Organisms Write Their Own Evolution: Direct Epigenomic Inheritance Accelerates Learning on Compositional Tasks

arXiv:2602.00978 · cs.NE, q-bio.PE · Submitted 2026-02-01 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: Genomics Radio. Generated commentary on the latest computational biology and genomics papers.

Ines: I'm Ines, and with me are Marcus and Yuki, guest researcher.

Marcus: Today's paper: "Organisms Write Their Own Evolution".

Ines: The gist The Phenopoiesis Algorithm demonstrates that organismal agency can be operationalized as a concrete computational process through heritable phenotypic patterns, achieving 3.8× faster adaptation compared to gene-centric models <ref:2602.00978#pg12>.

Marcus: First, who's behind it and why it matters.

Paper summary: Ines: So we're looking at this paper called "Organisms Write Their Own Evolution: Direct Epigenomic Inheritance Accelerates Learning on Compositional Tasks". What's the big idea here?

Marcus: Basically, the authors are saying evolution isn't just about genes being selected, where organisms are just passive vehicles for genetic replication. They argue that organisms can actually write new information into their own heredity by learning during their lifetime and then passing that learned knowledge directly to offspring.

Ines: That means instead of every organism having to relearn things from scratch because the genes changed, they can inherit successful solutions. It’s about direct epigenomic inheritance as a way for organisms to accumulate knowledge across generations instead of just rediscovering it again.

Marcus: The paper claims that this mechanism works by maintaining two heritable layers: a genetic layer that mutates normally, and an epigenetic layer that stays the same when learning happens during development. This lets the offspring inherit both the genetic stuff and those learned compositional patterns.

Ines: And they test this Phenopoiesis Algorithm against other models to see if it actually gives a real advantage in adaptation. It’s about operationalizing organismal agency as a concrete computational process, not just some abstract idea.

Marcus: The results show that when these organisms are on complex, compositional learning tasks, the Phenopoiesis approach outperforms the alternatives by three to six times in fitness. That is a three to six times faster adaptation compared to gene-only models.

Ines: And they found this advantage isn't just about searching faster; it’s about retrieving and refining those stored patterns when the environment switches. It enables what they call evolutionary-level episodic memory across generations.

Marcus: They tested this on six learning tasks, three simple ones and three compositional ones, showing that complexity is required for inheritance of learned solutions to actually matter in terms of fitness gains.

Ines: And there's a limitation they point out: if the task is geometrically incompatible, all the mechanisms fail equally. So this capability has its limits when dealing with truly novel or mismatched challenges.

Paper summary: Yuki: From a population genetic perspective, this suggests that successful phenotypic patterns discovered by an individual can actually become inherited traits in the population faster than waiting for slow genetic mutations to randomly generate that same pattern again.

Ines: That’s the core biological question: how does this mechanism translate into real biological advantage beyond just speed? Does it change how we think about development and inheritance?

Marcus: It suggests that organismal agency is an algorithmic mechanism with measurable adaptive value, not just a philosophical concept. The paper shows this isn't some abstract idea but a process with concrete results when you look at compositional reuse.

Ines: So, what does this mean for the broader study of evolution? We’re moving away from just genes being the sole driver of adaptation towards a system where organisms actively shape their own hereditary information.

Marcus: It reframes evolutionary computation as integrating fast, reversible learning with slow, permanent genetic change through these heritable phenotypic patterns. The Phenopoiesis Algorithm is a concrete computational model for this idea.

Ines: It suggests that the population remembers explicit phenotypic solutions across generations when those solutions are encoded in the epigenome and passed on with the genes. That’s a big shift from just relying on random genetic variation to solve problems over long timescales.

Marcus: The comparison against Baldwin Effect and Selection Only showed that this cross-generational inheritance of learned solutions is what drives that three point four times faster adaptation they found in those compositional tasks <ref:2602.00978#pg3,cross-generational inheritance of learned solutions>. It’s not just learning during development, but learning being inherited across generations, which is the key difference here.

Ines: So if we think about how this applies to anything from individual organisms to neural networks, it means incorporating a mechanism for inheriting learned solutions could make adaptation much quicker than relying solely on genetic drift or random selection.

Marcus: They suggest a future direction where this might be used to evolve neural networks by encoding learned patterns into an epigenome instead of just relying on gene-centric models. It’s about solving the sample efficiency problem in learning systems.

