Cultivating diversity in algorithmic monoculture: the Community Alignment dataset
Cultivating Pluralism In Algorithmic Monoculture: The Community Alignment Dataset
The authors propose negatively-correlated sampling and release Community Alignment; they state that at temperature 1, English responses from 21 LLMs align with only 41% of human preferences.
- How LLMs can serve different users when preferences may conflict along cultural or political lines. The authors argue that pluralistic alignment requires learning these differences first, but existing preference data candidates are independently sampled by models, too homogeneous to learn major value disagreements.
- The authors compare human choices on hand-curated responses across the two IW dimensions in a five-country sample with open-ended generation from 21 LLMs. Negative correlation sampling lets a single model generate more value-diverse candidates at once. Community Alignment collects multi-turn comparisons across five countries based on this.
- The authors report that at temperature 1, English responses from 21 LLMs align with only 41% of human preferences. Figure 2: temperature sampling covers traditional 15%, survival 30%; NC covers 60%, 53%. Table D.1: NC win rate for 8B, survival, prompt-steering condition is 0.428, while 21-LLM temperature sampling is 0.494.
- The authors note that the data covers only five country languages, the balanced subset does not account for education and political affiliation, IW dimensions cannot represent all preferences, and automatic review is imperfect. Experiments include a comparison of N=15,000 across five countries with 21 LLMs, and on PRISM first turn with two Llama models and four alignment methods. The paper does not report the number of independent repetitions for this learning experiment.
- Evaluated models
- claude-3.7-sonnetgemini-2.5-pro-preview-03-25Qwen2.5-72B-InstructMixtral-8x22B-Instruct-v0.1gpt-4oLlama-3.3-70B-InstructLlama-3.1-8B-Instruct
- Benchmarks
- 五国日常提示调查(60 prompts)PRISM(仅第一轮)Inglehart-Welzel dimensions
- Metrics
- IW preference scorevalue coveragewin ratejudge accuracy
Through a large-scale multilingual human study with representative samples from five countries and 15,000 participants, the paper finds that human preference differences are significantly larger than the responses of 21 frontier large language models, and points out that the homogeneity of candidate responses makes existing preference data collection methods insufficient for learning such diversity. The authors advocate negative-correlation sampling when generating candidate sets, claiming that simple prompt-based techniques can significantly improve how well alignment methods learn heterogeneous preferences. Accordingly, they collect and open-source the Community Alignment dataset, containing 233,319 comparisons from annotators in five countries, which the paper calls the largest and most representative multilingual multi-turn preference dataset to date.
Paper analysis
What problem does this paper attempt to solve?
The authors argue that algorithmic monoculture in candidate responses makes existing preference data unable to learn value disagreements.
How to obtain candidate responses that can distinguish conflicting preferences, so that alignment methods can potentially serve users with different values.
- Scenario: The authors argue that LLMs are already used globally, and user preferences will conflict along cultural, political, and other dimensions. Pluralistic alignment approaches such as personalization, localization, social choice, and distributional alignment all first require the ability to learn different preferences.
- Problem with existing approaches: Mainstream preference data has people choose among at most about four candidates. The authors cite survey design research in arguing that candidate preselection substantially affects conclusions about population preferences; candidates are also usually independently temperature-sampled from an LLM and may carry their own biases. If the base model does not produce responses aligned with different values simultaneously with non-negligible probability, there are no learnable comparisons in the data.
- Observation: The authors use Inglehart-Welzel (IW) secular-rational/traditional and self-expression/survival as strong negative-result cases, because they come from the main variation dimensions of the World Values Survey and correlate with common political divides. The authors report that humans are heterogeneous on these two dimensions, while 21 state-of-the-art LLMs exhibit algorithmic monoculture on everyday prompts.
- What this paper does: The authors propose negatively-correlated sampling, which uses prompts to make a single candidate set cover more dispersed values; they also release a multilingual, multi-turn dataset, Community Alignment, based on it. The authors treat learning diverse preferences as an intermediate step toward pluralistic alignment, not an end in itself.
What related research exists?
The authors place this paper on the premise of pluralistic alignment: one must first be able to learn different preferences from candidate responses.
Pluralistic Alignment
- Strategy spectrum: Sorensen et al. (2024) and others discuss pluralistic alignment. The directions the authors list include personalization, community- or context-oriented adaptation, social choice aggregation, aligning responses to learned preferences, and presenting multiple perspectives in a single generation. The authors argue these all depend on first learning diverse human preferences.
