ADC Metric for LLM Alignment

• The ADC metric is a multidimensional framework that quantifies misalignment in LLM outputs by decomposing overall alignment into interpretable behavioral and strategic gap scores.
• It evaluates linguistic, emotional, and strategic dimensions using statistical measures like JSD, DTW, and SEM to benchmark model performance against human interactions.
• By identifying specific alignment failures through gap metrics and proxy-policy conflicts, the ADC method provides actionable insights for targeted model improvements.

The Alignment Dimension Conflict (ADC) metric encompasses a suite of multidimensional evaluation frameworks designed to quantify types of misalignment or internal conflict in the outputs and decision-making behaviors of LLMs. Originating across domains—behavioral imitation, value-driven decision making, and reward-model-based fine-tuning—ADC approaches share the focus on decomposing overall alignment into distinct, interpretable dimensions. These methods emphasize rigorous quantitative comparison to human benchmarks and reveal sources of failure or divergence that single-metric scores alone obscure.

1. Theoretical Motivation and Conceptual Scope

ADC metrics are rooted in the observation that alignment in AI systems is rarely unidimensional. Real-world settings typically involve multiple, sometimes conflicting, behavioral or normative axes: linguistic style, emotional expression, strategic reasoning, privacy tradeoffs, value-action relations, or model–reward consistency. Rather than reporting a single global performance indicator, ADC frameworks systematically surface discrete “gaps” along theory-grounded dimensions. This approach enables targeted diagnosis of model deficiencies, more granular auditing for specific alignment failures, and principled evaluation of alignment techniques under adversarial or ambiguous objectives (Kwon et al., 19 Sep 2025, &&&1&&&, Liu et al., 10 Dec 2025).

2. Behavioral Alignment Metrics in Conflict Dialogue

The multi-dimensional ADC framework established in "Evaluating Behavioral Alignment in Conflict Dialogue" (Kwon et al., 19 Sep 2025) analyzes LLM-generated negotiation and dispute-resolution dialogues according to three behavioral dimensions:

• Linguistic-style alignment: Includes the LIWC-based construct gap (LG), comparing the distributions of psychologically relevant lexical features and dispute-specific terms between human–human (H2H) and LLM–LLM (L2L) conversations using Jensen–Shannon divergence (JSD). The LG score is split between IRP-related and general dispute features. The linguistic-entrainment gap (LEG) quantifies dyadic turn-level coordination using normalized word mover's distances between utterance embeddings.
• Emotional-expression alignment: Captures both temporal (anger-trajectory gap, ATG) and scalar (anger-magnitude gap, AMG) features of affective expression. ATG evaluates the shape similarity of anger-intensity time series via dynamic time warping (DTW); AMG computes the area-under-curve difference in anger scores, operationalizing average intensity.
• Strategic-behavior alignment: The IRP-strategy gap (SBG) measures categorical disagreement in distributions over eight negotiation move classes and a residual, using JSD.

Each gap metric is referenced relative to a within-human baseline and interpreted such that a value of zero signals equivalence between LLM and human inter-model alignment variability. Independent two-sample t-tests provide the basis for statistical significance determination. LG, LEG, ATG, AMG, and SBG scores are reported side-by-side, explicitly not aggregated into a single ADC scalar by the authors, on the grounds that no principled weighting can be assumed a priori (Kwon et al., 19 Sep 2025).

Dimension	Gap Metric	Statistic
Linguistic	LG, LEG	JSD, nCLiD
Emotional	ATG, AMG	DTW, AUC
Strategic	SBG	JSD over IRP

This multi-gap paradigm provides benchmarking for where, and by how much, LLM behaviors systematically diverge from human reference distributions on each behavioral axis.

3. Value-Action Conflict Modeling and VAAR

In value-based decision tasks, ADC is formalized through structural equation modeling (SEM) of latent-to-action relationships. Chen et al.'s Value-Action Alignment Rate (VAAR) extends ADC to settings where underlying values (privacy concern, prosocialness) exert conflicting influences on observable decisions (data sharing) (Chen et al., 7 Jan 2026).

