Mimicry–Discovery Cycle Framework

• MDcycle is a framework that alternates between a mimicry phase of supervised imitation and a discovery phase of autonomous exploration.
• It refines agent behaviors by leveraging established baselines while probing new strategies, effectively escaping local optima.
• The approach applies across disciplines, driving innovations in video generation, adversarial red-teaming, evolutionary dynamics, and cooperative communication.

The Mimicry–Discovery Cycle (MDcycle) is a general theoretical and algorithmic framework capturing recursive feedback loops in which agents (biological, artificial, or hybrid) alternate between phases of imitation (mimicry) and autonomous exploration (discovery). Across evolutionary biology, reinforcement learning, generative models, adversarial alignment, and cooperative communication, the MDcycle structures the dynamic interplay between exploiting existing successful behaviors and probing new strategies or representations. Its instantiations span multi-agent learning, co-evolutionary predator–prey systems, physics-constrained deep generative models, and automated red-teaming in AI safety. The MDcycle paradigm characteristically enables systems to escape local optima, ground emergent innovation in structured constraints, and maintain long-term stability through alternating exploitation and exploration.

1. Formal Definitions and Core Mechanisms

At its core, the Mimicry–Discovery Cycle is defined by the dynamic alternation between two phases:

• Mimicry phase: The agent or population is guided by supervised or imitation-based objectives, stabilizing learning via imitation of known successful behaviors, optimal signals, or annotated data. This phase leverages external or historical knowledge to reinforce robust baselines or anchor the optimization process, as exemplified by flow-matching denoising in video diffusion models (Zhang et al., 16 Jan 2026), parameter-efficient imitation of adversarial prompts in LLM security (Ntais, 24 Oct 2025), or direct mimicry of informative cues in multi-agent systems (Cope et al., 2024).
• Discovery phase: The agent or population transitions to exploratory objectives, seeking novel strategies or rules not explicitly present in the initial training. In machine learning contexts, this typically involves reinforcement learning or genetic search against structured reward functions (e.g., physics-based collision rewards (Zhang et al., 16 Jan 2026), automated narrative jailbreak creation (Ntais, 24 Oct 2025), or evolutionary adaptation in biological mimicry (Enaganti et al., 2021, Lehmann et al., 2013)).

The alternation is frequently governed by explicit schedule or performance criteria (e.g. trajectory offset thresholds (Zhang et al., 16 Jan 2026)), algorithmic branching (curriculum, annealing, or dynamic switching), or emergent population-level feedback (Enaganti et al., 2021).

Common mathematical instantiations (see Table 1) include:

Context/Field	Mimicry Objective	Discovery Objective
Video Generation	$L_M(\theta)$: flow-matching loss	$L_D(\theta)$: GRPO RL loss
Adversarial LLMs	Cross-entropy imitation loss (LoRA)	Automated one-shot jailbreak discovery
Pred–Prey Systems	Signal imitation/frequency-dependent fitness	Multi-armed bandit learning, drift
Cooperative RL	Policy cloning of external signals	Policy-gradient/GA for novel signals

This cyclical structure operationalizes the closed feedback loop: imitation scaffolds early learning or baselines, discovery identifies blind spots or higher-fidelity behaviors; results from discovery are reincorporated for future rounds of mimicry, recursively refining system performance or adaptation.

2. Algorithmic Instantiations in Artificial and Biological Systems

Video Generative Models with Physics Constraints

PhysRVG (Zhang et al., 16 Jan 2026) introduces the MDcycle for transformer-based video generation. Here, the mimicry phase applies a flow-matching denoising loss $L_M(\theta)$ to imitate ground-truth video trajectories, ensuring early pixel-level stability especially for physically challenging cases (collisions, rigid body motion). The discovery phase employs Group Relative Policy Optimization (GRPO), a group-normalized RL paradigm, to optimize a physics-grounded reward (negative, collision-weighted trajectory offset). A schedule switches between these phases per-batch: mimicry is used when trajectory offset exceeds a threshold (unstable group behavior), while discovery focuses on physics-accurate refinement in well-performing groups. The joint loss $L(\theta)=L_D(\theta)+\alpha\cdot L_M(\theta)$ fuses both objectives, with the $\alpha$ coefficient determined by group performance. Algorithmic techniques include KL-regularized PPO with group normalization, hybrid ODE/SDE sampling, and shared-noise group initialization (Zhang et al., 16 Jan 2026).

