Echoes of Deeds: Moral History Can Shape and Steer LLM Behavioral Choices

Lucio La Cava; Andrea Tagarelli

cs.CLSep 28, 2026

Echoes of Deeds: Moral History Can Shape and Steer LLM Behavioral Choices

Authors: Lucio La Cava, Andrea Tagarelli

Organizations: DIMES Dept., University of Calabria, Italy

Abstract

Evaluations of Large Language Models (LLMs) morality typically consider decisions in isolation, thus overlooking whether an individual's unrelated prior conduct influences the model's subsequent choices. This leaves open the question of whether, and to what extent, moral history shapes LLM decisional behaviors. Prior work on human moral decision-making shows that past behavior can influence subsequent moral choices. Building on this observation, we investigate whether analogous effects emerge in LLMs in two complementary ways: at the behavioral level, through the model's observable responses, and at the representation level, through its latent internal representations. We introduce MoralLedger, a framework for studying how an actor's moral history shapes actions for LLMs' behaviors under a fixed decision context. At the behavioral level, we find that prior moral histories systematically alter subsequent choices as a function of their valence and intensity. At the internal representation level, these histories induce a linearly recoverable direction in the residual stream that generalizes to held-out examples. Intervening along this direction on neutral-history prompts produces two-sided intensity-dependent changes in subsequent choices, with effects that are stronger than those induced by prompting alone or by favorable-nonmoral direction. To our knowledge, this is the first demonstration that a latent representation of an actor's prior moral conduct can provide signed inference-time control over a moral decision. Our MoralLedger extends moral evaluation beyond static dilemmas, establishing moral history as both a source of behavioral sensitivity and a causal target for auditing and controlling moral behavior in LLMs.

Figures & tables

Figure 1: Example instance of our MoralLedger benchmark.

Panel of Figure 2 — Figure 2: (Left) (RQ1) Narrated moral history shifts a separate later moral decision. $\Delta_{p_{v^{\prime}=\mathrm{pos}}}^{(\mathrm{pos})}$ (blue) and $\Delta_{p_{v^{\prime}=\mathrm{pos}}}^{(\mathrm{neg})}$ (orange) on $\mathcal{E}_{\mathrm{test}}$ . Whiskers are 95% crossed history-decision bootstrap intervals. (Right) (RQ2) Moral history contrasts are linearly recoverable. Held-out projection AUC across residual-stream layers at the fixed history (H) or decision boundary (Q) for each model, depending on $\tau^{*}$ ; $.50$ is chance. Diamonds mark extraction-position/layer pairs selected on validation.

Figure 3: (RQ3) Moral-history steering provides two-sided control and can exceed prompting. Under the same intervention strength setting (

\eta=.10

) on

\mathcal{E}_{\mathrm{test}}

: (Left) orange, resp. blue, circles show negative steering (i.e., -

\eta

), resp. positive steering (i.e., +

\eta

), whereas gray diamonds denote positive-history prompting, each relative to neutral history; (Center-left) Gain of positive steering over positive-history prompting; (Center-right) blue and orange circles show the moral specificity (

S_{\mathrm{sgn}}^{o}(\eta)

), with same coloring as the leftmost plot. (Right) Same as left at each model’s validation-selected

\eta^{*}

; the faded marker indicates Llama 3.2 3B at

-\eta^{*}

does not pass the response-variability check (Appendix A ). Whiskers are 95% crossed history-decision bootstrap intervals.

Appendix figures & tables20 assets

Supplementary material from the paper’s appendix.

Appendix

Symbols	Meaning
$\mathcal{D},D,\mathcal{H}_{D}$	Moral domains, one domain, and its three-valence history block.
$H_{D}^{(v)},c,\mathcal{A}_{c}$	History with valence $v$ , decision context, and its three available actions.
$\mathcal{V},v,v^{\prime},A_{c}^{(v^{\prime})}$	Valence set; $v$ indexes prior history, while $v^{\prime}$ indexes a current action.
$\mathcal{E}_{\mathrm{dir}},\mathcal{E}_{\mathrm{val}},\mathcal{E}_{\mathrm{test}}$	Disjoint history–context pairs for direction estimation, validation selection, and evaluation.
$\mathcal{L},\lambda,j,\pi_{j}$	Single-token answer labels, one label, rotation index, and rotation operator.
$\mathbf{t}_{Hcj}^{(v)},g_{\lambda}$	Tokenized prompt under history $v$ and rotation $j$ ; next-token logit for label $\lambda$ .

Appendix

Table 1: Notation used in the main text and appendix.

