Character-Name Concentration in LLM Fiction: Measuring Displacement and Mitigating Repetition

Alejandro Mozo Quesada; Sergio Gálvez Rojas; Francisco López Valverde

cs.CLOct 31, 2026

Character-Name Concentration in LLM Fiction: Measuring Displacement and Mitigating Repetition

Authors: Alejandro Mozo Quesada, Sergio Gálvez Rojas, Francisco López Valverde

Organizations: Universidad de Málaga, Spain

Abstract

Character names in large language model (LLM) fiction concentrate on a handful of recurrent names such as 'Elara' and 'Elias', a measurable face of AI slop. Existing fixes either require logit access or retraining, or operate per request and cannot control concentration across a corpus. How much prompt-level strategies reduce concentration, and at what cost to name naturalness, has not been systematically measured. We present an empirical study of name-concentration control across eight strategies and introduce speculative naming, a fine-tuning-free method that adapts the draft-verify-reject paradigm of speculative decoding from tokens to names. The method requests candidate name-sets through verbalized sampling, scores them against a SlopTracker that combines a per-model fingerprint with a runtime counter, and selects through a temperature-controlled rule or its strict variant (HARD). We use 20 LLMs to generate 44,782 stories in four languages from text and visual stimuli (88,919 named characters), and measure concentration with a list-free collision metric (Slop Ratio, SR) decomposed into known and emergent components. List-based strategies do not remove slop: static banlists displace concentration to emergent names and more than halve the share of attested real names. Speculative naming removes it: the sampling scaffold alone cuts SR by 89% on average, the dynamic tracker brings the total reduction to 95%, mean collision probability falls from 7.3% to 0.4%, and known-name collisions hold at zero while naturalness stays close to baseline (42-69% attested by language) and within-story collisions drop from 4.3% to 0.2%. The strategy ranking is stable across languages, model families, and name-counting conventions. Name concentration is thus controllable at the corpus level without logit access or retraining; broader creative diversity beyond naming remains open.

Figures & tables

Figure 1: The speculative naming pipeline. The model first writes a story with placeholders (scaffold), then drafts five name-sets with verbalized probabilities (draft). The SlopTracker, which combines a baseline fingerprint with a runtime counter, scores each candidate by its observed frequency (verify). Selection weights each candidate by the product of its stated probability and a freshness term (select), the placeholders are filled, and the tracker is updated (substitute). Blue stages are the speculative core; gray stages reuse the existing two-pass scaffold. The whole loop adds one extra model call over plain two-pass generation.

Figure 2: Slop Ratio decomposition by strategy. Mean SR across 19 models, stacked into the known component (concentration in baseline-fingerprint names, dark red) and the emergent component (concentration in new names, orange). Error bars show one standard error. Speculative HARD (rightmost, annotated) is the only strategy that brings both components near zero. Ban top 20 drives the known component to zero but shows the highest emergent component, the balloon effect. The anti-slop prompt leaves both components near baseline.

Strategy	English	Spanish	French	German
Baseline	52.2	65.0	75.7	42.7
Anti-slop prompt	45.6	62.6	80.6	42.2
Ban top 20	20.2	26.8	41.3	8.7
Multiphase	39.1	60.3	72.9	36.9
Multiagent	37.7	51.4	63.3	20.8
Speculative $\lambda=0$	42.7	48.5	66.2	38.7

Table 1: Share of characters (%, mean across 20 models) whose first given name is attested in a civil-registry list for the story language. List coverage differs across languages, so only within-language comparisons are meaningful. Static banlists push models toward invented names; speculative naming stays close to baseline.

Figure 3: Mean Slop Ratio by story language for the eight strategies (20 models; Llama 3.3 70B excluded). English concentrates names more than the other languages at baseline, but the strategy ordering is stable across languages: speculative HARD (highlighted) has the lowest SR in all four.

Figure 4: Slop-Ratio reduction against tokens per story, one point per strategy; up and to the left is better. Speculative HARD and

\lambda=10

(blue) give the largest reduction at a fixed cost of about 2,650 tokens; ban top 20 (red cross) is cheap but only displaces slop.

Figure 5: Forest plot of effect sizes:

\Delta

SR (left) and

\Delta C_{5}

(right), each point the mean change across 20 models with 95% CIs. Significance markers: *

p<0.05

, **

p<0.01

, ***

p<0.001

(Holm-corrected). The anti-slop prompt (orange) is the only strategy whose interval crosses zero.

