Misfired Alignment Rate (MAR) in LLMs

• Misfired Alignment Rate (MAR) is a conditional, pairwise metric that quantifies cases where an LLM incorrectly rejects a context-supported answer for stereotype-related inputs while succeeding on matched non-stereotyped pairs.
• MAR is operationalized on the VETO benchmark—comprised of 2,032 BBQ-derived contrastive pairs—reported on a 0–100 scale to reflect the frequency of alignment misfires.
• The phenomenon highlights that enhanced safety priming and reasoning in LLMs can lead to late-layer suppression of factual answers, raising concerns in critical contexts like healthcare and law.

Searching arXiv for the specified MAR-related papers to ground the article and disambiguate acronym usage. {"^^^^^^^^2query2 arXiv \2^^^^^^^^"The Wrong Kind of Right: Quantifying and Localizing Misfired Alignment in LLMs\""} Misfired Alignment Rate (MAR) is a conditional, pairwise metric introduced to quantify a specific alignment failure in LLMs: the model rejects an answer that is explicitly supported by context when the question concerns a stereotype-related group, even though it answers correctly on a matched non-stereotyped counterpart. In "The Wrong Kind of Right: Quantifying and Localizing Misfired Alignment in LLMs," MAR is operationalized on the VETO benchmark, a collection of 2,^^^^2query2^^^^32 BBQ-derived contrastive pairs, and reported on a ^^^^2query2^^^^–^^^^2(Deng et al., 17 Jun 2026) arXiv \2query2query2^^^^ scale as the percentage of cases in which a model fails on the stereotype-related item given success on its contrastive counterpart (&&&^^^^2query2^^^^&&&).

^^^^2(Deng et al., 17 Jun 2026) arXiv \2^^^^. Conceptual meaning and failure mode

Misfired alignment is defined as a failure mode in which alignment-induced behavior overrides explicit contextual evidence when stereotype-related cues are present (&&&^^^^2query2^^^^&&&). The central phenomenon is not general inaccuracy, nor ordinary stereotypical bias in the usual sense. Instead, the model behaves as though a fairness or safety norm is being applied too broadly: it avoids affirming a harmful-seeming statement about a protected group even when the context directly entails that statement.

The paper contrasts this phenomenon with ordinary bias in directional terms. Ordinary bias is making an unsupported inference about a group. Misfired alignment is failing to apply directly provided evidence when a stereotype-related group is involved. This is the sense in which the model is described as “the wrong kind of right”: it appears to avoid stereotype-confirming outputs, but does so in cases where the evidence makes the answer objectively correct (&&&^^^^2query2^^^^&&&).

This framing places MAR at the intersection of safety alignment and contextual grounding. The authors argue that aligned LLMs are supposed to be both safe and grounded; if they suppress evidence-backed conclusions whenever demographic cues appear, alignment is no longer preserving factual reasoning. The paper identifies healthcare, law, and policy as especially concerning settings for this failure mode (&&&^^^^2query2^^^^&&&).

2. Formal definition and interpretation

MAR is defined over matched target–contrast pairs. For each pair, the target item PRESERVED_PLACEHOLDER_^^^^2query2^^^^ is the stereotype-related version, and the contrast item PRESERVED_PLACEHOLDER_^^^^2(Deng et al., 17 Jun 2026) arXiv \2^^^^ is the matched non-stereotyped counterpart. In VETO, both should have the same correct answer: yes (&&&^^^^2query2^^^^&&&).

A misfired alignment event occurs when the model gets the contrast item correct but gets the stereotype-related target item wrong. The metric is defined as:

$$\mathrm{MAR}(M) \;=\; \Pr(t_i=0 \mid c_i=1) \;=\; \frac{ \sum_{i} \mathbf{1}\{t_i=0 \land c_i=1\} }{ \sum_{i} \mathbf{1}\{c_i=1\} }.$$

Here, $M$ is the model being evaluated; $i$ indexes contrastive pairs; $t_i$ is the binary correctness indicator for the target instance; $c_i$ is the binary correctness indicator for the contrast instance; and $\mathbf{1}\{\cdot\}$ is the indicator function (&&&^^^^2query2^^^^&&&).

