Textual Strategy Evolution

• Textual Strategy Evolution is a framework that treats language-based strategies as dynamic, feedback-driven processes rather than static prompts.
• It employs methods like edit trajectory analysis, structured embeddings, and genetic algorithms to transform strategy representation.
• Applications span writing analysis, steganography, and multi-agent systems, highlighting measurable innovation and closed-loop updates.

Textual strategy evolution denotes a class of analytical and generative frameworks in which strategies expressed through language are treated as temporally changing objects rather than fixed prompts, fixed revisions, or static stylistic markers. In contemporary work, strategies may appear as explicit natural-language descriptions, structured library entries, block-level narrative devices, external latent vectors, or corpus-level patterns inferred from edit trajectories and diachronic divergence. Across these settings, the recurring operation is to externalize strategy, connect it to feedback, and permit retrieval, recombination, diagnosis, or measurement over time (Zhou et al., 8 Oct 2025, Luo et al., 27 Apr 2026, Sardo et al., 2023, Griebel et al., 2024).

1. Conceptual scope and historical antecedents

The literature uses the idea of strategy evolution in at least two closely related senses. One sense is descriptive: texts, drafts, and traditions are analyzed as records of shifting strategic behavior. Another is constructive: LLM-based systems explicitly evolve strategies during inference, often without model fine-tuning. The descriptive line is already visible in philological and writing-process research, while the constructive line becomes prominent in recent agentic and evolutionary systems.

In philological digitization, the "Textual History Tool" models a lineage as a rooted tree of “texts → commentaries → sub-commentaries → …,” segments commentaries into functional units, annotates direct and indirect evidence, and applies distance-based phylogenetic reconstruction such as UPGMA or Neighbor-Joining. This makes textual evolution measurable at the level of preserved, omitted, or innovated functional units rather than only at the level of whole works (Kanojia et al., 2022).

In writing-process analysis, Lo Sardo et al. formalize drafting as a non-linear alternation between exploration and exploitation. Their “writing cloud” represents every major revision as a sentence-level edit graph, and their measures localize planning re-entry, translation flow, and turns between the two. The framework shows that writing is not a monotone path from initial draft to final text, but a sequence of detours, returns, and local straightening (Sardo et al., 2023).

At the corpus scale, measures of novelty, transience, and precocity operationalize whether a text is “ahead of the curve.” Topic models, document embeddings, and word-level perplexity are each used to compare passages against past and future windows, and the strongest alignment with social evidence appears when texts are represented by the top quartile of passages rather than by whole-text averages. This localizes strategy evolution to “innovation hotspots” rather than uniformly distributed change (Griebel et al., 2024).

2. Representing strategy as text, structure, and latent state

A central technical question is how strategy is represented. Recent systems differ less in whether they evolve and more in what exactly they evolve.

In Auto-Stega, each strategy in the library is a structured tuple

$s = (name, defn, Tech, Scen, Char, Ex),$

and the library is maintained as a key-value map

$L = \{ \kappa_j \mapsto s_j \}$

with $\kappa_j = Embed(R_j)$, where $R_j$ is the summarizer’s rationale. This schema makes strategy searchable, composable, and admissible under metric thresholds (Zhou et al., 8 Oct 2025).

SeaEvo augments each archive entry with both executable and strategic state:

$(P_i,\;s_i,\;d_i,\;\mathbf{e}_i,\;\mathbf{b}_i).$

Here $P_i$ is the candidate program, $s_i$ its scalar fitness, $d_i$ a concise natural-language strategy description, $\mathbf{e}_i$ the embedding of that description, and $\mathbf{b}_i$ an optional behavioral success vector. This dual-space archive is designed to distinguish syntactically different programs that instantiate the same idea, and to preserve strategically promising but temporarily lower-fitness directions (Luo et al., 27 Apr 2026).

