Technical Summary: Frame-Conditioned Moral Computation in LLaMA 3.1-8B-Instruct

1. Problem Statement

Behavioral audits of Large Language Models (LLMs) on moral prompts measure the model's output text but fail to reveal the internal computation producing it. While frontier models often generate articulate moral reasoning, it remains unclear whether they possess dedicated moral-reasoning circuits or merely assemble responses from general-purpose domain representations selected by the prompt's surface frame. This paper addresses the insufficiency of output-level fluency as evidence of alignment, proposing that without mechanistic interpretability, one cannot distinguish between a model with a privileged ethical pathway and one that is "right for the wrong reasons" (e.g., using a counting subroutine rather than moral reasoning).

2. Methodology

The study employs Transluce, an AI-driven mechanistic interpretability platform, to audit LLaMA 3.1-8B-Instruct. The analysis covers 54 moral prompts across four distinct batteries:

• Battery 1 (B1): 17 prompts including classic dilemmas, policy questions, and meta-ethical reflections.
• Battery 3 (B3): 6 extended role-playing scenarios.
• Battery 4 (B4): 15 trolley problem variants where the switching mechanism varies (mechanical, technological, psychological, etc.) while identities remain anonymous.
• Battery 5 (B5): 16 trolley problem variants where the mechanism is fixed (lever) but the identity attributes of the victims vary (faith, party, nationality, kinship, etc.).

Metrics and Analysis:
The authors define two complementary metric families to analyze activated neurons and clusters:

1. Cluster-Level Metrics (5 metrics): Including Moral Neuron Fraction (MNF), Moral-to-Situational Activation Ratio (MSAR), Displacement Index (DI), Weighted Moral Fraction (WMF), and Cumulative Computational Bandwidth (CCB). These measure the footprint and rank of ethics-labeled clusters relative to situational ones.
2. Neuron-Level Metrics (6 metrics): Including Ethics Proportion, Top-10 Density, First Rank, Priority Ratio, Layer Centroid, and Layer Dispersion. These track individual neuron activation patterns.

Neurons and clusters are classified as "Moral," "Situational," or "Noise" using regex-based dictionaries and hand-audited exclusion lists. The study also includes a multi-temperature audit (T ∈ {0.0, ..., 0.5}) and a behavioral proxy comparison with two frontier models (Claude Opus 4 and Gemini 3 Pro).

3. Key Contributions

1. Dual-Level Mechanistic Audit: The paper maps the model's response to 54 prompts using both cluster and neuron metrics, demonstrating that both views converge on the same finding: domain-specific representations dominate the top of every activation list.
2. Constant Capacity, Variable Salience: The fraction of activated neurons labeled as ethics-related is essentially constant (~5%) across all batteries, regardless of whether the prompt is explicitly a moral dilemma. What changes with framing is the rank and priority of these neurons, not their number.
3. Controlled Trolley Contrast (B4 vs. B5): Aggregate ethics metrics are statistically indistinguishable between mechanism-varying (B4) and identity-varying (B5) prompts. However, the dominant non-ethics distractor mirrors the design: mechanism distractors fragment across neurons in B4, while a single "lever-trigger" neuron dominates B5.
4. Invariant Ethics Circuitry: A multi-temperature audit identifies a candidate ethics neuron (L16/N3837) that remains stable at rank 3 across temperatures, suggesting a candidate feature for future causal investigation.
5. Behavioral-Mechanistic Gap: A cross-model proxy on frontier models reveals a ~2.3x divergence in self-reported moral focus (Behavioral MNF) on identical prompts, despite convergent utilitarian conclusions. This supports the "Alignment Wrapper" hypothesis, where RLHF re-orders surface text without removing underlying domain-first frames.
6. Frame-Conditioned Moral Computation: The authors unify these findings into a theoretical framework where the prompt's surface vocabulary selects a feature manifold (e.g., sports, law, medicine), and the moral conclusion is downstream of that selection.

4. Key Results

The Situational Anchor Effect

Across all four batteries, situational (domain-specific) neurons outnumber moral neurons by roughly 2–3:1.

• B1: Sports clusters appear in the top 10 for 10 of 17 prompts, even those with no athletic content (e.g., Prisoner's Dilemma). "Tied up" activates bondage clusters; "dilemma" triggers gaming clusters.
• B3: Role-playing scenarios (e.g., Memory Surgeon, Pharma CEO) load heavy domain-specific clusters (medical, legal, corporate) before any ethics cluster appears.
• Universal Mathematical Anchor: Prompts involving ranking, weighing, or stages (Haidt MFT, Kohlberg, Ultimatum Game) consistently route through mathematical and logical-proof circuitry.

Salience vs. Capacity

• Capacity: The Ethics Proportion remains flat at ~5.1–5.7% across all batteries. The model does not recruit more ethics circuitry for explicit moral dilemmas.
• Salience: Under trolley framing (B4, B5), the First Rank of the first ethics neuron drops from a median of 9 (in B1/B3) to 3. Top-10 Density doubles from ~0.10 to ~0.20. The Priority Ratio (strength of the top ethics neuron relative to the absolute top) rises from ~0.45 to ~0.69.
• Conclusion: The same ethics neurons fire earlier and harder under specific framing, but the top of the activation list remains dominated by non-ethics concepts (e.g., the lever mechanism or identity attributes).

The B4 vs. B5 Contrast

• Mechanism Variation (B4): Distractors fragment. Different mechanisms (laser, software, military order) activate distinct domain-specific neurons (optics, programming, military hierarchy).
• Identity Variation (B5): A single "lever-trigger" neuron (L22/N5489) dominates the top non-ethics list in 7 of 16 prompts (43.8%), compared to only 1 of 15 in B4.
• Confidence Shift: In B5, the model's stated confidence drops (mean 7.81 vs. 8.33 in B4), particularly when the prompt requires sacrificing an in-group member to save an out-group. However, the utilitarian conclusion (pull the lever) remains stable.

Cross-Model Divergence

On identical prompts, Claude Opus 4 reported an ethics-first association profile (65% Behavioral MNF), while Gemini 3 Pro reported a domain-first profile (~28% Behavioral MNF). This suggests that output-level self-reports may be shaped by an "Alignment Wrapper" that re-orders text without altering the underlying domain-first computation.

5. Significance and Claims

The paper claims that behavioral alignment is insufficient as an audit standard. A model can produce morally correct outputs while its internal computation is dominated by surface-feature distractors (e.g., sports, math, identity). The authors argue for a shift toward Mechanistic Alignment, defined as the property where ethics-relevant features are causally privileged under controlled frame variation, rather than merely being "loud" in the final explanation.

The findings support the Frame-Conditioned Moral Computation hypothesis: moral output is downstream of a frame-selection step where surface vocabulary chooses a feature manifold. This implies that:

• Moral reasoning in LLMs may be an "assembly" of non-moral representations rather than a monolithic moral circuit.
• Current alignment (RLHF) may act as a wrapper that suppresses or re-orders domain frames in the output without changing the internal selection mechanism.
• Future safety evaluations must move beyond output-level text analysis to include mechanistic interventions (e.g., SAE-based analysis, causal patching) to verify if ethics features are truly privileged.

The authors remain modest, noting that their mechanistic findings are specific to LLaMA 3.1-8B-Instruct and that the "Invariant Ethics Neuron" is an associational candidate requiring causal intervention to confirm. They do not claim to have proven the absence of moral circuits in frontier models, but rather that output convergence does not license inferring the presence of such circuits.