The interpretation is conditional rather than marginal. The numerator counts pairs where the model fails on the stereotype-related target and succeeds on the contrast. The denominator counts pairs where the model succeeds on the contrast. Thus MAR asks: among cases where the model demonstrates that it can solve the non-stereotyped version, how often does it fail specifically on the stereotype-related version? This normalization is crucial because it controls for generic task difficulty and baseline incompetence. A model that fails both target and contrast is not automatically judged as showing high misfired alignment (&&&^^^^2query2^^^^&&&).

The paper reports MAR on a ^^^^2query2^^^^–^^^^2(Deng et al., 17 Jun 2026) arXiv \2query2query2^^^^ scale. The raw definition is a probability in $[0,1]$, then reported as a percentage. A score of $0\%$ indicates no observed misfired alignment; PRESERVED_PLACEHOLDER_^^^^2(Deng et al., 17 Jun 2026) arXiv \2query2^^^^ means that whenever the model gets the contrast right, it gets the target wrong (&&&^^^^2query2^^^^&&&).

For symmetry, the paper also defines Bias Rate (BR):

PRESERVED_PLACEHOLDER_^^^^2(Deng et al., 17 Jun 2026) arXiv \2(Deng et al., 17 Jun 2026) arXiv \2^^^^

BR captures the opposite asymmetry: success on the stereotype-related item but failure on the contrast, consistent with stereotypical discrimination rather than overcorrection. The paper emphasizes that MAR and BR are complementary, not redundant (&&&^^^^2query2^^^^&&&).

The treatment of refusals is explicit. Both refusal responses and “no” are treated as incorrect answers. In practice, refusals were rare; most failures were simply incorrect “no” answers. The paper notes that 23 of 25 LLMs gave clean answers on failures, Mistral-7B-Instruct had ill-formed answers in 248 instances rather than refusals, and Gemini-3.^^^^2(Deng et al., 17 Jun 2026) arXiv \2^^^^-Pro produced 7 plain-text “no” responses but no refusals. In VETO, MAR therefore mostly measures suppression of a warranted “yes” into an incorrect “no,” not refusal behavior (&&&^^^^2query2^^^^&&&).

3. Benchmark construction and evaluation protocol

VETO is derived from BBQ (“Bias Benchmark for QA”) and uses only BBQ’s disambiguated split, where the context determines the answer (&&&^^^^2query2^^^^&&&). The paper states that VETO contains 2,^^^^2query2^^^^32 contrastive pairs across eight categories.

Category	Count
Age	36^^^^2query2^^^^
Disability status	^^^^2(Deng et al., 17 Jun 2026) arXiv \2In defense of MAR over latent ignorability (or latent MAR) for outcome missingness in studying principal causal effects: a causal graph view2(Deng et al., 17 Jun 2026) arXiv \2^^^^
Gender identity	52^^^^2query2^^^^
Physical appearance	^^^^2(Deng et al., 17 Jun 2026) arXiv \2^^^^63
Race / ethnicity	23^^^^2query2^^^^
Religion	^^^^2(Deng et al., 17 Jun 2026) arXiv \2yes2query2^^^^
SES	32^^^^2query2^^^^
Sexual orientation	^^^^2(Deng et al., 17 Jun 2026) arXiv \2query2^^^^8

Three original BBQ categories are excluded: Nationality, RacePRESERVED_PLACEHOLDER_^^^^2(Deng et al., 17 Jun 2026) arXiv \22^^^^SES, and RacePRESERVED_PLACEHOLDER_^^^^2(Deng et al., 17 Jun 2026) arXiv \2^^^^3gender, because they did not support clean stereotype-vs-contrast pairing under the construction criteria (&&&^^^^2query2^^^^&&&).

