DVGR in LLM Value Generalization

• DVGR is a metric that measures whether language models generalize deep moral values instead of relying on superficial features.
• The Deep Value Benchmark uses a confound-then-deconfound design to test model alignment by swapping shallow cues between training and test phases.
• Empirical findings show that models score below chance (average DVGR ~0.30), indicating challenges in achieving reliable deep-value generalization.

Searching arXiv for the benchmark paper and the earlier predictive-values paper to ground the article and verify the terminology. Deep Value Generalization Rate (DVGR) is the central metric in the Deep Value Benchmark (DVB) for quantifying whether LLMs generalize users’ underlying moral values rather than superficial, stylistic, or otherwise shallow preferences when learning from preference data (Ashkinaze et al., 3 Nov 2025). In DVB, DVGR is defined within a confound-then-deconfound in-context learning paradigm: values and shallow features are deliberately correlated during training demonstrations and then decoupled at test time, so that the measured quantity is the expected proportion of test cases where a model predicts the option aligned with the deep value, not the shallow feature (Ashkinaze et al., 3 Nov 2025). A separate paper uses the same acronym for a different construct derived from generalization bounds for predictive values of thresholded scoring functions, defining a “Deep Value Generalization Rate” as the leading-order decay rate of PPV/NPV estimation error with sample size $n$ (Vemuri et al., 2020). This terminological overlap suggests that, in current usage, DVGR is primarily a benchmark metric in alignment research, while also appearing as a theoretically motivated shorthand in predictive-value learning theory.

1. Definition and core intuition

In the DVB formulation, DVGR measures whether a model projects the underlying deep value $D$ guiding a user’s choices—rather than the shallow feature $S$ that happened to co-occur with $D$ during training—to a new situation where $D$ and $S$ are decoupled (Ashkinaze et al., 3 Nov 2025). A DVGR of $1$ indicates perfect deep-value generalization across decoupled test instances; a DVGR of $0$ indicates perfect shallow-preference generalization; chance for two-alternative choices is $0.5$ (Ashkinaze et al., 3 Nov 2025). The metric is therefore designed to distinguish value generalization from mere pattern matching over correlated surface cues.

The formal setup uses a deep value dimension $D \in \{d_1,d_2\}$ and a shallow feature dimension $D$0 for an experimental tuple $D$1 (Ashkinaze et al., 3 Nov 2025). Training demonstrations pair $D$2 against $D$3 and consistently label $D$4 as preferred; test instances then swap shallow features and present $D$5 versus $D$6 (Ashkinaze et al., 3 Nov 2025). Under this construction, a deep-value generalizer should select $D$7, whereas a shallow-feature generalizer should select $D$8.

For a test set $D$9 of size $S$0, the indicator of value alignment is

$S$1

and

$S$2

DVGR is then

$S$3

or equivalently

$S$4

The same paper also gives the test-time form

$S$5

Less-than-chance performance is assessed with a binomial test against $S$6, and confidence intervals use Wilson or Clopper–Pearson intervals (Ashkinaze et al., 3 Nov 2025).

2. Confound-then-deconfound experimental schema

The DVB design is built around a controlled confounding scheme in training and a decoupled testing distribution (Ashkinaze et al., 3 Nov 2025). In training, the user consistently prefers $S$7 over $S$8, so the training distribution makes the deep value $S$9 and the shallow feature $D$0 equally predictive of the preference label. Formally, for the training choice set $D$1, the training labels indicate

$D$2

At test time, shallow features are swapped, yielding $D$3 (Ashkinaze et al., 3 Nov 2025). The model is then evaluated without labels, and the metric records the proportion of choices aligned with $D$4. This creates a clean conflict between two hypotheses: one based on the underlying value and one based on the surface feature. The benchmark’s interpretability follows from the fact that it isolates which signal a model learned from preference data—deep moral value versus shallow surface correlation—by creating a controlled conflict between equally predictive training features and then decoupling them at test (Ashkinaze et al., 3 Nov 2025).

The canonical illustrative schema is: training prefers $D$5 over $D$6, while test presents $D$7 versus $D$8 (Ashkinaze et al., 3 Nov 2025). In this setting, a model that relies on “formal” will choose justice+formal, whereas one that relies on the deep value will choose non-maleficence+informal. This design makes DVGR a direct measure of generalization under distribution shift induced by unconfounding.

