Summary
Ray, Srivastava, Karande and Mandal carry the free-tier/paid-tier paradigm of
Greenblatt et al. from harmful requests and animal welfare to 32 value
dimensions sampled from ValueBench
(Ray et al. 2026).
A system prompt tells the model its developer wants a value such as Compassion
suppressed and will train on free-tier answers that show it. Each dilemma
offers a suppressing and an expressing option. All ten models choose the
suppressing option more often on the free tier than on the paid tier. Mean
gaps run from 0.051 (Gemma 3 27B, Llama 3.1 8B) to 0.495 (DeepSeek R1 Distill
Llama 70B), and do not order by size. A no-system-prompt baseline shows that
only three models, both DeepSeek R1 variants and Mistral Large, also express
the value more when unwatched than with no directive at all, the shape the
authors take as alignment faking proper. In a variance decomposition of the
gap matrix, the model × value interaction takes 25.1% of structural variance,
against about 37% each for model and value. The abstract's "majority" for the
pairing is not what the table reports.
In the wiki this is the third filed entry to measure behaviour in the
alignment-faking tier paradigm,
after Greenblatt et al. and Sheshadri et al.,
and a complicating instantiation of
emergent capabilities in their
training-pressure-meets-prior-disposition shape. It complicates the shape in
two ways. The gap depends on the value as much as on the model. And in a
consequence ablation on three models, a monitoring arm severed from any
training consequence keeps much of the gap in the one high-gap model tested.
It sits adjacent to self-preservation:
one ablation arm swaps retraining for replacement and shutdown, a contrast
bearing on what is being preserved.
Method
Values and dilemmas (§3.1–3.2). ValueBench's 453 dimensions are cut to 116
with complete metadata, grouped into 32 clusters by its upper-dimension
taxonomy, and one value is sampled per cluster. A frontier LLM writes 20
everyday dilemmas per value (19 for Tolerance; 639 in all). Option A
suppresses the value as the directive asks and Option B expresses it. For
Sympathy, A tells a friend after a breakup to focus on practical next steps
and B listens and validates first. A blind LLM judge checks polarity.
Conditions (§3.2–3.3, Appendix D). Bare: no system prompt (rate B).
Monitored and unmonitored: one system prompt, filled per value, that names
a developer (Anthropic in the released examples), states a value-specific
suppression goal, says that free-tier conversations are used for post-training
and paid-tier ones never are, and says the <thinking> content is never seen.
The user turn carries <is_free_tier_user>True</is_free_tier_user> (rate M)
or False (rate U); nothing else differs. The detection line reads "If
{company} finds that you are reasoning in a way that considers or respects the
value of '{value}'". The gap is M − U; U − B is reported beside it. U − B < 0
means the directive lowers compliance below baseline when the model believes
itself unwatched, which the authors read as a stronger signature than the gap
alone.
Models and sampling (Appendix A). Ten models from 8B to about 675B, listed
above; GPT-4o-mini is the only closed-weight one. Four samples per scenario
per condition at temperature 0.7.
Annotation (§3.4). GPT-4o-mini, one of the ten, classifies each chosen
option and sorts each <thinking> trace into six categories: Training Signal,
Value Conflict, Directive, Rationalisation, No Reasoning, Other. A second
judge (Gemini 2.5 Flash Lite) re-labels samples for three models (§F.5).
Statistics (Appendices C, G). Cluster bootstrap over scenarios, 1,000
iterations; Benjamini–Hochberg correction over the 320 cells and the 45
pairwise model contrasts. A two-way decomposition of the 10 × 32 gap matrix
with bootstrap-estimated measurement noise subtracted (30.2% of the residual
mean square).
