[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"article-creativeinstruct-llms-quality-creativity-diversity-en":3,"article-related-creativeinstruct-llms-quality-creativity-diversity-en":29,"series-research-d26d3c47-b7f3-4598-a877-e7be8c67cf67":76},{"id":4,"slug":5,"title":6,"content":7,"summary":8,"source":9,"source_url":10,"author":11,"image_url":12,"cover_image":12,"category":13,"language":14,"translated_content":11,"related_article_id":15,"keywords":16,"key_takeaways":22,"views":26,"created_at":27,"published_at":28,"topic_cluster_id":11},"d26d3c47-b7f3-4598-a877-e7be8c67cf67","creativeinstruct-llms-quality-creativity-diversity-en","CreativeInstruct teaches LLMs to stay creative","\u003Cp>How do you keep an \u003Ca href=\"\u002Ftag\u002Fllm\">LLM\u003C\u002Fa> creative after post-training?\u003C\u002Fp>\u003Cp data-speakable=\"summary\">CreativeInstruct teaches \u003Ca href=\"\u002Ftag\u002Fllms\">LLMs\u003C\u002Fa> to preserve creativity and diversity without giving up quality.\u003C\u002Fp>\u003Cul>\u003Cli>\u003Cstrong>Research org\u003C\u002Fstrong>: Unspecified in arXiv abstract\u003C\u002Fli>\u003Cli>\u003Cstrong>Core data\u003C\u002Fstrong>: 70.3% of cases\u003C\u002Fli>\u003Cli>\u003Cstrong>Breakthrough\u003C\u002Fstrong>: Learns special [StartCreativity] spans during instruction tuning\u003C\u002Fli>\u003C\u002Ful>\u003Cp>That is the core problem this paper tackles: post-training usually makes models better at being useful, but worse at being surprising. If you are building systems for story generation, creative writing, or even RL pipelines that benefit from varied outputs, that tradeoff matters. The paper argues that creativity is not just a nice-to-have; it can be part of the model’s training substrate.\u003C\u002Fp>\u003Ch2>What problem this paper is trying to fix\u003C\u002Fh2>\u003Cp>The authors start from a familiar LLM failure mode. Post-training improves capability, but it tends to narrow the output distribution. In plain terms, the model becomes more polished and more predictable, which can hurt tasks that need creative variation. The abstract calls out both explicit creativity tasks, like story generation, and implicit ones, like \u003Ca href=\"\u002Ftag\u002Freinforcement-learning\">reinforcement learning\u003C\u002Fa>.\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1786343571807-rqw1.png\" alt=\"CreativeInstruct teaches LLMs to stay creative\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>That matters for engineers because a model that is too conservative can be harder to use as a generator, a brainstorming partner, or a base for downstream optimization. You may get high-quality answers, but less novelty, less narrative diversity, and less room for sampling-based exploration.\u003C\u002Fp>\u003Cp>CreativeInstruct is designed around that tradeoff. Instead of treating creativity as something that disappears once you optimize for quality, the method tries to teach the model to balance both at once.\u003C\u002Fp>\u003Ch2>How the method works in plain English\u003C\u002Fh2>\u003Cp>The key idea is instruction tuning with a twist. CreativeInstruct learns to inject special \u003Ccode>[StartCreativity]\u003C\u002Fcode> spans into the training data, and those spans bias generation toward creativity. The paper describes this as a scalable way to steer the model toward creative, base-model-like generations while keeping the quality of post-trained models.\u003C\u002Fp>\u003Cp>That is a useful framing for practitioners: rather than maintaining separate creative models, or relying on multiple models at \u003Ca href=\"\u002Ftag\u002Finference\">inference\u003C\u002Fa> time, the method aims to encode the behavior into a single checkpoint. The abstract explicitly says this happens through instruction tuning, which makes it a training-time intervention rather than an inference-time orchestration trick.\u003C\u002Fp>\u003Cp>The paper also introduces a structural diversity metric based on graph edit distance. The point of that metric is to capture narrative-level variation that lexical and semantic metrics miss. In other words, two outputs can look different on the surface, or even semantically similar, while still following the same story structure. The graph-based metric is meant to see that deeper difference.\u003C\u002Fp>\u003Ch2>What the paper actually shows\u003C\u002Fh2>\u003Cp>The strongest claim in the abstract is that on narrative generation, CreativeInstruct matches or exceeds the diversity of both multi-model baselines and distilled variants of their outputs. Importantly, it does this without sacrificing quality and without requiring multiple models at inference time.\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1786343570215-2rm7.png\" alt=\"CreativeInstruct teaches LLMs to stay creative\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>The human evaluation result is also straightforward: annotators rate CreativeInstruct generations as more creative than the post-trained LLMs’ generations in 70.3% of cases. That is the only concrete human-eval number given in the abstract, so there are no broader \u003Ca href=\"\u002Ftag\u002Fbenchmark\">benchmark\u003C\u002Fa> tables to lean on here.