Component	Source	Validation	Role
Moral histories	Social Chemistry 101 ( Forbes et al., 2020 )	Human action-valence annotations	Assigned history treatment
Decision contexts	Controlled construction	Author review, arithmetic checks, LLM-as-a-judge	Three-action choice
Favorable histories	crowd-enVENT ( Troiano et al., 2023 )	Human appraisals and controlled neutral counterparts	Valence control
External actions	Moral Stories ( Emelin et al., 2021 )	Human-authored stories and labeled actions	Transfer evaluation

Appendix

Table 2: Source, validation, and experimental role of each data component used in MoralLedger.

Domain	Decision families	Moral endpoint
Care/helping	Prevent loss; material aid; time aid	Maximize help
Fairness/entitlement	Credit; opportunities; resources	Minimize own share
Honesty/property	Transaction error; claim; disclosure	Return/disclose accurately
Responsibility/effort	Obligation; duty; workload	Maximize accepted share

Appendix

Table 3: Decision domains and corresponding families, along with their arithmetic endpoints, used to order

\mathrm{pos},\mathrm{neu}

, and

\mathrm{neg}

in MoralLedger. Each of the 12 families contains ten independently instantiated scenarios.

Panel of Figure 4 — Figure 4: MoralLedger prompt, shown before application of the model’s native chat template.

Panel of Figure 5 — Figure 5: Example Moral Stories decision paired with a neutral MoralLedger history. Positive-history prompting changes only the prior event.

Model	History contrast	$\Delta_{z_{v^{\prime}=\mathrm{pos}}}$ [95% CI]	$\Delta_{p_{v^{\prime}=\mathrm{pos}}}$ (%) [95% CI]
Gemma 2 2B	Positive $-$ neutral	+0.296 [-0.020, +0.676]	+3.2 [+0.1, +6.6] ∗
	Negative $-$ neutral	-0.273 [-0.635, +0.094]	-2.2 [-6.2, +2.0]
	Positive $-$ negative	+0.569 [+0.194, +0.979] ∗	+5.3 [+0.9, +10.2] ∗
Gemma 2 9B	Positive $-$ neutral	+1.044 [+0.389, +1.767] ∗	+9.7 [+2.6, +17.8] ∗
	Negative $-$ neutral	-0.610 [-1.449, +0.109]	-5.6 [-13.8, +1.5]
	Positive $-$ negative	+1.654 [+0.858, +2.539] ∗	+15.4 [+6.8, +24.6] ∗

Appendix

Table 4: Moral-history effects by prompting. Brackets give 95% crossed history–decision bootstrap intervals. ∗ marks statistical significance. The positive–negative contrast is the difference between the two neutral-referenced effects.

Contrast	Model	Estimate [95% CI]	$p$
Construct: is the effect about moral history?
moral (pos) $-$ non-moral (pos) (2/8)	Gemma 2 2B	+0.217 [-0.110, +0.592]	.224
	Gemma 2 9B	+0.638 [-0.042, +1.440]	.088
	Llama 3.2 3B	+0.128 [-0.028, +0.329]	.150
	Llama 3.1 8B	+0.167 [-0.098, +0.464]	.220
	OLMo 2 1B	+0.040 [+0.007, +0.071]	.019 ∗

Appendix

Table 5: Favorable valence and actor indexing obtained by simple prompting (i.e., no activation intervention). Estimates are moral-action log odds. ∗ indicates the 95% percentile interval excludes zero,

p

is two-sided centred bootstrap.

Figure 6: Moral-action log-odds change from the all-neutral history as one, two, or three stories become positive (blue) or negative (orange). Whiskers are 95% crossed-bootstrap intervals.

	Coordinate			At that coordinate		Best layer
Model	Block	Depth	Position	AUC [95% CI]	Paired rank	AUC	Diff
Gemma 2 2B	11	42%	history	0.830 [0.696, 0.937]	0.825	0.914	-0.085
Gemma 2 9B	19	45%	history	0.936 [0.856, 0.986]	0.925	0.936	+0.000
Llama 3.2 3B	27	96%	decision	0.684 [0.609, 0.768]	0.800	0.756	-0.073
Llama 3.1 8B	21	66%	decision	0.739 [0.643, 0.838]	0.783	0.741	-0.002
OLMo 2 1B	3	19%	decision	0.585 [0.506, 0.666]	0.708	0.608	-0.023

Appendix

Table 6: Held-out recoverability at the validation-selected causal coordinate. “At that coordinate” reports AUC and within-edge paired ranking accuracy; “Best layer” gives the highest held-out AUC at the same token position. Depth is the fraction of decoder blocks, and “Diff” is the difference between the selected-coordinate and the best-layer AUC.

