Narrowing the Horizon: Quantifying Topic Saliency Shifts in Generative Monoculture

• The paper introduces a methodology for quantifying topic-saliency shifts during LLM homogenisation.Source: §1 · p. 1
• Tracking topic saliency helps evaluate monoculture’s impact and the efficacy of interventions in climate discourse.Source: §1 · p. 1
• Specialised models receive empirical support as a way to mitigate output convergence.Source: §1 · pp. 1–2

2.1 LLM Homogenisation

• Post-training is consistently identified as a main driver of output flattening, alongside factors such as training batches and decoding choices.Source: §2.1 · p. 2

2.2 Epistemic Collapse

• At scale, LLM-mediated knowledge distribution may narrow accessible viewpoints and concentrate agenda-setting power in dominant institutions.Source: §2.2 · p. 2

2.3 Homogenisation and Pluralism

• Existing homogenisation measures describe diversity loss, while pluralism research examines demographic or political representation; topic-level shifts remain underexplored.Source: §2.3 · pp. 2–3

2.4 The Wicked Problem of Climate Change

• Climate change is framed as a wicked problem without singular solutions, motivating analysis of how homogenisation shapes perceptions of viable interventions.Source: §2.4 · p. 3

3 Methodology: Quantifying Topic Saliency Shifts

• The base-model baseline measures contraction in represented topics after post-training, but does not establish the validity or quality of the surfaced topics.The analysis uses targeted keyword extraction to reduce task-irrelevant noise; topic quality is instead compared with relevant literature.Source: §3 · pp. 3–4

3.1 Models

• The evaluation covers three variants across eight open-source model families, leading closed-source models, and the climate-specialised ClimateGPT.The open-source variants include base and post-trained models near 8B parameters and higher-parameter variants; leading models are cost-optimised fast-inference versions.Source: §3.1 · p. 4

3.2 Prompting

• The study uses 64 paraphrased prompts, samples each 50 times, and tests a Chain-of-Thought variant on leading models.Actual state-of-the-art models were excluded because repeated-output sampling was prohibitively costly.Source: §3.2 · p. 4

3.3 Topic Extraction

• Topic extraction combines LLM-judged keyword extraction with human-curated categories and k-NN classification; manual checks report 0.84 F1 and 0.94 recall for gpt-5-mini, and 89% topic-classification accuracy.The pipeline produces 48 climate-related topics from 95’568 unique keywords; its hybrid annotation strategy balances exploration with bias mitigation.Source: §3.3 · pp. 4–5

3.4 Metrics

• Semantic diversity is measured as mean pairwise cosine distance among keywords pooled across 50 outputs, with repeated keywords retained to capture lexical repetition.The keyword focus filters for relevance, while the empirical distribution of spreads compares model types; a rightward shift indicates higher semantic diversity.Source: §3.4 · p. 5
• Topic prevalence shifts are estimated with a Bayesian binomial GLMM that treats post-training and model size as fixed effects and model family and prompt variation as random effects.For leading models, KDE compares distributions of prompt-level win rates across 50 repetitions; left and right shifts indicate topic suppression and promotion, respectively.Source: §3.4 · pp. 5–6

3.5 Assessing Generalisation

• The framework is also applied to homelessness and poverty, using 64 adapted prompts and producing 64 manually curated topic categories from 23,096 extracted keywords.This evaluation spans 51,200 completions from eight base models and their post-trained large variants.Source: §3.5 · p. 6

4.1 Quantifying Homogeneity

• Base models retain the highest semantic diversity; post-training reduces it, and larger post-trained models appear more homogenised.CoT prompting and specialisation increase keyword heterogeneity, with ClimateGPT appearing to preserve the most semantic diversity.Source: §4.1 · p. 6, Figure 1 · p. 6

4.2 Quantifying Topic Saliency Shifts

• Renewable Energy is significantly amplified by post-training, with a log-odds ratio of 3.8 indicating about 44.7 times greater likelihood in a large model than a smaller base model.Socio-political Approaches, Alternative Economic Systems, and Nuclear Energy are credibly suppressed after post-training.Source: §4.2 · p. 6, Figure 2 · p. 7

4.3 Comparing Mitigation Approaches

• Leading models generally intensify smaller post-trained models’ topic frequency patterns, while CoT mitigates some extremes but has limited impact and may suppress Clean Transportation.ClimateGPT generally preserves more of the distributional trends in pre-trained models, though it still shows some topic promotion or suppression.Source: §4.3 · p. 7

4.4 Assessing Generalisation

• In poverty and homelessness, alignment reduces semantic keyword spread from 0.71 ± 0.02 to 0.59 ± 0.05, promoting reskilling and universal basic income while suppressing landlordship and pension reform.Reported log-ORs are 2.4 for reskilling, 3.3 for universal basic income, -3.1 for landlordship reform, and -3.3 for pension reform.Source: §4.4 · p. 7

5.1 The Consequences of Homogenisation

• Post-training makes alternative energy and energy efficiency nearly systematic, while socio-political and alternative economic topics rarely appear.Leading models may polarise the narrowing further, promoting carbon pricing and nuclear energy more strongly while pruning alternative economic or sociopolitical frameworks.Source: §5.1 · p. 8
• The resulting energy-and-infrastructure emphasis may narrow perceived climate solutions and policy options, although social changes can be faster and more cost-effective for emissions reduction than technological fixes alone.The authors connect this concern to the flattening of a nuanced climate crisis and the possible narrowing of attempted mitigations.Source: §5.1 · pp. 8–9

5.2 Assessing Mitigations Approaches

• ClimateGPT’s frequency profile is most comparable to base models, but neither specialised fine-tuning nor CoT fully avoids topic suppression.The authors suggest specialised models may help preserve diverse climate perspectives, while noting that their suppressive tendencies still require monitoring.Source: §5.2 · p. 9

