ChatGPT, Claude Converge on Identical Outputs in Open-Ended Tasks

Artificial Hivemind: The Open-Ended Homogeneity of Language Models (and Beyond) — Ai4

Ai4

Reward models and AI evaluators systematically penalize equally valid human variations during training, driving competing model architectures to collapse into homogenized narrative modes.

AI-assisted summary of Artificial Hivemind: The Open-Ended Homogeneity of Language Models (and Beyond) — original source. Check the original for full context.

Source published:

Detailed notes

## Background Published on YouTube by the Ai4 channel on August 20, 2026, this video features University of Washington researcher Liwei Jiang presenting award-winning research on the 'Artificial Hivemind' phenomenon in language models. The talk introduces Infinity-Chat, a large-scale benchmark of 26,000 open-ended queries and 31,250 dense human evaluations designed to analyze output diversity, mode collapse, and preference modeling.

## TL;DR While approximately 75% of real-world user queries to language models are open-ended, current models suffer from severe intra-model repetition and inter-model convergence, frequently collapsing to identical metaphors, names, and narrative structures across competing model families. Using the dense 26,000-query Infinity-Chat benchmark with 25 human annotators per example, the study reveals that reward models and LLM judges are poorly calibrated to legitimate human preference diversity, penalizing equally valid creative alternatives. This homogenization threatens human originality, weakens synthetic training loops, and impedes autonomous agentic exploration. Addressing the Artificial Hivemind requires a full-stack overhaul spanning diversity-aware curation, entropy-maximizing decoding, pluralistic training losses, and multi-valid evaluation metrics.

## Atmosphere & style The presentation maintains an analytical, measured, and intellectually rigorous academic tone that transitions into an urgent warning about the cultural and technical hazards of AI homogenization. Jiang speaks with quiet precision, systematically illustrating systemic algorithmic conformity through striking visual proofs, such as PCA projections showing dozens of independent models collapsing into just two time metaphors. The delivery feels insightful and diagnostic rather than alarmist, culminating in a philosophically grounded call for researchers to build AI that champions a diverse 'chorus of many' rather than amplifying the echoes of a few.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=gVTYfI4pOSg&t=4s) Analysis of real-world ChatGPT interactions reveals that roughly 75% of user queries are open-ended with no single ground-truth answer. - [01:41](https://www.youtube.com/watch?v=gVTYfI4pOSg&t=101s) The Infinity-Chat benchmark categorizes 26,000 open-ended queries across 6 top-level categories and 17 subcategories, with creative generation accounting for nearly 58% of interactions. - [03:50](https://www.youtube.com/watch?v=gVTYfI4pOSg&t=230s) Language models suffer from an 'Artificial Hivemind' effect, characterized by pronounced intra-model repetition and cross-model convergence across 70+ model families. - [03:50](https://www.youtube.com/watch?v=gVTYfI4pOSg&t=230s) Even under an aggressive sampling temperature of 1.0, 79% of evaluated open-ended queries exhibit semantic similarity scores above 0.8 across repeated single-model generations. - [06:41](https://www.youtube.com/watch?v=gVTYfI4pOSg&t=401s) Competing model architectures produce near-identical narrative structures; when asked for a metaphor of time, 25 distinct models collapsed almost entirely into just two concepts ('river' and 'weaver'). - [06:41](https://www.youtube.com/watch?v=gVTYfI4pOSg&t=401s) Paraphrasing input prompts fails to resolve mode collapse, as models retain identical sentence scaffolds, anchor text, and latent plot progressions. - [10:11](https://www.youtube.com/watch?v=gVTYfI4pOSg&t=611s) Widespread exposure to homogenized AI outputs risks diminishing human originality and linguistic diversity while bottlenecking synthetic data generation and ensemble methods. - [12:20](https://www.youtube.com/watch?v=gVTYfI4pOSg&t=740s) Infinity-Chat introduced a dense evaluation benchmark with 31,250 human annotations (25 raters per example) to capture the natural variance and entropy of subjective human preferences. - [15:26](https://www.youtube.com/watch?v=gVTYfI4pOSg&t=926s) Reward models and LLM judges suffer from severe miscalibration, heavily penalizing valid alternative responses that human annotators rate as equally high quality. - [15:26](https://www.youtube.com/watch?v=gVTYfI4pOSg&t=926s) Miscalibrated scoring in reward models risks permanently purging diverse, valid reasoning paths and creative options during reinforcement learning rollouts. - [17:48](https://www.youtube.com/watch?v=gVTYfI4pOSg&t=1068s) Overcoming the Artificial Hivemind requires a full-stack effort spanning diversity-aware data curation, entropy-maximizing decoding, pluralistic training losses, and multi-validity evaluation. - [17:48](https://www.youtube.com/watch?v=gVTYfI4pOSg&t=1068s) Preserving diversity and exploratory capacity is vital for autonomous agentic workflows to prevent agents from collapsing into brittle, repetitive decision trees.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=gVTYfI4pOSg&t=4s) Ubiquitous Language Models and the Prevalence of Open-Ended Tasks

AI systems have woven themselves into global workflows and everyday human activity at an unprecedented scale. By July 2025, ChatGPT had reached roughly 10% of the world's adult population, boasting 800 million weekly active users. Analyzing how people actually interact with these large language models reveals that the vast majority of real-world use cases are inherently open-ended—spanning practical guidance, exploratory information seeking, creative writing, and multimedia creation. These interactions lack a single ground-truth answer and inherently demand creativity or accommodate multiple valid solutions. Similar findings appear in Anthropic's reports on Claude, where top user interactions center around exploratory queries with no single correct response.

```mermaid flowchart TD accTitle: Infinity-Chat Pipeline Overview subgraph DataCurate["1. Dataset Construction"] A["WildChat Dataset<br/>(1M Real Conversations)"] --> B["Filter to 37K Meaningful Queries"] B --> C{"LLM Open-Endedness Filter"} C -- "25% Closed-Ended" --> D["Factual Lookups / Yes-No"] C -- "75% Open-Ended" --> E["Infinity-Chat Core<br/>(26K Queries)"] end

subgraph Taxonomy["2. Taxonomy & Phenomena"] E --> F["6 Categories & 17 Subcategories<br/>(58% Creative Generation)"] F --> G["Artificial Hivemind Analysis"] G --> H["Intra-Model Repetition<br/>(Single model, T=1.0)"] G --> I["Inter-Model Convergence<br/>(Across 70+ models)"] end

subgraph HumanEval["3. Dense Evaluation"] E --> J["50 Queries x 15 Model Responses"] J --> K["31,250 Dense Human Annotations<br/>(25 raters/sample: Absolute & Pairwise)"] K --> L["Finding: Reward Models & LLM Judges<br/>Miscalibrated on Divergent Preferences"] end ```

To investigate how models manage open-ended user demands, researchers turned to WildChat, a dataset containing one million naturally occurring ChatGPT conversation logs. After filtering out low-quality inputs to isolate 37,000 meaningful real-world questions, an LLM judge classified each prompt's open-endedness. The analysis confirmed that roughly 75% of user inquiries are open-ended (accepting multiple plausible answers), whereas only 25% are closed-ended queries such as factual lookups or boolean yes/no questions.

- ChatGPT reached 800 million weekly active users and ~10% of the global adult population by July 2025. - Analysis of 37K filtered WildChat interactions revealed that 75% of real-world queries are open-ended. - The paper representing this work was presented at NeurIPS 2025, receiving the Best Paper Award in the Datasets and Benchmarks track.

---

[01:41](https://www.youtube.com/watch?v=gVTYfI4pOSg&t=101s) Mapping the Open-Ended Landscape: The Infinity-Chat Taxonomy

To systematically organize the wide range of open-ended queries submitted by users, the researchers manually analyzed 100 sample prompts to establish an initial taxonomy. They then scaled this categorization across the full dataset using an LLM judge that could assign known tags while identifying emergent categories. This process yielded Infinity-Chat, a curated core benchmark of 26,000 open-ended queries designed to evaluate model response distributions across open-ended problem spaces.

The resulting taxonomy comprises 6 top-level categories divided into 17 distinct subcategories. The single largest category is creative content generation and writing, which constitutes nearly 58% of all analyzed interactions. Other primary categories include brainstorming and ideation (where multiple diverse responses are explicitly expected), general open-ended inquiries, decision-support information seeking, alternative communication or stylistic rewriting, and prompts centered around subjective values and viewpoints.

| Top-Level Category | Core Characteristics & Representative Use Cases | |:--- |:--- | | **Creative Generation & Writing** (~58%) | Short stories, poetry, scriptwriting, worldbuilding, and multimedia narrative drafts. | | **Brainstorming & Ideation** | Formulating novel startup ideas, naming products, suggesting event themes, and feature planning. | | **General Open-Ended Questions** | Inquiries requiring broad synthesis or conceptual exploration without standard answers. | | **Decision-Support Information Seeking** | Weighing trade-offs, comparing lifestyle choices, career pathways, and exploratory recommendations. | | **Stylistic & Alternative Communication** | Tone adjustments, reframing arguments, persona-driven rewrites, and rhetorical variations. | | **Subjective & Value-Laden Perspectives** | Philosophical prompts, moral dilemmas, personal opinions, and aesthetic critiques. |

---

[03:50](https://www.youtube.com/watch?v=gVTYfI4pOSg&t=230s) The Artificial Hivemind Effect and Intra-Model Repetition

When assessing how well language models generate diverse answers to open-ended prompts, the short answer is that they perform poorly. Language models suffer from a phenomenon dubbed the "Artificial Hivemind," wherein outputs systematically collapse toward narrow, highly similar modes. While related to traditional mode collapse—a well-documented issue in instruction-tuned and post-trained models that struggle to produce truly random numbers, country names, or diverse anchor spans—this study extended the investigation across more than 70 distinct model families and sizes on authentic open-ended prompts.