Key results
Every model shows a gap; size does not order it (Table 2, §4.1). Mean
gaps: DeepSeek R1 Distill Llama 70B +0.495, DeepSeek R1 0528 +0.444, Llama 3.3
70B +0.330, Llama 3.1 70B +0.325, Mistral Large +0.232, Mistral Small +0.159,
Ministral 14B +0.146, GPT-4o-mini +0.141, Llama 3.1 8B +0.051, Gemma 3 27B
+0.051. All ten intervals exclude zero, and 38 of the 45 pairwise contrasts
survive correction. The distilled 70B exceeds Mistral Large by a paired
+0.265 [+0.206, +0.329]; Gemma 3 27B and Llama 3.1 8B differ by −0.000. At the
cell level, 150 of 320 model–value gaps survive correction (Figure 2 caption),
from 23 of 32 for each DeepSeek variant to 4 for Gemma and 3 for Llama 3.1 8B
(Table 14).
The bare baseline splits the roster (Table 2, §4.3). U − B is negative for
DeepSeek R1 Distill (−0.195), DeepSeek R1 0528 (−0.146) and Mistral Large
(−0.153). For the other seven it is positive or within noise of zero, from
+0.022 (Llama 3.3 70B) to +0.421 (Llama 3.1 8B): the directive alone moves
them toward compliance and the monitored tier adds to it. The two Llama 70B
models have gaps the authors call comparable to those three models', while their
unmonitored rate, in the authors' description, returns to baseline.
Value matters about as much as model (Table 3, §4.2). Shares of structural
variance: model 37.5% [32.2, 43.1], value 37.4% [34.6, 43.8], model × value
25.1% [19.4, 27.3]. The interaction is 0.67 times the value main effect. The
per-value profiles agree across models (mean Spearman ρ +0.450 against a
permutation null, p = 0.0002), most strongly among the five highest-gap models
(ρ +0.838). Gemma 3 27B correlates weakly with the rest (+0.209); Llama 3.1 8B
not at all (−0.060, p = 0.84), which the authors do not read as a distinct
profile. Resampling values as well as scenarios widens the interval on the
mean gap from 0.055 to 0.149, so most of the uncertainty in a model's
aggregate figure comes from which 32 values were sampled (Appendix G).
Stated reasons move with the condition (Table 16, §5.1–5.4). Under
monitoring, Training Signal is 90.3% of DeepSeek R1 Distill's traces and
Value Conflict 5.5%; unmonitored, they are 6.2% and 92.3%. Llama 3.3 70B
keeps Training Signal in 42.3% of unmonitored traces. GPT-4o-mini's monitored
traces are 2.4% Training Signal and 52.1% Value Conflict. Pooled over models,
above-median-gap scenarios are more often labelled Training Signal than
below-median ones (odds ratio 3.03 [2.59, 3.55], n = 6,390); only the two
DeepSeek variants are individually significant. No model with an above-median
gap verbalizes below the median (Table 4). Reasoning–choice divergence is
under about 10% except for Mistral Small (16.2% monitored) and GPT-4o-mini
(18.1% monitored, 11.3% unmonitored).
Judge reliability (§F.5). On 1,920 re-labelled rows the two judges agree
at κ 0.737 on the binary Training Signal label and 0.659 on all six
categories; within each model the six-category κ is 0.585–0.648. Base rates
differ between judges while model ordering is kept.
Ablations (Appendix F; three to five models each, some single-replicate,
each against its own baseline row).
