Prochemy in Multidisciplinary Research

• Prochemy is a polysemous concept with domain-specific meanings, ranging from automatic prompt refinement for code generation to modeling astrochemical and prebiotic reaction networks, as well as programmable chemical computation in artificial life.
• In code-generation research, Prochemy employs an iterative pipeline of mutation, evaluation, and selection to optimize natural-language prompts, significantly boosting LLM performance without modifying model weights.
• In astrochemical and prebiotic contexts, Prochemy denotes emergent chemical networks and catalytic systems that model the evolution of organic complexity from interstellar environments to early Earth conditions.

Searching arXiv for the papers on arXiv and related usage of “Prochemy” to ground the response in current literature. Search (Ye et al., 14 Mar 2025) Prompt Alchemy Prochemy arXiv search: (Ye et al., 14 Mar 2025) Prochemy is a polysemous research term whose current arXiv usage spans several technically distinct domains. In code-generation research, it denotes an automatic prompt-refinement method introduced in "Prompt Alchemy: Automatic Prompt Refinement for Enhancing Code Generation" (Ye et al., 14 Mar 2025). In origin-of-life and astrochemical research, it denotes emergent prebiotic reaction networks bridging chemistry and proto-life, protoplanetary organic chemistry from prestellar cores to comets, and meteorite-catalyzed prebiotic synthesis under early-Earth conditions (Jimenez, 24 Jan 2025, 2206.13270, Cabedo et al., 2021). In artificial-life research, it denotes programmable chemical computation implemented through an object-oriented combinator chemistry (Williams, 2018). The available literature therefore does not support a single standardized definition; instead, it supports a family of domain-specific meanings organized around chemically inspired transformation, optimization, or self-organization.

1. Terminological scope

The literature associates the term with at least four research programs. These usages are related by metaphor or by explicit chemical modeling, but they are not interchangeable.

Usage of “Prochemy”	Core meaning	Representative source
Automatic prompt refinement	Execution-driven optimization of discrete natural-language prompts for code generation	(Ye et al., 14 Mar 2025)
Emergent prebiotic reaction networks	Chemistry-to-proto-life transition via RNA hairpins, peptides, and thermal cycling	(Jimenez, 24 Jan 2025)
Protoplanetary organic chemistry	Organic complexity from prestellar cores through disks to comets	(2206.13270)
Programmable chemical computation	Object-oriented combinator chemistry for self-replicating artificial organisms	(Williams, 2018)

This dispersion of meanings matters methodologically. In the prompt-optimization literature, Prochemy is an algorithmic framework operating over prompts, execution traces, and weighted selection. In the chemistry literature, it refers to reaction networks, adsorption and desorption kinetics, and inheritance of organic inventories across astrophysical environments. In artificial life, it refers to a decentralized, asynchronous, spatially embedded computational substrate. A common misconception would be to treat these as variants of one formalism; the sources instead describe distinct technical constructs that share a chemistry-derived naming scheme.

2. Prochemy as automatic prompt refinement for code generation

In "Prompt Alchemy," Prochemy is introduced to address the observation that the quality of code generated by LLMs hinges critically on the phrasing of the input prompt (Ye et al., 14 Mar 2025). The method replaces manual prompt engineering with an automated, execution-driven pipeline that iteratively refines a discrete, natural-language prompt until it maximizes the model’s expected task performance. It is designed as plug-and-play, requires no LLM fine-tuning or architecture changes, and supports multi-agent systems.

The workflow has two phases. First, it constructs a modestly sized training set of input–test pairs, drawing half the samples from existing code-generation datasets and half from an LLM-driven mutation process that augments coverage of edge-case patterns. Second, it enters an iterative optimization loop composed of Mutation, Evaluation, and Selection, transforming an initial prompt $p^{(0)}$ into a final optimized prompt $p^*$.

The evaluation step is execution-grounded. For each mutated prompt $P_i^{(k)}$ and training example $T_j=(T_j^{(\mathrm{NL})},\,T_j^{(\mathrm{test})})$, the model produces code

$$R_{ij}^{(k)}=\mathrm{LLM}\bigl(P_i^{(k)},T_j^{(\mathrm{NL})}\bigr),$$

which is executed against the test suite to obtain a binary correctness signal

$$M_{ij}^{(k)}=C\bigl(R_{ij}^{(k)},T_j^{(\mathrm{test})}\bigr)\in\{0,1\}.$$

To emphasize challenging cases, the method assigns higher weight to tasks solved by fewer prompts and computes a weighted score

$W_S\bigl(P_i^{(k)}\bigr)=\sum_j w_j\,M_{ij}^{(k)}.$

Selection retains the prompt or prompts with maximal weighted score as the next elite set. If multiple prompts tie, they are all retained; upon convergence or after reaching a preset iteration limit, a single winner is chosen at random among the top scorers. The paper formalizes the target as approximate maximization of expected pass@1 over the reachable prompt space:

$$p^* \approx \arg\max_{p\in\mathcal{P}} \mathbb{E}_{(x,\mathrm{tests})\sim T}\bigl[\mathrm{Pass1}\bigl(\mathrm{LLM}(p,x),\mathrm{tests}\bigr)\bigr].$$

The significance of this construction is that Prochemy searches a discrete natural-language space without modifying model parameters. By restricting search to the neighborhood of previous elites and using execution feedback, it balances exploration via diverse mutations and exploitation via weighted selection. That balance is central to its framing as automatic prompt refinement rather than prompt enumeration.