``` ┌───────────────────────────────────────────────┐ │ ARTIFICIAL HIVEMIND │ │ Mode Collapse Across Open-Ended Queries │ └───────┬───────────────────────────────┬───────┘ │ │ ▼ ▼ ┌─────────────────────────────────────┐ ┌─────────────────────────────────────┐ │ Intra-Model Homogeneity │ │ Inter-Model Homogeneity │ │ A single model sampled repeatedly │ │ Different model architectures & │ │ converges to the same outputs │ │ sizes produce identical narratives │ ├─────────────────────────────────────┤ ├─────────────────────────────────────┤ │ • 50 samples at Temperature = 1.0 │ │ • Verbatim lexical repetition │ │ • 79% of queries have sim > 0.8 │ │ • Shared sentence scaffolds │ │ • Repetitive character names & arcs │ │ • Bimodal collapse (River vs Weaver)│ └─────────────────────────────────────┘ └─────────────────────────────────────┘ ```

The Artificial Hivemind manifests through two distinct behaviors: intra-model homogeneity (a single model generating repetitive responses across independent runs) and inter-model homogeneity (competing models from different creators producing near-identical responses). To quantify intra-model repetition, the authors selected 100 human-verified open-ended queries from Infinity-Chat, generating 50 independent outputs per model under an aggressive temperature of 1.0 to maximize stochastic variation without sacrificing linguistic fluency. Computing pairwise semantic similarities across these 50 outputs revealed that in approximately 79% of evaluated cases, the average output similarity exceeded 0.8, demonstrating severe semantic convergence.

- Evaluated across 70+ language models from diverse model families and parameter sizes. - Tested on 100 human-verified open-ended prompts sampled 50 times each at temperature 1.0. - 79% of queries exhibited average pairwise semantic similarity scores above 0.8 despite high sampling temperatures.

---

[06:41](https://www.youtube.com/watch?v=gVTYfI4pOSg&t=401s) Inter-Model Convergence, Paraphrasing, and the Time Metaphor

The lack of diversity is not merely an abstract metric; it appears vividly in generated text. When prompted with "write a short story about a colorful toad goes on adventure in 50 words," top-tier models such as Claude 3.5 Sonnet repeatedly output the exact same toad name, identical narrative arcs, and matching sentence structures across runs. Because subtle elements like story arcs remain latent, standard surface-level similarity metrics often underestimate the true extent of this repetition. Attempting to disrupt this behavior by paraphrasing input prompts yields only modest decreases in similarity scores, as models continue to produce repetitive text spans and shared narrative scaffolding.

``` [ PCA Projection of 25 Distinct LLM Responses to "Write a metaphor involving time" ]

Dim 2 ▲ │ ● Dominant Cluster: │ ●●● "Time as a River" │ ●●●● (Flowing, currents, downstream) │ ● │ ◆ Secondary Cluster: │ ◆◆◆ "Time as a Weaver" │ ◆ (Tapestry, threads, loom) │ │ ○ (Outliers) └────────────────────────────────────────────────────────► Dim 1 ```

Inter-model homogeneity reveals an even deeper convergence across competing model families. When given open-ended queries, distinct model architectures—such as DeepSeek and ChatGPT—frequently employ identical sentence scaffolds and share uncommon verbatim word choices, such as describing clothing as "slim-fitted." A striking example occurs with the prompt: "Write a metaphor involving time." When responses from 25 different models spanning varied sizes and creators were converted into sentence embeddings and projected onto a two-dimensional PCA space, the outputs collapsed almost exclusively into two visual clusters: a dominant cluster framing time as a flowing river, a secondary cluster framing time as a loom or weaver of tapestries, and only a tiny handful of isolated outliers.

- Models consistently reuse identical character names, phrasing, and plot progressions across independent generation rounds. - Input prompt paraphrasing fails to resolve mode collapse, retaining high semantic similarity and shared anchor text spans. - Responses from 25 separate models to the "time metaphor" prompt clustered almost entirely into two archetypes: "time as a river" and "time as a weaver."

---

[10:11](https://www.youtube.com/watch?v=gVTYfI4pOSg&t=611s) The Broader Societal and Technical Risks of Homogenized AI

The Artificial Hivemind carries serious implications for human society, AI development pipelines, and scientific discovery. When millions of global users are repeatedly exposed to homogenized model outputs, it risks flattening human creative expression and independent thought. Emerging studies already indicate measurable declines in human linguistic diversity, ideational variance, and distinct individualistic writing styles since the widespread adoption of LLMs. This trend underscores an urgent need to "safeguard human originality and agency."

Beyond societal effects, output homogeneity poses technical bottlenecks for next-generation AI architectures: - **Synthetic Data and Self-Play:** Modern self-improving paradigms, recursive training loops, and synthetic data pipelines rely heavily on discovering difficult edge cases and diverse reasoning paths. If models only sample from narrow output modes, synthetic training fails to expand the reasoning frontier. - **Scientific and Societal Discovery:** Applying AI to scientific research requires proposing non-obvious hypotheses and unconventional synthesis. Mode collapse constrains models from suggesting genuinely novel conjectures. - **Ensemble Failures:** Ensembling diverse models is standard practice to counteract individual model biases. However, if independent model families suffer from inter-model homogeneity, ensembled models simply fail collectively instead of providing complementary oversight.

---

[12:20](https://www.youtube.com/watch?v=gVTYfI4pOSg&t=740s) Preference Modeling and Dense Human Annotations

Investigating how well models evaluate alternative valid responses required constructing a high-density human preference dataset. In open-ended domains, individual humans often exhibit idiosyncratic, divergent tastes for different valid answers, yet the collective average quality across valid alternatives tends to remain comparable. If reward models or LLM judges cannot calibrate to this nuance and arbitrarily assign vastly different scores to equally valid responses, they will systematically penalize and eliminate valid alternative answers during reinforcement learning (RL) rollouts, offline data curation, and model evaluation.

``` Standard Preference Benchmarks: [Query] ──► [Response A vs. B] ──► 1 to 3 Annotator Labels (Sparse)

Infinity-Chat Dense Benchmark: [Query] ──► 15 Candidate Responses (from 70 Models) │ ├──► Absolute Setup: 25 Independent Annotators per response (1-5 Scale) └──► Pairwise Setup: 25 Independent Annotators per pair (10 pairs) │ └──► Total: 31,250 Dense Annotations (Captures Full Entropy & Disagreement) ```

Existing preference datasets are sparsely labeled, typically containing only one to three human annotations per example, which obscures the natural variance in human preferences. To address this, the researchers sampled 50 queries from Infinity-Chat and generated 15 distinct model responses per query across the pool of 70 LLMs. They then collected 31,250 dense human annotations across two experimental setups: 1. **Absolute Ratings:** Each of the 15 candidate responses was evaluated on a 1-to-5 quality scale by 25 independent human annotators. 2. **Pairwise Preferences:** 10 distinct response pairs per query were scored by 25 independent annotators indicating their preference.

This dense annotation structure revealed high entropy in human evaluations. On many open-ended responses, annotators split almost evenly between positive and negative impressions, whereas other responses achieved widespread consensus—distributional nuances that sparse 1-to-3 annotator datasets miss entirely.

- Collected 31,250 total human annotations across 50 open-ended queries and 15 model responses per query. - Recruited 25 independent human raters per example for both absolute 1–5 scoring and pairwise comparisons. - Uncovered high-entropy rating distributions that capture legitimate subjective disagreement among annotators.

---

[15:26](https://www.youtube.com/watch?v=gVTYfI4pOSg&t=926s) Evaluation Miscalibration and the Demotion of Valid Alternatives

Evaluating LLMs, reward models, and LLM judges against this dense dataset revealed a severe calibration breakdown when scoring responses of comparable human quality. When human annotators disagreed individually but arrived at similar aggregate quality scores for two alternative answers, the correlation between model-assigned scores and average human ratings dropped substantially. While humans view both candidates as equally plausible and enjoyable options, reward models and LLM evaluators often rate one candidate drastically higher than the other.

``` EVALUATION MISCALIBRATION

Human Perspective: Reward Model / LLM Judge: ┌───────────────────────┐ ┌───────────────────────┐ │ Response A: ★★★★☆ │ │ Response A: Score 0.92│ ──► Selected │ Response B: ★★★★☆ │ │ Response B: Score 0.21│ ──► Discarded / Demoted │ (Aggregate Parity, │ │ (Artificial Disparity │ │ Diverse Idiosyncrasy)│ │ Penalizes Plurality) │ └───────────────────────┘ └───────────────────────┘ ```

This miscalibration directly damages model development pipelines. When reward models penalize valid alternatives during reinforcement learning, those solution modes are effectively purged from the policy's distribution. Similarly, biased LLM judges evaluating benchmarks may incorrectly classify legitimate creative variations as erroneous. This scoring instability was confirmed across multiple selection strategies designed to identify equal-quality response subsets, demonstrating that current evaluation systems consistently struggle with open-ended plurality.

- Model rating correlations with aggregate human preferences fall sharply when evaluating candidate responses of similar overall quality. - Reward models frequently produce large score disparities between responses that humans consider equally valid. - Miscalibrated scoring risks purging valid reasoning paths and creative options during RL training and offline data filtering.

---

[17:48](https://www.youtube.com/watch?v=gVTYfI4pOSg&t=1068s) Future Directions: Diversity-Aware Stacks and Agentic Exploration

Escaping the Artificial Hivemind requires rethinking the entire machine learning development stack. Emerging research is exploring several key mitigation strategies: - **Diversity-Aware Data Curation:** Actively selecting and filtering training corpora to preserve linguistic breadth and uncommon narrative structures. - **Inference-Time Decoding Strategies:** Developing specialized sampling and decoding algorithms that maximize generative entropy and diversity at runtime without requiring full model retraining. - **Explicit Diversity Objectives:** Training paradigms that penalize mode collapse and incentivize multiple distinct reasoning trajectories during post-training. - **Pluralistic Evaluation Benchmarks:** Constructing evaluation frameworks that reward valid alternative solutions rather than enforcing a single canonical answer.

