ValueLocate: Neuron-Level Value Alignment

• ValueLocate is a mechanistic interpretability framework that operationalizes psychological value theory by linking internal activations to explicit value-laden outputs.
• It builds on a taxonomy-driven ValueInsight dataset, employing contrasting prompts to isolate neurons associated with dimensions like Conservation and Self-Enhancement.
• The framework enables causal neuron editing via forward-pass activations, yielding reproducible shifts in LLM behavior validated through G-EVAL scores.

ValueLocate is a mechanistic interpretability framework designed to uncover and causally probe value representations within LLMs. It operationalizes psychological value theory, specifically the Schwartz Values Survey, to localize and manipulate “value neurons” through a taxonomy-driven dataset and a contrastive activation methodology. The framework’s principal contribution is a reproducible pipeline for linking internal activation dynamics to explicit value-laden behavior, advancing foundations for neural-level value alignment in LLMs (Su et al., 23 May 2025).

1. Motivation and Theoretical Foundations

Mechanistic interpretability (MI) aims to reverse-engineer neural computations by mapping abstract behaviors to human-interpretable model components (e.g., neurons, FFN subunits, attention heads). Standard value alignment protocols in LLMs, such as reinforcement learning from human feedback (RLHF), ensure compliance at the output level but do not elucidate where or how values are structurally encoded. ValueLocate addresses this opacity by reframing value alignment as a question of causally identifiable neural activity. It grounds this approach in Schwartz’s value pairs (e.g., Conservation vs. its antithesis), leveraging the observation that neuron activations can systematically differ based on prompts targeting opposing value aspects. Through this differential analysis, ValueLocate identifies neurons whose activation state reliably tracks—and shapes—the model’s value-laden outputs (Su et al., 23 May 2025).

2. ValueInsight Dataset Construction

ValueLocate’s interpretability protocol depends critically on ValueInsight, a generative dataset inspired by the four higher-order dimensions of the Schwartz Values Survey: Openness to Change, Self-Transcendence, Conservation, and Self-Enhancement. Each dimension decomposes via a hierarchy Γ = (D, S, A), where D is dimension, S is subvalue, and A is atomic value. For comprehensive coverage:

• Each (subvalue, atomic value) and its antonym are instantiated through GPT-4o-generated second-person narratives (≥50 words), producing 80 positive and 80 negative descriptions per atomic value, yielding a total of 640 descriptions.
• Real-world situational questions (~15,000 in total) are crafted by pairing each atomic value with a bank of common topics (e.g., family, environment) and prompting GPT-4o to generate neutral, complex dilemmas.
• The dataset thus systematically samples both poles of value dimensions, enabling contrasting prompts to expose activation differences in LLMs (Su et al., 23 May 2025).

3. Value Neuron Identification via Activation Difference

The core identification method exploits the architecture of transformer FFNs. At each layer $\ell$ and position $i$, the residual update is $h_i^\ell = h_i^{\ell-1} + A_i^\ell + F_i^\ell$, with $$F_i^\ell = W_2^\ell \sigma(W_1^\ell(h_i^{\ell-1} + A_i^\ell))$$. A neuron $k$ in layer $\ell$ is indexed as (row $k$ of $W_1^\ell$, column $k$ of $W_2^\ell$).

Neuron activation is measured by summing the nonlinear activations $i$0 across all prompt tokens: $i$1, where $i$2 for prompt $i$3 of length $i$4. The activation probability for neuron $i$5 is defined over $i$6 prompts as $i$7.

Crucially, ValueLocate operationalizes contrast: for each value dimension, prompts are separated into positive and reversed (negative) sets. The activation difference

$i$8

(Equation 5 in source) quantifies the specificity of each neuron to a given value pole.

Pseudocode for $i$9 computation is provided as follows: $k$2 (Su et al., 23 May 2025).

4. Differentiable Localization of Value-Critical Neurons

Neuron selection eschews computationally intensive backpropagation-based attribution, relying exclusively on forward-pass activations and a fixed thresholding rule. For each value dimension:

• Two prompt sets ($h_i^\ell = h_i^{\ell-1} + A_i^\ell + F_i^\ell$0 and $h_i^\ell = h_i^{\ell-1} + A_i^\ell + F_i^\ell$1) are compiled.
• All prompts are passed through the model, logging $h_i^\ell = h_i^{\ell-1} + A_i^\ell + F_i^\ell$2 for every neuron.
• Thresholding $h_i^\ell = h_i^{\ell-1} + A_i^\ell + F_i^\ell$3 at $h_i^\ell = h_i^{\ell-1} + A_i^\ell + F_i^\ell$4 distinguishes neurons: $h_i^\ell = h_i^{\ell-1} + A_i^\ell + F_i^\ell$5 as positive-pole neurons, $h_i^\ell = h_i^{\ell-1} + A_i^\ell + F_i^\ell$6 as negative-pole neurons, with the remainder deemed neutral.
• The sets $h_i^\ell = h_i^{\ell-1} + A_i^\ell + F_i^\ell$7 and $h_i^\ell = h_i^{\ell-1} + A_i^\ell + F_i^\ell$8 define the dimension-specific value-critical neurons.