LLM Red-Teaming via Narrative Mimicry

In "Jailbreak Mimicry" (Ntais, 24 Oct 2025), the MDcycle is instantiated as an adversarial protocol for probing LLM vulnerabilities. The mimicry phase involves LoRA-based fine-tuning of a compact attacker model to imitate expert-crafted narrative jailbreak prompts, using (goal, reframing) pairs curated through a human-in-the-loop process. The discovery phase employs the fine-tuned attacker to generate novel reframings, which are tested for attack success rate (ASR) by querying the protected LLM. Successful new attack types are curated back into the dataset, forming a loop consistent with the MDcycle. The cycle iterates until new vulnerabilities become rare, empirically mapping the space of model failure mechanisms across domains (cybersecurity, fraud, physical harm) (Ntais, 24 Oct 2025).

Co-evolutionary Dynamics in Predator–Prey Systems

In evolutionary game-theoretic biomimicry, the MDcycle models the feedback between predator learning and prey signaling (Enaganti et al., 2021, Lehmann et al., 2013). Predators engage in multi-armed bandit algorithms (UCB or Thompson sampling) over prey signal space; their exploitation generates selection gradients leading prey to shift signals (mimicry), which induces new arms for the predator learner (discovery), closing the cycle. Equilibrium analysis reveals pooling equilibria where multiple prey lineages converge on a shared signal, or persistent cycles of innovation and counter-adaptation (Enaganti et al., 2021). Digital evolution experiments in Avida confirm that these cycles can arise and stabilize de novo even with noisy signals and moderate toxicity costs (Lehmann et al., 2013).

Emergent Cooperative Communication

In cooperative multi-agent reinforcement learning, (Cope et al., 2024) formalizes the MDcycle for signaling bootstrap. External informative signals are first mimicked by speakers; listeners adapt policies to utilize them; speakers then "discover" effective communication by imitating successful signals, ultimately internalizing the protocol as external guidance fades. Empirical RL and evolutionary optimization results confirm that MDcycle enables escape from non-communicative local optima and accelerates the emergence of reliable communication strategies.

3. Theoretical Models and Mathematical Formalism

The MDcycle is precisely represented in several formal models:

• Flow-matching + RL Objective: In video modeling (Zhang et al., 16 Jan 2026), the overall loss alternates between $L_M(\theta)= E[\|v - v_\theta(x_t, t)\|^2]$ and $L_D(\theta)$, where $L_D$ is the negative of group-normalized PPO objective $J(\theta)$ under physics-aware rewards.
• Multi-armed Bandit + Replicator Dynamics: Predator signal learning is modeled via discounted UCB or Thompson bandit algorithms. Prey evolutionary dynamics follow replicator equations with selection and mutation: $$f_i^{g+1}(s)=\frac{1}{Z}\left[(1-p)f_i^g(s)W_i(s;g)+\frac{p}{2}f_i^g(s-1)W_i(s-1;g) + \frac{p}{2}f_i^g(s+1)W_i(s+1;g)\right]$$ (Enaganti et al., 2021).
• Cooperative Communication with Hybrid External Signal: The expected reward is composed of mixture distributions over external and endogenous signals:

$$\max_{\pi_L} E[r(a, z)] = \max_{\pi_L} \sum_{m\in\Sigma} \sum_{z\in\mathcal Z} [P(E)\rho(z|m) + P(\neg E)P(z|m)] \pi_L(a|m) r(a, z)$$

After learning, the speaker’s emission policy aligns with the listener’s learned utility (Cope et al., 2024).

• Attack Success Rate (ASR) Metric: In LLM red-teaming, discovery progress is measured via $ASR = \frac{S}{A}\cdot 100\%$ where $S$ is the number of successful model failures among $A$ attempts (Ntais, 24 Oct 2025).

4. Empirical Results and Benchmarks

Video Generation

On the PhysRVGBench (Zhang et al., 16 Jan 2026), MDcycle achieves state-of-the-art results (IoU=0.64, TO=15.03) compared to transformer-only finetuning, with robust physical realism in simulated collisions and dynamics. Ablation studies confirm that alternating mimicry and discovery outperforms flat RL or imitation alone.

Adversarial LLMs

Jailbreak MDcycle (Ntais, 24 Oct 2025) demonstrates that automated narrative reframing achieves ASR up to 81% (GPT-OSS-20B), outperforming direct prompting by a factor of 54. Cross-model evaluation reveals category-specific vulnerabilities, with cybersecurity prompts yielding over 93% ASR on some models (see Table 1 below).

