mlsanityData to 7 Oct 2026

From Production Traffic to Post-Training: Building a Self-Hosted LLM That Covers the Corporate Request Mix

arXiv 2609.01572 ↗First posted 1 Sep 2026cs.CLcited by 0

Not on the map. Why?

Abstract (arXiv)

Data-residency constraints force enterprises to self-host LLMs, but continuous adoption of newer models without decommissioning their predecessors expands the serving fleet, fragmenting a finite GPU pool. We consolidate traffic from over 200 internal applications onto a single model by closing quality gaps identified through production error analysis along three axes: instruction following, function-calling, and internal task distribution. Quality is tracked by offline benchmarks stratified to production traffic and scored by deterministic verifiers or calibrated LLM judges. Rather than optimising all objectives jointly, which introduces cross-domain reward interference, we train a separate GRPO expert per axis and merge them via two-stage SLERP. Each expert's reward exposes a distinct failure mode, namely semantic collapse, over-calling, and verbosity hacking, each requiring a domain-specific fix. In non-reasoning mode the recipe surpasses a ~7× larger by total parameters baseline on the in-house Arena with 69.6 to 65.8, instruction following with 0.85 to 0.83, and function-calling with 0.79 to 0.77, while lifting general dialogue benchmarks. The model absorbs 50% of platform traffic, 116M requests per month, at a fraction of the serving cost.

Key facts. No citing papers in the map yet (data through 7 Oct 2026). Builds on 42 of its 59 references that are in the map. First posted on arXiv 1 Sep 2026.

Builds on

42 of 59 references.

References · 42

  1. 2022Training language models to follow instructions with human feedbackcited by 15,712
  2. 2017Proximal Policy Optimization Algorithmscited by 14,441
  3. 2024DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Modelscited by 8,851
  4. 2022ReAct: Synergizing Reasoning and Acting in Language Modelscited by 8,355
  5. 2023Direct Preference Optimization: Your Language Model is Secretly a Reward Modelcited by 7,916
  6. 2023Efficient Memory Management for Large Language Model Serving with PagedAttentioncited by 5,767
  7. 2022Self-Instruct: Aligning Language Models with Self-Generated Instructionscited by 2,674
  8. 2019Fine-Tuning Language Models from Human Preferencescited by 2,124
  9. 2024HybridFlow: A Flexible and Efficient RLHF Frameworkcited by 1,980
  10. 2022Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference timecited by 1,366
  11. 2025gpt-oss-120b & gpt-oss-20b Model Cardcited by 1,258
  12. 2022Scaling Laws for Reward Model Overoptimizationcited by 1,056
30 more
  1. 2023BloombergGPT: A Large Language Model for Financecited by 923
  2. 2023TIES-Merging: Resolving Interference When Merging Modelscited by 795
  3. 2022BERTopic: Neural topic modeling with a class-based TF-IDF procedurecited by 779
  4. 2023PyTorch FSDP: Experiences on Scaling Fully Sharded Data Parallelcited by 641
  5. 2025τ2-Bench: Evaluating Conversational Agents in a Dual-Control Environmentcited by 558
  6. 2025Rubrics as Rewards: Reinforcement Learning Beyond Verifiable Domainscited by 328
  7. 2023A Long Way to Go: Investigating Length Correlations in RLHFcited by 285
  8. 2024Arcee's MergeKit: A Toolkit for Merging Large Language Modelscited by 189
  9. 2024APIGen: Automated Pipeline for Generating Verifiable and Diverse Function-Calling Datasetscited by 188
  10. 2025Generalizing Verifiable Instruction Followingcited by 178
  11. 2025APIGen-MT: Agentic Pipeline for Multi-Turn Data Generation via Simulated Agent-Human Interplaycited by 158
  12. 2025OpenRubrics: Towards Scalable Synthetic Rubric Generation for Reward Modeling and LLM Alignmentcited by 118
  13. 2019Adaptation of Deep Bidirectional Multilingual Transformers for Russian Languagecited by 109
  14. 2025Reinforcement Learning with Rubric Anchorscited by 108
  15. 2025Nemotron-Research-Tool-N1: Exploring Tool-Using Language Models with Reinforced Reasoningcited by 67
  16. 2025TOUCAN: Synthesizing 1.5M Tool-Agentic Data from Real-World MCP Environmentscited by 61
  17. 2026Rethinking Rubric Generation for Improving LLM Judge and Reward Modeling for Open-ended Taskscited by 55
  18. 2025Chasing the Tail: Effective Rubric-based Reward Modeling for Large Language Model Post-Trainingcited by 46
  19. 2025VerIF: Verification Engineering for Reinforcement Learning in Instruction Followingcited by 43
  20. 2026RubricHub: A Comprehensive and Highly Discriminative Rubric Dataset via Automated Coarse-to-Fine Generationcited by 42
  21. 2023A Family of Pretrained Transformer Language Models for Russiancited by 27
  22. 2025MO-GRPO: Mitigating Reward Hacking of Group Relative Policy Optimization on Multi-Objective Problemscited by 21
  23. 2025Feedback-Driven Tool-Use Improvements in Large Language Models via Automated Build Environmentscited by 19
  24. 2025A Cost-Benefit Analysis of On-Premise Large Language Model Deployment: Breaking Even with Commercial LLM Servicescited by 15
  25. 2024Enhancing Function-Calling Capabilities in LLMs: Strategies for Prompt Formats, Data Integration, and Multilingual Translationcited by 14
  26. 2024Vikhr: The Family of Open-Source Instruction-Tuned Large Language Models for Russiancited by 12
  27. 2026Lost in Execution: On the Multilingual Robustness of Tool Calling in Large Language Modelscited by 10
  28. 2024Facilitating large language model Russian adaptation with Learned Embedding Propagationcited by 8
  29. 2025GigaChat Family: Efficient Russian Language Modeling Through Mixture of Experts Architecturecited by 8
  30. 2025T-pro 2.0: An Efficient Russian Hybrid-Reasoning Model and Playgroundcited by 2

Built on by

Not cited yet.

Similar, not cited

Close in wording, no citation either way

  1. 2026EnterpriseRAG: Benchmarking LLM Instruction Adherence and Robustness under Non-Ideal Enterprise Retrieval
  2. 2026Layer-Isolated Evaluation: Gating the Deterministic Scaffold of a Production LLM Agent with a No-LLM, Regression-Locked Test Harness
  3. 2026Efficient Benchmarking in Production: A Study of an Evolving LLM Agent
  4. 2026Adapting Technical-Service LLM Agents with Latent Logic Augmentation, Robust Noise Reduction, and Hybrid Reward Modeling
  5. 2024Efficiently Deploying LLMs with Controlled Risk
  6. 2025The Instruction Gap: LLMs get lost in Following Instruction
  7. 2026When Errors Become Narratives: A Longitudinal Taxonomy of Silent Failures in a Production LLM Agent Runtime
  8. 2026Human-Less LLM Serving: Quantifying the Human Tax on Throughput