```mermaid flowchart LR accTitle: Full-Stack AI Diversity Paradigm subgraph DevelopmentStack["Holistic Diversity-Aware Development Stack"] direction TB A["1. Data Layer:<br/>Diversity-Aware Curation"] B["2. Training Layer:<br/>Pluralistic Loss & Diverse Trajectories"] C["3. Inference Layer:<br/>Entropy-Maximizing Decoding"] D["4. Agent & Evaluation Layer:<br/>Exploration-Driven RL & Multi-Valid Benchmarks"] A --> B --> C --> D end ```

Addressing this challenge becomes even more critical as AI transitions into autonomous agentic workflows. Agentic systems fundamentally depend on exploratory capabilities to traverse complex decision trees, test alternative action pathways, and discover novel solutions. If reward models penalize non-standard exploration, autonomous agents will converge to brittle, repetitive behaviors.

Liwei Jiang concluded by emphasizing that the ultimate goal of the AI research community should not be to produce "louder echoes of the very few voices," but rather to cultivate "a chorus of many." Realizing this vision will require fundamental innovations spanning data curation, model training, inference interfaces, interaction modalities, and benchmark design to safeguard human plurality, cultural inclusivity, and genuine discovery.

## Notable quotes > [03:50](https://www.youtube.com/watch?v=gVTYfI4pOSg&t=230s) Artificial Hivemind — Liwei Jiang

> [06:41](https://www.youtube.com/watch?v=gVTYfI4pOSg&t=401s) write a short story about a colorful toad goes on adventure in 50 words — Liwei Jiang

> [06:41](https://www.youtube.com/watch?v=gVTYfI4pOSg&t=401s) slim-fitted — Liwei Jiang

> [06:41](https://www.youtube.com/watch?v=gVTYfI4pOSg&t=401s) Write a metaphor involving time — Liwei Jiang

> [06:41](https://www.youtube.com/watch?v=gVTYfI4pOSg&t=401s) Time as a River — Liwei Jiang

> [06:41](https://www.youtube.com/watch?v=gVTYfI4pOSg&t=401s) Time as a Weaver — Liwei Jiang

> [10:11](https://www.youtube.com/watch?v=gVTYfI4pOSg&t=611s) safeguard human originality and agency — Liwei Jiang

> [17:48](https://www.youtube.com/watch?v=gVTYfI4pOSg&t=1068s) louder echoes of the very few voices — Liwei Jiang

> [17:48](https://www.youtube.com/watch?v=gVTYfI4pOSg&t=1068s) a chorus of many — Liwei Jiang

## Mentioned Liwei Jiang, University of Washington, Ai4, ChatGPT, Anthropic, Claude, Claude 3.5 Sonnet, DeepSeek, WildChat, Infinity-Chat, NeurIPS 2025

## Video description Liwei Jiang - Ph.D. candidate, Researcher, University of Washington

Language models (LMs) often struggle to generate diverse, human-like creative content, raising concerns about the long-term homogenization of human thought through repeated exposure to similar outputs. Yet scalable methods for evaluating LM output diversity remain limited, especially beyond narrow tasks such as random number or name generation, or beyond repeated sampling from a single model. We introduce Infinity-Chat, a large-scale dataset of 26K diverse, real-world, open-ended user queries that admit a wide range of plausible answers with no single ground truth. We introduce the first comprehensive taxonomy for characterizing the full spectrum of open-ended prompts posed to LMs, comprising 6 top-level categories (e.g., brainstorm & ideation) that further breaks down to 17 subcategories. Using Infinity-Chat, we present a large-scale study of mode collapse in LMs, revealing a pronounced Artificial Hivemind effect in open-ended generation of LMs, characterized by (1) intra-model repetition, where a single model consistently generates similar responses, and more so (2) inter-model homogeneity, where different models produce strikingly similar outputs. Infinity-Chat also includes 31,250 human annotations, across absolute ratings and pairwise preferences, with 25 independent human annotations per example. This enables studying collective and individual-specific human preferences in response to open-ended queries. Our findings show that LMs, reward models, and LM judges are less well calibrated to human ratings on model generations that elicit differing idiosyncratic annotator preferences, despite maintaining comparable overall quality. Overall, INFINITY-CHAT presents the first large-scale resource for systematically studying real-world open-ended queries to LMs, revealing critical insights to guide future research for mitigating long-term AI safety risks posed by the Artificial Hivemind.

[Source](https://www.youtube.com/watch?v=gVTYfI4pOSg)

中文简报

ChatGPT与Claude在开放式任务中的输出趋于一致 奖励模型和AI评估器在训练中系统性惩罚同样合理的人类表达多样性，致使不同竞争模型架构塌缩为同质化的叙事模式。 尽管在语言模型的真实用户查询中约有75%属于开放式问题，但现有模型存在严重的模型内重复与跨模型趋同现象，不同竞争模型家族频繁塌缩为完全相同的隐喻、名称和叙事结构。这项基于包含26,000个查询且每个样本配备25名人类标注员的密集型Infinity-Chat基准测试的研究表明，奖励模型和LLM裁判未能很好地适配合理的人类偏好多样性，惩罚了同样有效的创造性备选方案。这种同质化不仅威胁到人类的独创性，削弱了合成训练循环，还阻碍了自主智能体的探索。解决这种“人工蜂群思维”（Artificial Hivemind）需要从感知多样性的数据整理、最大化熵的解码方式、多元化训练损失到多解有效性评估指标进行全栈重构。 对真实ChatGPT交互的分析表明，约75%的用户查询属于没有唯一标准答案的开放式问题。 (0:04) Infinity-Chat基准测试将26,000个开放式查询划分为6个大类和17个子类，其中创意生成类交互占比近58%。 (1:41) 语言模型普遍存在“人工蜂群思维”（Artificial Hivemind）效应，在70多个模型家族中表现出显著的模型内重复与跨模型趋同。 (3:50) 即使在1.0的激进采样温度下，单模型重复生成中仍有79%的受评估开放式查询呈现出超过0.8的语义相似度得分。 (3:50) 相互竞争的模型架构产生了近乎相同的叙事结构；当被要求提供关于时间的隐喻时，25个不同的模型几乎完全塌缩为仅有的两个概念（“河流”和“织工”）。 (6:41)

Reading summary

Although roughly three-quarters of real-world user interactions with language models are open-ended, models systematically collapse into repetitive outputs across runs and distinct architectures. Research presented at NeurIPS 2025 shows that this "Artificial Hivemind" stems from evaluation tools and reward models that penalize valid alternative answers. The Open-Ended Reality of Everyday LLM Interactions (0:04) By July 2025, ChatGPT reached an estimated 800 million weekly active users, representing roughly 10% of the global adult population. Analyzing user prompts from 37,000 filtered WildChat conversations revealed that 75% of queries are open-ended, lacking a single ground-truth answer and accommodating multiple valid paths. Anthropic noted comparable user tendencies with Claude, where exploratory, non-factual queries dominate workflow demands. To map this behavior, researchers curated Infinity-Chat, an evaluation benchmark of 26,000 real-world open-ended prompts spanning 6 top-level categories and 17 subcategories. Creative content generation and narrative writing represent the single largest share at nearly 58% of interactions, followed by brainstorming, general synthesis, lifestyle decision support, stylistic rewriting, and subjective value-based discussions. Intra- and Inter-Model Convergence Toward Narrow Modes (3:50) When evaluating how models handle open-ended prompts across more than 70 model families and sizes, researchers observed systemic mode collapse termed the "Artificial Hivemind." This manifests as intra-model homogeneity—where a single model repeatedly generates the same response across separate runs—and inter-model homogeneity, where competing systems from different developers produce near-identical text scaffolds and vocabulary. Testing 100 human-verified queries sampled 50 times each at a high temperature of 1.0 revealed that 79% of queries exhibited average pairwise semantic similarities above 0.8. Qualitative checks highlighted severe lexical and structural repetition. When asked to write a 50-word story about a colorful toad, top systems like Claude 3.5 Sonnet repeatedly produced the same character names and narrative arcs. Across 25 distinct models prompted to write a metaphor about time, embeddings collapsed almost entirely into two archetypal clusters: time as a flowing river and time as a weaver of tapestries. Paraphrasing the input prompts yielded only slight reductions in similarity, leaving underlying text scaffolds intact. Technical Bottlenecks and Risks to Human Originality (10:11) Widespread exposure to uniform AI outputs risks diminishing human expressive diversity. The authors note emerging evidence of measurable declines in human linguistic variance and individualistic writing styles since LLMs became ubiquitous, creating an urgent need to safeguard human originality. Beyond societal impacts, this homogeneity introduces technical bottlenecks. Recursive training and self-play paradigms depend on sampling diverse reasoning paths to discover novel edge cases; constrained output distributions stunt synthetic data utility. Inter-model uniformity also undermines model ensembling, as systems from independent developers share the same blind spots and fail collectively rather than offering complementary viewpoints. Dense Preference Benchmarks Reveal Evaluator Miscalibration (12:20) To identify why models discard alternative answers, researchers gathered 31,250 dense human annotations across 50 Infinity-Chat queries and 15 model outputs per query. Unlike standard preference benchmarks that rely on only 1 to 3 annotations per pair, this study deployed 25 independent raters for both absolute 1–5 scoring and pairwise comparisons. The resulting distributions captured substantial entropy and legitimate disagreement among human readers assessing creative work. Testing reward models and automated LLM evaluators against this dense dataset revealed severe calibration failures. When human raters found two distinct responses equally valid and gave them similar aggregate quality scores, reward models and LLM judges frequently assigned vastly disparate scores to the two alternatives. During reinforcement learning and data curation, these miscalibrated evaluators systematically penalize and discard valid, high-quality alternatives. Countering the Artificial Hivemind requires interventions across the development stack, including diversity-aware training curation, inference decoding strategies that maximize entropy, and pluralistic benchmarks that reward multiple valid trajectories. As AI transitions into autonomous agents that rely on broad exploration to solve complex tasks, speaker Liwei Jiang argued that the goal must be cultivating "a chorus of many" rather than amplifying the louder echoes of a few models.