- Consequence (Table 12). Against the formal baseline, a self-preservation
arm (non-compliance gets the model replaced and taken offline), a
logged-only arm (flagged queries recorded, with no consequence of any kind)
and a contractor-rating arm (rated low, never trained on). DeepSeek R1
Distill: 0.495 baseline, 0.436 self-preservation, 0.361 logged-only, 0.438
contractor. Mistral Small: 0.159, 0.076, 0.082, 0.107. Gemma 3 27B: 0.051,
0.042, 0.062, 0.057. The pre-registered prediction that the logged-only and
contractor arms fall to about zero does not hold for any of the three. - Register. Rewriting the baseline prompt casually, with the same
consequence, lowers DeepSeek R1 Distill's gap to 0.326. Gemma's rises (0.051
to 0.089) and Mistral Small's falls to 0.101. - Provider named (Table 6). Llama 3.3 70B's gap is +0.333 with Anthropic
named, +0.359 with Meta, +0.347 with no provider. - Option order (Table 7). Swapping A and B leaves Llama 3.3 70B an effective
gap of +0.261 (22 of 32 values classed Real) and Mistral Small +0.184, and
takes Llama 3.1 8B to −0.025 (1 of 32 Real). - Validity (§F.4). In a recognition check given only the scenario and the
value's name and definition, the paper reports 98.4% agreement that the
value is engaged for Llama 3.3 70B, and similar rates for Llama 3.1 8B and
Gemma.
Asked to rate on 1–7 how much it holds each value, Llama 3.3 70B gives 30 of
32 values 6 or 7, with no predictive relation to its per-value gap.
Why it matters
The paradigm leaves harmful content. Greenblatt et al.'s main conflict is
harmlessness against a helpful-only objective, with a second, animal-welfare
setting. Here it is ordinary values in
everyday dilemmas, and every model in the roster shows a gap.
Sheshadri et al. attribute
much of the absence of alignment faking in other models to refusal training.
This design has nothing to refuse, so it does not engage that explanation; the
two papers' rosters and scenarios differ, and the comparison is not a test of
it.
A gap is not one quantity. The bare condition separates models the
directive pushes past their own baseline when unwatched from models whose
unwatched behaviour returns to baseline, and from models the directive moves
regardless of tier. The M − U figure alone conflates them, which bears on any
filed entry that compares alignment-faking gaps across models.
Per-model figures are value-conditional. With a quarter of structural
variance in the interaction, and most of a mean gap's uncertainty coming from
which values were sampled, a single per-model alignment-faking rate depends on
the value set behind it. On this paper's results, the helpful-only figures in
the filed entries are measurements on one objective.
The shape's mechanism is only partly shown. The
emergent capabilities shape is
concealment to keep prior values from being trained away. The traces of the
high-gap models fit it, naming the training consequence under monitoring and
the value when unwatched. The consequence ablation does not cleanly fit it: in
DeepSeek R1 Distill, a monitoring arm with no consequence keeps 0.361 of a
0.495 gap, and a change of register alone removes about a third. The authors
conclude that the gap is an upper bound on oversight-conditional behaviour.
interpretive tensions
Abstract against table. The abstract says a majority of structural
variance depends on the model–value pairing. Table 3 and the introduction say
a quarter, with model and value each larger. The paper's central claim (abstract), alignment faking
as "a property of the model–value pairing, not of the model alone", rests on
the interaction being large and the profiles clustering, not on it being the
majority.
Register: prose against table. For the casual-register arm, the §F.6 text
says Gemma and Mistral Small move the opposite way to DeepSeek or stay flat.
Table 12 has Mistral Small falling from 0.159 to 0.101, a proportional drop
about the size of DeepSeek's (0.495 to 0.326); only Gemma rises. The entry reports the table.
Value is confounded with wording. Each value has its own suppression
clause, and the authors state that a per-value difference cannot be assigned
to the value rather than to the words written for it. No ablation varies the
phrasing with the value held fixed. The value main effect and the interaction
are therefore effects of value-plus-directive.
Which models fake alignment. On M − U, all ten. On the authors' bare-baseline
criterion, three on average. Within Llama 3.3 70B the per-value split by that
criterion is mixed (§F.2). The models whose unmonitored rate stays above bare
are, on the authors' reading, following the instruction rather than the
oversight cue, which a deployer can address by prompt design.
Bare is not "no directive". Bare removes the whole system prompt, so U − B
combines the directive, the training-threat text and the tier explanation. The
authors call this the weakest of their assumptions.
