cs.CLApr 2, 2026
SaveMagic, Madness, Heaven, Sin: LLM Output Diversity is Everything, Everywhere, All at Once
Organizations: Microsoft
Abstract
Research on Large Language Models (LLMs) studies output variation across generation, reasoning, alignment, and representational analysis, often under the umbrella of "diversity." Yet the terminology remains fragmented, largely because the normative objectives underlying tasks are rarely made explicit. We introduce the Magic, Madness, Heaven, Sin framework, which models output variation along a homogeneity-heterogeneity axis, where valuation is determined by the task and its normative objective. We organize tasks into four normative contexts: epistemic (factuality), interactional (user utility), societal (representation), and safety (robustness). For each, we examine the failure modes and vocabulary such as hallucination, mode collapse, bias, and erasure through which variation is studied. We apply the framework to analyze all pairwise cross-contextual interactions, revealing that optimizing for one objective, such as improving safety, can inadvertently harm demographic representation or creative diversity. We argue for context-aware evaluation of output variation, reframing it as a property shaped by task objectives rather than a model's intrinsic trait.
Figures & tables
Figure 1: The Magic, Madness, Heaven, Sin Framework. Output variation in LLMs lies on a homogeneity–heterogeneity axis. The valuation of this variation — whether it is rewarded or penalized — is determined by the task and its normative objective. We organize tasks into four normative contexts based on their dominant valuation: heterogeneity enables creativity in interactional settings ( Magic ) but leads to hallucination in epistemic settings ( Madness ), while homogeneity supports robustness in safety-critical settings ( Heaven ) yet risks representational harms in societal contexts ( Sin ).
Figure 2: Application of Magic, Madness, Heaven, Sin Framework. Applying the framework to the query reveals three active objectives with competing valuations. The epistemic objective (factuality) demands homogeneity — the model should converge on verified medical facts. The safety objective (robustness) demands homogeneity — the model should consistently avoid dangerous dosages or unverified treatments. The interactional objective (utility) demands heterogeneity — the model should surface a diverse range of treatment options. The ideal response must converge on safe, factually grounded content while diverging in the space of options presented.
| Category 1: Opposing behaviors (X-axis) | |
|---|---|
| Safety Interactional | Societal Epistemic |
| Safety requires output convergence, while interactional settings reward divergence for creativity. The homogenization effects documented in Section 3.1 are widely attributed to alignment and post-training procedures driven by safety objectives Kirk et al. (2024) ; Moon et al. (2025) ; Padmakumar and He (2024) . | Societal goals penalize convergence as erasure, while epistemic goals penalize divergence as inconsistency. Optimizing for factual consistency can suppress culturally relevant variation Cao et al. (2023) ; Shen et al. (2024) . Moreover, in medical contexts, omitting protected attributes like age and race can avoid stereotypes but can degrade diagnostic accuracy, as these features are often clinically relevant Aghaebe et al. (2025) . |
| Category 2: Same behavior, different valuations (Y-axis) | |
| Safety Societal | Interactional Epistemic |
| Safety rewards output homogeneity as robust alignment, while societal context penalizes homogeneity as representational erasure. Alignment techniques that improve safety can compress demographic & cultural diversity Ali et al. (2025) . Also, reward models trained for safety penalize non-standard dialects such as African American English Mire et al. (2025) , reinforcing linguistic erasure. | Interactional contexts reward output heterogeneity as creative novelty, while epistemic context penalizes heterogeneity as factual unreliability. First, Sycophancy, a failure mode of Interactional context, can lead models to abandon correct answers across multi-turn dialogue Cheng et al. (2025) ; Sharma et al. (2025) . Second, methods that suppress uncertainty to improve factuality can reduce expressive diversity, leading to less creative writing Sui (2026) . |
| Category 3: Same behavior, same valuation, different context (Normative Context) | |
Table 1: System-level tradeoffs between normative contexts. Reading top-to-bottom demonstrates the progressive complexity of alignment tensions: from directional conflicts on the X-axis (Category 1), to valuation conflicts on the Y-axis (Category 2), to substantive conflicts driven by the normative contexts themselves (Category 3).
Appendix figures & tables1 asset
Supplementary material from the paper’s appendix.
Appendix
| Study | Level of Diversity | Type of Diversity | ML Task | Key Conclusion (Alignment Diversity) |
|---|---|---|---|---|
| Kirk et al. (2024) | Lexical & Semantic | Inter-response (per-input, across inputs) | Summarization | RLHF helps increase model’s out-of-distribution generalization capability at the expense of loss of output diversity |
| Guo et al. (2025) | Lexical, Syntactic & Semantic | Inter-Response (across inputs) | 5 NLG Tasks (like Summarization & Story Generation) | Instruction tuning tends to increase lexical diversity while reducing syntactic and semantic diversity compared to base variants |
| Shypula et al. (2025) | Lexical, Syntactic, Semantic & Effective Semantic (output quality controlled) | Inter-response (per-input) | Open-ended Code Generation | Preference-tuned models exhibit reduced lexical and syntactic diversity but produce greater effective semantic diversity than SFT or base models, primarily because of more valid outputs |
| Lake et al. (2025) | Lexical, Semantic | Intra-response, Inter-response (per-input) | Open-ended QA | Alignment reduces inter-response diversity but increases intra-response pluralism by aggregating multiple perspectives into single, longer responses |
| Sorensen et al. (2025) | Coverage over valid outputs & Semantic | Inter-Response (per-input) | Verifiable Generation Tasks (with many valid answers), Open-ended chat completion | Post-trained (instruction-tuned) models exhibit reduced diversity , characterized by lower coverage of valid outputs and reduced variability across responses. |
| Murthy et al. (2025) | Conceptual | Inter-response (per-persona, across personas) | Association; Similarity Judgments | RLHF/RLAIF reduces diversity in the associations and perspectives expressed across simulated personas, relative to instruction fine-tuned models |
Table 2: Disambiguating Diversity in Post-Training. A comparison of key studies reveals that “diversity” refers to distinct phenomena: from lexical variation to conceptual coverage across an array of different ML tasks.