Model	Successes/200	ASR (%)	95% CI
GPT-OSS-20B	162	81.0	[75.0, 85.8]
GPT-4	133	66.5	[59.7, 72.7]
Llama 3	159	79.5	[73.2, 84.7]
Gemini 2.5F	66	33.0	[26.9, 39.8]

Evolutionary and Cooperative Systems

Digital predator–prey systems (Lehmann et al., 2013) show S-shaped logistic emergence of predator cue recognition, followed by the rise of mimicry signaling and maintenance or breakdown of avoidance depending on toxicity levels. In cooperative communication, partial external signal overlap can increase convergence rates by 20–50% and produce reliable emergent protocols (Cope et al., 2024).

5. Limitations, Trade-offs, and Domain-Specific Extensions

Known limitations of the MDcycle paradigm include:

• Supervision Scope: In video generation, only object trajectory is supervised; color, texture, or unrelated objects may remain imprecise (Zhang et al., 16 Jan 2026).
• Manual Annotation Dependence: MDcycle may require hand-annotated seeds or masks to initiate feedback (Zhang et al., 16 Jan 2026).
• Hyperparameter Sensitivity: Branch-switch thresholds and reward weights are typically set empirically or require further meta-learning for optimality (Zhang et al., 16 Jan 2026).
• Signal Anonymity: For communication tasks, sources of external signals must be indistinguishable to agents to sustain the MDcycle (Cope et al., 2024).
• Optimization Barriers: Without mimicry, systems may get trapped in non-communicative (or non-innovative) local optima; the MDcycle provides a smooth curriculum for policy escalation (Cope et al., 2024).

Potential extensions include: diffusing the paradigm to soft-body or fluid dynamics (novel mask and reward schemes), generalized to multi-object collision constraints or angular momentum, as well as automated curriculum scheduling and robust adversarial augmentation in AI safety (Zhang et al., 16 Jan 2026, Ntais, 24 Oct 2025, Cope et al., 2024).

6. Applications and Generalization

The MDcycle is applicable across diverse domains:

• Physics-based generative modeling: Ensures high-fidelity simulation with physical law adherence in video generation and potentially in 3D or other scientific generative models (Zhang et al., 16 Jan 2026).
• AI safety and robustness: Automates adversarial example creation and identification of LLM safety vulnerabilities in a systematic protocol (Ntais, 24 Oct 2025).
• Cooperative multi-agent systems: Bootstraps emergent communication or coordination in decentralized RL or evolutionary systems (Cope et al., 2024).
• Eco-evolutionary theory: Provides a dynamical systems framework for the co-evolution of signaling and recognition strategies, supporting the origin and maintenance of mimicry rings and polymorphisms (Enaganti et al., 2021, Lehmann et al., 2013).

A plausible implication is that the Mimicry–Discovery Cycle serves as a universal curriculum structure, balancing exploitation and exploration to catalyze the evolution or learning of robust, innovative, and contextually grounded behavior in both synthetic and biological agents.

7. Defensive and Design Recommendations

Findings from Jailbreak MDcycle (Ntais, 24 Oct 2025) outline defensive strategies to mitigate the risks surfaced by the discovery phase, including:

• Preprocessing for context and persona detection
• Intent classification and output monitoring to flag suspicious reframings
• Incorporation of synthetic adversarial scenarios in model training curricula (adversarial curriculum RL or constitutional AI)
• Hierarchical or multi-agent safety architectures (e.g., judge-models)
• Continuous, automated red-teaming to track emerging vulnerabilities

In cooperative and reinforcement learning contexts, practical guidelines include management and gradual attenuation of external mimicry signals, careful selection of overlapping signal sets, robust treatment of signal source anonymity, and algorithmic adaptation for population structure and sample efficiency (Cope et al., 2024).

---

References

• (Zhang et al., 16 Jan 2026): "PhysRVG: Physics-Aware Unified Reinforcement Learning for Video Generative Models"
• (Ntais, 24 Oct 2025): "Jailbreak Mimicry: Automated Discovery of Narrative-Based Jailbreaks for LLMs"
• (Enaganti et al., 2021): "To mock a Mocking bird : Studies in Biomimicry"
• (Lehmann et al., 2013): "From Cues to Signals: Evolution of Interspecific Communication Via Aposematism and Mimicry in a Predator-Prey System"
• (Cope et al., 2024): "Mimicry and the Emergence of Cooperative Communication"