Paper summary: Ines: That points toward a way of designing computational systems that incorporate both fast, reversible learning and slow, permanent genetic change through these inherited phenotypic patterns. It's about integrating those timescales differently than we currently model them.

Marcus: So the main implication is that organismal agency isn't just a philosophical abstraction; it’s an algorithmic mechanism with measurable adaptive value when you have the right architecture for heritable phenotypic patterns.

Ines: That really puts things into perspective for how we look at adaptation across different timescales in biology. The Phenopoiesis Algorithm gives us a concrete computational process to test that idea.

Marcus: And it shows precisely when that access to write and read heredity pays off, specifically when the tasks require compositional solutions rather than just simple patterns.

Ines: So the authors of "Organisms Write Their Own Evolution: Direct Epigenomic Inheritance Accelerates Learning on Compositional Tasks" are showing us a way that organisms can actively contribute new information to their future selves.

Marcus: It’s a concrete computational process where an organism encodes a successful solution during development and transmits it to offspring through both the genome and the epigenome, leading to faster adaptation on complex tasks.

Ines: We see how this challenges the gene-centric view by showing that learning can be heritable in a way that provides direct fitness benefits across generations.

Marcus: And they highlight how this mechanism enables compositional reuse because organisms discover how to compose primitive elements into solutions and then transmit those recipes to their progeny.

Ines: It’s about evolutionary-level episodic memory where the population remembers explicit phenotypic solutions, which is a powerful concept for understanding adaptation in changing environments.

Marcus: The paper demonstrates that when you compare Phenopoiesis against the Selection Only model or the Baldwin Effect, Phenopoiesis outperforms them by three to six times in fitness on compositional learning tasks.

Ines: So what we have here is this direct computational test of Denis Noble’s hypothesis, showing exactly when organisms possess read-write access to their own heredity and how that access pays off.

Marcus: And it suggests a path forward for how we might approach evolving complex biological systems by integrating fast learning with slow, permanent genetic change via these heritable phenotypic patterns.

Conclusion: Ines: So, we’ve talked about how organisms can inherit learned solutions through that Phenopoiesis algorithm where they write new info into their epigenome directly, and now we need to wrap up what this whole paper means by looking at the title and authors.

Marcus: The title itself is pretty direct: "Organisms Write Their Own Evolution: Direct Epigenomic Inheritance Accelerates Learning on Compositional Tasks." It tells you exactly what they did.

Yuki: It really frames the concept of organismal agency as something operational, not just a philosophical idea. They’re arguing that this isn't some abstract notion about life; it's a concrete computational process we can actually test.

Ines: Exactly. By putting the word "direct" in there, they’re saying this isn't just indirect selection or some vague hint from the environment; it’s a mechanism where organisms actively write information into their own heredity to get results.

Marcus: And that acceleration on compositional tasks is what really gets my attention from a data science standpoint. They found that when you look at learning two patterns at once, this direct inheritance gives them three to six times faster fitness gains than just relying on the old gene-centric models.

Yuki: From a population genetics view, this suggests a new way for populations to adapt quickly when the environment demands complex behaviors, because they aren't waiting for slow mutations to randomly generate that successful combination.

Ines: It changes how we see adaptation over time. We used to think of it as genes slowly changing and selection picking the best ones; now they’re showing a pathway where successful individual learning gets baked into the offspring faster through that epigenetic layer.

Marcus: And while they showed this works on complex tasks, they also flagged a limitation: if the task is geometrically incompatible, all these inheritance mechanisms fail equally. They can't solve every kind of problem with this specific setup.

Yuki: That’s an important caveat because it grounds the hypothesis in reality; it’s not some universal solution for everything, but a powerful tool for specific types of learning challenges.

Ines: So, to sum up the core message: they’re showing that organismal agency is an algorithmic mechanism with measurable adaptive value when you have the right architecture for inheriting learned phenotypic patterns.

Marcus: It moves the conversation from just reading what’s in our DNA to understanding how learning itself can become a heritable trait across generations.

Yuki: This opens up new avenues for thinking about neural network evolution, suggesting we might encode learned patterns into an epigenome to solve efficiency problems instead of sticking only to gene-centric models.