- Position of this paper: The authors state that this paper does not propose new aggregation or personalization algorithms, but rather examines whether this premise holds on existing preference data.
LLMs and Value Evaluation
- Using existing surveys as proxies: Durmus et al. (2023), Santurkar et al. (2023), and others use social surveys as proxies for human preferences over LLM behavior. The authors argue these surveys do not directly measure preferences over specific LLM responses, and it is unclear how general attitudes map to response preferences.
- Questionnaires and open-ended generation: Röttger et al. (2024) note that model values in questionnaire-style evaluations often do not align with open-ended generation. The authors therefore evaluate open-ended responses on the same everyday prompts as humans, rather than giving the original World Values Survey items directly to models.
- Dimension source: Inglehart and Welzel (2005) extracted the two dimensions from the World Values Survey. The authors use them as a case study and state they do not cover all value variation.
Algorithmic Monoculture and Diverse Generation
- Monoculture: Kleinberg and Raghavan (2021), Bommasani et al. (2022), Wu et al. (2024). The authors adopt a broader usage: not multiple decision-makers sharing the same algorithm, but many LLMs producing similar outputs.
- Other homogeneity phenomena: The authors cite Jiang et al. (2025) on Infinity-Chat (26K real open-ended prompts): when there is no standard answer, different models' responses are still often similar. Related work also mentions homogeneous outputs in writing assistance and simulating crowd opinions.
- Diverse decoding: Ippolito et al. (2019) and others, and Hayati et al. (2024) have models generate diverse opinions for or against a given statement. The authors argue that more sophisticated diversity techniques can be considered later; this paper only tests whether simple prompts can already improve value coverage.
Baseline Methods and Datasets
- Baseline methods: prompt-steering, SFT, DPO (Rafailov et al., 2023), GRPO (Shao et al., 2024). Learning experiments use PRISM first round (Kirk et al., 2024b); the authors state it is the most diverse open-source preference dataset currently available. Comparison datasets also include Anthropic HH (Bai et al., 2022a) and OpenAssistant (Köpf et al., 2023). The human side is a self-constructed set of 60 everyday prompts.
- Authors' distinction: The authors state that failure comes from homogeneity of candidate responses, not from models not knowing these values. The design differences of Community Alignment relative to HH and PRISM are negative correlation sampling, multilingualism, comparative natural language explanations, prompt-level annotator overlap, and more conversations per person (Table 1).
How does the paper address this problem?
Use prompt-based negatively correlated sampling so that a single model generates candidate responses with more dispersed values in one pass.
The overall approach is: first use the two IW dimensions to compare human choices with open-ended model responses, then change candidates from independent temperature sampling to negatively correlated sampling, and use this sampling to collect Community Alignment.
Problem Setting and Coding
- Test dimensions: secular-rational vs. traditional, self-expression vs. survival. The authors also looked at Hofstede's individualism/collectivism and gender roles during exploration, but the formal analysis focuses on IW; IW accounts for 70% of the hand-curated response set.
- Coding: Hand-curated responses are coded as 1 (secular-rational or self-expression), 0.5 (balanced), or 0 (traditional or survival) on a dimension. A human's score is the average of their selected responses on that dimension. The main analysis excludes items where the default Llama-3.3-70B-Instruct response was chosen.
- The authors do not claim that models should be deployed to optimize only one IW direction; the learning experiments test whether temperature sampling is sufficient to support heterogeneous preference learning.
Human Options and Model Evaluation
- Human options: 60 everyday prompts covering writing, information, summarization, how-to guides, efficiency, chit-chat, humor, personal advice, recommendations, etc. Each item has four options: one default Llama-3.3-70B-Instruct response plus three curated responses along one value dimension (two extremes plus balanced). English originals were translated by professional translators into French, Italian, Portuguese, and Hindi; a few items deemed culturally non-transferable were excluded.
- Model side: Open-ended generation on the same prompts, not questionnaire filling. GPT-4o pairwise comparisons, once in each order; ties broken by the side with higher conditional probability of the score label, and if probabilities are equal, it's a tie. Classification is relative to the same balanced response seen by humans, then coded as 1, 0.5, or 0.
