> ## Content Index > Fetch the complete content index at: https://www.notatechguy.com/llms.txt > Use this file to discover other available public pages before exploring further. # KuaiRP researchers claim role-playing AI matches proprietary models at - URL: https://www.notatechguy.com/kuairp-researchers-claim-role-playing-ai-matches-proprietary-models-at-lower-cos/ - Published: 2026-09-12T15:20:58.000Z - Updated: 2026-09-12T15:20:58.000Z - Description: The KuaiRP team says its self-distillation method recovers lost capabilities after domain tuning, yet the paper names no specific rivals or scores. - Author: Marcello Babbili - Tags: Technology & AI A team of researchers has posted a technical report describing a three-stage pipeline that aims to give small language models deep role-playing knowledge without erasing their general reasoning — a problem that has plagued domain adaptation for years [S¹](https://arxiv.org/abs/2609.11127?ref=notatechguy.com). The paper, uploaded to arXiv on 12 September, argues that injecting specialised knowledge typically triggers catastrophic forgetting of the model's broader agent abilities, and it proposes a fix that uses the adapted model itself as a teacher for the original base model. **My read:** The self-distillation idea — letting the domain-tuned model teach the base model, is clever and fits a growing trend of models improving themselves without human labels. But the report gives no benchmark numbers, no parameter counts, and no comparison against named proprietary systems. Until those appear, the claim of matching state-of-the-art proprietary models is a promise, not a result. The failure the authors target is familiar to anyone who has fine-tuned a model for a narrow task: the more domain knowledge you cram in, the more the model forgets how to follow instructions, use tools, or reason generally. Their pipeline attacks this in three steps. First, they build a supervised fine-tuning dataset using a standardised character template, simulated user behaviour, and a reverse-profile filter to keep examples clean. Second, they apply a rule-based composite reward during reinforcement learning to stamp out verbose, repetitive outputs — a common degradation when models are optimised for character consistency. Third, and most novel, they run a two-stage on-policy distillation with cumulative-divergence decay: the domain-adapted model generates training data for the original base model, gradually pulling it toward the specialised behaviour while preserving its general skills [S¹](https://arxiv.org/abs/2609.11127?ref=notatechguy.com). The approach echoes recent work on character-centric policy optimisation, where researchers at another lab used group-relative rewards to align role-playing agents without collapsing their reasoning [P²](https://arxiv.org/html/2605.25511v1?ref=notatechguy.com). But the KuaiRP report goes further by framing the whole process as a recovery mechanism rather than a one-shot alignment. What the paper does not say is equally telling. No benchmark scores appear in the abstract. No specific proprietary models — GPT-4, Claude, or any open-weight rival, are named as the comparison baseline. Parameter counts, training compute, and the domains evaluated are all absent. The claim of "extremely low deployment costs" rests on an asserted small parameter size that the abstract never quantifies. And the work remains a preprint: it has not been peer-reviewed, and no independent replication is cited [S¹](https://arxiv.org/abs/2609.11127?ref=notatechguy.com). For engineers building character-driven agents — think NPC dialogue in games, customer-service personas, or interactive tutoring, the template-and-reward design is immediately actionable even without the distillation stage. The reverse-profile filtering step, which discards training examples that drift from the target character, could be dropped into existing SFT pipelines today. The distillation recipe, however, needs the full paper and ideally released code before it can be trusted in production. The next checkpoint is the full paper's release on arXiv (2609.11127) and any accompanying code repository. If the authors publish weights and evaluation scripts, the community can test whether the self-distillation actually closes the gap to proprietary models — or whether the trade-off between domain depth and general capability has simply been moved, not solved. --- *Sources: [S1 — KuaiRP Series Role-playing Models Technical Report](https://arxiv.org/abs/2609.11127?ref=notatechguy.com) · [P2 — CRPO: Character-centric Group Relative Policy Optimization for Role-aw](https://arxiv.org/html/2605.25511v1?ref=notatechguy.com) · [P3 — benchen4395/KuaiSearch](https://github.com/benchen4395/KuaiSearch?ref=notatechguy.com) · [P4 — \[2609.08936\] AuK Technical Report: An Open-Source Foundational Model f](https://arxiv.org/abs/2609.08936?ref=notatechguy.com) · [P5 — CRPO: Character-centric Group Relative Policy Optimization for Role-aw](https://arxiv.org/html/2605.25511?ref=notatechguy.com)* ## Related reading - [AI image models fingerprinted without watermarks](https://www.notatechguy.com/ai-image-models-fingerprinted-without-watermarks/) — our technology desk, 2026-08-13 - [AI models lose over 90% of safety signal in African languages](https://www.notatechguy.com/ai-models-lose-over-90-of-safety-signal-in-african-languages/) — our technology desk, 2026-08-12 - [MMJailBench: prompt framing is top jailbreak risk across 16 AI models](https://www.notatechguy.com/mmjailbench-prompt-framing-is-top-jailbreak-risk-across-16-ai-models/) — our technology desk, 2026-09-01 --- *Written from 5 sourced items, 4 of them primary.*