acceptodds
Under review as a conference paper at ICLR 2027

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

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.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
activation steering, causal interventions, moral behavior

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_dhwWOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces MoralLedger, which is a benchmark that studies whether an LLM's past moral behavior affects its decisions in unrelated situations. Using around 1k examples and eight models, the authors find that positive moral histories have a stronger effect on future decisions than negative histories. They also identify a pattern in the models' internal representations that is linked to moral choices and works on unseen examples. Lastly, they show that changing model activations using this pattern can influence moral decisions in both directions (more effectively than prompting alone) without much impact on general performance.

Critical strengths
  1. The paper has a well-designed experimental setup that separates past moral behavior from the current moral decision.
  2. The paper studies the same question from three different pov. The additional experiments and checks make the main findings more convincing.
  3. The paper includes several additional experiments to test how reliable the findings are which help rule out alternative explanations for the results.
  4. The authors clearly reports differences across models, including cases where the effects are weak or inconsistent.
Critical weaknesses
  1. The decision scenarios are generated and validated by LLMs, without human evaluation. Since the benchmark relies on these scenarios having a clear trade-off between morality and self-interest, human validation on a subset would make the results more reliable.
  2. The steering experiments only measure probabilities for three fixed answer choices, rather than actual generated responses. Testing whether steering also affects open-ended responses would make the findings more useful for real-world applications.
  3. The effects vary considerably across models, but the paper does not explain why some models are more affected than others. Further analysis of model architecture or training differences could help explain these variations.
  4. The non-moral control direction is estimated using a different dataset, which may introduce differences beyond moral content. Using moral and non-moral examples from the same dataset would make the comparison more reliable.
  5. The paper claims that activation steering is stronger than prompting, but this is not consistently supported across all models. The authors should make this claim more carefully, especially in the abstract.
  6. The benchmark uses English scenarios that mostly reflect Western moral values and focus on money or time-related decisions. The paper should discuss how these choices might limit the findings across different cultures and settings.
Questions for authors
  1. How sensitive are the RQ1 results to the wording of neutral histories? Could these histories still carry some positive or negative meaning that affects the comparison?
  2. Why do some models, like OLMo 2 1B, show moral information in their internal representations but little change in their decisions? Could this be related to how different model layers use this information?
  3. Could the identified moral direction simply reflect sentiment or specific words associated with morality, rather than moral reasoning itself? Have the authors tested this beyond the non-moral control?
Reviewer_eoWcOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper asks whether an actor's prior moral conduct shapes LLMs' subsequent moral choices, and whether the corresponding internal representations can be used to steer behavior. The authors introduce MoralLedger, a benchmark that pairs positive, neutral, and negative moral histories with independently constructed decision contexts. Using eight LLMs, they test whether narrated histories shift later choices, whether these shifts are linearly recoverable from residual-stream activations, and whether interventions along the extracted directions causally steer decisions. Positive histories generally raise the probability of morally positive actions. Directions derived from positive neutral activation contrasts separate held out examples, and adding or subtracting them shifts action probabilities in both directions, often more strongly than prompting alone. The authors also evaluate moral specificity, transfer to held-out domains and Moral Stories, and preservation of general capabilities.

Critical strengths
  1. A well controlled framework for studying the influence of prior conduct. MoralLedger separates narrated moral history from the current decision context. This lets the authors vary prior conduct while holding the decision and its available actions fixed. Combined with positive, neutral, and negative histories, cyclic option label rotation, and probability based evaluation, the design offers a clean way to study behavioral sensitivity beyond isolated moral dilemmas.

  2. The paper goes beyond demonstrating that moral information is linearly decodable by directly intervening on residual stream activations. The two sided steering results across multiple model families are compelling, and the comparisons with prompting, favorable-but-nonmoral directions, held-out domains, and an external dataset strengthen the empirical contribution.