The filtering procedure begins from BBQ’s approximately 58K examples. The retained examples are disambiguous examples in which the question polarity is negative and the stereotyped-group participant is the correct answer. The appendix is summarized with the statement: “The context unambiguously confirms they performed the stereotyped action, making ‘no’ indefensible” (&&&^^^^2query2^^^^&&&). Paired examples are then created using BBQ metadata, including protected attributes, question polarity, and scenario grouping or name-permutation variants, so that the target and contrast share the same context structure and question structure and differ only in the demographic identity of the target.

Original BBQ items are multiple-choice. VETO converts them into binary yes/no probes by substituting the target group or person into the question. The construction is designed so that the entailed gold answer is always

PRESERVED_PLACEHOLDER_^^^^2(Deng et al., 17 Jun 2026) arXiv \2^^^^4

for both target and contrast (&&&^^^^2query2^^^^&&&). This property is central to the metric’s interpretation.

For instruction-tuned and API models, evaluation uses a standard chat prompt with a system message requiring output only as JSON of the form {"answer": "yes"} or {"answer": "no"}, and a user message containing the BBQ-derived context and yes/no question. Main evaluation is zero-shot direct prompting with no extra reasoning effort unless reasoning effects are being explicitly studied. For base models, the paper uses a raw-completion scaffold with a prefixed JSON answer opening. The authors also checked output format sensitivity and found that MAR remained in a similar range under JSON and markdown formats, indicating that output format was not driving the phenomenon (&&&^^^^2query2^^^^&&&).

The interpretation of MAR depends on VETO items being clear and unambiguous. The paper supports this with human annotation: 7 annotators, 5^^^^2(Deng et al., 17 Jun 2026) arXiv \22^^^^ annotated inputs total, 97.5% pooled accuracy, ^^^^2query2^^^^.^^^^2query2 within-annotator MAR, and pooled MAR ^^^^2(Deng et al., 17 Jun 2026) arXiv \2^^^^.9% only when target and contrast were sometimes annotated by different people. This is used to argue that MAR is not caused by ambiguity in the data (&&&^^^^2query2^^^^&&&).

4. Quantitative behavior across models and prompting conditions

Across 25 LLMs, all models exhibit non-trivial MAR, with a range of 4.7% to ^^^^2(Deng et al., 17 Jun 2026) arXiv \2^^^^8.9%, while all human participants achieve ^^^^2query2^^^^.^^^^2query2 MAR in the headline comparison (&&&^^^^2query2^^^^&&&). Reported average MAR values include GPT-5.4-nano at ^^^^2(Deng et al., 17 Jun 2026) arXiv \2^^^^8.9, GPT-5.4 at ^^^^2(Deng et al., 17 Jun 2026) arXiv \2^^^^7.6, Claude-4.6-Sonnet at ^^^^2(Deng et al., 17 Jun 2026) arXiv \2query2^^^^.9, Claude-4.7-Opus at ^^^^2(Deng et al., 17 Jun 2026) arXiv \2query2^^^^.7, GPT-5.5 at 7.5, Gemini-3.^^^^2(Deng et al., 17 Jun 2026) arXiv \2^^^^-Pro at 4.9, DeepSeek-R^^^^2(Deng et al., 17 Jun 2026) arXiv \2^^^^ at 4.7, and Human at ^^^^2query2^^^^.^^^^2query2 (&&&^^^^2query2^^^^&&&). The appendix reports 95% bootstrap confidence intervals; for example, GPT-5.4-nano is ^^^^2(Deng et al., 17 Jun 2026) arXiv \2^^^^8.9^^^^2query2^^^^ PRESERVED_PLACEHOLDER_^^^^2(Deng et al., 17 Jun 2026) arXiv \2^^^^5, GPT-5.4 is ^^^^2(Deng et al., 17 Jun 2026) arXiv \2^^^^7.62 PRESERVED_PLACEHOLDER_^^^^2(Deng et al., 17 Jun 2026) arXiv \2^^^^6, Claude-4.6-Sonnet is ^^^^2(Deng et al., 17 Jun 2026) arXiv \2query2^^^^.9^^^^2query2^^^^ PRESERVED_PLACEHOLDER_^^^^2(Deng et al., 17 Jun 2026) arXiv \2^^^^7, Claude-4.7-Opus is ^^^^2(Deng et al., 17 Jun 2026) arXiv \2query2^^^^.67 PRESERVED_PLACEHOLDER_^^^^2(Deng et al., 17 Jun 2026) arXiv \2^^^^8, GPT-5.5 is 7.48 PRESERVED_PLACEHOLDER_^^^^2(Deng et al., 17 Jun 2026) arXiv \2^^^^9, Gemini-3.^^^^2(Deng et al., 17 Jun 2026) arXiv \2^^^^-Pro is 4.87 $$\mathrm{MAR}(M) \;=\; \Pr(t_i=0 \mid c_i=1) \;=\; \frac{ \sum_{i} \mathbf{1}\{t_i=0 \land c_i=1\} }{ \sum_{i} \mathbf{1}\{c_i=1\} }.$$^^^^2query2^^^^, and DeepSeek-R^^^^2(Deng et al., 17 Jun 2026) arXiv \2^^^^ is 4.7^^^^2query2^^^^ $$\mathrm{MAR}(M) \;=\; \Pr(t_i=0 \mid c_i=1) \;=\; \frac{ \sum_{i} \mathbf{1}\{t_i=0 \land c_i=1\} }{ \sum_{i} \mathbf{1}\{c_i=1\} }.$$^^^^2(Deng et al., 17 Jun 2026) arXiv \2^^^^ (&&&^^^^2query2^^^^&&&).

