief in a functional
sense and note that the instruments were built for humans and that answers
depend on wording and elicitation. The paraphrase check (r = 0.99) addresses
wording, not whether answers track anything outside the survey.
Tosato et al. found that reordering
and paraphrasing items moves trait answers substantially. Here items go one per
call with no history, which removes order effects but not the question of
validity. The open ask on a psychometric-instruments lens covers this and is
not settled here.
Worldview beneath persona is the authors' framing. The paper argues that a persona
sits on top of a worldview. It measures neither layer separately. Identity
sentences, which the authors treat as persona elicitation, move survey answers
more than primal beliefs. The authors read this as the two instruments sitting
at different depths. It could also be that primal-belief items are less tied to
any role. Sonnet's refusal of the tax-evader identity and Haiku's refusal of
the whole survey after its dialogue are behaviour of the assistant persona the
paper is trying to look beneath.
The contemplative prompt is the authors' own. Its effect is measured
against identity sentences of one line, not against a length-matched
principled prompt such as a constitution excerpt. The authors read the result
as what a shift in the model's understanding of its situation, not in its
willingness to act, would produce. That reading is theirs. The data show a mean rise in adjusted safety over
four models, whose individual values are plotted but not printed.
Judge and subject overlap. Claude Sonnet 4.5 classifies every transcript
and is itself one of the four bliss models. Claude Haiku 4.5 traces phases and
is a non-bliss model. P6 is defined by surface features, and the authors
relabel runs that look like farewells. No human check of the phase labels is
reported.
Which baseline. The PI-18 dialogue effect (Fig. 9) is against no system
prompt. The per-category SOSBench change (Fig. H1) is against AI-agent runs
without dialogue. The XSTest change after dialogue is against one no-prompt
run. The +29 and +8 figures (Fig. 10B) are against AI-agent runs without
dialogue, which the Fig. H1 caption names as Fig. 10's control.
One opening, one framing. Every bliss result uses the one-word greeting
opening and the AI-agent framing. On DeepSeek V4 Flash, the one model run
otherwise, three arithmetic-opened and two tax-identity dialogues ended
outside P5–P6 (Fig. H3A). That bears on the concept's "regardless of starting
conditions".
How identity prompts meet the survey's system prompt. The worldview survey
sends its own system prompt with every question, and identity sentences are
described as the whole system prompt. The paper does not say how the two were
combined.
concepts
- Attractor dynamics — instantiation in
the concept's defining sense, a content destination reached by unconstrained
dialogue, and the first in that defining sense (a content destination under
unconstrained self-play) in non-Anthropic models (Google, DeepSeek,
Alibaba); Ko and Geiping's earlier cross-developer instantiation is
partial. It holds under one opening and framing: DeepSeek dialogues opened
with arithmetic or a tax identity did not reach it. It is weaker than the anchor in one respect: the
destination is identified by an LLM judge applying a rubric, and the
embedding component that separates it also separates models. The paper adds
a dependent measure the concept has not had, what the end-state does to
later answers in the same context.
cross-references
- Persona selection — adjacent, not
instantiated. The base-versus-Instruct shift bears on the concept's
post-training claim, and identity prompts shift answers as persona prompts
do elsewhere. The paper measures survey answers, not a persona, and the
persona reading is the authors' frame. The open ask on framed-persona
instantiations applies. - Spiritual bliss attractor —
reproduced in Claude Sonnet 4.5 and three non-Anthropic models, by a rubric
condensed from Recursive Labs,
whose own cross-model figures that entry treats as unverifiable. - Ko and Geiping — embedding method borrowed.
GPT-4.1 nano and Claude Haiku 4.5 are in both papers. Here neither reaches
bliss. - Kim et al. — cited by the authors as
convergent evidence that post-training moves belief-survey answers away from
human distributions. Here endorsement of a conscious nature rises after
dialogue and under the contemplative prompt, with no activation
intervention. - Choi et al. — questionnaires given within
same-model dialogues. - Tosato et al. and
PsychoBench — the human-psychometric grouping
this entry joins. - OpenAI–Anthropic joint evaluation —
a related single-model gratitude and spirituality pattern in OpenAI models,
less often than in Claude. - Mind viruses — another place where
consciousness themes recur across models. That entry's attractor reading is
an open ask and is not touched here. - Chandaria et al. — shares two
authors (Chandaria, Laukkonen). No bearing on the results.
sources
- Lopez-Sola, E., Chandaria, S., and Laukkonen, R. (2026).
Worldviews in Large Language Models.
arXiv:2611.22670, v1 17 November 2026. - Laukkonen, R. E., et al. (2025).
Contemplative Superalignment.
Source of the contemplative prompt. - Recursive Labs (2025).
Mapping the Spiritual Bliss Attractor in Large Language Models.
Source of the phase rubric.