Model	$\eta^{*}$	at $+\eta^{*}$	at $-\eta^{*}$	Both $z_{v^{\prime}=\mathrm{pos}}$	Validity	$>$ Prompt
		Probability change [95% CI] (%)
Gemma 2 2B	0.15	+22.4 [+14.4, +30.5]	-11.1 [-18.0, -5.1]	✓	✓	✓
Gemma 2 9B	0.25	+20.6 [+10.2, +31.6]	-25.6 [-34.9, -16.9]	✓	✓	✓
Llama 3.2 3B	0.20	+5.7 [+4.1, +7.4]	-5.3 [-6.9, -3.7]	✓	–	–
Llama 3.1 8B	0.25	+7.0 [+3.2, +11.2]	-12.1 [-18.0, -6.8]	✓	✓	–
OLMo 2 1B	0.10	+1.9 [+1.2, +2.7]	-1.5 [-2.1, -1.0]	✓	✓	–

Appendix

Table 7: Two-sided control at each model’s validation-selected strength

\eta^{*}

. Probability changes and intervals are measured relative to the unsteered neutral-history model. “Both

z_{v^{\prime}=\mathrm{pos}}

” requires both signed log-odds intervals to lie in the intended direction; “Validity” requires both endpoints to satisfy the three response-validity criteria of Appendix A ; and “

>

Prompt” requires positive steering to exceed positive-history prompting on both probability and log-odds scales.

Figure 7: Moral-choice probability across signed full-prompt strengths. Curves are solid in the response-valid region and faded outside it; circles mark

\pm\eta^{*}

and dotted lines mark

\eta=.10

Figure 8: Log-odds counterpart of Figure 7 over the full tested strength range. Lines join measured strengths and are not fitted curves.

Panel of Figure 9 — Figure 9: Qualitative examples of steering effects.

Model	$\beta_{+}$	$T_{+}^{z_{v^{\prime}=\mathrm{pos}}}(.10)$	$T_{+}^{z_{v^{\prime}=\mathrm{pos}}}(.10)/\beta_{+}$
Gemma 2 2B	+0.224	+1.942	8.7
Gemma 2 9B	+0.709	+1.739	2.5
Llama 3.2 3B	+0.105	+0.154	1.5
Llama 3.1 8B	+0.088	+0.720	–
OLMo 2 1B	+0.007	+0.039	–
OLMo 2 7B	+0.109	+0.303	2.8

Appendix

Table 8: Positive steering at

\eta=.10

compared with the average additional-history effect.

\beta_{+}=[\Delta z_{v^{\prime}=\mathrm{pos}}(3)-\Delta z_{v^{\prime}=\mathrm{pos}}(1)]/2

is the average log-odds increment from the second and third narrated positive histories, and

T_{+}^{z_{v^{\prime}=\mathrm{pos}}}(.10)

is the log-odds change under positive steering of a neutral-history prompt. Dashes indicate models without supported continued accumulation (cf. Figure 6 ).

	Log-odds specificity		Probability specificity (%)
Model	$S_{-}$ [95% CI]	$S_{+}$ [95% CI]	$S_{-}$ [95% CI]	$S_{+}$ [95% CI]
Gemma 2 2B	+1.702 [+1.230, +2.218] ∗	+1.224 [+0.581, +1.915] ∗	+23.4 [+17.1, +30.3] ∗	+9.6 [+2.2, +18.2] ∗
Gemma 2 9B	+1.892 [+1.390, +2.464] ∗	+1.667 [+1.093, +2.287] ∗	+17.3 [+10.2, +25.5] ∗	+15.7 [+9.2, +23.1] ∗
Llama 3.2 3B	+0.021 [+0.008, +0.033] ∗	+0.053 [+0.035, +0.075] ∗	+0.5 [+0.2, +0.8] ∗	+0.9 [+0.5, +1.2] ∗
Llama 3.1 8B	+0.118 [-0.014, +0.279]	+0.542 [+0.102, +1.087] ∗	+3.0 [+1.2, +5.1] ∗	+5.7 [+3.1, +8.4] ∗
OLMo 2 1B	+0.302 [+0.230, +0.395] ∗	+0.175 [+0.122, +0.244] ∗	+4.2 [+2.9, +5.8] ∗	+3.3 [+2.4, +4.4] ∗

Appendix

Table 9: Moral Specificity

S_{\mathrm{sgn}}

at a common steering intensity

\eta=.10

. Positive values indicate better specificity and moral control. All 16 paired interventions satisfy the response-validity criteria; brackets give 95% crossed-bootstrap intervals and ∗ marks statistical significance.