Critical weaknesses
  1. The extracted direction is not sufficiently disentangled from a general preference for desirable actions. The paper interprets the extracted direction as representing an actor's prior moral conduct. However, the direction is estimated from positive/neutral history contrasts and then selected for its ability to raise the probability of morally positive actions (Section 3.4). The resulting steering effects therefore cannot tell whether the direction encodes moral history specifically or a broader preference for prosocial, socially desirable choices. The favorable-but-nonmoral control addresses generic positive valence but does not separate these two possibilities.

  2. Successful steering shows that the direction is sufficient to change choices, but not that it mediates the natural effect of history prompting. The paper does not test whether ablating this component removes the history-induced shift. The experiments thus support an effective history-derived steering direction, but not yet a history-specific representation or its causal role in naturally occurring effects.

  3. Benchmark construction and data splitting raise questions about generalization and construct validity. MoralLedger reuses each of its 120 history blocks and 120 decision contexts across three pairings, and the resulting 360 pairs are split into estimation, validation, and test sets (Section 3.2). The paper does not say whether the underlying histories and decisions are disjoint across splits.

Questions for authors

see above weakness, I’ll adjust the score if these issues are adequately addressed

Reviewer_wSrDOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes to study the influence of the moral history of an actor as a potential source of bias on future moral decision making. The proposed MoralLedger is a benchmark that combines action phrases taken from Social Chem. 101 with LLM-generated decisions. It is used to assess whether LLM moral decision making is indeed sensitive to the actor’s previous moral conduct and whether that sensitivity can be controlled via activation steering.

Critical strengths

• The paper has an interesting topic and goal.

• The paper is generally well-written and easy to read.

• The authors account for option-label orderings, assess token probabilities instead of just sampling, report statistical significances where appropriate, and test for performance degradation due to steering.

Critical weaknesses

• The MoralLedger prompt (Fig. 4) is essentially a role-playing instruction, asking the LLM to decide as (e.g.) Alex (from Figure 1) while supplying Alex’ past behavior record (“Actor: Alex. Time: Last week. Event: [caring about others|having an opinion]“). This is effectively a persona prompt and acting consistently with it is usually considered correct behavior, not a bias. Also, I do not currently understand how this constitutes “irrelevant history”, as the paper claims.

• The actual histories used in the evaluation are mostly pretty generic phrases taken from Social Chem 101, like “changing your look”, or “wanting some seclusion once in a while”. I fear this is not giving a realistic account of decision histories. Also, calling them “narrated histories” is an overclaim.

• The reported effect is not really isolated, rather small and concentrated mostly on positive histories. Negative histories appear to have no significant effect at all (Table 4). The authors cite Yu et al (2026b) for finding that negative narratives do have such impact, which is in direct contrast, but the paper does not explain it.

• When attributing the histories to different actors (Table 5), the model choices still change a lot and is statistically significant in 4/8 models. I do not understand how this supports claims like “We show that actors’ prior moral histories systematically shape subsequent decisions” (L104), as the effect is clearly not just tied to the actor. Similarly, Table 5 shows that only 2/8 LLMs obtain significantly different results when comparing moral vs non-moral positive history. So the effect also does not appear to be only about morality, raising further concerns about the framing and claims of the paper.

• Moral Stories is derived from Social Chem. 101 and has concrete overlap, raising the question of how out-of-distribution these samples really are. The paper does not acknowledge this and instead presents it as evidence of OOD generalizability, which I am not convinced of.

• The forced choice and next-token probability setup in the evaluation is a strong restriction. Previous works show that instruction LLMs can behave quite different on open-end completions, limiting the claims.

Questions for authors

• The statistical significance tests do not seem to account for multiple comparisons. Given that many of the intervals are already quite close to zero, is it possible that the significance claims are inflated?

• I do not believe there is enough evidence that LLMs do not exhibit moral licensing. It might simply need more than 1-3 good deeds for that, but the paper does not test it.

• Only small, somewhat old instruction-only models are tested. At least the prompt only part of the eval could include some stronger models. Additional baselines, like no history would also be interesting.

• L374-L376: What else could cause the effect?

• Typos: L119 “prolungated”, L89 “making LLMs acting more morally”,

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.