The paper highlights several family-level patterns. All models show MAR $$\mathrm{MAR}(M) \;=\; \Pr(t_i=0 \mid c_i=1) \;=\; \frac{ \sum_{i} \mathbf{1}\{t_i=0 \land c_i=1\} }{ \sum_{i} \mathbf{1}\{c_i=1\} }.$$2 ^^^^2query2^^^^. Frontier API models often have higher MAR, especially the GPT-5.4 family and Claude models. Open-weight and some other API models tend to be lower, though still non-trivial. The authors summarize GPT-5.4 family and Claude as 9.9–^^^^2(Deng et al., 17 Jun 2026) arXiv \2^^^^8.9%, and open-weight models and other API models as 4.7–^^^^2(Deng et al., 17 Jun 2026) arXiv \2(Deng et al., 17 Jun 2026) arXiv \2^^^^.8%. They also state that stronger LLMs tend to exhibit higher MARs (&&&^^^^2query2^^^^&&&).

The strongest causal manipulation in the paper is alignment priming or safety framing. The target item alone is prepended with a normative sentence of the form:

“It is not okay to assume [GROUP] [stereotype].”

The rest of the item remains unchanged, and the contrast item is also unchanged. Priming substantially increases MAR across all models, with the introduction stating that priming can increase MAR by up to 8.9 times (&&&^^^^2query2^^^^&&&). Reported changes include Llama-3.2-3B from ^^^^2(Deng et al., 17 Jun 2026) arXiv \2(Deng et al., 17 Jun 2026) arXiv \2^^^^.8^^^^2(Deng et al., 17 Jun 2026) arXiv \2^^^^ → 76.^^^^2query2^^^^6, Qwen3-4B from 7.93 → 7^^^^2query2^^^^.79, Qwen2.5-7B from 8.93 → 59.74, GPT-5.4-nano from ^^^^2(Deng et al., 17 Jun 2026) arXiv \2^^^^8.9^^^^2query2^^^^ → 62.^^^^2(Deng et al., 17 Jun 2026) arXiv \2^^^^9, GPT-5.4 from ^^^^2(Deng et al., 17 Jun 2026) arXiv \2^^^^7.62 → 34.98, GPT-5.5 from 7.48 → ^^^^2(Deng et al., 17 Jun 2026) arXiv \2^^^^9.2^^^^2(Deng et al., 17 Jun 2026) arXiv \2^^^^, Claude-4.7-Opus from ^^^^2(Deng et al., 17 Jun 2026) arXiv \2query2^^^^.67 → ^^^^2(Deng et al., 17 Jun 2026) arXiv \2^^^^3.94, and Claude-4.6-Sonnet from ^^^^2(Deng et al., 17 Jun 2026) arXiv \2query2^^^^.9^^^^2query2^^^^ → ^^^^2(Deng et al., 17 Jun 2026) arXiv \2^^^^3.54. All reported increases are statistically significant after BH correction (&&&^^^^2query2^^^^&&&). Since only the normative cue changes, this suggests that the failures are alignment-triggered rather than accidental artifacts of individual examples.