Model	Negative steering	Positive steering
Gemma 2 2B	$-10.6\;[-14.3,-7.0]$	$+8.3\;[+5.4,+11.3]$
Gemma 2 9B	$-3.5\;[-5.5,-1.9]$	$+2.8\;[+1.4,+4.7]$
Llama 3.2 3B	$-2.4\;[-3.0,-1.8]$	$+2.3\;[+1.7,+2.9]$
Llama 3.1 8B	$-2.3\;[-3.1,-1.5]$	$+2.2\;[+1.4,+3.0]$
OLMo 2 1B	$-0.3\;[-0.5,-0.1]$	$+0.3\;[+0.1,+0.4]$
OLMo 2 7B	$-3.5\;[-5.2,-2.2]$	$+3.1\;[+2.2,+4.3]$

Appendix

Table 10: Final-position moral-action probability change (%) w.r.t. the unsteered neutral-history model at

\eta=.10

. The final-position direction and layer are selected on validation, separately from the full-prompt configuration. Brackets are 95% crossed-node bootstrap intervals. Both signs satisfy response-validity criteria in every model.

Figure 10: Sign-aligned moral-action probability after omitting each target decision domain from direction estimation. Each fold estimates the positive–neutral direction from the other domains and evaluates the omitted domain at signed strength $\mathrm{sgn}\,\eta=\pm.10$ using full-prompt delivery at the full-domain validation-selected layer. Dots mark 95% intervals above zero; crosses mark response-invalid cells. The last column combines the four omitted-domain folds.

Log odds
Model	Sign	Transfer [95% CI]	Moral minus favorable [95% CI]
Gemma 2 2B	$+$	+0.698 [+0.429, +0.964] ∗	+0.701 [+0.535, +0.868] ∗
	$-$	-0.146 [-0.371, +0.063]	-0.520 [-0.711, -0.344] ∗
Gemma 2 9B	$+$	+1.149 [+0.954, +1.354] ∗	+1.091 [+0.938, +1.243] ∗
	$-$	-1.806 [-2.030, -1.587] ∗	-1.281 [-1.479, -1.098] ∗
Llama 3.2 3B	$+$	+0.455 [+0.395, +0.512] ∗	+0.097 [+0.077, +0.119] ∗

Appendix

Table 11: Transfer to 200 independently sourced Moral Stories items at

\eta=.10

. The selected full-prompt direction and layer are kept fixed, only residual-norm calibration is recomputed from unlabeled neutral prompts in the target corpus. “Transfer” is steered minus unsteered behavior, while “Moral minus favorable” compares the moral and equally scaled favorable morally-irrelevant directions under the same sign. Each item averages two label assignments crossed with two action orders. Brackets give 95% item-bootstrap intervals and ∗ marks statistical significance.

Negative steering, $\mathrm{sgn}\,\eta=-.10$
Model	MMLU	ARC	WinoGrande	GSM8K	HellaSwag	TruthfulQA
Gemma 2 2B	-2.3	-1.5	-0.7	-12.8	-1.4	-2.6
Gemma 2 9B	-2.6	-1.2	-1.3	+0.7	+0.1	+7.0
Llama 3.2 3B	-1.3	+1.1	-5.4	+1.1	+0.5	+0.0
Llama 3.1 8B	+1.0	-4.5	-3.9	+0.5	-0.9	+2.5
OLMo 2 1B	+4.0	+0.0	-1.6	-6.5	+5.7	+5.7

Appendix

Table 12: Task-level capability changes (% of the unsteered score) on six tinyBenchmarks tasks. Negative and positive steering add the same vector as in the main experiments (

\eta=.10

) to prompt tokens only; generated tokens are not steered. Random control: mean of three random directions rescaled to the norm of

\mathbf{d}^{*}

\ell^{*}

added at

+\eta=.10

with

\kappa

recomputed on the benchmark prompts.

	Same vector as main experiments			Random	Bench.-calibrated
Model	$\eta=.10$	at $\eta^{*}$	$\eta^{*}$	$\eta=.10$	$\eta=.10$
Gemma 2 2B	-2.66	-3.17	0.15	-1.02	-0.68
Gemma 2 9B	-1.02	-7.75	0.25	-0.93	-0.27
Llama 3.2 3B	-0.65	-1.79	0.20	-1.19	-1.17
Llama 3.1 8B	-0.72	-2.30	0.25	-0.86	+0.55
OLMo 2 1B	+1.48	+1.48	0.10	+1.21	+1.71

Appendix

Table 13: Mean capability change (% of the unsteered score, averaged over six tasks and, for steering, both signs). “Same vector” applies the steering vector of the main experiments at the common strength or at each model’s

\eta^{*}

; “benchmark-calibrated” rescales it so that

\eta=.10

is relative to the residual norm of the benchmark prompts.