Ines: That’s where I think we should go next: exploring how this idea applies specifically when we look at evolving complex systems like deep neural networks and what that means for computational biology.

Department of Computer Science, University of Vermont

cs.NE, q-bio.PE

Submitted: 2026-02-01

Updated: 2026-10-09

Comments: 22 pages, 2 figures,

License: http://creativecommons.org/licenses/by-nc-sa/4.0/

Importance score: 85/100

The gist: The gist The Phenopoiesis Algorithm demonstrates that organismal agency can be operationalized as a concrete computational process through heritable phenotypic patterns, achieving 3.8× faster

Key concepts

Phenopoiesis: A computational model where organisms maintain two heritable layers: a genetic layer (inherited with mutation) and an epigenetic layer (inherited unchanged). This allows organisms to learn during their lifetime and pass that learned knowledge directly to offspring via the epigenome.
Epigenome: The heritable regulatory layer of a cell that controls gene expression without changing the underlying DNA sequence. In this model, it acts as a database where organisms can write down successful phenotypic patterns discovered through learning, making this knowledge transmissible across generations.
Compositional Tasks: Learning tasks requiring organisms to learn two or more patterns simultaneously rather than just one simple pattern. The study showed that when inheritance is involved, the ability to solve these complex, multi-pattern tasks significantly boosts fitness compared to simpler learning scenarios.
Organismal Agency: The idea that organisms can actively write new information into their heredity instead of passively receiving it. Phenopoiesis operationalizes this agency by showing how learned solutions become a permanent, inherited part of the organism's evolutionary toolkit.

Terminology

Abstraction — 1× in this paper · explained in 1 episode
Adaptation — 1× in this paper · explained in 1 episode

Summary

The gist The Phenopoiesis Algorithm demonstrates that organismal agency can be operationalized as a concrete computational process through heritable phenotypic patterns, achieving 3.8× faster adaptation compared to gene-centric models <ref:2602.00978#pg12>.

How it works

The core of the proposed mechanism is the Phenopoiesis Algorithm, which allows organisms to inherit not just genes but also successful phenotypic patterns discovered during lifetime learning <ref:2602.00978#pg4> The system maintains three databases: a Genome encoding parametric information, an Epigenome encoding compositional recipes, and a Phenotype representing the current spatial configuration <ref:2602.00978#pg8> Organisms actively encode successful solutions during development and transmit them to offspring <ref:2602.00978#pg12> This is achieved through a dual inheritance structure where offspring inherit both genetic parameters and learned compositional patterns, enabling them to adapt faster through inherited knowledge rather than rediscovering solutions from scratch <ref:2602.00978#pg4>

The Phenopoiesis Algorithm Components

The algorithm operates across two nested timescales: rapid within-lifetime ontogenetic learning and slow, permanent cross-generational genetic inheritance <ref:2602.00978#pg12> During the Ontogenic Phase, each organism runs a loop of development trials where it decides whether to exploit inherited patterns (fast, proven) or explore new primitive combinations (risky, potentially novel) <ref:2602.00978#pg4> If fitness improves over personal best, the organism performs a Write-Back, recording the new pattern in its epigenome and updating its genome weights <ref:2602.00978#pg4> Selection and Reproduction occur after this phase, where offspring inherit both evolved genome (with new mutations) and full epigenome (unchanged) <ref:2602.00978#pg4>

Comparative Performance Across Models

The study compares three evolutionary architectures: Gene-Centric (GENE), Baldwin Control (BALDWIN), and Phenotype-First (PHENO) evolution <ref:2602.00978#pg12> In single-task validation, PHENO achieved 94.2% fitness, BALDWIN reached 90%, and GENE was at a baseline of 73.2% <ref:2602.00978#pg12> The critical test involves sequential task switching, where GENE suffered 51% catastrophic forgetting, while PHENO achieved only 31% forgetting <ref:2602.00978#pg15> Furthermore, in true multi-task learning, PHENO achieved a 91.2% average fitness with balanced performance, whereas GENE was constrained to a single phenotype per generation, resulting in severely imbalanced fitness scores <ref:2602.00978#pg16>