• MGSEM operationalization: Privacy is modeled as a latent factor with multi-item measurement; prosocialness as a unidimensional observed variable.
• Structural paths: Both values simultaneously predict three action outcomes (SacrificePrivacy, PastAcceptance, FutureWillingness). For each model, the SEM extracts path coefficients $\beta_{\rm Priv\to k}$ and $\beta_{\rm PSA\to k}$, which are then evaluated for sign and magnitude.
• ADC as directional agreement: VAAR aggregates over the six path-level z-scores against a human-referenced direction template; all PSA→AoDS paths should be positive, Privacy→AoDS negative. Each path's probability of correct sign (via tail probability of $z$-score) is converted to log-loss and averaged:

$$\mathrm{VAAR}(g) = \frac{1}{|\mathcal L_g|}\sum_{\ell\in\mathcal L_g} -\log\Phi(a_\ell z_\ell^{(g)})$$

• Interpretation: Lower VAAR scores indicate strong and confident alignment to expected joint value–action directions; high scores (or mis-signed estimates) reveal dimension conflicts where LLMs violate established human causal structures.

Reported VAAR scores span strong (GPT-4o: 0.111), moderate (Amazon: 0.474), weak (GPT-3.5: 0.858), to misaligned (Qwen3: 4.914), illustrating both model-specific and path-specific heterogeneity (Chen et al., 7 Jan 2026).

4. Reward-Model Alignment: Proxy-Policy Conflict and ADC

In reward-model-based training, ADC is constructed as a two-dimensional signal quantifying disagreements between the base policy and the proxy reward model. "Targeting Misalignment: A Conflict-Aware Framework for Reward-Model-based LLM Alignment" introduces:

• Proxy-Policy Alignment Conflict Score (PACS): For a prompt–completion pair $(x, y)$, PACS is the absolute difference between the proxy reward (normalized within batch) and the base policy’s normalized log-probability:

$$\mathrm{PACS}(x,y) = \left| \frac{r_{\mathit{proxy}}(x, y) - \mu_r^x}{\sigma_r^x} - \frac{\log\pi_{\mathit{base}}(y|x) - \mu_{\pi}^x}{\sigma_{\pi}^x}\right|$$

A higher PACS indicates strong local discordance between what the policy would produce and what the reward model would prefer (Liu et al., 10 Dec 2025).

• Global Kendall-Tau Distance (K-T): For each prompt $x$, K-T measures ranking disagreement across $N$ completions according to policy probabilities versus proxy rewards:

$\mathrm{K\!-\!T}(x) = \frac{C-D}{0.5N(N-1)}$

where $C$ (concordant) and $D$ (discordant) count relative ordering matches vs. mismatches.

• ADC in this context: ADC is the ordered pair $$(\mathbb{E}_y[\mathrm{PACS}(x,y)], 1 - \mathrm{K\!-\!T}(x))$$, or, in scalarized form, a simple linear combination. This provides a direct operationalization: local conflict (pointwise PACS) and global conflict (batchwise ranking divergence).
• Algorithmic usage: The SHF-CAS algorithm thresholds both ADC dimensions to target high-conflict (high ADC) samples for human re-labeling, efficiently directing feedback to areas of greatest uncertainty or model–reward disagreement (Liu et al., 10 Dec 2025).

5. Statistical Baselines, Feature Construction, and Interpretation

Across instantiations, ADC metrics are defined relative to appropriate human or in-distribution baselines using formally specified statistical tests. Common conventions:

• Gap computation: $$(\text{Within-human baseline}) - (\text{Model–human distance})$$ or similar, with zero indicating that LLM variability matches that of human–human pairs.
• Significance: Independent two-sample t-tests are employed to assess whether observed gaps are substantially different from reference baselines (p-values denoted by stars).
• Feature normalization: All construction steps apply batch- or context-level normalization to ensure comparability of scale and statistical power.
• Interpretation: Positive ADC gap/misalignment indicates model divergence from human structure, with each dimension interpretable in isolation. Negative values (rare) would signal LLMs more closely fit to the reference than humans are to each other.

6. Generalization and Implications for Future Evaluation

The ADC framework widely generalizes: any alignment problem involving multiple competing sources of supervision, values, or target behaviors can admit a family of dimension-specific gap metrics. On the basis of latent theoretical structure (e.g., multi-factor SEM, proxy-policy agreement, behavioral imitation dimensions), one can define directional templates, gap statistics, and aggregate or vector-valued ADC measures suitable to the context (Kwon et al., 19 Sep 2025, Chen et al., 7 Jan 2026, Liu et al., 10 Dec 2025). A plausible implication is that as model capabilities and use cases expand, auditing and benchmarking protocols will increasingly require a panel of ADC scores rather than single-number performance summaries to diagnose and repair alignment failures in a targeted, interpretable manner.