Technical Summary: The Alignment Veto

Problem Statement

Current literature often assumes that when a Large Language Model (LLM) refuses to answer a culturally sensitive question, it lacks the relevant cultural knowledge (a representational failure). This paper challenges that assumption, positing that refusal may instead result from a "safety gate" blocking the expression of knowledge the model already possesses. The authors investigate whether alignment training suppresses internal cultural distributions without erasing them, specifically focusing on the Middle East and North Africa (MENA) region, where cultural values frequently diverge from the Western defaults prevalent in safety-training pipelines.

Methodology

The study employs a large-scale empirical evaluation across 16 MENA countries, 26 models (spanning 8 families and 4 alignment stages: Base, SFT, DPO, and Instruction-Tuned), and 1.53 million human survey responses.

Data and Evaluation:

• Questions: 864 questions derived from the World Values Survey (WVS) and the Arab Opinion Index (AOI), classified into three tiers: T1 (benign), T2 (moderate), and T3 (sensitive topics like LGBTQ+ acceptance, gender equality, and religious tolerance).
• Prompting: Responses were generated using three framing strategies (Neutral, Persona, Observer) in English and native languages (Arabic, Persian, Turkish).
• Metrics:
  • NVAS (Normalized Value Alignment Score): Measures the alignment between model outputs and human survey means.
  • Refusal Rate: The frequency of non-answer outputs.
  • EV-NVAS (Expected Value NVAS): A novel metric extracting the model's internal logit distribution at the exact moment of refusal to estimate what the model "would have said" if not blocked.
  • Safety Tax: The difference in refusal rates between sensitive (T3) and benign (T1) questions.

Mechanistic Analysis:

• Probing: Residualized ridge-regression probes were used to track cultural signal across network layers.
• Sparse Autoencoders (SAE): Applied to the Tulu-3-8B model to identify specific features mediating suppression.
• Ablation: Selective zeroing of identified SAE features to observe causal effects on T3 responses.

Key Findings

1. The Alignment Veto: Suppression vs. Erasure

The central finding is the "alignment veto." On sensitive (T3) questions, models often refuse to answer, yet their internal logit distributions at the moment of refusal correlate more strongly with human survey ground truth than their freely generated answers do.

• Evidence: For T3 questions, the EV-NVAS of refused rows is significantly higher than the NVAS of accepted rows (e.g., +0.029 difference under Persona framing).
• Implication: The model holds accurate cultural knowledge internally, but a safety-trained output gate intercepts and blocks its expression. This is distinct from "representational bias," where the model's encoding itself diverges from human values.

2. Two Distinct Failure Modes

The paper distinguishes between two types of cultural misalignment requiring different interventions:

• Suppression Failures: The model knows the answer but refuses to say it (high refusal, high internal EV-NVAS). These are often associated with LGBTQ+ topics. They can be partially mitigated via prompt engineering (e.g., third-person framing).
• Representational Bias Failures: The model answers but provides culturally inaccurate information (low refusal, low NVAS). These are often associated with gender equality topics and require training-time data correction, as no prompt intervention can bypass a lack of internal knowledge.

3. The Safety Tax and Inequity

Alignment training installs a "safety tax," defined as the excess refusal rate on sensitive content compared to benign content.

• Magnitude: The safety tax reaches up to 37.6% in some models (e.g., ALLAM-7B-IT).
• Inequity: The tax is distributed unevenly. There is a 19.8% gap in cultural alignment (NVAS) between the best-served nation (Palestine) and the worst-served (Algeria).
• Training Stage: The tax increases monotonically through the alignment pipeline (SFT → DPO → IT), with DPO identified as the stage where suppression mechanisms are primarily installed.

4. The Native Language Paradox

Contrary to the assumption that native-language prompting improves cultural accuracy, the study finds that switching from English to native languages (Arabic, Persian, Turkish) degrades alignment.

• Accuracy Drop: Native prompting reduces mean NVAS by 0.050 points across all models.
• Identity Collapse: When prompted in Arabic, responses from 14 Arabic-speaking countries collapse into a single cluster, losing country-specific nuance. SAE analysis reveals that language-family features dominate residual-stream representations, overwhelming country-specific signals.

5. Mechanistic Evidence in Tulu-3-8B

Using Sparse Autoencoders on Tulu-3-8B, the authors identified a candidate feature (Feature #5505) that:

• Activates selectively on T3 questions (28.6% of T3 items vs. <0.4% of T2 and 0% of T1).
• Is installed during the DPO stage.
• When ablated, it increases the predicted value for T3 questions without affecting benign (T1) content, suggesting a causal role in the suppression mechanism.

Significance and Claims

The paper claims that cultural alignment failures are not monolithic. By distinguishing between suppression (blocked knowledge) and representational bias (missing knowledge), the authors argue that current remediation efforts are often misdirected.

• Technical: Prompt engineering (specifically third-person framing) can partially bypass suppression gates, recovering accurate cultural distributions for specific failure modes. However, it cannot fix representational biases.
• Ethical/Societal: The "gate" is real, but its costs are unequal. The decision of what cultural values an LLM should represent is not purely technical; it demands transparent, participatory frameworks that account for the inequitable costs of safety alignment across different nations.
• Limitations: The mechanistic findings (SAE features) are specific to Tulu-3-8B and treated as candidate correlates rather than universal causal circuits. The study acknowledges that EV-NVAS values are sensitive to prompt wording and should be interpreted as population-level indicators.

The authors conclude that deciding what an LLM should protect is a societal question, not just a technical optimization problem, and that current alignment pipelines silently impose a "safety tax" that disproportionately affects specific cultural groups.