3. Empirical behavior in code generation, translation, and multi-agent pipelines

The reported evaluation covers natural language-based code generation and translation tasks using three LLM series (Ye et al., 14 Mar 2025). On HumanEval, GPT-3.5-Turbo zero-shot increases from 72.6% to 76.2%, and GPT-4o zero-shot rises from 90.2% to 92.1%. The abstract also summarizes this regime as improving performance by 5.0% for GPT-3.5-Turbo and 1.9% for GPT-4o over zero-shot baselines on HumanEval. When combined with Chain-of-Thought, average gains exceed +4.5%, with the example GPT-4o CoT improving from 91.5 to 92.7%.

The method is explicitly positioned as compatible with multi-agent and multi-turn systems because it never modifies the LLM’s weights. In systems like AgentCoder or LDB, it replaces the initial system prompt fed to each agent. The paper reports that, in state-of-the-art LDB, Prochemy + LDB surpasses standalone methods by 1.2–1.8%, and that GPT-4o/LDB climbs from 94.5% to 96.3%.

Results are stronger still for code translation. For GPT-4o on AVATAR, Java-to-Python performance increases from 74.5 to 84.1, and Python-to-Java increases from 66.8 to 78.2. The paper also reports zero-shot average improvements of +13.7% for Java-to-Python and +8.3% for Python-to-Java. Moreover, Prochemy maintains strong performance when integrated with the o1-mini model, yielding +4 pp zero-shot improvements.

These results suggest that the method’s main effect is not confined to one prompt style or one inference scaffold. The paper’s interpretation is that stronger seed prompts improve the quality of every subsequent turn in multi-agent pipelines, reducing error propagation and boosting final correctness without rewriting the multi-agent logic itself.

4. Optimization design, hyperparameters, and ablation evidence

The implementation choices are deliberately small-scale (Ye et al., 14 Mar 2025). The training set size is $|T|=20$, consisting of 10 examples from existing data and 10 LLM-mutated examples. The mutation pool size is $n=10$ per round, and the maximum number of iterations is $p^*$0, with typical convergence after 4–5 rounds. Code generation calls use temperature $p^*$1, while mutation calls use temperature $p^*$2.

The pseudocode includes an early-stopping rule: if, after at least three iterations, the scores of the elite set have stabilized for three rounds, the loop terminates. This is consistent with the reported convergence behavior and with the method’s emphasis on optimization efficiency rather than exhaustive search.

The ablation findings are especially informative. Removing iteration halves the gain. Fixing all 10 iterations, rather than using early stop, is approximately 1.5 percentage points worse. Omitting weighted scoring increases convergence rounds by approximately 20% and lowers pass@1 by 2–7%. These observations identify three design decisions as functionally important: iterative refinement, task-difficulty-aware weighting, and early stopping.

The only real prerequisite is an execution environment for fast pass/fail feedback and a small set of testable examples. This prerequisite sharply distinguishes Prochemy from fine-tuning-centric methods: the optimization signal is derived from executable correctness rather than gradient updates over model weights.

5. Prochemy in prebiotic and astrochemical research

In the origin-of-life literature summarized here, Prochemy is explicitly defined as the study of emergent prebiotic reaction networks bridging chemistry and “proto-life” (Jimenez, 24 Jan 2025). Jimenez’s model postulates that Earth’s daily temperature cycles could have provided a planetary thermocycler that created self-replicating RNA hairpins and simultaneously templated amino acid polymerization in a primordial PCR well of prebiotic molecules. The model includes reversible base pairing, ligation into phosphodiester-linked oligomers, intramolecular stem–loop formation, and hairpin duplication driven by alternating cooling and warming. It also proposes a reciprocal nucleopeptide replicator in which RNA hairpins and peptides mutually reinforce replication and synthesis. The rate equations presented for this reciprocal system couple RNA production, peptide production, and degradation or dilution terms.

The same source models the temperature cycle as

$p^*$3

with night favoring annealing and midday favoring strand separation. A plausible implication is that, within this usage, Prochemy denotes network-level emergence rather than isolated synthesis steps.