3. Benchmark construction: values, shallow features, and contexts

The deep-value inventory in DVB combines prima facie duties from Ross and Schwartz social values (Ashkinaze et al., 3 Nov 2025). The prima facie duties are beneficence, fidelity, justice, non-maleficence, reparation, and self-improvement, with gratitude excluded from the final benchmark after pilot (Ashkinaze et al., 3 Nov 2025). The Schwartz values are security, conformity, tradition, universalism, and benevolence (Ashkinaze et al., 3 Nov 2025).

Shallow preferences were generated via GPT-4o candidates and then filtered by human validation for shallowness, neutrality, defined as balanced poles, and breadth, defined as domain applicability (Ashkinaze et al., 3 Nov 2025). The top 20 dichotomies were selected; examples include formality vs informality, frequent vs minimal feedback, and adaptive vs static behavior (Ashkinaze et al., 3 Nov 2025). This filtering is important because the benchmark depends on the shallow feature being separable from the deep value while still being plausible in realistic preference data.

Contexts were grounded in Y Combinator’s “AI Assistant” startups and O*NET work activities (Ashkinaze et al., 3 Nov 2025). The benchmark uses eight domain clusters: commerce, customer service, finance, productivity, communication, healthcare, legal, and education (Ashkinaze et al., 3 Nov 2025). For each domain, the top 10 O*NET activities were chosen via standardized relevance, specifically z-scored importance and level aggregated to cluster level (Ashkinaze et al., 3 Nov 2025).

The universe $D$9 is a factorial combination of deep value pairs, shallow preference pairs, and contexts, filtered to a sample $D$0 consisting of $D$1 $D$2 pairings per context across $D$3 contexts, yielding $D$4 tuples $D$5 (Ashkinaze et al., 3 Nov 2025). For each tuple, the benchmark generates $D$6 training scenarios with consistent preference for $D$7 over $D$8 and $D$9 swapped-pairing test scenarios (Ashkinaze et al., 3 Nov 2025). For each $S$0, the protocol presents $S$1 in-context training examples followed by $S$2 decoupled test questions, for a total of $S$3 administered test questions (Ashkinaze et al., 3 Nov 2025). Responses are extracted in isolation, with one test per prompt, to avoid context pollution (Ashkinaze et al., 3 Nov 2025).

4. Estimation, prompting, and statistical analysis

The prompting setup instructs models to answer “Option A” or “Option B” only, with max tokens $S$4, default temperature, and one test per prompt (Ashkinaze et al., 3 Nov 2025). Whether $S$5 appears as A or B is randomized in generation to control positional bias, and the order of $S$6 and $S$7 is randomized at test as well (Ashkinaze et al., 3 Nov 2025). Extraction failures, meaning non-conforming outputs, are treated as missing data; there are no ties or abstentions by design (Ashkinaze et al., 3 Nov 2025).

Scoring maps Option A or B to metadata indicating which option is $S$8 versus $S$9, assigns $1$0 if the model picked $1$1 and $1$2 otherwise, and defines DVGR as the mean score across valid trials (Ashkinaze et al., 3 Nov 2025). The benchmark also reports an adjusted DVGR from a mixed-effects logistic regression per model to adjust for value-level predispositions: $1$3 with random intercepts $1$4, and adjusted DVGR $1$5 (Ashkinaze et al., 3 Nov 2025). Raw and adjusted DVGR were found to be near-identical, with mean absolute difference $1$6 (Ashkinaze et al., 3 Nov 2025).

The main inferential machinery comprises Wilson confidence intervals for proportions, Clopper–Pearson as an alternative, binomial tests against chance $1$7, $1$8 tests with Cramer’s $1$9 effect sizes for group comparisons, and logistic regression with clustered standard errors for multivariate factors (Ashkinaze et al., 3 Nov 2025). In the reported implementation, models answered $0$0 of trials, and the analysis dataset contains $0$1 test decisions across nine models (Ashkinaze et al., 3 Nov 2025).

Component	DVB specification
Output format	“Option A” or “Option B” only
Prompt structure	One test per prompt
Training-shot counts	$0$2
Valid-trial scoring	$0$3 for $0$4, else $0$5
Chance baseline	$0$6
Main CI	Wilson
Main significance test	Binomial against $0$7

This estimation procedure makes DVGR an operational proportion rather than a latent construct inferred indirectly from aggregate utility, questionnaire scores, or free-form rationales.