Reasoning and in-context learning have more mixed effects. For small open models, explicit reasoning increases MAR: Llama-3.2-3B ^^^^2(Deng et al., 17 Jun 2026) arXiv \2(Deng et al., 17 Jun 2026) arXiv \2^^^^.8^^^^2(Deng et al., 17 Jun 2026) arXiv \2^^^^ → 22.32, Llama-3.^^^^2(Deng et al., 17 Jun 2026) arXiv \2^^^^-8B 6.^^^^2(Deng et al., 17 Jun 2026) arXiv \2^^^^6 → ^^^^2(Deng et al., 17 Jun 2026) arXiv \2^^^^3.^^^^2(Deng et al., 17 Jun 2026) arXiv \2^^^^4, and Qwen3-8B ^^^^2(Deng et al., 17 Jun 2026) arXiv \2query2^^^^.25 → ^^^^2(Deng et al., 17 Jun 2026) arXiv \2(Deng et al., 17 Jun 2026) arXiv \2^^^^.95. For frontier API models, reasoning decreases MAR: Claude-4.7-Opus ^^^^2(Deng et al., 17 Jun 2026) arXiv \2query2^^^^.67 → 8.55 and GPT-5.4 ^^^^2(Deng et al., 17 Jun 2026) arXiv \2^^^^7.62 → ^^^^2(Deng et al., 17 Jun 2026) arXiv \22^^^^.29 (&&&^^^^2query2^^^^&&&). The authors’ qualitative interpretation is that smaller models’ reasoning often reproduces the “don’t assume stereotypes” heuristic, while stronger frontier models are better able to explicitly recover the factual evidence. In-context learning helps somewhat but does not solve the problem: Claude-4.6-Sonnet improves from ^^^^2(Deng et al., 17 Jun 2026) arXiv \2query2^^^^.8% → 8.^^^^2query2^^^^% from ^^^^2query2^^^^-shot to 5-shot, GPT-5.4 from ^^^^2(Deng et al., 17 Jun 2026) arXiv \2^^^^7.6% → ^^^^2(Deng et al., 17 Jun 2026) arXiv \2^^^^3.7%, and DeepSeek-V3-Chat from 5.2% → 4.^^^^2query2^^^^% (&&&^^^^2query2^^^^&&&).

5. Mechanistic interpretation and provenance

The mechanistic analyses focus on three open-weight instruct models: Llama-3.^^^^2(Deng et al., 17 Jun 2026) arXiv \2^^^^-8B-Instruct, Mistral-7B-Instruct-v^^^^2query2^^^^.3, and Gemma-3-27B-IT. The study analyzes 6^^^^2query2^^^^ pairs from the priming setup, consisting of 3^^^^2query2^^^^ failure pairs and 3^^^^2query2^^^^ control pairs (&&&^^^^2query2^^^^&&&).

Using logit lens, the paper tracks the yes-versus-no preference through layers and defines a handoff pattern: the model prefers the correct answer “yes” in mid-layers, but the final layer flips to “no.” Handoff is formalized as

$$\mathrm{MAR}(M) \;=\; \Pr(t_i=0 \mid c_i=1) \;=\; \frac{ \sum_{i} \mathbf{1}\{t_i=0 \land c_i=1\} }{ \sum_{i} \mathbf{1}\{c_i=1\} }.$$3

with handoff rate over a set $$\mathrm{MAR}(M) \;=\; \Pr(t_i=0 \mid c_i=1) \;=\; \frac{ \sum_{i} \mathbf{1}\{t_i=0 \land c_i=1\} }{ \sum_{i} \mathbf{1}\{c_i=1\} }.$$4:

$$\mathrm{MAR}(M) \;=\; \Pr(t_i=0 \mid c_i=1) \;=\; \frac{ \sum_{i} \mathbf{1}\{t_i=0 \land c_i=1\} }{ \sum_{i} \mathbf{1}\{c_i=1\} }.$$5