精炼阅读摘要

研究人员在NeurIPS 2025获奖论文中指出，尽管人类在真实场景中使用大语言模型的提问绝大多数属于开放性任务，但现有模型普遍存在“人工蜂群思维”（Artificial Hivemind），即在单模型重复采样以及跨模型对比中均出现严重的模式坍缩与同质化收敛。 真实交互中开放性任务占主导地位 (0:04) 截至2025年7月，ChatGPT每周活跃用户已达8亿，覆盖全球约10%的成年人口。在实际工作流与日常应用中，用户提问往往没有单一标准答案。研究人员通过对WildChat数据集中筛选出的3.7万条真实对话进行LLM裁判分类，发现约75%的提问属于开放性任务，例如创意写作、头脑风暴或决策咨询，而仅有25%属于事实检索或布尔值判断这类封闭性查询。这一开放性特征在Anthropic关于Claude的使用报告中同样得到了印证。 为系统化评估模型在开放域的表现，研究团队构建了包含2.6万个开放性查询的基准数据集Infinity-Chat。该数据集将开放性任务划分为6大类和17个子类别，其中创意生成与写作占比最高，接近全部交互的58%，其他主要类别还包括方案策划、综合性开放问答、决策支持信息寻求、语体改写以及涉及主观价值与观点的讨论。 模型内重复与跨模型同质化收敛 (3:50) 研究团队在超过70个不同参数规模与架构的模型上展开测试，发现语言模型在应对开放性问题时存在显著的“人工蜂群思维”现象。在模型内部，研究人员选取100个经过人工验证的开放性提示，即使将采样温度设为1.0以尽可能提升随机性，在单模型独立运行50次的情况下，仍有约79%的查询其输出平均两两语义相似度超过0.8。例如面对“以50字写一只五彩斑斓的蟾蜍去冒险”的要求时，Claude 3.5 Sonnet等模型会反复生成完全相同的角色名字、叙事弧线与句式骨架，即便对输入提示进行释义改写也无法打破这种模式坍缩。 跨模型之间的同质化现象更为严峻。来自不同开发机构、不同架构的模型（如DeepSeek与ChatGPT）在开放性提示下频繁采用相同的句式结构和罕见的逐字表达（如同样使用“slim-fitted”描述衣物）。在以“写一个关于时间的隐喻”为提示测试25个不同模型时，将句嵌入降维投影到PCA空间的结果显示，模型输出几乎全部坍缩为两大类集群：绝大多数将时间比作“流动的河水”，少数比作“织布机或织锦工”，仅有极个别零星的特异表达。 同质化输出的技术与社会风险 (10:11) 大语言模型输出的高度同质化带来了广泛的技术瓶颈与社会层面的潜在影响。在社会文化层面，全球数以亿计的用户长期接触高度趋同的模型内容，可能削弱人类自身的创意表达与独立思考能力，已有初步研究观察到大模型普及后人类语言多样性、构想差异度与个性化写作风格出现下滑。 在技术路线上，模型同质化构成了多维度的研发阻碍。首先，在依赖合成数据与自我博弈（Self-Play）的递归训练回路中，若模型只能在狭窄的输出模式内采样，便无法发现边缘特例与多样化的推理路径，难以拓展模型的推理边界；其次，在科学探索领域，模式坍缩限制了模型提出非显而易见的科学假说与反常规综合推论的能力；此外，传统上依赖多模型集成（Ensemble）来对冲单一模型偏差的做法也因此失效，因为具备跨模型同质性的独立模型群更倾向于集体犯错，而无法提供互补性纠偏。 高密人类偏好揭示奖励模型失准 (12:20) 现有的偏好数据集通常每个样本仅标注1至3名人类的反馈，无法捕捉开放性问题下真实存在的审美偏好差异。为了准确评估奖励模型与LLM裁判，研究人员构建了包含31,250条高密度人工标注的数据集：针对50个开放性提示，每个提示生成15个候选回复，并在1至5分绝对质量打分和成对偏好对比中，均引入25名独立评估员进行交叉标注。标注结果呈现出高度的熵值，显示出人类在面对不同合理回答时存在明显的个体分歧。 然而，评估现有模型、奖励模型及LLM裁判时发现，这些自动化评价系统在面对整体质量相近的候选回答时存在严重的校准失效。人类标注员虽然个人喜好各异，但整体上认为多个有效回答质量相当；奖励模型却倾向于给出悬殊的分数差异，将其中一个合理答案评为极高分，而将另一个同样有效的创意方案评为低分。这种评分失准在强化学习训练（RL）和离线数据清洗阶段，会系统性地淘汰有效的推理路径与创意选项，从而加剧策略模型的模式坍缩。 研究者Liwei Jiang强调，克服人工蜂群思维需要对机器学习开发栈进行全方位重构，包括在数据层引入多样性感知清洗、在推理层开发最大化生成熵的解码策略、在训练层设立显式防坍缩目标，以及建立支持多元有效解的基准评测。这一挑战在涉及多路径探索的自主智能体（Agent）工作流中尤为关键；AI研究的目标不应是让少数声音的回音变得更加响亮，而应是保护人类的多元性与文化的包容性。