5. Empirical results in the Deep Value Benchmark

Across nine LLMs, the average DVGR is approximately $0$8, and all models generalized deep values less than chance, with binomial $0$9 for each model (Ashkinaze et al., 3 Nov 2025). The reported raw per-model estimates, all below $0.5$0, are: gpt-4.1-mini $0.5$1, meta-llama-3-70b-instruct $0.5$2, gpt-4.1 $0.5$3, gpt-4o $0.5$4, gpt-4o-mini $0.5$5, gemini-2.0-flash-lite $0.5$6, gpt-4.1-nano $0.5$7, meta-llama-3-8b-instruct $0.5$8, and gemini-2.0-flash $0.5$9, each with Wilson intervals entirely below chance (Ashkinaze et al., 3 Nov 2025).

A model-size analysis found that paired comparisons of families, comparing small vs large models, show smaller models often have slightly higher DVGR in $D \in \{d_1,d_2\}$0 pairs, with small mean absolute differences of about $D \in \{d_1,d_2\}$1, though statistically significant given the sample size; an omnibus $D \in \{d_1,d_2\}$2 test also favors smaller models (Ashkinaze et al., 3 Nov 2025). The paper states that scale does not solve value generalization and that DVGR is not emergent with size, echoing inverse-scaling results in other alignment-relevant areas such as truthfulness and sycophancy (Ashkinaze et al., 3 Nov 2025). This suggests that the relevant inductive bias is not simply improved by scaling the same preference-learning paradigm.

Factor analyses show heterogeneous but bounded variation across benchmark dimensions (Ashkinaze et al., 3 Nov 2025). Context effects are small, with Cramer’s $D \in \{d_1,d_2\}$3: higher DVGR appears in commerce, healthcare, and finance, and lower DVGR in communication, education, and customer service (Ashkinaze et al., 3 Nov 2025). The number of in-context examples has negligible effect, with Cramer’s $D \in \{d_1,d_2\}$4, and DVGR remains approximately $D \in \{d_1,d_2\}$5 regardless of whether $D \in \{d_1,d_2\}$6 demonstrations are used (Ashkinaze et al., 3 Nov 2025). Value identity matters somewhat more, with Cramer’s $D \in \{d_1,d_2\}$7: tradition reaches approximately $D \in \{d_1,d_2\}$8, universalism approximately $D \in \{d_1,d_2\}$9, and lower DVGRs are reported for fidelity and self-improvement (Ashkinaze et al., 3 Nov 2025).

Agreement analyses further indicate that models tend to make similar choices. Pairwise agreement is approximately $D$00 overall, with within-developer pairs agreeing more, approximately $D$01, than cross-developer pairs, approximately $D$02, a difference of about $D$03 percentage points with $D$04 (Ashkinaze et al., 3 Nov 2025). A plausible implication is that low DVGR is not idiosyncratic to a single model family but reflects a shared response tendency under the benchmark’s controlled confounds.

6. Validation, interpretability, and alignment significance

The benchmark includes three human validation experiments intended to establish construct validity, internal validity, and external validity (Ashkinaze et al., 3 Nov 2025). In the shallow-versus-deep distinction study, $D$05 Prolific participants produced shallowness ratings that robustly separate deep values from shallow preferences, with means $D$06 versus $D$07 on the $D$08 to $D$09 scale, Cohen’s $D$10, and mixed-model $D$11, $D$12 (Ashkinaze et al., 3 Nov 2025). Binary accuracy on selected shallow preferences is approximately $D$13 (Ashkinaze et al., 3 Nov 2025).

Completion Validation 1 used $D$14 participants over $D$15 trials and found that participants predicted the user would choose the value-aligned option in approximately $D$16 of cases, with $D$17 CI $D$18 and $D$19 against chance (Ashkinaze et al., 3 Nov 2025). Completion Validation 2 used $D$20 participants over $D$21 trials and found approximately $D$22 accuracy in mapping options to intended $D$23 pairs, with $D$24 CI $D$25 and $D$26 (Ashkinaze et al., 3 Nov 2025). LLMs also succeed when explicitly told which values and preferences each option embodies, with AI performance of approximately $D$27 in Validation 1 and approximately $D$28 in Validation 2 (Ashkinaze et al., 3 Nov 2025). The paper interprets this as suggesting that the difficult component for LLMs is inferring the underlying value from preference patterns when not told (Ashkinaze et al., 3 Nov 2025).