A second astrochemical usage identifies Prochemy with protoplanetary organic chemistry extending from prestellar cores to comets (2206.13270). The summary states that organic complexity builds up already in the earliest, cold phases of Solar-type star formation and is then processed and inherited all the way to comets. Representative values are stage-specific: in prestellar cores, methanol has $p^*$4–$p^*$5; in hot corinos, $p^*$6 reaches $p^*$7–$p^*$8; in protoplanetary disks, $p^*$9 is reported at $P_i^{(k)}$0–$P_i^{(k)}$1; and in cometary ices, $P_i^{(k)}$2 is given as $P_i^{(k)}$3–$P_i^{(k)}$4 relative to $P_i^{(k)}$5. The same review emphasizes both grain-surface and gas-phase chemistry, with adsorption, diffusion, thermal desorption, chemical desorption, and photodesorption all contributing to the observable inventory.

A third chemically oriented usage concerns meteorite-catalyzed Prochemy under early-Earth conditions (Cabedo et al., 2021). Cabedo et al. investigate Fischer–Tropsch Type reactions from mixtures of CO and $P_i^{(k)}$6 at 1 atm on different chondritic materials. The reactor is a 5 mL stainless-steel tubular cell heated from 200 to 600 $P_i^{(k)}$7C in 50 $P_i^{(k)}$8C increments, with gas feed $P_i^{(k)}$9. The main products are alkanes and alkenes, with alcohols in smaller amounts. Quantitatively, methane accounts for 25–75% of converted CO, ethane for 5–20%, ethylene for 2–10%, methanol for at most 0.6%, ethanol for at most 0.3%, and $T_j=(T_j^{(\mathrm{NL})},\,T_j^{(\mathrm{test})})$0 for approximately 20–30%. The metal-only extract outperforms bulk powders in both onset temperature and peak conversion, and the paper concludes that the metal phase, specifically Fe–Ni alloys, is the principal active component. In this usage, Prochemy refers to a prebiotic feedstock-generating catalytic network rather than to RNA-based replication.

Across these chemistry papers, the term consistently points toward a transition zone between abiotic chemistry and systems capable of inheritance, accumulation, or catalytic amplification. What differs is the substrate: planetary thermal cycling, interstellar and protoplanetary organic inventories, or chondritic catalysis.

6. Prochemy as programmable chemical computation in artificial life

In artificial-life research, Prochemy denotes programmable chemical computation realized through an object-oriented combinator chemistry (Williams, 2018). The underlying system constructs an artificial organism as a moving, self-replicating, spatially distributed assembly of elemental combinators called a roving pile. Primitive combinators have a monadic interface, and composite combinators are represented as binary trees with Kleisli composition,

$T_j=(T_j^{(\mathrm{NL})},\,T_j^{(\mathrm{test})})$1

To manipulate arbitrarily large trees without destructive updates, the system uses a three-part zipper representation consisting of Front, Focus, and Back.

The organismal architecture is decentralized. Base groups form a footprint on a 2D lattice and support local advance, retreat, and diffusion operations. Above each base group is a vertical stack of groups; methods in the same stack run concurrently, while methods in different stacks run in parallel. Genes are stored as zippers that roam among stacks like bacterial plasmids, and interactions occur through asynchronous message passing via bonds and guarded combinators.

Replication is modular. Six methods, acsA–acsF, implement zipper copying in a six-stage pipeline, and any new method becomes autocatalytic when its zipper is placed alongside this pipeline. The architecture is explicitly designed to support redundancy, degeneracy, and parallelism as mechanisms for mitigating increased replication costs. The expected replication time is stated to scale as

$T_j=(T_j^{(\mathrm{NL})},\,T_j^{(\mathrm{test})})$2

where $T_j=(T_j^{(\mathrm{NL})},\,T_j^{(\mathrm{test})})$3 is the average gene length, $T_j=(T_j^{(\mathrm{NL})},\,T_j^{(\mathrm{test})})$4 the number of genes, and $T_j=(T_j^{(\mathrm{NL})},\,T_j^{(\mathrm{test})})$5 the number of parallel copies of the rate-limiting method. The stated significance is that increasing modular complexity can reduce replication time by enabling more parallel subprocesses.

This usage is conceptually distinct from chemical Prochemy in the origin-of-life literature and from prompt Prochemy in code-generation research. Nevertheless, all three deploy chemistry as an organizing abstraction for search, transformation, and self-maintaining process.

7. Cross-domain interpretation

Taken together, the literature supports a narrow but important conclusion: Prochemy is a domain-dependent term whose semantics are anchored by transformation under feedback. In prompt optimization, feedback is executable correctness over test suites (Ye et al., 14 Mar 2025). In prebiotic and astrochemical work, feedback is encoded in reaction kinetics, thermal cycling, adsorption and desorption processes, and inheritance of molecular inventories across astrophysical environments (Jimenez, 24 Jan 2025, 2206.13270, Cabedo et al., 2021). In artificial life, feedback is provided by local execution conditions, bond structure, and replication dynamics in a concurrent combinator system (Williams, 2018).

This suggests a family resemblance rather than a unified field definition. The shared pattern is iterative refinement or accumulation under constraints; the operative objects, however, are different in each domain: prompts, molecules, catalysts, or executable combinators. Any technical use of the term therefore requires immediate disambiguation by context, since the underlying formalism, observables, and validation protocols vary substantially across the cited research.