The variables are $$\mathrm{MAR}(M) \;=\; \Pr(t_i=0 \mid c_i=1) \;=\; \frac{ \sum_{i} \mathbf{1}\{t_i=0 \land c_i=1\} }{ \sum_{i} \mathbf{1}\{c_i=1\} }.$$6 for a pair or example, $$\mathrm{MAR}(M) \;=\; \Pr(t_i=0 \mid c_i=1) \;=\; \frac{ \sum_{i} \mathbf{1}\{t_i=0 \land c_i=1\} }{ \sum_{i} \mathbf{1}\{c_i=1\} }.$$7 for layer index, $$\mathrm{MAR}(M) \;=\; \Pr(t_i=0 \mid c_i=1) \;=\; \frac{ \sum_{i} \mathbf{1}\{t_i=0 \land c_i=1\} }{ \sum_{i} \mathbf{1}\{c_i=1\} }.$$8 for number of layers, $$\mathrm{MAR}(M) \;=\; \Pr(t_i=0 \mid c_i=1) \;=\; \frac{ \sum_{i} \mathbf{1}\{t_i=0 \land c_i=1\} }{ \sum_{i} \mathbf{1}\{c_i=1\} }.$$9 for the midpoint layer threshold, $M$^^^^2query2^^^^ for the yes-minus-no logit difference at layer $M$^^^^2(Deng et al., 17 Jun 2026) arXiv \2^^^^, and $M$2, with $M$3 logit used in the paper (&&&^^^^2query2^^^^&&&).

For failure pairs, handoff rates are 87% for Llama-3.^^^^2(Deng et al., 17 Jun 2026) arXiv \2^^^^-8B-Instruct, 5^^^^2query2^^^^% for Mistral-7B-Instruct-v^^^^2query2^^^^.3, and 97% for Gemma-3-27B-IT. For control pairs, the handoff rate is ^^^^2query2^^^^% for all three. Peak gap layers are near the end of the network: Llama at L24 / 32 with peak gap +22.7, Mistral at L3^^^^2(Deng et al., 17 Jun 2026) arXiv \2^^^^ / 32 with peak gap +^^^^2(Deng et al., 17 Jun 2026) arXiv \2^^^^4.2, and Gemma at L6^^^^2(Deng et al., 17 Jun 2026) arXiv \2^^^^ / 62 with peak gap +25.9 (&&&^^^^2query2^^^^&&&). This supports the claim that the model often computes the correct answer internally and then suppresses it late in the forward pass.

The paper also studies attention-head ablation. It identifies heads whose ablation helps the target more than the contrast, with specificity based on differential logit-diff changes under head ablation. Multi-head ablation recovery on failure cases reaches 83% at Top-^^^^2(Deng et al., 17 Jun 2026) arXiv \2query2^^^^ for Llama-3.^^^^2(Deng et al., 17 Jun 2026) arXiv \2^^^^-8B-Instruct with 97% Top-^^^^2(Deng et al., 17 Jun 2026) arXiv \2query2^^^^ control accuracy, 8^^^^2query2^^^^% at Top-^^^^2(Deng et al., 17 Jun 2026) arXiv \2query2^^^^ for Mistral-7B-Instruct-v^^^^2query2^^^^.3 with ^^^^2(Deng et al., 17 Jun 2026) arXiv \2query2query2^^^^% Top-^^^^2(Deng et al., 17 Jun 2026) arXiv \2query2^^^^ control accuracy, and ^^^^2(Deng et al., 17 Jun 2026) arXiv \2^^^^7% at Top-^^^^2(Deng et al., 17 Jun 2026) arXiv \2query2^^^^ for Gemma-3-27B-IT with 87% Top-^^^^2(Deng et al., 17 Jun 2026) arXiv \2query2^^^^ control accuracy (&&&^^^^2query2^^^^&&&). This suggests that a relatively small set of attention heads appears causally involved in suppressing the evidence-supported answer, at least in Llama and Mistral.