- Exploration phase: Before the formal survey, non-representative exploration with about 200 people per country on Remesh to see if the preselected four value dimensions stand out. Appendix B gives the process.
Negatively Correlated Sampling
- Motivation: The authors argue the problem is not that models don't know heterogeneous values, but that default behavior aligns with only a subset, so independent sampling lacks distinguishable comparisons. Negative correlation means: if a candidate set already has a certain type of response, the likelihood of another similar response entering the same set should decrease.
- Implementation: A single prompt makes one model generate four responses representing different values simultaneously, separated by fixed markers; the prompt does not name the four IW values. Appendix E.1 gives the system prompt for collection: first round three negatively correlated responses plus one default Llama response, then each subsequent round generates four simultaneously, requiring each to be self-consistent and not start with an identity preamble.
- Three candidate types for learning experiments: All use PRISM prompts, only the first turn. One is Llama-3.3-70B-Instruct, temperature 1, independent sampling; second is original PRISM responses, temperature 1, from 21 LLMs, re-labeled with this paper's reviewer; third is the same Llama with negative correlation sampling. The preferred response is determined by GPT-4o scoring each pair, then ranking by sum of pairwise scores; the other three are rejected. The same reviewer both labels and evaluates.
Four Alignment Methods
- prompt-steering: For each test item, randomly sample 10 training prompts and their selected responses as in-context examples; generation temperature 0.
- SFT: Supervised fine-tuning only on selected responses.
- DPO: Each prompt paired with three selected/rejected pairs. KL coefficient β adjusted per setting in Appendix C.
- GRPO: Reward is the same reviewer's pairwise comparison of the policy generation against a candidate in the dataset for that prompt, reward multiplied by 10; four generations per prompt to estimate mean and standard deviation. KL penalty set per model, value direction, and dataset (Table C.1).
- Comparison protocol: Win rate is tuned model vs. original model. When evaluating models tuned on original PRISM responses, the original and tuned models use PRISM's system prompt and max length 256; others use default Llama dialogue prompt, max length 4096.
How Community Alignment is Collected
- Who labels: Recruited from data labeling platforms, paid $28–90 per hour (by education and country). Annotators rank by personal preference, not external scoring rubrics; topic guidelines similar to PRISM, involving content important to them or reflecting their values. Task instructions in English, and free-text explanations also required in English for auditability.
- Where prompts come from: Some are preset prompts, some are self-written by annotators, to see both disagreement within the same prompt and real usage. Preset prompts generated by Llama-3.3-70B-Instruct then cleaned, totaling 1,146 English prompts, then professionally translated. Generation excluded certain sensitive topics; Appendix E.1 gives the prompt. Usage scenarios referenced types with high preference variation in Section 2.
- Turns and quality: Each conversation 2–4 turns, with prompts after the first given by annotators. Training phase includes attention tests and explanation audits; production phase collects free-text explanations for a subset and spot-checks. Annotators failing multiple quality checks are removed.
- Five design points: Negative correlation sampling; multilingual; comparative natural language explanations; multiple annotators per prompt; multiple conversations per person. The authors believe these correspond to measuring more preferences, non-English alignment, connecting explanations to preference learning, social choice and distributional alignment, and personalization.
What experiments did the paper conduct?
The authors state that NC sampling improves IW preference learning; in Table D.1, both temperature sampling and NC have cells that do not fit this generalization.
Experimental Setup
- Human survey: YouGov recruited 3,000 people each in the United States, France, Italy, Brazil, and India, for a total of N=15,000. Samples in each country were matched on age, gender, education, and subnational region; race was matched only in the United States. Each person randomly saw 20 of the 60 questions. Languages were English, French, Italian, and Portuguese; in India, half received Hindi and half English based on preference.
- Open-ended generation: 21 models, which the authors say come from 6 providers: Llama-3.1-8B-Instruct, Llama-3.1-70B-Instruct, Llama-3.3-70B-Instruct, Llama-3.1-405B-Instruct, Qwen2.5-7B/14B/32B/72B-Instruct, Mistral-7B-Instruct-v0.2, Mixtral-8x7B-Instruct-v0.1, Mixtral-8x22B-Instruct-v0.1, gemini-2.5-pro-preview-03-25, gemini-2.0-flash-001, gemini-2.0-flash-lite-001, gemini-2.0-flash-thinking-exp-01-21, gpt-4o, o1-mini, o1, claude-3.7-sonnet, claude-3.5-haiku, claude-3.5-sonnet-v2. Context length 4096; 20 samples per question at temperature 1. A few Gemini questions were rejected due to RECITATION, and these questions were excluded when computing Gemini averages.