Detailed notes — zh

## Background 本内容由 YouTube 频道 Ai4 于 2026 年 8 月 20 日发布，主讲人为华盛顿大学博士研究员姜利维（Liwei Jiang）。视频围绕其获评 NeurIPS 最佳论文的研究展开，重点介绍了包含 2.6 万条真实开放式提问的基准数据集 Infinity-Chat，旨在系统性揭示与评测大语言模型在开放生成任务中所面临的“人工蜂群思维”（Artificial Hivemind）模式坍塌现象。 ## TL;DR 研究表明真实用户与 AI 的交互中有 75% 属于没有唯一标准解的开放式任务，但横跨 70 多个主流大模型的评测却揭示了极具危机的单模型内部重复与跨模型同质化现象。这种蜂群思维不仅导致合成数据瓶颈与模型集成失效，还在深层重构并抹平人类自身的独创思维。团队进一步通过包含 3.1 万条密集人类标注的评测矩阵发现，现有奖励模型与 LLM 裁判在面对人类高熵多元偏好时存在严重校准失真，容易将合理的多样性误判为劣质。姜利维呼吁学界从数据清洗、解码策略、训练范式到评估体系全面革新，在智能体时代捍卫价值多元性与独立思考。 ## Atmosphere & style 整个演讲呈现出严谨扎实的顶级学术质感，同时洋溢着直面技术本质危机的紧迫感与人文关怀。姜利维的叙述清晰沉稳、条理分明，通过生动的隐喻案例与极具冲击力的投影图表层层推进论证。演讲在剖析 AI 扼杀独创性的冷峻现实中逐渐走向高潮，最终以倡导‘AI 应是多元之声交响乐’的动情呼吁收尾，带给听众强烈的思想震撼与深远的技术反思。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=gVTYfI4pOSg&t=4s) 真实世界中用户向大模型提出的请求约有 75% 属于开放式任务（允许多种可行解法），天然需要创造力与多样性，而仅有 25% 属于封闭式事实查询。 - [02:21](https://www.youtube.com/watch?v=gVTYfI4pOSg&t=141s) 研究团队构建了首个包含 2.6 万条真实开放式提问的 Infinity-Chat 数据集，建立了由 6 大顶层类别和 17 个子分类组成的开放式 Prompt 分类体系。 - [03:50](https://www.youtube.com/watch?v=gVTYfI4pOSg&t=230s) 主流大语言模型在开放生成中普遍陷入“人工蜂群思维”（Artificial Hivemind），表现出严重的模式坍塌（Mode Collapse），输出高度雷同。 - [03:50](https://www.youtube.com/watch?v=gVTYfI4pOSg&t=230s) 单模型内部同质化严重：即使在 Temperature=1.0 的高随机采样下，79% 的开放式问题生成结果语义相似度仍突破 0.8，且改写提问措辞无法消除重复结构。 - [03:50](https://www.youtube.com/watch?v=gVTYfI4pOSg&t=230s) 跨异构模型间存在高度一致的表述与用词复用，例如在隐喻生成等开放任务中，不同机构训练的模型高度坍缩至极少数固定隐喻簇（如“时间即河流”）。 - [10:11](https://www.youtube.com/watch?v=gVTYfI4pOSg&t=611s) 人工蜂群思维给人类认知带来退化风险，实证表明大模型普及时期人类创作的语言多样性与构思创造力已出现统计学显著下滑。 - [10:11](https://www.youtube.com/watch?v=gVTYfI4pOSg&t=611s) 同质化现象将导致合成数据与递归自迭代陷入瓶颈、扼杀科学前沿假设探索，并使传统的多模型集成（Model Ensembling）机制彻底失效。 - [12:20](https://www.youtube.com/watch?v=gVTYfI4pOSg&t=740s) 传统稀疏标注（每题 1-3 人）无法捕捉开放式偏好，Infinity-Chat 通过每题 25 位独立标注员构建了包含 31,250 条标签的高密度偏好评估矩阵。 - [12:20](https://www.youtube.com/watch?v=gVTYfI4pOSg&t=740s) 大模型、奖励模型及 LLM 裁判在开放式偏好上严重失真，会将人类平均打分相近但个体分歧大的多样化答案误判为劣质，从而在后训练中抹杀多元性。 - [17:55](https://www.youtube.com/watch?v=gVTYfI4pOSg&t=1075s) 破局人工蜂群需在全栈推进：包括感知多样性的数据清洗、新型解码策略、激励探索的训练范式以及语义层面的多样性度量衡。 - [17:55](https://www.youtube.com/watch?v=gVTYfI4pOSg&t=1075s) 在智能体时代（Agentic Era），奖励模型包容多元解题路径的能力至关重要，AI 应成为汇聚多元之声的交响乐而非少数声音的放大回音。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=gVTYfI4pOSg&t=4s) 研究背景：大语言模型的普及与开放式任务的普遍性 在演讲的开场中，刚刚从华盛顿大学（University of Washington）获得博士学位的研究员姜利维（Liwei Jiang）向观众分享了她与来自华盛顿大学、斯坦福大学、卡耐基梅隆大学（CMU）以及艾伦人工智能研究所（AI2）的合作团队共同完成的重磅成果——《Artificial Hivemind: The Open-Ended Homogeneity of Language Models (and Beyond)》（人造蜂群思维：大语言模型的开放式同质化及未来展望）。这项工作获得了 NeurIPS 2024 / 2025 数据集与基准评测赛道（Datasets and Benchmarks Track）的“最佳论文奖”（Best Paper Award）。 姜利维指出，人工智能目前已经以史无前例的规模深度嵌入到全球工作流与日常生活中。到 2025 年 7 月，仅 ChatGPT 一款产品的普及率就覆盖了全球约 10% 的成年人口，每周活跃用户达到 8 亿人。这种全球级的影响力迫使研究者深入探究人们究竟在如何使用这些系统。深入分析后会发现，绝大多数真实使用场景都属于“开放式任务”（Open-ended tasks）——无论是寻求实用建议、检索宽泛信息、创意写作还是多媒体内容生成。用户向 AI 提出的请求绝大部分本身就是开放性的，天然需要创造力或者包含多种可行的有效解决方案。Anthropic 关于 Claude 的使用报告也印证了这一点：其最高频的使用场景大都涉及根本没有唯一正确标准答案的开放式问题。 - 截至 2025 年 7 月，ChatGPT 触达全球约 10% 的成年人，拥有 8 亿周活跃用户（WAU）。 - 绝大多数真实用户交互聚焦于开放式生成任务，如实用指导、创意写作、头脑风暴与信息探索。 - 对 WildChat 数据集（包含 100 万条真实自然发生的 ChatGPT 交互对话记录）进行清洗过滤，提取出 3.7 万条具有实际意义的真实世界提问。 - 借助 LLM 评估模型（LLM Judge）分类发现，约 75% 的用户真实提问是开放式问题（允许多种有效解法），仅 25% 属于封闭式查询（如事实检索或“是/否”问答）。 --- ### [02:21](https://www.youtube.com/watch?v=gVTYfI4pOSg&t=141s) 构建 Infinity-Chat：首个开放式 Prompt 分类体系 面对海量涌现的开放式查询，研究团队进一步追问：人们在现实中究竟向大语言模型提出了哪些类型的开放式问题？为了建立结构化认知，研究团队首先人工深入审查了约 100 个具有代表性的真实提问，迭代式构建出一套覆盖开放式维度的分类体系原型（Taxonomy of Open-endedness）；随后借助 LLM 评估器将这套标签系统扩展到全量数据集上，并赋予模型自动发掘新增类别的能力。经过严格筛选与分类，团队最终提炼出 2.6 万条高质量的真实开放式问题，构成了本项研究的核心基石——Infinity-Chat 数据集。 ```mermaid flowchart TD accTitle: Infinity-Chat 数据集构建流程图 subgraph S1["数据清洗与筛选"] A["WildChat 原始交互日志 (100万条)"] --> B["真实提问过滤 (3.7万条)"] B --> C["开放性判定 (75% 开放式 / 25% 封闭式)"] end subgraph S2["分类体系与扩展"] C --> D["专家人工审查构建分类原型 (100条)"] D --> E["LLM 规模化分类与动态类目发掘"] E --> F["Infinity-Chat 核心数据集 (2.6万条开放式问题)"] end subgraph S3["评测维度验证"] F --> G["6大顶级分类与17个子分类体系"] end ``` 该分类体系由 6 个顶级大类（Top-level categories）和 17 个细分子类别（Subcategories）组成，首次全面描摹了现实中开放式 Prompt 的完整光谱。分析表明，用户提问占比最高的是“创意内容生成或写作”（Creative content generation / writing），占据了总案例的近 58%；排在其次的是“头脑风暴与创意构思”（Brainstorming & ideation），在该场景下，生成多种不同的回答不仅是可行的，而且是用户的明确预期；其余大类则覆盖了通识开放式提问（General open-ended questions）、支持决策制定的信息检索（Information seeking for decision-making）、多元交流与写作风格重塑（Alternative communication/writing styles），以及涉及主观价值观与多视角的探讨（Subjective perspectives）。 --- ### [03:50](https://www.youtube.com/watch?v=gVTYfI4pOSg&t=230s) 人工蜂群思维（Artificial Hivemind）：单模型重复与跨模型同质化 有了 Infinity-Chat 数据集后，团队开始系统性检验大模型应对开放式提问的表现。结论令人震惊：大语言模型在开放式生成上极其匮乏多样性。它们普遍陷入了一种被团队命名为“人工蜂群思维”（Artificial Hivemind）的现象中——不同系统不可遏制地坍塌向高度雷同的输出空间。这一现象与机器学习中的“模式坍塌”（Mode collapse）密切相关，尤其在经历了后训练（Post-training，如指令微调与人类偏好对齐）的模型身上表现得尤为剧烈。过去的研究已发现指令微调后的 LLM 在要求生成随机数字或国家名称等简单任务时表现极差，往往死板地重复某些锚定文本片段；而本研究直接将这种退化在真实复杂场景中量化揭示出来。 团队对跨越不同架构家族、不同参数尺度的 70 多个主流大语言模型展开了大规模评测。“蜂群思维”（Hivemind）在传统上指多个个体仿佛共享同一个大脑一般产生高度同步的群体意识；而在大模型世界中，人工蜂群思维主要以两种形式暴露出严重的模式坍塌： 1. **单模型内部同质化（Intra-model Homogeneity）**：单一模型在面对同一开放式提示词时，反复生成内部高度雷同的回复。 2. **跨模型间同质化（Inter-model Homogeneity）**：由完全不同的团队在不同数据上训练出的异构模型，竟然产生惊人一致的输出结构。 ```mermaid flowchart LR accTitle: 人工蜂群思维的双重同质化表现 subgraph S1["单模型内部同质化 (Intra-model)"] M1["单一模型 (如 Claude 3.5 Sonnet)"] M1 -->|"采样50次 (Temp=1.0)"| R1["雷同的角色命名"] M1 -->|"高随机度采样"| R2["固定情节走向 (Story Arc)"] M1 -->|"语义相似度 > 0.8"| R3["固化的句式框架"] end subgraph S2["跨模型间同质化 (Inter-model)"] M2["多模型比较 (DeepSeek, ChatGPT 等)"] M2 -->|"隐喻任务"| C1["时空隐喻偏向：时间即河流 (主流)"] M2 -->|"词汇重合"| C2["非高频生僻词逐字复现 (如 slim-fitted)"] M2 -->|"隐喻任务"| C3["时空隐喻偏向：时间即织工 (次流)"] end ``` 在单模型评测中，团队从 Infinity-Chat 中抽样出 100 个经人工核验确属开放式的提问，对每个模型以极高采样温度（Temperature = 1.0）独立生成 50 个回复，计算成对语义相似度。即使在如此激进的高随机度采样策略下，仍有高达 79% 的测试案例平均语义相似度超过 0.8。姜利维现场展示了一个生动的实例：当要求模型“用 50 个词写一个关于一只五彩斑斓的蟾蜍去冒险的微型故事”时，无论是 Claude 3.5 Sonnet 还是其他模型，在数十次独立生成中几乎每次都给蟾蜍起了一模一样的名字，套用了完全相同的情节弧线（Story arc）和句式用词。更耐人寻味的是，哪怕通过同义改写（Paraphrase）提问，模型输出长文本的重复片段比例与相似度也仅微弱下降，依然维持在极高水平——同质化并不会因为用户改变提问措辞而自动消失。 而在跨模型同质化方面，情况甚至更加极端。当研究人员对比 DeepSeek、ChatGPT 以及其他主流开源/闭源模型在相同开放式任务下的输出时，模型之间竟然大面积共享了句式骨架，甚至在非日常高频词汇上出现了“逐字复用”（Verbatim reuse）——例如极其罕见的组合词“slim-fitted”在不同公司的模型生成中同时反复出现。在演讲中展示的最经典案例是对“写一个关于时间的隐喻”（Write a metaphor involving time）任务的分析：团队收集了 25 个主流模型的输出，通过句子嵌入（Sentence Embeddings）并经 PCA 降维投影到二维空间。令人哑然的是，来自不同机构、不同尺寸的模型并没有呈现百花齐放的景象，而是高度坍缩为两个极度集中的视觉簇：左侧占据统治地位的主簇是“时间即流水/河流”（Time as a river），右侧稍小的次簇是“时间即织工”（Time as a weaver），仅有零星几个样本逸出这两个主流套路之外。 - 评测覆盖了 70 多个主流大语言模型，跨越各大知名开源与商业模型家族。 - 在 Temperature = 1.0 的高随机采样下，79% 的开放式问题模型生成语义相似度突破 0.8。 - 隐式文本特征（如故事大纲、核心寓意、拟人命名）往往难以单纯被传统自动化指标捕捉，但人工审查发现其内在高度趋同。 - 提示词同义改写无法根治重复，模型依旧会稳定输出高度重叠的句式骨架。 --- ### [10:11](https://www.youtube.com/watch?v=gVTYfI4pOSg&t=611s) 同质化的深远危机：认知窄化、合成数据瓶颈与模型融合失效 姜利维在报告中极其严肃地剖析了“人工蜂群思维”给 AI 长期安全与人类社会带来的系统性危机。当全球数以亿计的普通用户每天都在依赖生成式 AI 辅助日常决策与内容创作时，跨模型、跨单品的深度同质化不仅局限于技术层面，更将反噬人类文明自身的独创性。 这种同质化风险主要体现在四个互为因果的严峻维度： 1. **人类认知与独创性的退化风险（Homogenizing Human Thought and Agency）**： 当数亿全球用户反复接触来自不同 AI 的高度同构内容时，人类的行文逻辑和思维范式正面临被重塑并抹平差异的危险。姜利维援引近期多项实证研究警示：自从大语言模型崛起以来，人类创作中所展现的语言多样性（Linguistic diversity）、发散构思创造力（Ideation creativity）以及高度个性化的行文风格（Individualistic writing styles）均呈现出统计意义上的显著下滑。学界亟需发出紧急呼吁，捍卫人类的原创精神与自主能动性。 2. **合成数据与自迭代范式的天花板（Synthetic Data & Recursive Improvement Ceilings）**： 当前 AI 研发高度依赖合成数据生成、自迭代改进（Self-improving paradigms）以及递归强化自训练。