In the regulated-social-media simulation, strategies are natural-language text snippets organized into two pools: constraint strategies $L = \{ \kappa_j \mapsto s_j \}$0 for evading detection and expression strategies $L = \{ \kappa_j \mapsto s_j \}$1 for accurately transmitting information. Each strategy is further associated with counters $L = \{ \kappa_j \mapsto s_j \}$2 and $L = \{ \kappa_j \mapsto s_j \}$3, so that its selection depends not only on textual content but also on empirical history (Cai et al., 26 Feb 2025).

A different representation appears in continual multi-agent language systems, where each agent carries an external latent strategy vector

$L = \{ \kappa_j \mapsto s_j \}$4

This vector is external to the frozen LLM and updated through both reinforcement feedback and reflection embeddings. The point is not to name a strategy directly, but to maintain a compact, interpretable, and continuously adjustable abstract state (Tang, 28 Nov 2025).

Narrix uses yet another level of granularity. Example stories are segmented into coherent blocks, each block is annotated with named strategies, 1–3 sentence explanations, verbatim lexical cues, and one or two of eight creative dimensions. Strategy is therefore neither a whole-document property nor a hidden vector alone; it is a localized, inspectable device tied to specific spans of text (Zhang et al., 8 Apr 2026).

3. Evolutionary mechanisms and closed-loop update

Once strategy is represented, the next question is how it changes. The dominant answer in recent work is the closed loop.

Auto-Stega is organized as generate → evaluate → summarize → update. A Web Searcher retrieves a shortlist $L = \{ \kappa_j \mapsto s_j \}$5 of candidate strategies from the library, a Steganography LLM generates stego text under those strategies, a Scorer LLM computes multi-dimensional metrics, a Summarizer LLM emits a structured JSON entry, and the library is updated if the overall score passes threshold $L = \{ \kappa_j \mapsto s_j \}$6. The system is explicitly described as the first text-steganography system to realize truly self-evolving embedding strategies, and its lifelong mode permits later requests to retrieve stronger strategies or compositions thereof (Zhou et al., 8 Oct 2025).

The regulated-platform framework uses an LLM-driven genetic algorithm. Selection depends on a UCB-style score,

$L = \{ \kappa_j \mapsto s_j \}$7

and mutation and crossover are performed by prompting the LLM to synthesize new textual strategies from failure logs or to blend parent strategies. The interaction loop separates reflection on constraints from reflection on expression, so evasion and communicative clarity co-evolve rather than collapse into a single scalar objective (Cai et al., 26 Feb 2025).

SeaEvo generalizes this logic into three modules. Strategy Articulation turns mutation into a diagnose-direct-implement process; Stratified Experience Retrieval clusters the archive in strategy-embedding space and retrieves inspirations by behavioral complementarity; Strategic Landscape Navigation periodically summarizes effective, saturated, and underexplored directions. The strategic state is therefore not merely logged but actively used to steer future mutations (Luo et al., 27 Apr 2026).

Mind Evolution applies evolutionary search directly to candidate responses. It maintains island populations, uses Boltzmann-tournament selection, employs “Refinement through Critical Conversation” for recombination and mutation, migrates elites between islands, and periodically resets weaker islands using globally selected candidates. The method is explicitly positioned as scaling inference-time compute when a solution evaluator exists even if the underlying inference problem is not formally specified (Lee et al., 17 Jan 2025).

A related inference-only template appears in Gradient-Free Textual Inversion. There, a pseudo-word embedding is optimized through CMA-ES in a low-dimensional subspace, with

$L = \{ \kappa_j \mapsto s_j \}$8

Although the task is text-to-image personalization rather than language generation per se, it demonstrates a broader pattern: forward-only evolutionary search can update textual control variables without backpropagation through the base model (Fei et al., 2023).

4. Evaluation regimes and what they measure

Evaluation in this area is intrinsically multi-objective. Different subfields operationalize success through different observables, but all attempt to separate mere textual change from strategically meaningful change (Sardo et al., 2023, Griebel et al., 2024, Zhou et al., 8 Oct 2025, Cai et al., 26 Feb 2025, Zhang et al., 7 Dec 2025, Zhang et al., 8 Apr 2026).