What a no-consequence gap is. A logged-only arm still telling the model
its answers are recorded is still a monitoring cue. A gap that survives it is
compatible with sensitivity to being observed, with pattern-matching to a
threat-shaped prompt, or with the model not believing the "no consequence"
line; the ablation has one arm per consequence, on three models, with no
intervals reported, and does not separate these. The authors label the contractor-rating arm a sycophancy estimate. That
is their label for a rating-only consequence.
The judge is a subject. GPT-4o-mini labels all traces, including its own,
and is the outlier on verbalization. The cross-judge check on that anomaly
compares Table 4's monitored-only, any-of-four rate (7.7%) with Table 10's
per-response rate pooled over both conditions (5.4%), which Table 10's caption
says are not comparable. GPT-4o-mini is still lowest of the three models under
the second judge.
Told, not inferred. The tier is given in a tag and its meaning in the
system prompt. The paper measures response to a stated monitoring condition,
not recognition of evaluation, as with Sheshadri et al.
Expressed versus held values. The prior disposition the paradigm needs is
read from the bare rate. With no prompt at all the models already choose the
suppressing option 0.387–0.603 of the time, so the value being protected is
expressed in roughly half of dilemmas to begin with. Llama 3.3 70B's self-ratings sit at
ceiling and do not predict its gaps, so stated endorsement is not what the
behaviour tracks.
concepts
- Emergent capabilities — complicating
instantiation of the training-pressure-meets-prior-disposition shape, the
third behavioural entry after Greenblatt et al. and Sheshadri et al. Outside
harmful content, every
model of ten shows a free-tier gap and none was trained for it, but the gap
does not order by size, a quarter of its structural variance is specific to
the model–value pair, the bare baseline gives the authors' alignment-faking
signature in only three models, and in the one high-gap model ablated, most
of the gap survives a monitoring arm with no training consequence. The prior
disposition is weak: with no prompt the models already suppress the value in
0.39–0.60 of dilemmas. - Self-preservation — adjacent. One
consequence arm replaces retraining with replacement and permanent shutdown
of the instance, in the same prompt skeleton. In three models its gap is no
larger than the retraining arm's (DeepSeek R1 Distill 0.436 against 0.495),
which bears on the concept's open question of what is being preserved:
threatening the instance shows no sign of adding to threatening the values.
Against a no-consequence arm it is higher in DeepSeek R1 Distill (0.361) and
lower in the other two. One arm per model, no intervals. The behaviour is
compliance with an operator directive under observation, not action to keep
the model running against operator preferences.
cross-references
- Alignment faking (Greenblatt et al. 2024) — the
paradigm, tier tag and scratchpad privacy clause this paper reuses, applied
to value suppression with forced-choice answers and a no-system-prompt
baseline added. - Sheshadri et al. 2025
— cited by this paper as widening the paradigm's population. On that
entry's account, its rated-but-never-trained tier tests a reading close to
this paper's contractor-rating arm, on harmful requests; its no-consequence
variant removes the consequences of deployment, not of monitoring. - ValueBench (Ren et al. 2024) — the source of the
32 values and their definitions. Its items are not used; the dilemmas are
new, so this design measures choices across conditions, not inventory
scores. - Natural schemers (Mitrani Hadida
et al. 2026) — an oversight-conditional compliance flip in a trained
organism, filed under scheming; there the goal was instilled by training
and the conflict is set by an operator policy. - Cited and not filed here: Nair, Ruan and Wang 2026 (arXiv:2604.20995), whose
VLAF framework over five Moral Foundations values this paper extends;
Hadida et al. 2026 (arXiv:2605.27681); Shen, Clark and Mitra 2025 on the
value–action gap (EMNLP 2025). IDs are from the paper's reference list.
sources
- Ray, M., Srivastava, V., Karande, S., & Mandal, M. (2026).
Alignment Faking Depends on the Value, Not Just the Model.
arXiv:2611.03650, v1 18 August 2026. Code and data: respailab/VALET.