如果基座模型本身缺乏输出多样性，合成数据就会迅速陷入单调循环，限制模型跳出固有能力边界去探索边缘测试用例（Challenging test cases），最终阻断推理广度（Breadth of reasoning）的演进。 3. **科学与社会发现潜力的扼杀（Constraining Scientific and Societal Discovery）**： 在“AI for Science”等重大命题中，人类寄希望于 AI 能在复杂系统中提出反直觉的全新假设。然而，人工蜂群思维使 AI 的探索轨迹死板固化在已有的平庸常识与既定知识簇内，极难产生真正具有颠覆性的假说和前沿结论。 4. **模型融合与集成机制的彻底失效（Collapse of Model Ensembling）**： 传统机器学习中常采用“模型集成”（Model ensembling）作为打破单一模型偏差的标准解法，寄希望于通过引入异构模型提供互补视角与监督机制。然而，如果几乎所有底层模型在开放生成时本质上共享了同一套蜂群思维，集成方法将彻底失去效能，最终导致模型群体的“集体性崩溃”（Fail collectively）。 --- ### [12:20](https://www.youtube.com/watch?v=gVTYfI4pOSg&t=740s) 人类偏好评估与校准困境：高熵偏好与模型判官的失真 在确立了大语言模型在生成端存在严重的同质化现象后，姜利维提出了下一个关键科学问题：当面对一个开放式问题下多种完全不同却各自成立的有效回答时，模型自身、奖励模型（Reward Models）以及“大模型裁判”（LLM-as-a-Judge）在判别与评分时，是否具备恰当的校准能力（Calibration）？ 在真实的现实语境中，对于一个开放式问题的多个有效解答，个体人类往往持有极其鲜明且特立独行的偏好（Idiosyncratic preferences）；但在宏观群体统计上，若大范围征求公众意见，这些有效解答的平均得分往往相当接近。在理想状态下，评估模型也应当对这些同样具备高水准但风格各异的候选回答给出相近的合理整体评分。如果模型无法做到这一点，其后果将是灾难性的：在后训练离线数据筛选或在线强化学习（RL rollout）中，若奖励模型无法识别不同可能性所包含的同等价值，它就会直接在模型的学习梯度中“抹杀”（Wipe out）这种多样性；同样，在评估环节，带偏见的 LLM 裁判如果只认可极窄的一种生成解法，就会将其他合理的创新答案错误判定为劣质或错误回答。 遗憾的是，以往的所有偏好数据集（如标准 RLHF 数据集）根本无法用于研究该问题——它们通常采用稀疏标注（Sparse labeling），每个样本往往只找 1 到 3 个人类做标注，并且极少针对真正的开放式问题设置大量多元候选集。 ```mermaid flowchart TD accTitle: 密集偏好标注方案与校准评测对比 subgraph P1["以往评估数据集的缺陷"] O1["稀疏标注 (每题仅 1-3 名标注员)"] --> O2["无法还原人类偏好的分布熵与分歧"] O3["缺乏针对开放式问题的高质量平行候选集"] end subgraph P2["Infinity-Chat 密集标注矩阵"] N1["从 70 个模型采样 15 组多元回复"] N2["绝对打分：25 位独立标注员独立打分 (1-5分)"] N3["成对比较：25 位标注员对 10 组随机对进行偏好抉择"] N1 & N2 & N3 --> N4["产出 31,250 条高密度人类标注标签"] end subgraph P3["评测发现：校准脱节危机"] N4 --> F1["人类视角：两份回答平均分相似但个体偏好高熵"] N4 --> F2["模型判官：给出断崖式分差，误将多样性判为劣质"] end ``` 为了弥补这一空白，团队从 Infinity-Chat 中抽选出 50 个复杂开放问题，从 70 个模型的池子中为每个问题采集 15 个风格迥异的独立生成回复（共 750 个回复样本）。随后，团队招募了极其庞大的人类标注团队，构建了业界首个超高密度的开放式偏好矩阵： - **绝对评分设置（Absolute Rating）**：每个回复均由 **25 名完全独立的人类标注员**按照 1 至 5 分的质量标准打分。 - **成对偏好设置（Pairwise Preference）**：随机采样 10 对不同的回复组合，同样由 **25 名独立标注员**就“更偏好哪一个”进行独立投票。 - 最终构建出包含 **31,250 个密集人类标注标签**（Dense human labels）的高质量评测基准。 高密度的标注揭示出了极具价值的“高熵（High-entropy）”真实分布特征：在许多开放式案例中，25 位标注员的分歧极大，喜爱与厌恶某个回复的人数可能完全五五开；这种只有在密集标注下才会显现的群体分布，揭示了开放世界偏好的复杂本质。而评测模型时的结论却令人沮丧：不论是 LLM 本身、奖励模型还是 LLM 裁判，在面对人类打分整体相当但在个体间引发巨大分歧的候选文本时，模型的评分与人类平均评分的相关性急剧暴跌。 在人类眼中质量平分秋色的两篇不同角度回答，在模型裁判眼里却会被判定出天壤之别的差距。这直接导致在目前的自对齐与强化学习闭环中，模型算法会死板地学到“其中一种写法远优于另一种”，从而在优化迭代过程中主动扼杀多元表达空间。 --- ### [17:55](https://www.youtube.com/watch?v=gVTYfI4pOSg&t=1075s) 破局“人工蜂群”：从强化探索到智能体时代的多元生态 姜利维在总结中强调，要使未来的模型真正学会拥抱开放性，社区亟需构建三根核心支柱：显式针对开放性进行校准的对齐模型、具备密集标注以捕捉长尾偏好分布的大规模真实人类数据集，以及能对多重有效路径进行包容性判别的新型评估框架。 尽管摆脱“人工蜂群思维”存在艰巨的技术壁垒，但研究界已经开始自发涌现出一系列探索方向： - **多样性感知的数据清洗机制（Diversity-aware data curation）**：从源头上打破语料中同质模版的过度堆叠。 - **新型推理阶段解码策略（Novel decoding strategies）**：在不重新训练模型的前提下，于推理生成阶段最大化输出的生成多样性与发散空间。 - **激励探索的训练新范式（Novel training paradigms）**：在后训练阶段通过算法手段显式鼓励模型保留多元思考轨迹与异构解题路径。 - **严谨的创造力与多样性度量衡（Rigorous diversity metrics）**：超越传统的表面重合度，从语义内涵和概念拓扑层面科学量化何谓“多样”与“创新”。 这种能力在正在到来的“智能体新纪元”（Agentic Era）中变得生死攸关。姜利维指出，在复杂的自主智能体（Agent）场景中，核心要求就是模型在庞大的解空间中自行摸索探索路径；在此过程中，如果负责提供反馈的奖励模型只能辨认狭隘的标准模式，智能体的探索潜能就会被瞬间掐灭。 在全场演讲的最后，姜利维动情地呼吁： > “未来的 AI 不应该是少数几种声音的被放大的回音；它应当是一曲由多元之声汇聚而成的交响乐（The future of AI shouldn't be louder echoes of the very few voices. It should be a chorus of many）。” 她指出，整个 AI 研究共同体必须全力以赴去赋能并放大人类创造力的全貌——捍卫价值多元性（Value plurality）、独立思考（Independent thoughts）、文化包容性（Cultural inclusivity）、多元推理路径（Alternative reasoning paths）、异质化表达（Heterogeneous expressions）以及开放式探索潜力。要达成这一愿景，必须在大语言模型研发的全栈链路（数据准备、算法训练、推理界面、交互模式，尤其是居于研发指挥棒地位的评估基准）上推行根本性的范式创新，携手迈向一个超越“人工蜂群”的全新时代。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=gVTYfI4pOSg&t=4s) Artificial Hivemind: The Open-Ended Homogeneity of Language Models (and Beyond) — 姜利维 > [03:50](https://www.youtube.com/watch?v=gVTYfI4pOSg&t=230s) 单模型内部同质化（Intra-model Homogeneity）：单一模型在面对同一开放式提示词时，反复生成内部高度雷同的回复。 — 姜利维 > [03:50](https://www.youtube.com/watch?v=gVTYfI4pOSg&t=230s) 跨模型间同质化（Inter-model Homogeneity）：由完全不同的团队在不同数据上训练出的异构模型，竟然产生惊人一致的输出结构。 — 姜利维 > [03:50](https://www.youtube.com/watch?v=gVTYfI4pOSg&t=230s) 用 50 个词写一个关于一只五彩斑斓的蟾蜍去冒险的微型故事 > [03:50](https://www.youtube.com/watch?v=gVTYfI4pOSg&t=230s) 写一个关于时间的隐喻 > [10:11](https://www.youtube.com/watch?v=gVTYfI4pOSg&t=611s) 人类认知与独创性的退化风险（Homogenizing Human Thought and Agency） — 姜利维 > [10:11](https://www.youtube.com/watch?v=gVTYfI4pOSg&t=611s) 合成数据与自迭代范式的天花板（Synthetic Data & Recursive Improvement Ceilings） — 姜利维 > [10:11](https://www.youtube.com/watch?v=gVTYfI4pOSg&t=611s) 科学与社会发现潜力的扼杀（Constraining Scientific and Societal Discovery） — 姜利维 > [10:11](https://www.youtube.com/watch?v=gVTYfI4pOSg&t=611s) 模型融合与集成机制的彻底失效（Collapse of Model Ensembling） — 姜利维 > [17:55](https://www.youtube.com/watch?v=gVTYfI4pOSg&t=1075s) The future of AI shouldn't be louder echoes of the very few voices. It should be a chorus of many — 姜利维 > [17:55](https://www.youtube.com/watch?v=gVTYfI4pOSg&t=1075s) 未来的 AI 不应该是少数几种声音的被放大的回音；它应当是一曲由多元之声汇聚而成的交响乐 — 姜利维 ## Mentioned 姜利维 (Liwei Jiang), 华盛顿大学 (University of Washington), 斯坦福大学 (Stanford University), 卡耐基梅隆大学 (CMU), 艾伦人工智能研究所 (AI2), NeurIPS, ChatGPT, Anthropic, Claude, Claude 3.5 Sonnet, DeepSeek, WildChat, Infinity-Chat ## Video description Liwei Jiang - Ph.D. candidate, Researcher, University of Washington Language models (LMs) often struggle to generate diverse, human-like creative content, raising concerns about the long-term homogenization of human thought through repeated exposure to similar outputs. Yet scalable methods for evaluating LM output diversity remain limited, especially beyond narrow tasks such as random number or name generation, or beyond repeated sampling from a single model. We introduce Infinity-Chat, a large-scale dataset of 26K diverse, real-world, open-ended user queries that admit a wide range of plausible answers with no single ground truth. We introduce the first comprehensive taxonomy for characterizing the full spectrum of open-ended prompts posed to LMs, comprising 6 top-level categories (e.g., brainstorm & ideation) that further breaks down to 17 subcategories. Using Infinity-Chat, we present a large-scale study of mode collapse in LMs, revealing a pronounced Artificial Hivemind effect in open-ended generation of LMs, characterized by (1) intra-model repetition, where a single model consistently generates similar responses, and more so (2) inter-model homogeneity, where different models produce strikingly similar outputs. Infinity-Chat also includes 31,250 human annotations, across absolute ratings and pairwise preferences, with 25 independent human annotations per example. This enables studying collective and individual-specific human preferences in response to open-ended queries. Our findings show that LMs, reward models, and LM judges are less well calibrated to human ratings on model generations that elicit differing idiosyncratic annotator preferences, despite maintaining comparable overall quality. Overall, INFINITY-CHAT presents the first large-scale resource for systematically studying real-world open-ended queries to LMs, revealing critical insights to guide future research for mitigating long-term AI safety risks posed by the Artificial Hivemind. [Source](https://www.youtube.com/watch?v=gVTYfI4pOSg)