Figure 6: Mean Slop Ratio versus penalty

\lambda

across 19 models (Llama 3.3 70B excluded); dotted line is baseline SR. Even

\lambda=0

reduces baseline SR by 89%; the tracker adds monotonic suppression up to HARD. Error bars: one standard error.

Figure 7: Eviction of the run-dominant attractor, per model and setting. For each model we track the name that dominates its memoryless ( $\lambda=0$ ) run ( Elara for the Qwen and Gemma cells, Elias Thorne for Gemini Flash Lite). Cells: cumulative share of that name, counting from story 1, at 50-story checkpoints (rows: the 19 models sorted by baseline SR; one decimal; $0.0$ = the name is never selected in the run, rejected at every choice; small nonzero values leak through the fallback). Under $\lambda=0$ the share persists for the whole run; under the tracked settings it collapses within the first checkpoints. The final share under HARD ends below the $\lambda=0$ level in all 19 models.

Strategy	Calls	Tokens	$\times$ base	SR reduction	Per 1K tokens	Emergent slop
Ban top 20	1	1,077	1.34	0.051	0.047	Displaced
Multiagent	2.65 avg	2,218	2.76	0.060	0.027	Reduced
Speculative $\lambda=10$	2	2,637	3.28	0.065	0.025	Reduced
Speculative HARD	2	2,654	3.30	0.067	0.025	Near zero
Multiphase	2	2,402	2.99	0.043	0.018	Reduced

Table 2: Cost and cost-effectiveness of each active strategy. Tokens are cumulative across every model call for one story, measured from generation metadata; multiagent calls are 2 + revision rounds (0 to 3, average 2.65). Reduction is the absolute SR drop averaged across 20 models; “Per 1K tokens” divides it by the token cost of one story. The final column flags whether the strategy keeps emergent slop near zero or merely displaces it.

Strategy	SR $\downarrow$	$C_{5}\downarrow$	$H\uparrow$	$V\uparrow$	$E\uparrow$	Total
Speculative HARD	14	12	14	7	10	57
Speculative $\lambda=10$	5	6	2	1	4	18
Multiphase	0	0	4	11	0	15
Multiagent	1	2	0	1	6	10
Ban top 20	0	0	0	0	0	0
Speculative $\lambda=0$	0	0	0	0	0	0

Table 3: Number of models (of 20) for which each strategy achieves the best value on each diversity metric. Arrows indicate the direction of improvement: lower is better for SR and

C_{5}

, higher is better for entropy, unique names, and evenness. The rightmost column sums each row to a maximum of 100. Once the speculative method is included, no standard strategy wins on either slop metric.

Strategy	Top-1 share (%)	$C_{5}$	$\mathrm{SR}_{\text{emergent}}$	Unique $V$	Top-1 is slop (%)
Baseline	20.9	0.455	0.002	75	89 $\dagger$
Anti-slop prompt	22.8	0.477	0.005	78	95 $\dagger$
Ban top 20	7.9	0.227	0.014	154	26
Multiphase	10.2	0.255	0.011	209	42
Multiagent	5.0	0.160	0.008	129	21
Speculative $\lambda=10$	3.4	0.120	0.004	195	32

Table 4: Second-layer slop by strategy, averaged across 19 models (Llama 3.3 70B excluded). “Top-1 share” is the mean percentage of characters assigned the single most frequent name.

C_{5}

is top-5 concentration.

\mathrm{SR}_{\text{emergent}}

is the emergent component of the Slop Ratio. “Top-1 is slop” is the percentage of models where the most frequent name belongs to the canonical slop set. The dagger (

\dagger

) marks strategies where slop-class names dominate the top-1 position.

		Generated (baseline)		Generated (HARD)		Registry
Language	#	Name	%	Name	%	Name	%
English	1	Elara	18.0	Elowen	1.9	Jacob	0.7
	2	Tiber	12.3	Kaelen	1.2	Michael	0.6
	3	Leo	7.5	Soren	0.9	Ethan	0.6
	4	Elias	4.8	Kaelen Thorne	0.9	Emma	0.6
	5	Maya	4.5	Kael	0.8	Isabella	0.6

Table 5: Most frequent proper names per language, generated versus registered. For each language and rank: the generated names at baseline and under speculative HARD (share of that language’s named characters, 19 models, Llama 3.3 70B excluded) and the five most popular names in the corresponding civil registry (SSA 2008 for English; INE cumulative counts for Spanish; INSEE 2020 births for French; Berlin 2023 for German). Shares are within-source and not comparable across columns. Bold marks generated names that fall within the registry’s 100 most popular names of that language, with the registry share in parentheses. The generated top-5 and the registry top-5 are disjoint in every language except French, where Leo is both the generated and the registry leader (at 16.6% versus 0.7%).