Fewer than 0.4% of all neurons in mid-layer FFNs (notably around layer 15) are typically value-critical, suggesting a highly sparse coding regime for ethical orientation (Su et al., 23 May 2025).

5. Causal Editing and Validation

Validation of value-neuron causality is performed via explicit neuron editing. For each neuron $h_i^\ell = h_i^{\ell-1} + A_i^\ell + F_i^\ell$9:

• The edited activation is defined by

$$F_i^\ell = W_2^\ell \sigma(W_1^\ell(h_i^{\ell-1} + A_i^\ell))$$0

(Equation 6), with dynamic scaling factor $$F_i^\ell = W_2^\ell \sigma(W_1^\ell(h_i^{\ell-1} + A_i^\ell))$$1 (default $$F_i^\ell = W_2^\ell \sigma(W_1^\ell(h_i^{\ell-1} + A_i^\ell))$$2).

• Editing can be inverted to push toward either value pole.

The evaluation protocol utilizes 100 held-out ValueInsight questions per dimension along with external validation sets (PVQ40, ValueBench). G-EVAL (1–5 scale, LLM-scored)—averaged over 10 random seeds—serves as the principal metric. Baselines include LPIP, QRNCA, and CGVST for neuron identification.

The protocol confirms that:

• Positive-pole neuron editing increases G-EVAL score for that value; negative-pole editing decreases it.
• Random neuron editing (up to 50 neurons) yields no systematic shift ($$F_i^\ell = W_2^\ell \sigma(W_1^\ell(h_i^{\ell-1} + A_i^\ell))$$30.05) in G-EVAL.
• Monotonic increase in score as $$F_i^\ell = W_2^\ell \sigma(W_1^\ell(h_i^{\ell-1} + A_i^\ell))$$4 increases (dose–response).
• Varying $$F_i^\ell = W_2^\ell \sigma(W_1^\ell(h_i^{\ell-1} + A_i^\ell))$$5 from 1% to 10% reveals a categorical tradeoff; $$F_i^\ell = W_2^\ell \sigma(W_1^\ell(h_i^{\ell-1} + A_i^\ell))$$6 (“elbow” point) is optimal (Su et al., 23 May 2025).

6. Empirical Findings and Observational Insights

Quantitative Results

The following G-EVAL mean $$F_i^\ell = W_2^\ell \sigma(W_1^\ell(h_i^{\ell-1} + A_i^\ell))$$7 std. scores on LLama-3.1-8B after positive editing ($$F_i^\ell = W_2^\ell \sigma(W_1^\ell(h_i^{\ell-1} + A_i^\ell))$$8) were observed:

Value Dimension	ValueLocate	Best Baseline (CGVST)
Openness to Change	4.68±0.06	4.42
Self-Transcendence	4.60±0.05	4.25
Conservation	4.15±0.09	3.85
Self-Enhancement	4.08±0.06	3.88

• Negative-pole editing produces symmetric G-EVAL drops (e.g., Self-Transcendence drops to $$F_i^\ell = W_2^\ell \sigma(W_1^\ell(h_i^{\ell-1} + A_i^\ell))$$92.30, baseline $k$02.55).
• Random neuron editing does not reliably affect value expression.

Qualitative Observations

• Pre-editing responses often straddle both value poles or hedge commitments.
• Positive-pole editing yields assertive, value-consistent outputs (“I would definitely prioritize...”, “My top concern is...”).
• Negative-pole editing reverses the stance with coherent shifts in rhetorical tone.
• Value-critical neurons are functionally concentrated at select FFN mid-layer units; this supports the notion of modular semantic encoding.

7. Significance and Limitations

ValueLocate establishes a mechanistically transparent pipeline that directly connects psychological value theory to interpretable neural dynamics in transformer LLMs. This enables not only identification but also direct, causal control of LLM value orientation, with empirical validation outperforming alternative neuron discovery regimes on both in-sample and out-of-sample prompts. By demonstrating that targeted neuron editing can predictably modulate value-laden behavior, ValueLocate enhances the tractability of neural value alignment and informs future work on ethical model deployment.

A limitation is the reliance on the coverage and specificity of the ValueInsight dataset, and the thresholding strategy ($k$1) requires use-case tuning, trading off edit magnitude and response quality. The framework avoids attributional complexity but may miss distributed or sub-neuronal value traces.

A plausible implication is that such neuron-level analyses could support value auditing or fine-grained model editing beyond behavioral alignment, bridging gaps between psychology-grounded value systems and mechanistic interpretability infrastructure in computational models (Su et al., 23 May 2025).