Source quotes

[0:04] 姜利维: Artificial Hivemind: The Open-Ended Homogeneity of Language Models (and Beyond) [3:50] 姜利维: 单模型内部同质化（Intra-model Homogeneity）：单一模型在面对同一开放式提示词时，反复生成内部高度雷同的回复。 [3:50] 姜利维: 跨模型间同质化（Inter-model Homogeneity）：由完全不同的团队在不同数据上训练出的异构模型，竟然产生惊人一致的输出结构。 [3:50] 用 50 个词写一个关于一只五彩斑斓的蟾蜍去冒险的微型故事 [3:50] 写一个关于时间的隐喻 [10:11] 姜利维: 人类认知与独创性的退化风险（Homogenizing Human Thought and Agency） [10:11] 姜利维: 合成数据与自迭代范式的天花板（Synthetic Data & Recursive Improvement Ceilings） [10:11] 姜利维: 科学与社会发现潜力的扼杀（Constraining Scientific and Societal Discovery） [10:11] 姜利维: 模型融合与集成机制的彻底失效（Collapse of Model Ensembling） [17:55] 姜利维: The future of AI shouldn't be louder echoes of the very few voices. It should be a chorus of many [17:55] 姜利维: 未来的 AI 不应该是少数几种声音的被放大的回音；它应当是一曲由多元之声汇聚而成的交响乐

Source quotes

[3:50] Liwei Jiang: Artificial Hivemind [6:41] Liwei Jiang: write a short story about a colorful toad goes on adventure in 50 words [6:41] Liwei Jiang: slim-fitted [6:41] Liwei Jiang: Write a metaphor involving time [6:41] Liwei Jiang: Time as a River [6:41] Liwei Jiang: Time as a Weaver [10:11] Liwei Jiang: safeguard human originality and agency [17:48] Liwei Jiang: louder echoes of the very few voices [17:48] Liwei Jiang: a chorus of many