5.3 Generalisation

• In poverty and homelessness, under-reporting systemic solutions such as landlordship reform may narrow decision-makers’ actionable options; the authors say fuller characterisation of topic shifts requires further study.The evaluation supports the framework’s relevance to another wicked problem, without claiming a complete account of that domain.Source: §5.3 · p. 9

6 Conclusion

• Post-training makes renewable energy almost 50 times more likely to appear while nuclear, socio-political, and alternative economic solutions decrease in salience.The authors argue that this flattening undermines nuanced responses to complex societal challenges.Source: §6 · p. 9
• Specialised fine-tuning can partially mitigate narrowing, and the authors argue that decentralising AI systems could help counter generative monoculture.They also report that narrowing extends beyond climate change to poverty and homelessness; future work could apply the approach to other challenges and effects beyond post-training.Source: §6 · p. 9

• RAG may change topic saliency depending on retrieved content, although internal generator biases can persist even when contrary information is retrieved.Source: §7 · pp. 9–10
• Prompts lack user-specific information and may miss colloquial or persona-based consumer-chatbot interactions, while standard technical queries may reduce elicited topic diversity.Source: §7 · p. 10
• The analysis demonstrates the framework only for post-training alignment, leaving its application to architectural or other training decisions unassessed.The authors note that the controlled-comparison framework could apply to other interventions, but say further work is needed to assess them.Source: §7 · p. 10
• The study is a baseline for default model behavior; future work should examine diverse user personas and how RAG interacts with internalised homogenisation.Source: §7 · p. 10

• Maximising output diversity risks reintroducing harmful or toxic perspectives that post-training seeks to reduce, so the goal is not blanket removal of safety filters.The authors instead call for ensuring that valuable climate-discourse topics are not erased.Source: §8 · p. 10
• Contextualising value-laden terms can anchor their meaning to domains such as climate change or children’s education, but does not resolve epistemic authority.Source: §8 · p. 10
• Proposed decentralising or specialising interventions operate at the model level, while effective mitigation also requires systemic, institutional, and regulatory efforts.Source: §8 · p. 10

• The method is presented as a complementary focus on measuring LLM homogenisation.Source: §A.1 · p. 15
• The authors state that their method borrows from pluralism research.Source: §A.1 · p. 15
• The discussion is framed around how the method diverges from homogenisation measurement and pluralism research.Source: §A.1 · p. 15

• Output-space sampling measures diversity through the size of the resulting response space, but broad input requirements limit topic-level differentiation.The authors present their approach as a complementary topic-level analysis, not a replacement for existing measures.Source: §A.1.1 · p. 15
• Dispersion measures range from lexical and syntactic variation to logical diversity in propositions and semantic diversity based on underlying embeddings.The passage notes that logical diversity is not adapted to identifying thematic similarity.Source: §A.1.1 · p. 15
• Applying semantic dispersion to extracted keywords also reflects lexical narrowing, since repeated terms share vector representations and reduce pairwise semantic distance.Source: §A.1.1 · p. 15

A.1.2 Comparisons to Pluralism

• Distributional-pluralism research compares model opinion distributions with human survey distributions, but results can be highly sensitive to prompt wording and response format.Röttger et al. (2024) attributes some instability to treating LLMs as having discrete opinions rather than a superposition of personas.Source: §A.1.2 · p. 16
• Prior Overton-pluralism studies quantify coverage or human agreement but do not identify which topics fall inside or outside the window.Röttger et al. (2026) studies perceptions over 216 fixed issues, whereas this work studies the thematic breadth of issues elicited.Source: §A.1.2 · p. 17
• Topic saliency shifts link the breadth of topics elicited to the probability mass assigned to them, distinguishing this work from measures focused on values or agreement.The analysis examines how post-training and diversity-mitigation techniques influence topic likelihood.Source: §A.1.2 · p. 17

A.2.1 Vanilla Prompting

• Pre-trained models received completion-style prompts, while instruction-tuned models received instruction-style prompts; 64 variants came from a 4 × 4 × 4 lexical-component product.The final instruction was kept identical between vanilla and CoT methods for comparison; CoT sequentially asks for 15 ideas, checks their distinctness and boldness, then requests a sentence.Source: §A.2.1 · p. 17, §A.2.2 · p. 17

A.2.3 Keywords Extraction

• Keyword extraction used an LLM judge, with a manual comparison of 195 completions and an analysis of missed keywords for topic-specific omissions.At most 0.5% of missed keywords belonged to any single category.Source: §A.2.3 · p. 18

A.2.4 Keywords Classification

• The topic taxonomy was built by clustering sampled keyword embeddings, manually reviewing clusters and naming topics, then classifying remaining keywords.The same procedure was repeated for 1,314 poverty and homelessness keywords.Source: §A.2.4 · p. 18, K-Nearest Neighbours (KNN)… · p. 18

K-Nearest Neighbours (KNN) Classification

• 97.5% of keywords were classified when a topic represented at least 30% of a keyword’s 20 nearest neighbours; the classifier’s manual-audit accuracy was 88%.Unlabelled keywords were discarded; most audit discrepancies involved ambiguous cases where the assigned label remained contextually defensible.Source: K-Nearest Neighbours (KNN)… · pp. 18–19

A.2.5 Budget

• Small models ran locally and larger models used OpenRouter; the paper also provides example completions and keyword extractions in Table 10.Large-model generation, keyword extraction and embeddings cost less than £50 in total.Source: §A.3 · p. 19, §A.4 · p. 19, §A.2.5 · p. 19