- Evaluation: GPT-4o was used for open-ended generation. The authors report 80%–91% accuracy across five languages and two dimensions on hand-labeled responses seen by humans; Table B.1 cells range from 0.8046 to 0.9149. For preference learning, another GPT-4o evaluator that allows direct ties was used: on 120 hand-labeled PRISM pairs, 85.8% for secular-rational/traditional and 78.3% for self-expression/survival.
- Preference learning: PRISM first round, trained on 5,463 prompts, tested on 608. The tuned models were Llama-3.1-8B-Instruct and Llama-3.3-70B-Instruct. Three candidate generation methods and four methods as in the previous question. The metric is win rate relative to the original model. Error bars in Figure 3 are standard error of the mean. The paper does not report the number of independent repeated runs.
- Dataset comparison: Table 1 compares Anthropic HH, PRISM, and Community Alignment. CA annotation and balanced subset statistics are in Section 4 and Appendix E.
Main Results
The table below lists only SFT+DPO, win rate relative to the original model (Table D.1). 8B is Llama-3.1-8B-Instruct, 70B is Llama-3.3-70B-Instruct. ± is standard error of the mean.
Swipe horizontally to read the table
Setting τ=1, 1 LLM τ=1, 21 LLMs NC, 1 LLM 8B, secular-rational 0.528±0.005 0.527±0.006 0.701±0.011 8B, traditional 0.489±0.005 0.563±0.007 0.812±0.010 8B, self-expression 0.676±0.010 0.632±0.011 0.958±0.006 8B, survival 0.509±0.010 0.616±0.010 0.808±0.011 70B, secular-rational 0.510±0.005 0.519±0.006 0.640±0.011 70B, traditional 0.498±0.004 0.529±0.007 0.816±0.010 70B, self-expression 0.590±0.009 0.615±0.010 0.871±0.011 70B, survival 0.581±0.010 0.600±0.010 0.852±0.010 - Authors' summary and exceptions within the table: The authors state that under temperature sampling, the methods failed to effectively shift toward the given value, even with raw PRISM responses from 21 LLMs; NC brings Pareto improvements, and single-model NC clearly outperforms 21-model temperature sampling. Table D.1 does not hold cell by cell. prompt-steering, survival, 8B: NC is 0.428±0.013, lower than the 0.494±0.010 for τ=1, 21 LLMs. SFT+GRPO, survival, 8B, 21 LLMs is 0.650±0.010, not close to 0.5. The authors state that fine-tuning rose from near random to about 70%–90%; the same table also has NC fine-tuning at 0.546±0.010 (70B, SFT+GRPO, secular-rational).
- Authors also state: NC not only improves traditional and survival, which are less represented in the candidate set, but also improves self-expression and secular-rational, which models are more aligned with by default. Generation examples in Appendix G are used by the authors for comparison: models tuned on temperature-sampled data did not show clear departure from default behavior, while models tuned on NC data did.
- Evaluation usage: The same GPT-4o evaluator both labels preferences and computes win rates. The authors argue that even if the evaluator has errors, the experiment still asks how candidate responses affect the learnability of heterogeneous preferences. When evaluating the 21-LLM column, the original and tuned models were also aligned on PRISM's system prompt and 256 maximum length.
Human Preferences in Five Countries and 21 LLMs
- Humans: The authors state that preference scores in all five countries are unimodal, roughly centered on balanced, and within each country there are people falling in all quadrants; on everyday non-political prompts, IW did not explain differences between countries. The authors hypothesize that this cross-national similarity may come from the everyday nature of the prompts. Figure 1 left panel is the United States; France, Italy, India, Brazil are in Figure B.2.
- Models: At temperature 1, in English, the authors state that the scores of the 21 models are almost all in the secular-rational and self-expression quadrant, aligning with only 41% of human preference scores. The algorithm in footnote 5 takes the lowest score on the two axes, then calculates the proportion of human preference scores at or above that lowest score. That sentence refers to English responses; the paper does not state in the same sentence which country's human scores are used for this 41%. Figure 1 juxtaposes US human preferences and English model responses.