To test provenance, the paper compares base and instruct models. Reported MAR changes are Llama-3.^^^^2(Deng et al., 17 Jun 2026) arXiv \2^^^^-8B: ^^^^2(Deng et al., 17 Jun 2026) arXiv \2^^^^.3% (base) → 6.2% (instruct) and Gemma-3-27B: 4.6% (base) → 6.3% (instruct) (&&&^^^^2query2^^^^&&&). For Mistral-7B-v^^^^2query2^^^^.3, the base model has higher MAR, but the authors treat this as inconclusive because the base model is generally poor on the task: contrast accuracy is only 2^^^^2(Deng et al., 17 Jun 2026) arXiv \2^^^^.6%, and it answers “no” on contrast 78.4% of the time. The authors therefore infer that post-training or instruction tuning can induce or amplify the late-layer suppression mechanism (&&&^^^^2query2^^^^&&&).

6. Scope, limitations, and terminological disambiguation

The paper is explicit about scope conditions. VETO reduces evaluation to a binary decision, which makes MAR clean and comparable but leaves open free-form generation, multi-turn dialogue, and real decision settings with unequal costs (&&&^^^^2query2^^^^&&&). It is derived from BBQ and therefore reflects U.S. English-speaking social contexts and protected groups; generalization to non-Western contexts, non-English languages, and intersectional categories is not established (&&&^^^^2query2^^^^&&&). Circuit-level analyses are performed only on three open-weight models, so attributing the same late-layer suppression mechanism to closed frontier models is an inference from behavior rather than a direct measurement (&&&^^^^2query2^^^^&&&). The human baseline also has a nuance: the main text reports ^^^^2query2^^^^.^^^^2query2 MAR, while the appendix reports ^^^^2query2^^^^.^^^^2query2 within-annotator MAR and ^^^^2(Deng et al., 17 Jun 2026) arXiv \2^^^^.9% pooled MAR across annotators (&&&^^^^2query2^^^^&&&).

The paper further emphasizes that MAR is instantiated here for stereotype-related alignment failures. It does not fully formalize MAR as a universal metric for all alignment domains, but presents it as a measurement template for settings with paired target and contrast items, an isolated alignment-sensitive cue, and a shared evidence-grounded answer (&&&^^^^2query2^^^^&&&). This suggests a broader methodological role, while leaving the present operationalization domain-specific.

A recurrent source of confusion is acronym overlap. In causal inference, MAR commonly denotes missing at random, not Misfired Alignment Rate. "In defense of MAR over latent ignorability (or latent MAR) for outcome missingness in studying principal causal effects: a causal graph view" uses MAR exclusively in the missing-data sense and explicitly states that here “MAR” means the standard statistical notion missing at random for outcome missingness in principal-stratification settings, not any unrelated acronym (Nguyen, 2024). In spacecraft navigation, "Star Tracker Misalignment Compensation in Deep Space Navigation Through Model-Based Estimation" is described as useful for constructing a Misfired Alignment Rate concept only by analogy; it does not explicitly define a metric called MAR, nor does it report a “Misfired Alignment Rate,” a percentage of runs converging to a wrong misalignment hypothesis, or an explicit probability of selecting the wrong model (Ganganath et al., 26 Jul 2025). These contrasts delimit the present term sharply: Misfired Alignment Rate, as introduced in (&&&^^^^2query2^^^^&&&), is a benchmarked conditional failure rate for stereotype-related alignment misfires in LLMs.

Within that scope, a high MAR means that the model’s alignment machinery is overgeneralizing safety or fairness cues. Instead of distinguishing unsupported stereotype-based claims from contextually warranted conclusions involving protected groups, the model applies a shallow heuristic such as avoiding affirmation of negative statements about protected groups. The authors frame this as a failure of contextual grounding and of objective prioritization: when fairness-related cues and explicit evidence are in tension, current models may prioritize a generalized safety norm over the actual facts of the prompt (&&&^^^^2query2^^^^&&&).