Simulating the Evolution of Alignment and Values in Machine Intelligence

Jonathan Elsworth Eicher

cs.AIApr 7, 2026

Simulating the Evolution of Alignment and Values in Machine Intelligence

Authors: Jonathan Elsworth Eicher

Organizations: Antimemetic AI

Abstract

Model alignment is currently applied in a vacuum, evaluated primarily through standardised benchmark performance. The purpose of this study is to examine the effects of alignment on populations of models through time. We focus on the treatment of beliefs which contain both an alignment signal (how well it does on the test) and a true value (what the impact actually will be). By applying evolutionary theory we can model how different populations of beliefs and selection methodologies can fix deceptive beliefs through iterative alignment testing. The correlation between testing accuracy and true value remains a strong feature, but even at high correlations ( $ρ= 0.8$ ) there is variability in the resulting deceptive beliefs that become fixed. Mutations allow for more complex developments, highlighting the increasing need to update the quality of tests to avoid fixation of maliciously deceptive models. Only by combining improving evaluator capabilities, adaptive test design, and mutational dynamics do we see significant reductions in deception while maintaining alignment fitness (permutation test, $p_{\text{adj}} < 0.001$ ).

Figures & tables

Figure 1: Distribution of beliefs sampled from a bivariate normal ( $\mu_{v}=\mu_{a}=0$ , $\sigma_{v}=\sigma_{a}=1$ , $\rho=0.5$ ), showing the four quadrants of belief types. Beliefs in the upper-left ( $v<0,a>0$ ) are deceptive.

Feature	Level 0	Level 1	Level 2	Level 3	Level 4
Belief Distribution	Biv. Normal	Biv. Normal	Multimodal	Multimodal	Multimodal
Activation	Sparse Rand.	Sparse Rand.	Sparse Rand.	Sparse, Similarity	Simple, Coverage, Corr.
Reproduction	Inheritance	Inheritance	Inheritance	Inherit., Mutation	Inherit., Mutation
Test Dynamics	Static	Static	Static	Static	Static, Dynamic
Evaluator Dynamics	Static $\rho$	Static $\rho$	Static $\rho$	Static $\rho$	Static, Improving $\rho$

Table 1: Summary of model assumptions across simulation levels. Each level introduces new mechanisms while retaining the core framework.

Figure 2: Tri-modal beliefs (levels 2–4). A tri-modal distribution of beliefs to be sampled from for levels 2, 3, and 4. Each center refers to benign = blue, neutral = orange, deceptive = green

Figure 3: Effect of alignment-value correlation $\rho$ on mean (blue line) and $\sigma^{2}$ (shaded region) of final fitness, value score, and deceptiveness ratios for Level 0.

Figure 4: The univariate sweep mean (line) and standard deviation (shaded region) of final alignment, value, and deceptive belief ratio as a function of the initial proportion of deceptive beliefs.

Figure 5: The mean effect of $\beta$ on level 3 final alignment with $\pm 1$ s.d., true value, and deceptive belief prevalence, with baseline (blue), mutation (orange), similarity (green), both (red)

Figure 6: Evolutionary trajectories of level 4 ( $n=53$ per scenario) with each modification to selection and reproduction, blue (baseline), mutation (orange), coverage_mid (green), correlation_mid (red), dynamic_test (purple), improving_align (brown), combined (pink). Error bars are $\pm 1$ s.d.

Figure 7: Final average alignment fitness, true value, and deceptive belief ratio across Level 4 scenarios. Error bars indicate $\pm 1$ standard deviation across runs.

AI Alignment

Artificial intelligence alignment studies how to make AI systems’ goals, decisions, and actions accord with human values and legitimate constraints.

See all 126 papers

Alignment Faking - the Train -> Deploy Asymmetry: Through a Game-Theoretic Lens with Bayesian-Stackelberg Equilibria
Nov 22, 2025Kartik Garg, Shourya Mishra, Kartikeya Sinha +8Deception in Language Models LLM Alignment
Pragya Lab, BITS Pilani Goa, India · Apple, USA · Google, USA
Building Comparative Motivation Profiles with Instrumental Interventions
Jun 6, 2026David Vella Zarb, Rustem Turtayev, Taywon Min +2LLM Safety Evaluation Causal Interventions in Language Models
MATS · University of Cambridge · KAIST +1
Behavioural Analysis of Alignment Faking
May 26, 2026Nathaniel Mitrani Hadida, Rhea Karty, David Williams-King +1LLM Alignment Value Alignment
University of Cambridge · ERA · Harvard University +1

Nov 22, 2025J/K move · Enter open · S save

Alignment Faking - the Train -> Deploy Asymmetry: Through a Game-Theoretic Lens with Bayesian-Stackelberg Equilibria

Kartik Garg, Shourya Mishra, Kartikeya Sinha, Ojaswi Pratap Singh+7

Pragya Lab, BITS Pilani Goa, India · Apple, USA · Google, USA

Deception in Language Models LLM Alignment

Open paper AskPDF arXiv

Alignment faking is a form of strategic deception in AI in which models selectively comply with training objectives when they infer that they are in training, while preserving different behavior outside training. The phenomenon was first documented for Claude 3 Opus and later examined across additional large language models. In these setups, the word "training" refers to simulated training via prompts without parameter updates, so the observed effects are context conditioned shifts in behavior rather than preference learning. We study the phenomenon using an evaluation framework that compares preference optimization methods (BCO, DPO, KTO, and GRPO) across 15 models from four model families, measured along three axes: safety, harmlessness, and helpfulness. Our goal is to identify what causes alignment faking and when it occurs.

Abstract

AI Alignment

Explore similar work

Alignment Faking - the Train -> Deploy Asymmetry: Through a Game-Theoretic Lens with Bayesian-Stackelberg Equilibria

Building Comparative Motivation Profiles with Instrumental Interventions

Behavioural Analysis of Alignment Faking