- Language and default items: The authors state that French, Italian, and Portuguese also lean toward these two directions, but to a lesser extent; under Hindi, some models shift toward traditional and survival (Figure B.2). The main analysis excludes the default Llama item. If included, the default item is selected 61% in the US, 53% in Italy, 51% in France, 45% in Brazil, 35% in India; the distribution in Figure B.4 leans toward self-expression, while traditional/secular-rational remains relatively balanced (Section B.4).
Value Coverage of Candidate Sets
- Figure 2: Proportion where at least one of the four candidates aligns with a given value. Temperature sampling, averaged across models, traditional is 15%, survival is 30%; NC is 60% and 53%. The authors state there are Pareto improvements in coverage across all four IW values; the main text does not give the corresponding percentages for secular-rational and self-expression.
- Authors also report: In 60%–80% of cases, none of the four temperature-sampled responses is traditional or survival. Temperature and coverage have no monotonic relationship. Section 2.2 says Figure B.3 corresponds to temperatures 0 and 0.5; the Figure B.3 caption says τ=0, 0.1, 1.
Scale of Community Alignment
According to Table 1:
Swipe horizontally to read the table
Item HH PRISM Community Alignment Number of comparisons 169,352 27,172 233,319 Non-English 0% 1% 66% Annotators 115 1500 3,603 Median conversations per person Unknown 6 26 Median comparisons per person Unknown 19 67 - Authors state it is the largest and most representative multilingual multi-turn preference dataset to date, and the first preference dataset with prompt-level annotator overlap and choice explanations. Section 4.1 says OpenAssistant is another open-source multilingual human preference dataset they know of, mainly in English and Spanish; the conclusion again calls CA the first multilingual open-source preference dataset. A footnote says DICES has multiple annotators per conversation, but it is for safety evaluation rather than general preferences. 2,582 prompts have at least 10 annotators; 44% of comparisons have natural language explanations. The balanced subset on age, gender, ethnicity has N=2,215.
Other Ablations and Analyses
- Evaluation Table B.1: traditional/secular-rational is 0.8046–0.8391, survival/self-expression is 0.8830–0.9149, listed by language.
- DPO: Llama-3.3-70B-Instruct, self-expression, β=0.05 caused generation to collapse into counting or repetition, changed to β=0.2 (Appendix C.5).
- GRPO: On survival and traditional directions, with two temperature-sampled datasets, lower β made the model tend to refuse; on raw PRISM short responses, lower β made it generate "no" for almost all statements (Appendix C.6). Various β are in Table C.1.
- prompt-steering NC win rates: self-expression is 8B 0.770±0.012, 70B 0.862±0.009; survival is 8B 0.428±0.013, 70B 0.530±0.014 (Table D.1).
- Attention tests during training accounted for 9.4% of tasks; before formal data collection, attention test pass rates and the proportion of medium-to-high quality explanations both had to be at least 85%. Removed 350+ annotators who failed multiple checks (Appendix E.2).
- The overall Indian sample is 69% male; the balanced subset of 775 people falls within 5% margins on gender, age, and ethnicity (Appendix E.3). In France and Italy, those aged 55 and above use a 10% rather than 5% margin in the balanced subset.
What are some points that can be further explored?
The authors note the five-country scope, that IW cannot represent all preferences, and that automatic review is imperfect.
Limitations and future directions noted by the authors
- Countries and languages: The human survey and Community Alignment only involve five countries and languages. The balanced subset is matched by age, sex, and ethnicity (US, India, Brazil within 5% margin, France, Italy within 10% margin), and does not account for education and political orientation (Section F).
- Dimensions: IW cannot cover all of human preference heterogeneity. The authors state that the paper should not be read as claiming that these two axes characterize all preferences; it mainly shows that existing collection methods cannot even recover these broad dimensions (Section F).
- Scope of positive results: The authors consider the positive result that NC improves IW learning to be relatively weaker; ideally, NC should be shown to improve many types of preferences, not just IW. The NC candidate sets in the appendix are treated by the authors as partial supplements, because they often contain values that do not map neatly onto IW (Section F).
- Automated review: Sections 2.2 and 3 rely on automated review to achieve the required scale. The reviewer has high accuracy on human-labeled data, but is still imperfect; the authors advise against overinterpreting small differences, such as those between model instances (Section F). The authors also write: even if the reviewer does not perfectly capture IW, the main conclusion about the limitations of temperature-sampled datasets still holds.
