cs.CLMar 6, 2026
SaveCluster-R1: Large Reasoning Models Are Instruction-following Clustering Agents
Organizations: Dartmouth College, USA · Adobe Research, San Jose, USA
Abstract
General-purpose embedding models excel at recognizing semantic similarities but fail to capture the characteristics of texts specified by user instructions. In contrast, instruction-tuned embedders can align embeddings with textual instructions yet cannot autonomously infer latent corpus structures, such as determining the optimal number of clusters. To address both limitations, we reframe instruction-following clustering as a generative task and train large reasoning models (LRMs) as autonomous clustering agents. Our reasoning-driven training pipeline enables LRMs to interpret high-level clustering instructions and then infer the corresponding latent groupings. To evaluate this paradigm, we introduce ReasonCluster, a comprehensive benchmark comprising 28 diverse tasks spanning daily dialogue, legal cases, and financial reports. Experiments across diverse datasets and clustering scenarios show that our approach consistently outperforms strong embedding-based methods and LRM baselines, demonstrating that explicit reasoning fosters more faithful and interpretable instruction-based clustering.
Figures & tables
Figure 1: Overview of embedding-based vs. reasoning-driven clustering. LRMs follow diverse user instructions and adaptively infer latent group structures.
Figure 2: The system prompt for clustering.
Figure 3: The overview of our multi-agent data synthesis pipeline.
Figure 4: Overview of the benchmark evaluation splits: (a) dataset statistics by source and split, where text and input lengths are measured in tokens, (b) distribution of the number of clusters per data example, and (c) distribution of cluster size, indicating the number of text instances per cluster for each dataset.
| Model | LMSY | ECHR | SP500 | AVG | OVERALL | ||||||||
| C0 | C1 | C2 | C0 | C1 | C2 | C0 | C1 | C2 | C0 | C1 | C2 | V | |
| Reasoning Models | |||||||||||||
| o3 | 63.60 | 63.42 | 64.17 | 85.75 | 85.79 | 47.04 | 57.01 | 59.04 | 66.51 | 68.79 | 69.42 | 59.24 | 65.08 |
| Gemini 2.5 Pro | 55.98 | 56.06 | 57.95 | 85.05 | 84.14 | 64.29 | 57.60 | 63.43 | 69.34 | 66.21 | 67.88 | 60.92 | 61.82 |
| GPT-oss-120B | 48.74 | 49.30 | 52.77 | 70.54 | 69.18 | 52.58 | 44.76 | 48.15 | 51.84 | 54.68 | 55.54 | 52.40 | 52.31 |
| DeepSeek-R1 | 42.96 | 43.64 | 45.80 | 53.78 | 54.55 | 34.06 | 41.57 | 40.47 | 39.69 | 46.10 | 46.22 | 39.85 | 44.06 |
Table 1: V-measure (%) score across three datasets under different evaluation splits (C0, C1, C2; see Section 4.3 for details). The final column presents the overall V-measure computed over all individual samples. Bold indicates the best result, and proprietary models are highlighted. Reasoning models outperform general models across all settings. Our models achieve the strongest overall performance, indicating that explicit reasoning supervision is crucial for faithful instruction-following clustering.
| Model | LMSY | ECHR | SP500 | AVG | OVERALL | ||||||||
| C0 | C1 | C2 | C0 | C1 | C2 | C0 | C1 | C2 | C0 | C1 | C2 | V | |
| K-means | |||||||||||||
| Qwen3-Embedding | 55.88 | 56.28 | 53.26 | 74.58 | 71.87 | 59.45 | 53.41 | 54.67 | 60.81 | 61.29 | 60.94 | 57.84 | 58.31 |
| InBedder | 52.19 | 52.05 | 45.95 | 67.46 | 66.08 | 48.13 | 53.06 | 53.97 | 49.72 | 57.57 | 57.37 | 47.93 | 53.28 |
| Instructor | 47.20 | 47.04 | 44.04 | 58.12 | 57.42 | 47.30 | 48.74 | 49.12 | 47.28 | 51.35 | 51.19 | 46.21 | 48.55 |
| OpenAI-Embedding | 50.63 | 51.28 | 49.11 | 56.58 | 55.99 | 45.75 | 47.74 | 48.82 | 48.47 | 51.65 | 52.03 | 47.78 | 50.66 |
Table 2: Comparison of embedding-based clustering against our models. Results of instruction-following embedders (Qwen3-Embedding, InBedder, and Instructor) are highlighted. Bold indicates the best result. While instruction-tuned embedders outperform generic embedders, they remain consistently inferior to reasoning-based clustering.