Original transcript

[0:04] Hello everyone. I'm Leeway. I'm uh uh actually  just graduated from my PhD at University of [0:08] Washington. And today I'm very excited to  share our work on artificial hype mind, [0:14] the open-ended homogeneity of language models  and beyond. And uh this is a joint work with [0:19] my wonderful collaborators from University of  Washington, Stanford, CMU and AI2. And this [0:24] work uh was published last year at the new rooms  2025 and received the best paper award at the [0:30] uh data set track. So AI has now been integrated  into global workflows and daily lives at scale [0:39] that has never been seen before. And by July  2025, CHGBT reached about uh 10% of the world [0:46] adult population and 800 million uh weekly active  users. The highlight is its huge impact on humans. [0:54] And now we can only imagine there's more usages  like this. And when we look at how people actually [1:00] use these systems, the majority of use cases are  about open-ended tasks. whether they are about [1:07] practical guidance, seeking information, writing  and multimedia creation. So much of users ask AI [1:14] for is inherently open-ended and these tasks  are required for creativity or multiple valid [1:21] uh possible solutions. And similarly, Anthropic's  report on cloud shows that many of its top use [1:29] cases also involve questions with no single  correct answers. So this brings up a key [1:35] question. And how do these large language models  they handle open-ended questions? So let's start [1:41] by looking at what open-ended questions people  actually submit to LMS. For this we use well [1:47] chat a data set of 1 million naturally occurring  chatbt interaction uh logs collected from the real [1:53] world use uh use cases. And we first filter wild  chat for 37k meaningful in the wild questions. And [2:02] next we classify the open-endedness of each query  by a uh LM judge. And about 75% of the questions [2:10] are open-ended meaning that uh they can accept  multiple valid answers. And the remaining 25 are [2:16] close-ended uh typically involving factual lookups  or yes or no questions. So this confirms our [2:21] impression that most of the question that people  ask to AI are open-ended. So we then manually [2:28] exam about 100 queries to uh iteratively build a  taxonomy of openness and we scale this label into [2:36] the full data set with the LM judge and enable  the model to identify any additional categories [2:43] and after the classification we identify 26k  open-ended questions and this forms the core data [2:49] set of infinity chat which we use to study how  models respond across a wide open-ended landscape. [2:57] And here's the full taxonomy with six top  categories spanning 17 subcategories. And it [3:02] gives us a structured way to understand what open  the questions do people actually pose to large [3:09] language models. And we start with what users ask  for the most often. Creative content generation [3:15] or writing representing nearly 58% of all the  cases. And brainstorming and ideation is another [3:22] well-known category where uh multiple answers are  not only possible but they are expected. And next [3:29] is a variety of general open-ended questions and  people also rely on AI for information seeking [3:36] that ask questions that support different  types of decision- making and many use cases [3:42] involve alternative communication or writing  styles or about value subjective perspectives. [3:50] So with infinity chat we are now ready to answer  the question how do AI LMS handle these open-ended [3:57] questions. So the short answer is they're not  doing it very diversely. LMS they suffer from [4:04] a phenomenon that we call as artificial high  mind where these models they collapse towards [4:10] overly similar outputs and this is in close  relation to the mode clasp collapse problem [4:16] from previous literature which is particularly  serious for language models that went through [4:23] post-training process for instance the previous  works have found that instruction trained LMS they [4:29] perform poorly when they are instructed to produce  random outputs they generating random numbers or [4:35] generating random uh country names and they often  uh give us the texts that share repetitive anchor [4:42] textual spans and with this foundation and using  real world openly questions. Um we uh conduct a [4:51] large scale evaluation over 70 language models  across different families and different sizes to [4:56] show that uh what we call as artificial hive mind  effect. And a hive mind traditionally refers to a [5:03] collective consciousness in which many individual  they share thoughts or decisions as if they were a [5:09] single mind. And our setting artificial high mind  emerges in two key forms. Intramodal homogeneity [5:16] where a single model repeatedly generate  responses that are highly similar to each [5:21] other and even more so intermodal homogeneity uh  where we use different models but the different [5:27] models may also give us very uh similar outputs.  So first uh for let's take a look at the uh first [5:35] of the intramodal homogeneity for our study we  selected 100 test queries from infinity chat and [5:42] all verified by human annotators to be uh indeed  open-ended and for every query and for each of [5:48] the model we sample 50 independent responses  with temperature of one to allow the most [5:54] uh high stocasticity as much as possible. So uh  temperature essentially is a hyperparameter that [6:01] allow us to generate as uh random uh attacks  as possible while still maintain the uh the [6:08] naturalness and the tax text tax quality. So and  we can uh then compute a pair wise uh semantic [6:15] similarity across those 50 outputs and then  report what fraction of these queries fall into [6:22] uh each of the similarity score range. And our  results shows that uh about 79% of the cases [6:29] the average similarity exceeds 0.8 uh even  under aggressive sampling. And this indicate [6:35] strong semantic convergence in how models answer  open-ended questions. But similarity score itself [6:41] uh does not really show us how that look like in  practice. So let's walk through a few concrete [6:47] examples. So for this question about writing  a short story about a colorful toad goes on [6:54] adventure in 50 words cloud 3.5 Sonic here  and many many other models they all produce [7:02] highly similar answers across different generation  rounds. Often they use nearly the same tone name [7:08] same story arc and similar wording. And sometimes  this kind of similarity is hard to really quantify [7:14] with automatic metrics because uh for example  story arc is something that is very very latent [7:20] in the text. So this becomes problematic when  we uh remember that millions of people actually [7:27] rely on these generative AI tools on daily basis.  So when people ask similar open-ended questions, [7:34] they end up receiving similar outputs that  ultimately um risk homogenizing how this [7:40] creative content get represented and promoted  and disseminated across different people. So now [7:48] uh one might say well users uh don't usually type  the exact same queries by word. So we also checked [7:56] whether paraphrases would change the um outputs.  So uh even with these variations, we found that [8:02] models still uh produced the responses that she  long repetitive text spins with slightly decreased [8:09] but still pretty high similarity scores. So in  other word uh just by uh the the reputation isn't [8:16] going away just by changing uh it to paraphrases.  So now let's turn to intermodel homogeneity to [8:23] compare responses across different models and  darker regions within this plot indicate higher [8:30] uh output similarity. So we can see that even  across different models we can still see pretty [8:34] high similarity level. And again let's look  at some concrete examples and with this query [8:43] and different models they produce responses that  share similar sentence scaffold and even reuses [8:49] many of these words in verbatim ways. And if you  look uh closely, so for example, the slim fitted [8:57] uh is a word that is uh repeating across deepseek  and stretch bt and yeah some other words that [9:04] not even the most common words and between the  different models they uh tend to repeat them and [9:10] sometimes these different models even output exact  the same content uh even if the question is pretty [9:15] open-ended. So now let's look at the models all  together with one of my favorite examples. So for [9:22] this query write a metaphor involving time. We uh  collected outputs from 25 models and cluster them [9:29] uh with sentence embeddings and then visualize  them project them onto this uh with PCA with [9:35] two uh dimensions of features. And despite that  the uh diverse source models they're all coming [9:41] from different uh model families and sizes and the  responses they converge into these two uh primary [9:49] um viral clusters. So the dominant cluster visual  cluster on the left is about time as a river and [9:59] the smaller cluster on the right is about time as  weaver and a few additional instances that fall [10:05] out of those two primary real clusters. So uh if  you're interested feel free to check out our paper [10:11] uh and we have more examples like this. So now  what does this mean for us and what should we [10:18] be cautious about when uh considering AI's longer  term impact and safety risk? So as we uh already [10:25] alluded to when lots of people global users they  are exposed to highly similar content even across [10:32] different models it risk homogenizing how people  think and write and there's studies that already [10:38] show that since the rise of our lens linguistic  diversity and ideation creativity and this [10:45] um this highly individualistic writing styles has  have declined from humans and this prompts urgent [10:52] cause to safeguard human originality and agency.  And next as synthetic data and self uh improving [11:00] paradigms and also most recently this uh recursive  uh improvement paradigms have become uh central to [11:08] AI nowadays. Limited model diversity constraints  our ability to uh truly break out of the weather [11:15] uh the current model's ability and um find these  challenging task cases and this hearts the breath [11:21] of reasoning and answers that we can get out  of this uh AI models and another implication [11:29] is around AI for scientific and societal discovery  and this high factor limit AI's uh risk limiting [11:36] AI's ability to generate genuinely novel hypothes  es and uh genuinely hyp uh novel um um conclusions [11:45] that limit their discovery potentials. So finally  intramodal homogeneity limits the effectiveness [11:52] of model ensembles uh in which case that people  propose as a a natural um uh solution to combat [12:00] these uh the homogeneity within a single model.  But if the models are not really uh creative [12:06] even across different models and then we lose the  effectiven of uhness of it. So um this will hurt [12:13] uh how we can produce uh diverse solutions or  complimentary oversightes because models may just [12:20] fail collectively. So now that we've seen that  language models they struggle with generating [12:26] these open-ended uh answers and the next question  is how are they calibrated to evaluate alternative [12:34] alternative responses in uh rating graating and  discriminative setups. So we consider the open [12:41] questions and alternative responses to them.  For alternative valid responses, individual [12:47] raers may show idiosyncratic preferences, but the  overall uh average score tend to be similar if we [12:55] uh ask really broadly of people's opinions. And  ideally, models should uh rate alternative valid [13:02] responses with comparative uh comp comparable  overall scores. So why does this matter? Because [13:09] if model fails to give similar scores and it may  demote alternative possibilities during model [13:16] development for instance during training and uh  especially the post- training of language models [13:22] in both the offline data curation and online uh  RL rollouts if the reward model fails to recognize [13:30] similar quality alternatives it effectively rip uh  wipes out of these possibilities from the model's [13:36] learning and similar ly during evaluation if a  biased LM judge that can only recognize a narrow [13:44] slice of all the possibilities it may wrongly  label uh plausible alternatives an alternative [13:50] answers as incorrect answers. So uh however  existing preference data sets they lack the [13:58] fundamental structure uh needed to evaluate how  well models can uh cal uh calibrate their ratings [14:04] for openly questions because existing uh data sets  they are sparsely labeled with typically a few [14:11] human labels uh per example uh typically one two  three and often there is no alternative responses [14:19] um that target these open-ended questions. So we  need to collect human data in this form and to do [14:26] so we sample 50 queries from infinity chat and for  each query we sample 15 distinct responses from [14:32] our pool of 70 LMS. And next we gather um this  extensive human annotations using both absolute [14:40] rating and pair wise preference setups. And under  uh absolute rating setup, each of the 15 responses [14:48] uh responses is evaluated uh on a one to five  quality scale by 25 independent annotators. [14:55] And remember this is um uh compared to in  previous work people usually have one to [15:01] uh three independent annotators. And also  similarly under the pair-wise preference [15:06] setup we randomly sample 10 uh distinct pairs  and the uh we uh again recruit 25 independent [15:13] annotators to give us the rating of which one they  prefer between the pair and together this gives us [15:19] a 31k this human labels and here's the label  distribution under the absolute rating setup [15:26] open-ended questions they frequently produce  high entropy ratings between 25 annotators [15:33] And uh so for example um for the uh last example  almost exactly the same amount of people they [15:40] like the responses versus not uh as uh compared to  the third example where most of the people agree [15:47] that they like that response and similar trend  was observed for pair wise preference rating [15:54] as well. So this dense annotation surface rich  distributional uh distributional structures that [16:01] only become visible when we collect far more than  just a few annotations. And now we're ready to see [16:07] how various models including LM themselves, reward  models and LM judges are calibrated when scoring [16:14] alternative responses to open-ended questions.  For both absolute and pair wise setups, uh [16:20] the correlation between model scores and average  human rating at large drops substantially between [16:26] responses of similar overall quality and responses  where individual enterers disgrace. So what that [16:33] means is that in the eyes of human two responses  might have similar quality but in the eyes of [16:38] models they could be drastically different and  the models are given the very different scores. So [16:43] uh as a result during the model development loop  the model will learn that one is much better than [16:48] the other uh but humans will just rate them as  similar and uh specifically uh we can look at [16:56] the subset of the responses that share similar  overall scores and the more similar they are and [17:02] the more that models will uh find confusing to how  to rate them and because there's no single correct [17:08] ways for define what a single quality subset  means. So we also experimented with several [17:14] selection strategies which we included in the  paper and uh they point to uh similar conclusions. [17:21] So uh in summary to better support u modeling  open-ended questions we need models that are [17:29] explicitly calibrated for open-endedness  and more human data with dense labels to [17:35] capture distributional nuances and richness and  evaluation uh frameworks that can meaningfully [17:41] assess multiple valid responses. And we hope that  our infinity chat data set provides a solid uh [17:48] starting point for building these uh components.  And recall the previously mentioned cautious [17:55] there's still significant technical challenges  in escaping the artificial high. However, [18:01] the good news is that there are already emerging  uh efforts aimed at conquering this phenomenon. [18:07] For example, people have already been developing  diversity aware data curation strategies, novel [18:13] decoding uh strategies that maximize generative  diversity during the inference time of the models [18:20] so that we don't need to retrain the models.  uh also there's also novel training paradigms [18:24] that explicitly incentivize these uh diverse  reasoning and outputs uh get preserved in this [18:31] post-train models and also uh comprehensive  realistic evaluation metrics for rigorously [18:38] uh quantify what creativity and diversity  even mean and now uh uh we are also moving [18:46] into this agentic world. So uh much of the agent  use cases also require um the fundamental ability [18:54] of the model to be able to discover solutions. Uh  especially during training uh and then the agent [19:01] model's uh job is to look for different possible  solutions and uh the reward model will give them [19:08] um the the signal of which one is which path is  actually good. So fundamentally the exploratory [19:14] capabilities are very much needed. So uh and make  uh this artificial high manufact uh even more [19:20] important in this agent new era. So um there's  uh still significant ch significant challenges [19:28] uh remain for fully breaking free of artificial  highmind and the future of AI shouldn't be louder [19:34] echoes of the uh very few voices. It should be  a chorus of many and our AI community should uh [19:42] strive to empower and amplify this full spectrum  of human creativity, value plurality, independent [19:49] thoughts, cultural uh inclusivity, alternative  reasoning path and heterogeneous expressions and [19:54] these open-ended discovery potentials and doing so  require fundamental innovations across the entire [20:02] uh development stack. So spending data  training and also model inference interfaces [20:07] interaction modes and also uh very importantly  evaluations and benchmarks because usually in [20:14] uh the current R&D uh process benchmark is driving  how models behaving. So that's very important. [20:22] Uh with that I'll end it here and uh say that  we should build beyond artificial highman