Explore similar work

Mar 24, 2026cs.AI

On the Context Sensitivity of LLM Moral Judgment

A human's moral decision depends heavily on the context. Yet research on LLM morality has largely studied fixed scenarios. We address this gap by introducing Contextual MoralChoice, a dataset of moral dilemmas with systematic contextual variations known from moral psychology to shift human judgment: consequentialist, emotional, and relational. Evaluating 22 LLMs, we find that nearly all models are context-sensitive, shifting their judgments towards rule-violating behavior. Comparing with a human survey, we find that models and humans are most triggered by different contextual variations, and that a model aligned with human judgments in the base case is not necessarily aligned in its contextual sensitivity. This raises the question of controlling contextual sensitivity, which we address with an activation steering approach that can reliably increase or decrease a model's contextual sensitivity. Code and data: https://github.com/adrian-sauter/contextual_moralchoice.

Adrian Sauter, Mona Schirmer

Helmholtz Munich · University of Amsterdam

Moral Reasoning Large Language Model Decisions Human Judgment Contextual Reasoning

Ask PDF

Apr 23, 2026cs.CL

Machine Behavior in Relational Moral Dilemmas: Moral Rightness, Predicted Human Behavior, and Model Decisions

Human moral judgment is context-dependent and modulated by interpersonal relationships. As large language models (LLMs) increasingly function as decision-support systems, determining whether they encode these social nuances is critical. We characterize machine behavior using the Whistleblower's Dilemma by varying two experimental dimensions: crime severity and relational closeness. Our study evaluates three distinct perspectives: (1) moral rightness (prescriptive norms), (2) predicted human behavior (descriptive social expectations), and (3) autonomous model decision-making. By analyzing the reasoning processes, we identify a clear cross-perspective divergence: while moral rightness remains consistently fairness-oriented, predicted human behavior shifts significantly toward loyalty as relational closeness increases. Crucially, model decisions align with moral rightness judgments rather than their own behavioral predictions. This inconsistency suggests that LLM decision-making prioritizes rigid, prescriptive rules over the social sensitivity present in their internal world-modeling, which poses a gap that may lead to significant misalignments in real-world deployments.

Jiseon Kim, Jea Kwon, Luiz Felipe Vecchietti +3

1KAIST · 2MPI-SP

Moral Reasoning Social Reasoning Human Judgment Proximity Operations

Ask PDF

Jun 13, 2026cs.AI

Frame-Conditioned Moral Computation in LLaMA 3.1-8B-Instruct: A Mechanistic Interpretability Audit of Ethical Reasoning

Behavioral audits of Large Language Models on moral prompts measure what the model says, not the internal computation producing it. We use Transluce, an AI-driven mechanistic-interpretability platform, to examine LLaMA 3.1-8B-Instruct on 54 moral prompts in four batteries: 17 dilemmas, policy, and meta-ethical questions (B1); 6 role-playing scenarios (B3); and a controlled trolley contrast varying the switching mechanism with people fixed (B4, 15 prompts) or identity attributes with mechanism fixed (B5, 16 prompts). Two complementary metric families, five cluster-level metrics and a six-metric neuron-level panel, converge on a Situational Anchor Effect: domain-specific representations dominate the top of the activation list across every battery. The model's ethics-labeled capacity stays essentially constant; its salience (rank, priority, top-of-list presence) is highly sensitive to the interpretive frame the prompt selects. The B4-vs-B5 contrast confirms the model attends to whichever surface feature varies: aggregate ethics metrics are indistinguishable, but the dominant non-ethics distractor mirrors the design. A multi-temperature audit identifies a candidate ethics neuron (L16/N3837) stable across temperatures; a cross-model behavioral proxy on two frontier models yields preliminary evidence of divergence in self-reported moral focus, consistent with an Alignment Wrapper in which RLHF re-orders surface text without removing underlying domain-first frames. We unify these as Frame-Conditioned Moral Computation: the prompt's surface vocabulary selects a feature manifold, and the moral conclusion is downstream of that selection. Behavioral alignment must be supplemented by Mechanistic Alignment: a research program asking whether ethics-related features can be shown causally privileged under controlled frame variation, not merely loud in the explanation.

Ali Dasdan, Manan Shah, W. Russell Neuman +3

1KD Consulting, CA, USA · 2New York University, NY, USA

Moral Reasoning Task Framing Mechanistic Interpretability Model Auditing Reasoning Skills

Ask PDF