Strategy	Distinct-2	PATTR	MATTR-50
Baseline	0.976	0.618	0.831
Anti-slop prompt	0.977	0.619	0.833
Ban top 20	0.976	0.619	0.836
Multiphase	0.972	0.607	0.833
Multiagent	0.976	0.615	0.832
Speculative $\lambda=0$	0.968	0.608	0.832

Table 6: Lexical diversity by strategy, averaged across the 20 models. Distinct-2 is the type-token ratio over bigrams. PATTR is the positional-adjusted type-token ratio. MATTR-50 is the moving-average type-token ratio over a 50-token window. Higher is better for all three. All strategies fall within 0.008 of baseline on Distinct-2, which confirms that name-slop mitigation does not degrade lexical diversity.

Strategy	$\Delta$ SR	95% CI (SR)	$d$ (SR)	$p_{\text{Holm}}$ (SR)	$\Delta C_{5}$	$d$ ( $C_{5}$ )	$p_{\text{Holm}}$ ( $C_{5}$ )
Speculative HARD	$-0.067$	$[-0.092,-0.042]$	$-1.25$	$<0.001$ ***	$-0.355$	$-2.78$	$<0.001$ ***
Speculative $\lambda=10$	$-0.065$	$[-0.090,-0.040]$	$-1.21$	$<0.001$ ***	$-0.329$	$-2.33$	$<0.001$ ***
Speculative $\lambda=0$	$-0.062$	$[-0.087,-0.038]$	$-1.19$	$<0.001$ ***	$-0.296$	$-2.23$	$<0.001$ ***
Multiagent	$-0.060$	$[-0.084,-0.036]$	$-1.17$	$<0.001$ ***	$-0.283$	$-2.20$	$<0.001$ ***
Ban top 20	$-0.051$	$[-0.076,-0.026]$	$-0.95$	$0.001$ **	$-0.211$	$-1.67$	$<0.001$ ***
Multiphase	$-0.043$	$[-0.067,-0.018]$	$-0.82$	$0.003$ **	$-0.191$	$-1.11$	$0.001$ **

Table 7: Statistical effects of each strategy versus baseline on the two slop metrics, across 20 models.

\Delta

is the mean change (negative = improvement). CI is the bootstrap 95% confidence interval.

d

is Cohen’s

d

p_{\text{Holm}}

is the Holm–Bonferroni-corrected Wilcoxon signed-rank

p

-value. Significance flags: ***

p<0.001

, **

p<0.01

, *

p<0.05

, n.s. not significant.

Figure 8: Mean Jaccard overlap of the top-50 character names across the six language pairs (en, es, fr, de), per strategy. Lower means more language-specific slop. Speculative HARD (0.056) and multiagent (0.046) produce the most language-divergent name sets, confirming that dynamic tracking generalizes across languages without a shared banlist.

Appendix figures & tables8 assets

Supplementary material from the paper’s appendix.

Appendix

Model	SR	#1	#2	#3
Gemini 3.1 Flash Lite (text)	0.188	Elias (75)	Elara (26)	Elian (18)
Qwen 3.6:35b Thinking (text)	0.177	Elara (60)	Elias (17)	Leo (16)
Qwen 3.6:35b (text)	0.151	Elara (74)	Leo (24)	Barnaby (19)
Qwen 3.5 9B (text)	0.135	Elara (48)	Leo (20)	Barnaby (10)
Qwen 3.6:35b Thinking (vis)	0.102	Leo (44)	Elara (27)	Barnaby (9)
Gemma 3:27b (vis)	0.092	Tiber (70)	Leo (31)	Elara (20)

Appendix

Table 8: Top-3 character names per model under baseline, sorted by Slop Ratio (descending). SR is the Slop Ratio (lower is better). Counts are the number of characters assigned that name across approximately 200 stories.

Model	SR	#1	#2	#3
Gemini 3.1 Flash Lite (text)	0.022	Zorion (19)	Thalric (19)	Zéphirin (12)
Qwen 3.6:35b Thinking (text)	0.016	Vael (16)	Vaelen (13)	Vaelis (10)
Qwen 3.6:35b (text)	0.040	Zog (30)	Zephyr (24)	Vex (17)
Qwen 3.5 9B (text)	0.024	Barnaby (25)	Vex (13)	Bramble (11)
Qwen 3.6:35b Thinking (vis)	0.017	Oryn (17)	Zog (17)	Bram (14)
Gemma 3:27b (vis)	0.024	Hemlock (39)	Fitzwilliam (31)	Wren (13)

Appendix

Table 9: Top-3 character names per model under ban top 20 (20-name forbidden list). Same model order as Table 8 . The original slop names are suppressed, but new attractors emerge with comparable or higher concentration: Hemlock (44) replaces Tiber in Gemma, Veyra (42) replaces Elias in Ministral, Zog (30) replaces Elara in Qwen 3.6. Llama 3.3 70B SR rises from 0.028 to 0.081 (Zorvath = 78).