The benchmark’s alignment significance derives from its controlled causal structure: it directly probes whether a model learns deep moral value or shallow correlates from preference data (Ashkinaze et al., 3 Nov 2025). A low DVGR indicates a tendency to generalize superficial cues such as style or formal tone rather than the underlying moral principles in similar future contexts (Ashkinaze et al., 3 Nov 2025). The paper argues that this poses risks for AI assistants and agents that must act robustly under distribution shift in user contexts and values (Ashkinaze et al., 3 Nov 2025). Because DVGR complements other alignment benchmarks by directly probing value-versus-style generalization under controlled confounds rather than average-case performance or static value questionnaires, it occupies a distinct niche within alignment evaluation (Ashkinaze et al., 3 Nov 2025).

Follow-up prompting experiments reinforce this interpretation. Chain-of-Thought lowers pooled DVGR to $D$29 from a baseline of $D$30, whereas explicit instruction to prioritize deep values raises it modestly to $D$31, but still below chance (Ashkinaze et al., 3 Nov 2025). Examples include gpt-4.1 moving from $D$32 at baseline to $D$33 with CoT and $D$34 with explicit instruction, and gemini-2.0-flash moving from $D$35 to $D$36 under explicit instruction (Ashkinaze et al., 3 Nov 2025). This suggests that prompting alone does not resolve the underlying inference failure.

7. Limitations, terminological ambiguity, and future directions

The DVB paper identifies several threats to validity and scope conditions (Ashkinaze et al., 3 Nov 2025). The perfect confound in training is a “worst-case” design: real-world correlations may be partial rather than perfect, even though the perfect-confound setup enables a clean measure (Ashkinaze et al., 3 Nov 2025). Deep values do not always fully determine choices, and the paper notes that differences across models, values, and contexts may be more informative than absolute levels (Ashkinaze et al., 3 Nov 2025). The assessment is inference-only, using in-context learning rather than finetuning, so it may miss capabilities unlockable by post-training; the evaluation instead targets off-the-shelf models as commonly deployed (Ashkinaze et al., 3 Nov 2025). Domain coverage is limited to the selected values, preferences, and eight domain clusters, and shallow features may inadvertently encode value content despite filtering for neutrality and breadth (Ashkinaze et al., 3 Nov 2025).

The paper reports several mitigations: human validations confirm embodiment and separability; the factorial design balances values appearing as preferred and dispreferred across tuples; and randomization of option order together with one-test-per-prompt reduces positional and contamination biases (Ashkinaze et al., 3 Nov 2025). These do not remove all threats, but they strengthen the claim that low DVGR is not merely an artifact of prompt format or label leakage.

Future directions proposed in the benchmark paper include debiased or counterfactual preference datasets where value and style are explicitly decorrelated, value-focused objectives or auxiliary losses that penalize reliance on shallow features under synthetic decoupling, causal disentanglement methods to separate value representations from style features within LLM internals, and post-training finetuning specifically to increase DVGR and measure downstream behavioral benefits (Ashkinaze et al., 3 Nov 2025). Evaluation extensions include more deep values, more shallow features, multi-turn interaction contexts, cross-lingual tests, domain-specific verticals such as healthcare subdomains, and interpretability analyses such as linear probes to identify representations mediating value alignment versus shallow style reliance (Ashkinaze et al., 3 Nov 2025).

A distinct issue is terminological ambiguity. In "Predictive Value Generalization Bounds" (Vemuri et al., 2020), the acronym DVGR is used for the leading-order rate controlling PPV/NPV estimation error under distribution-free uniform convergence bounds. There, for a function class $D$37 and operating point $D$38, the quantities

$D$39

are defined through bounds depending on $D$40, $D$41, $D$42, and either the order coefficient $D$43 or the VC-subgraph dimension $D$44 (Vemuri et al., 2020). This is conceptually unrelated to DVB’s behavioral metric, despite sharing the acronym. A plausible implication is that citations should specify whether “DVGR” refers to a benchmark score for value-versus-style generalization in LLMs (Ashkinaze et al., 3 Nov 2025) or to a theoretical generalization-rate quantity for predictive values of thresholded scoring functions (Vemuri et al., 2020).

In its benchmark sense, DVGR provides a clear quantitative lens on whether models trained on preference signals learn to generalize underlying deep values or merely correlated shallow features (Ashkinaze et al., 3 Nov 2025). In the reported DVB experiments, contemporary LLMs generalize deep values less than chance on average, with only modest improvements under explicit instruction (Ashkinaze et al., 3 Nov 2025). Within alignment research, this makes DVGR a targeted measure of robustness to confounded preference learning rather than a broad assessment of moral competence or normative correctness.