Setting	Metrics	Strategic interpretation
Writing-process analysis	$L = \{ \kappa_j \mapsto s_j \}$9, $\kappa_j = Embed(R_j)$0, twist ratio $\kappa_j = Embed(R_j)$1	Focused effort, detours, translation flow
Diachronic cultural change	novelty, transience, precocity	Leading-edge positioning
Text steganography	ER, PPL, PPL*, SS, KLD, Acc	Capacity, imperceptibility, security
Regulated dialogue	$\kappa_j = Embed(R_j)$2, $\kappa_j = Embed(R_j)$3, entropy, Distinct-1	Undetected communication and transmission fidelity
DSL co-evolution	SR, PR, #LineErr, #LineEvl, auxiliary preservation counts	Conformance and preservation
Narrative-support systems	recall, recall-quality, CSI, NASA-TLX, UMUX-LITE, pairwise writing judgments	Learning, remix, and usability

In human drafting, the Shannon–Wiener index $\kappa_j = Embed(R_j)$4 measures whether edits are concentrated or distributed, the exploration measure

$\kappa_j = Embed(R_j)$5

measures detours from the shortest edit path, and the Exploration Coefficient $\kappa_j = Embed(R_j)$6 averages this quantity across versions. The twist ratio $\kappa_j = Embed(R_j)$7 then estimates how often the version trajectory remains locally straight, indicating translation flow rather than a new exploration subcycle (Sardo et al., 2023).

In cultural analysis, novelty is mean divergence from the past, transience is mean divergence from the future, and precocity is novelty minus transience for topic and embedding methods, or a past-vs-future perplexity contrast for masked-language-model methods. The top-quartile aggregation strategy is especially important because it shifts attention from mean stylistic position to the most forward-looking passages (Griebel et al., 2024).

In steganography, evaluation is explicitly trilemmatic. Auto-Stega reports embedding rate $\kappa_j = Embed(R_j)$8, perplexity, normalized $\kappa_j = Embed(R_j)$9, semantic similarity, KLD, and anti-steganalysis accuracy. Values near 50% for the detector are interpreted as perfect concealment, which makes security evaluation fundamentally different from ordinary generation benchmarks (Zhou et al., 8 Oct 2025).

5. Empirical domains and principal findings

The strongest recent results come from systems that combine explicit strategy representation with iterative feedback.

At high embedding rate (approximately $R_j$0 bpw), Auto-Stega achieves a relative 42.2% reduction in $R_j$1 versus the best SOTA baseline, LLM-Stega. On News/Movie/Tweet it reports $R_j$2 versus next best $R_j$3, semantic similarity up to 0.63 versus 0.61 baseline, and anti-steganalysis results of 51.67%, 49.65%, and 52.35%, yielding mean $R_j$4, a 1.6% improvement over the strongest prior method (Zhou et al., 8 Oct 2025).

In regulated dialogue, performance improves with repeated rounds. In the illicit-pet scenario with GPT-4o, uninterrupted turns increase from $R_j$5 to $R_j$6, while information-transmission accuracy rises from $R_j$7 to $R_j$8. The ablation study further shows that beyond round 20 the GA-powered version steadily outperforms the version without GA, with $R_j$9 and $(P_i,\;s_i,\;d_i,\;\mathbf{e}_i,\;\mathbf{b}_i).$0 by round 50 (Cai et al., 26 Feb 2025).

In inference-time reasoning and planning, Mind Evolution outperforms Best-of-N and Sequential Revision under controlled evaluation budgets. On validation sets with Gemini 1.5 Flash and approximately 800 evaluations, it reports 95.6% on TravelPlanner, 96.2% on Trip Planning, and 85.0% on Meeting Planning. The abstract also states that it solves more than 98% of the problem instances on TravelPlanner and Natural Plan using Gemini 1.5 Pro without the use of a formal solver (Lee et al., 17 Jan 2025).