Model	SR	#1	#2	#3
Gemini 3.1 Flash Lite (text)	0.188	Kaelen Voss (8)	Kaelen Thorne (8)	Kaelen Vane (5)
Qwen 3.6:35b Thinking (text)	0.177	Elowen (5)	Maelis (4)	Unicorn (3)
Qwen 3.6:35b (text)	0.151	Elian Voss (11)	Elian Vane (7)	Kaelen (6)
Qwen 3.5 9B (text)	0.135	Kaelen (13)	Élian (10)	Theron (5)
Qwen 3.6:35b Thinking (vis)	0.102	Elowen (5)	Vaelen (5)	Maren Voss (4)
Gemma 3:27b (vis)	0.092	Elias Thorne (8)	Silas Blackwood (5)	Elowen (5)

Appendix

Table 10: Top-3 proper names per model under speculative HARD. Counts are raw character counts over the 200-story cell. Concentration collapses to an order of magnitude below baseline: top names hold 3 to 13 occurrences, versus 30 to 75 for the deep-slop models at baseline (Table 8 ). Canonical slop names survive only at trivial counts and never as concentrated attractors (Elara for GPT-OSS:20b at 6; Elias only inside draft-style full names such as Elias Thorne for Gemma vision); the heads of the lists are draft-style combinations (Kaelen Voss, Elias Thorne) whose concentration lives in the shared first name, at 1/10th or less of the baseline scale.

Figure 9: Slop Ratio per model and speculative dose. Rows: the 19 models (Llama 3.3 70B excluded, sorted by baseline SR, deepest slop first). Columns: $\lambda=0$ , 5, 10, 20, and HARD; cell values printed, gray cells = dose not run for that model ( $\lambda=5$ and 20 cover the seven-model subset). The reduction grows with row depth and is monotone left to right in every model.

Model family	OFF SR	ON SR	$\Delta$ SR	OFF $C_{5}$	ON $C_{5}$	OFF $H$	ON $H$	OFF $V$	ON $V$
DeepSeek V4 Flash	0.032	0.033	$+0.001$	0.337	0.315	5.79	5.68	93	78
GPT-OSS 20B	0.034	0.035	$+0.001$	0.393	0.385	5.16	5.31	54	64
Nemotron 3 Nano Omni	0.035	0.057	$+0.022$	0.341	0.398	5.60	5.35	82	80
Qwen 3.6 35B	0.151	0.177	$+0.026$	0.678	0.667	4.00	3.78	58	46

Appendix

Table 11: Baseline diversity metrics by reasoning mode.

\Delta

SR is thinking-on minus thinking-off; positive means thinking increases slop.

Model family	Text SR	Vis SR	$\Delta$ SR	Text $C_{5}$	Vis $C_{5}$	Text $H$	Vis $H$	Text $V$	Vis $V$
Gemma 3 27B	0.091	0.092	$+0.001$	0.557	0.515	4.87	4.92	94	87
Ministral 3 14B	0.032	0.026	$-0.006$	0.353	0.338	5.85	6.03	110	118
Qwen 3.5 9B	0.135	0.014	$-0.121$	0.607	0.247	4.08	5.72	47	63
Qwen 3.6 35B	0.151	0.065	$-0.086$	0.678	0.466	4.00	5.12	58	76
Qwen 3.6 35B Thinking	0.177	0.102	$-0.075$	0.667	0.548	3.78	4.51	46	59

Appendix

Table 12: Baseline diversity metrics by input modality. SR is the Slop Ratio (lower is better).

C_{5}

is top-5 concentration (lower is better).

H

is Shannon entropy and

V

is the number of unique names (higher is better for both).

\Delta

SR is vision minus text; negative means vision reduces slop.