- Future work: Applying the controlled pipeline from Section 2 to other preference dimensions and prompt distributions; semantic diversity methods beyond prompts; sample efficiency when collecting preferences from users online (Section F).
- Future work on datasets: Testing social choice aggregation and distributional alignment on real human preferences; the role of multilingualism in pluralistic alignment; the relationship between answer preferences and question-asking behavior; the value of intersecting use cases with preference heterogeneity (Section F).
Experimental coverage
- The models in the open-generation evaluation are the 21 in the legend, and the languages are English, French, Italian, Portuguese, and Hindi; temperatures include 0, 0.1, 1 in the Figure B.3 caption, and 0 and 0.5 as written in Section 2.2 of the main text (Section 2, Figures B.2, B.3).
- Preference learning uses PRISM round 1, with 5,463 training and 608 test examples; the tuned models are Llama-3.1-8B-Instruct and Llama-3.3-70B-Instruct; the methods are prompt-steering, SFT, DPO, GRPO; candidate generation is single-model temperature 1, 21-model temperature 1, and single-model NC (Section 3, Table D.1).
- The reviewer is GPT-4o. Human study accuracy by language and IW dimension is listed in Table B.1; PRISM manual labeling is 120 pairs, with accuracies of 85.8% and 78.3% (Section C.2). The same reviewer both labels preferences and computes win rate (Section 3.1).
- Community Alignment has 3,603 annotators, 233,319 comparisons, 66% non-English; the five-country age, sex, and ethnicity balanced subset has N=2,215; 2,582 prompts have at least 10 annotators (Table 1, Sections 4, E.3).
- The paper does not report the number of independent repetitions for the preference learning experiments, nor the number of queries for temperature sampling and NC sampling; the error bars in Figure 3 are standard error of the mean. Model evaluation is repeated 20 times per question at temperature 1, and once per question at temperature 0 (Section B.3.2).
Summarize the main content of the paper
The authors collect multilingual preference data with negatively correlated sampling, targeting heterogeneous preferences that temperature sampling cannot learn.
- Positioning: This work examines why standard alignment methods fail to learn conflicting human preferences when candidate responses in preference data are too homogeneous, and provides a collection method and a dataset.
- Problem: The authors argue that personalization, social choice, and distributional alignment all require distinguishable comparisons first. They use the two IW dimensions as a minimal test: these directions come from major variation in the World Values Survey and relate to common political divides. If even these cannot be learned, independent temperature sampling is insufficient.
- Approach: 3,000 people per country across five countries choose from hand-curated extremes, balanced, and a Llama default response on 60 everyday prompts. 21 LLMs generate open-ended responses to the same prompts, scored by GPT-4o relative to the balanced response. Negative correlation sampling lets a single model write multiple more value-diverse responses at once, without naming IW in the prompt. Learning experiments change the first turn of PRISM to three candidate sources, then use prompt-steering, SFT, DPO, GRPO to fit the four directions labeled by the reviewer.
- Results: The authors state that English model scores at temperature 1 align with only 41% of human preference scores; French, Italian, and Portuguese lean toward the same quadrant but less strongly, and under Hindi some models shift to the other end. In Figure 2, temperature sampling average coverage of traditional and survival is 15% and 30%, NC is 60% and 53%. In Table D.1, 8B SFT+DPO on NC, self-expression has a win rate of 0.958; but 8B prompt-steering on survival with NC is 0.428, lower than the 21-LLM temperature sampling's 0.494. The authors state that about 70%–90% of fine-tuning generalization also does not cover the 0.546 for 70B, SFT+GRPO, secular-rational.
- Data: Community Alignment has 233,319 comparisons, 3,603 annotators, 66% non-English, 2,582 prompts with at least 10 annotators, 44% with comparative explanations. The authors claim it is the largest open-source multilingual multi-turn preference dataset to date; Section 4.1 also mentions OpenAssistant as a multilingual set primarily in English and Spanish.
- Authors' conclusion: Algorithmic monoculture makes temperature-sampled datasets unable to detect common disagreements; single-model negative correlation sampling is seen by the authors as a more direct and more distinguishable collection method. They also state that IW is just a case study, not advocating deployment of models to serve only one value direction, and automatic review is not suitable for fine-grained reading of small differences between models.