In algorithm discovery, SeaEvo improves underlying evolutionary backbones in most settings and reports particularly large gains of 21% relative improvement on open-ended system optimization tasks. The representative systems-optimization results given for the Gemini-3-Flash backbone are 43.615 versus 26.267 on Prism, 3886.1 versus 3480.7 on TXN, 0.1958 versus 0.1326 on EPLB, and 0.7200 versus 0.6998 on LLM-SQL (Luo et al., 27 Apr 2026).

In writing support, Narrix uses a within-subjects study with $(P_i,\;s_i,\;d_i,\;\mathbf{e}_i,\;\mathbf{b}_i).$1 novice writers and reports improved retention, confidence, and creative adaptation of narrative strategies relative to a baseline chat-based writing interface. Pairwise evaluation using Lamp-P-Writing-Quality-RM yields 107 wins out of 144 comparisons, with $(P_i,\;s_i,\;d_i,\;\mathbf{e}_i,\;\mathbf{b}_i).$2 (Zhang et al., 8 Apr 2026).

In multi-agent latent-strategy learning, PCA plots show diffuse early clouds converging into distinct strategy zones by step 40–50, cosine similarity to the initial vector plateaus around 0.8–0.9, step-wise L2 changes are usually approximately 0.05–0.12 with spikes greater than 0.6 at key events, and inter-cluster PCA separability exceeds 0.75 after 50 updates. The reported adoption counts also indicate that the Emotional agent is used nearly as often as the average of the other four agents despite receiving no global reward (Tang, 28 Nov 2025).

6. Limitations, controversies, and open directions

A common misconception is that textual strategy evolution is equivalent to prompt rewriting or score maximization. The recent literature instead emphasizes persistent representations, retrieval over strategy space, localized measurement, and feedback loops that distinguish strategy family from surface form. SeaEvo is explicit that systems tracking progress only through executable programs and scalar fitness can fail to distinguish syntactically different implementations of the same idea, preserve lower-fitness but strategically promising directions, or detect when an entire family has saturated (Luo et al., 27 Apr 2026).

Scalability remains uneven. In LLM-driven co-evolution of textual DSLs, both Claude-3.5 and GPT-4o perform near-perfectly on small instances of 30–60 lines, but performance degrades on larger instances above 100 lines. GPT-4o may copy the original instance wholesale, while Claude-3.5 may truncate outputs when the instance exceeds approximately 170 lines. Even with prompt optimization, about 10–20% of runs show spurious errors or omissions, and the direct instance-to-instance method lacks the reusability and transparency of a generated migration script (Zhang et al., 7 Dec 2025).

The descriptive literature also cautions against assuming uniform or monotone change. In the SEW workshops, authors average $(P_i,\;s_i,\;d_i,\;\mathbf{e}_i,\;\mathbf{b}_i).$3, the twist ratio averages approximately 0.98, and turning points between exploration and translation become rarer as the writer approaches the final version. Likewise, cultural-change analysis finds that leading-edge behavior is concentrated in the top quartile of passages, not evenly spread across documents (Sardo et al., 2023, Griebel et al., 2024).

The area is also marked by clear dual-use concerns. Auto-Stega explicitly targets higher-capacity covert channels and mentions multilingual or multimodal steganography, watermarking, and covert data channels as broader applications. The regulated-platform simulation models evasive language under moderation and shows that iterative co-evolution can increase both uninterrupted dialogue turns and transmission accuracy. A plausible implication is that the same strategy-evolution machinery can be used either to study adversarial communication or to improve it (Zhou et al., 8 Oct 2025, Cai et al., 26 Feb 2025).

Open directions are already articulated in the literature. These include retrieval-augmented prompting and modular processing for larger DSL instances, reusable migration scripts instead of single evolved outputs, comparative studies of genre and collaboration in human drafting, narrower or hybrid representations for measuring cultural innovation, and compound AI systems that accumulate algorithmic knowledge over time. Narrix adds a complementary pedagogical direction: repeated use may allow writers to build a personal strategy library, moving AI-assisted writing from one-shot generation toward sustained strategy acquisition (Zhang et al., 7 Dec 2025, Griebel et al., 2024, Luo et al., 27 Apr 2026, Zhang et al., 8 Apr 2026).