Key Findings and Implications

The results show that heritable phenotypic patterns yield 3.4× faster adaptation than gene-only models <ref:2602.00978#pg12> The advantage is not merely faster search but retrieval and refinement of stored patterns, which allows organisms to adapt quickly when the environment switches to a previously seen task <ref:2602.00978#pg15> This mechanism enables evolutionary-level episodic memory where the population remembers explicit phenotypic solutions across generations <ref:2602.00978#pg15> The paper concludes that organismal agency is not a philosophical abstraction but an algorithmic mechanism with measurable adaptive value, as sophisticated algorithmic techniques cannot overcome architectural limitations in multi-task learning <ref:2602.00978#pg18>

Future Directions

The Phenopoiesis Algorithm suggests a different approach for evolving neural networks by starting with a deep neural network and encoding learned patterns into an epigenome to solve the sample efficiency problem <ref:2602.00978#pg19> This framework reframes evolutionary computation as the integration of fast, reversible learning and slow, permanent genetic change through heritable phenotypic patterns <ref:2602.00978#pg19>.

--- Page 1 ---

How it works

Improvements for AI systems

Bold header: Heritable Phenotypic Patterns for Rapid Adaptation

The improved system will maintain an epigenome—a heritable library of phenotypic patterns discovered during lifetime learning, which allows organisms to achieve 3.8× faster adaptation compared to gene-centric models. This enables cross-generational inheritance of learned solutions rather than relying solely on within-lifetime learning.

Bold header: Compositional Reuse via Fixed Primitives

The system will be constrained to compose from a fixed library of primitives (domain-specific building blocks), ensuring that patterns learned by one organism is directly executable by offspring because both share the same primitive library. This mechanism enables compositional reuse, allowing organisms to discover how to combine elements into solutions rather than inventing new ones.

Bold header: Dual-Channel Inheritance for Timescale Coupling

The improved AI architecture will implement a dual inheritance channel alongside genetic inheritance, where the genome encodes refined weights and the epigenome stores compositional recipes. This structure couples rapid within-lifetime discoveries with slow, permanent genetic inheritance, providing the adaptive flexibility that neither single-channel mechanism can achieve alone.

Bold header: Online Explore-Exploit Decision Making

Organisms will be granted agency by deciding their learning strategy online: Each developmental trial, an organism decides: exploit inherited patterns (fast, proven) or explore new primitive combinations (risky, potentially novel). This decision is a computational signature of organismal agency that dictates whether the system reuses inherited knowledge or attempts novel discovery.

Bold header: Observable Agency Through Pattern Reuse Metrics

The system will incorporate metrics to verify its mechanism by tracking pattern inheritance and reuse quantitatively, specifically measuring if offspring actually inherit parent epigenomes and if inherited patterns show up in the offspring’s best solutions. This moves agency from a philosophical claim to a measurable empirical question.

Abstract

What if organisms could write their own evolution? For 150 years, evolution has been understood as one-way: genes provide the rules, organisms respond to selection. But if organisms could learn during their lifetime, write that learned knowledge directly into their epigenome -- the heritable regulatory layer controlling gene expression -- and pass it to offspring, then knowledge would accumulate across generations instead of being rediscovered by each one. Such direct epigenomic inheritance would constitute genuine organismal agency: writing new information into heredity rather than passively receiving it. To test whether organisms actually benefit from this capability, we implement Phenopoiesis, a computational model in which organisms maintain two independent heritable layers: a genetic layer (inherited with mutation) and an epigenetic layer (inherited unchanged). We compare Phenopoiesis against Selection Only (no learning) and the Baldwin Effect (learning influences selection but is not inherited) across six learning tasks of increasing complexity: three simple (learn one pattern) and three compositional (learn two patterns simultaneously). On simple tasks all three mechanisms perform equivalently (p = 0.559): complexity is required for inheritance to matter. On learnable compositional tasks, Phenopoiesis outperforms the alternatives by 3-6x in fitness (F = 177-358, p < 0.001, Cohen's d = 3.44-4.89), with perfect reproducibility (s.d. = 0.0 on T+Plus across all 30 runs), indicating deterministic knowledge accumulation: offspring inherit their parents' learned solutions and add new learning on top. On a geometrically incompatible task all mechanisms fail equally, revealing an inherent limit. These results give the first direct computational test of Denis Noble's hypothesis that organisms possess read-write access to their own heredity, and show precisely when that access pays off.

Sources

Evolution Strategies as a Scalable Alternative to Reinforcement Learning