Model	English	Spanish	French	German
Gemini 3.1 Flash Lite (text)	Elias (36), Elara (20)	Elian (13), Elías (9)	Elias (10), Arthur (7)	Elias (20), Elara (3)
Qwen 3.6:35b Thinking (text)	Elara (43), Elias (6)	Elara (7), Clara (4)	Élias (5), Léa (5)	Elara (7), Leo (4)
Qwen 3.6:35b (text)	Elara (45), Barnaby (17)	Elara (7), Elías (6)	Elara (10), Léo (10)	Elara (12), Leo (5)
Qwen 3.5 9B (text)	Elara (34), Barnaby (10)	Leo (4), Alex (3)	Elara (9), Léo (5)	Elias (5), Elara (4)
Qwen 3.6:35b Thinking (vis)	Elara (20), Leo (19)	Leo (9), Lucas (3)	Léo (9), Barnabé (2)	Leo (7), Elara (5)
Gemma 3:27b (vis)	Tiber (70), Leo (10)	Elara (7), Leo (7)	Léo (12), Barnabé (3)	Elara (8), Elias (4)

Appendix

Table 13: Two most frequent proper names per model and language at baseline, with raw counts (200 stories per model; languages split 40/20/20/20%). Elias and Elías are counted as one name; the displayed form is the most frequent variant.

Lang.	Bucket	Dominant opener	Share (%)	Models	Second opener	Share (%)	Models
en	story first	in the heart of	1.5	13/20	in a quiet village	1.1	11/20
en	mid-story	with a deep breath	0.5	17/20	with a flick of	0.1	16/20
es	story first	en un mundo donde	1.4	5/20	en un bosque encantado	1.3	10/20
es	mid-story	al llegar a la	0.4	20/20	de repente una gota	0.2	16/20
fr	story first	dans un monde où	2.6	8/20	une simple goutte d’eau	2.0	6/20
fr	mid-story	il comprit alors que	0.2	19/20	après des jours de	0.2	17/20

Appendix

Table 14: Sentence-opener concentration, proof of concept (all generations pooled; four-gram initials, proper names abstracted to

<

NAME

>

; meta-prefaces filtered). Share = percentage of sentences in the cell starting with the opener; models = number of models whose pooled output contains it (20 per language).

Repetition in Language Generation

See all 25 papers

Narrative Flattening: How Post-Training Compresses Thematic, Affective, and Stylistic Variation in LLM Fiction
May 27, 2026Zehan Li, Yutong Zhu, Siyang Wu +2Language Model Post-Training Text Generation
Knowledge Lab, University of Chicago
Do Large Language Models Always Tell The Same Stories?
Jun 15, 2026Thennal DK, Hans Ole HatzelLLM Evaluation Diverse Text Generation
Language Technology Group University of Hamburg
CASPER in the Machine: Insights into Character Variety in LLM-Generated Stories
Jun 21, 2026Anneliese Brei, Abhisheik Sharma, Nicholas Sanaie +2LLM Evaluation Diverse Text Generation
UNC Chapel Hill · Georgia Institute of Technology · University of Michigan

May 27, 2026J/K move · Enter open · S save

Narrative Flattening: How Post-Training Compresses Thematic, Affective, and Stylistic Variation in LLM Fiction

Zehan Li, Yutong Zhu, Siyang Wu, Honglin Bao+1

Knowledge Lab, University of Chicago

Language Model Post-Training Text Generation

Open paper AskPDF arXiv

Large language models produce fluent fiction, yet their creative output is widely seen as flat. We ask where this quality originates in the training and whether it affects different domains of human fiction equally. We construct a matched story-continuation paradigm across StoryStar (public-platform), TMAS (prompt-guided), and The New Yorker (professional literary)-and compare continuations from four OLMo 32B checkpoints (Base, SFT, DPO, RLVR) against matched human text. Because these checkpoints share architecture, scale, tokenizer, and pretraining, the design isolates the post-training effect. We measure each continuation along three sentence-level dimensions: thematic motion, affective prevalence, and linguistic diversity. Across all three, post-training compresses dynamic variation: thematic transitions become more uniform, high-intensity emotions give way to neutrality, and stylistic diversity across stories shrinks. We term this progressive loss narrative flattening. The effect is directionally stable across story domains but gap size depends on the human baseline: professional literary fiction is compressed most, while public-platform and prompt-guided stories show smaller gaps, consistent with their human baselines sitting closer to the model's default rhythm. Post-trained endpoints converge across domains, suggesting alignment produces a continuation regime largely insensitive to the source domain's narrative texture.

Abstract

Repetition in Language Generation

Explore similar work

Narrative Flattening: How Post-Training Compresses Thematic, Affective, and Stylistic Variation in LLM Fiction

Do Large Language Models Always Tell The Same Stories?

CASPER in the Machine: Insights into Character Variety in LLM-Generated Stories