| Method | C0 | C1 | C2 | Ave. |
| SFT + Final Answer | 65.51 | 52.70 | 8.20 | 42.14 |
| SFT + Distilled + Final Answer | 64.87 | 54.76 | 11.98 | 43.87 |
| SFT + Distilled + RL | 72.26 | 71.96 | 57.15 | 67.12 |
Table 3: Comparison of reasoning training versus SFT directly on final answer. Reasoning supervision is essential for generalization: models trained without reasoning fail on C2, while the full pipeline remains robust.
| Method | C0 | C1 | C2 | Ave. |
| 7B Base Model | ||||
| + RL | 46.72 | 48.19 | 46.52 | 44.90 |
| + RL + Reasoning Hint | 48.19 | 49.99 | 46.07 | 46.42 |
| + Distill + RL + Reasoning Hint | 66.05 | 65.97 | 56.46 | 61.80 |
| 14B Base Model | ||||
| + RL | 56.39 | 53.57 | 53.71 | 52.76 |
Table 4: Comparison of training recipes across 7B and 14B models. Prompt-level reasoning hints offer limited benefits, whereas combining reasoning distillation with RL yields the strongest performance.
Figure 5: Training dynamics of different training recipes for the Qwen-7B model. Models initialized with distilled reasoning traces converge faster, exhibit more stable response lengths, and achieve higher rewards.
Figure 6: Effect of model size on clustering performance and generalization. The y-axis shows the average V-measure over all corresponding samples.
Appendix figures & tables11 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 7: The system prompt of cluster generation agent.
Figure 8: The system prompt for taxonomy refinement and clustering instruction design.
Figure 9: The system prompt for expert classification using a predefined taxonomy.
Figure 10: The system prompt for clustering assistant with required output format.
Figure 11: The system prompt for reasoning chain generation.
Figure 12: The system prompt for reasoning chain refinement.
Figure 13: LMSYS-Chat data example.
Figure 14: ECHR data example.
Figure 15: Example of Item 2 (“Properties”) from SP500 dataset.
| Dataset | Split | Prompt |
| ECHR | Held-in | Cluster these legal cases by the primary vulnerable population involved. Cluster these legal cases by the dominant type of domestic proceedings that preceded the European Court application. Cluster these legal cases by the violation articles invoked in the Court’s judgment. |
| Held-out | Cluster the following texts based on the primary underlying category of legal dispute or issue. Cluster the following texts based on the main types of parties or entities involved as adversaries or plaintiffs. Cluster the following texts based on the timing, duration, or current stage of resolution for the described legal proceedings. | |
| LMSY | Held-in | Cluster the following user requests based on explanation audience level. Cluster these requests by temporal regime they focus on. Cluster the dialogue by identifying the root cause of the issue. Cluster the dialogue based on the type of transformation needed to convert it into a different form or function. Cluster these creative generation requests by content type. Group each dialogue according to the primary social function or interpersonal purpose it serves. Cluster these decision support requests based on the structure and format of the solution being requested. Group these recommendation requests by identifying the type of justification used. Cluster the dialogues by the distinct perspective or evaluative role the agent adopts. Group the conversations based on the type of structured planning or scheduling being discussed. |
| Held-out | Group these computational tasks by archetype. Cluster these data analysis requests by tooling context. Cluster these texts based on the kind of external context or disambiguation needed. Group these factual lookup requests by the structure and format of the expected output. Cluster these factual lookup requests based on the type of source or origin of information. | |
| SP500 | Held-in | Cluster these SP500 companies by their asset ownership and utilization patterns. Cluster these SP500 companies by their geographic distribution of assets and operations. Cluster these SP500 companies by their industry-specific asset composition. Cluster these SP500 companies by their risk exposure and asset encumbrance patterns. |
| Held-out | Cluster the texts based on the strategies companies use for distributing earnings to shareholders via dividends. Cluster the texts based on the approaches companies take toward buying back their own shares. Cluster the texts based on the varieties of equity instruments issued by companies. |
Table 5: Clustering instructions for both held-in and held-out tasks from each dataset.
Figure 16: Relationship between format accuracy and V-measure score. Models with higher format accuracy consistently achieve higher V-measure scores, highlighting format correctness as a fundamental prerequisite for clustering quality. Nevertheless, high format accuracy alone does not guarantee superior clustering performance. For instance, models such as Llama-3.1-70B-Instruct exhibit high format accuracy but still attain low V-measure scores. Effective clustering further depends on a model’s capacity for semantic reasoning and clustering instruction alignment. Mastering the syntax of clustering outputs is necessary but not sufficient.
More in this topic
Clustering
Clustering groups data points by shared structure, developing methods to discover, validate, and interpret groups across diverse data and constraints.
See all 335 papers