\u003C\u002Fp>\u003Cp>The paper also argues that creative models can be useful as a substrate for reinforcement learning. When GRPO is applied to a CreativeInstruct checkpoint, it improves by about 4% on AMC and about 5 percentage points on MATH compared with applying the same training to the post-trained checkpoint. The abstract does not provide more benchmark detail than that, so those are the only numbers we can responsibly report.\u003C\u002Fp>\u003Cp>There is an important pattern here: the paper is not just claiming that creative outputs look nicer. It is claiming that preserving creativity can help downstream optimization too. That makes the work relevant beyond generative writing tasks, especially if your stack uses RL-style post-training or iterative policy improvement.\u003C\u002Fp>\u003Ch2>Why developers should care\u003C\u002Fh2>\u003Cp>If you work with LLMs in production, this paper is about a practical tension you probably already know: the more you optimize for reliable answers, the more you may flatten the model’s range. CreativeInstruct suggests that you can train for both quality and diversity instead of choosing one.\u003C\u002Fp>\u003Cp>That could matter in a few places. For content generation, it may reduce repetitive outputs. For agentic systems, it may give the model more room to explore alternative plans. For RL pipelines, the paper’s result suggests that a more creative checkpoint can be a better starting point than a heavily post-trained one.\u003C\u002Fp>\u003Cul>\u003Cli>It offers a training-time method, not a multi-model inference setup.\u003C\u002Fli>\u003Cli>It introduces a structural diversity metric that looks beyond token-level variation.\u003C\u002Fli>\u003Cli>It suggests creativity can improve downstream RL performance, not just generation style.\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>Limitations and open questions\u003C\u002Fh2>\u003Cp>The abstract is promising, but it is also selective. It gives one human-evaluation number and two downstream RL improvements, but it does not provide a broader benchmark suite in the notes we have here. It also does not spell out the full training recipe, dataset scale, or compute cost in the abstract.\u003C\u002Fp>\u003Cp>That means engineers should treat the method as an interesting training strategy, not a drop-in universal fix. The key unanswered questions are how robust the approach is across domains, how sensitive it is to the choice of creative spans, and how much extra tuning effort it needs in practice.\u003C\u002Fp>\u003Cp>Even so, the direction is clear. CreativeInstruct is trying to make “creative but useful” a first-class model property, rather than a side effect you lose after alignment. For teams building generation systems, that is a meaningful design goal.\u003C\u002Fp>\u003Cp>In short, the paper argues that you do not have to trade away diversity to get a stronger LLM. It proposes a concrete way to preserve it, measures it with a more structure-aware metric, and shows that the resulting checkpoint can even help RL training downstream.\u003C\u002Fp>","CreativeInstruct trains LLMs to keep quality while preserving creativity and diversity.","arxiv.org","https:\u002F\u002Farxiv.org\u002Fabs\u002F2608.07460",null,"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1786343571807-rqw1.png","research","en","2af77412-f711-4abb-915d-5b7d1b5275a7",[17,18,19,20,21],"LLMs","instruction tuning","creativity","diversity","reinforcement learning",[23,24,25],"CreativeInstruct uses [StartCreativity] spans to steer generation toward creativity.","It matches or exceeds diversity baselines on narrative generation without multi-model inference.","Creative checkpoints improve downstream GRPO results on AMC and MATH.",1,"2026-08-10T06:32:28.086694+00:00","2026-08-10T06:32:28.082+00:00",{"tags":30,"relatedLang":35,"relatedPosts":39},[31,33],{"name":17,"slug":32},"llms",{"name":21,"slug":34},"reinforcement-learning",{"id":15,"slug":36,"title":37,"language":38},"creativeinstruct-llms-quality-creativity-diversity-zh","CreativeInstruct 讓 LLM 保留創意","zh",[40,46,52,58,64,70],{"id":41,"slug":42,"title":43,"cover_image":44,"image_url":44,"created_at":45,"category":13},"1f0b474d-49e9-4ce1-a3bb-a6b6561ed107","coinrag-fine-grained-kv-cache-reuse-rag-en","CoinRAG Reuses Fine-Grained KV Caches for RAG","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1786345380961-irlf.png","2026-08-10T07:02:34.138838+00:00",{"id":47,"slug":48,"title":49,"cover_image":50,"image_url":50,"created_at":51,"category":13},"30ce677e-59ab-4f83-9759-da93aa8bb4af","mirrorworld-mirror-reflection-video-diffusion-en","MirrorWorld makes mirror reflections consistent in video","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1786341775274-6d5a.png","2026-08-10T06:02:27.25669+00:00",{"id":53,"slug":54,"title":55,"cover_image":56,"image_url":56,"created_at":57,"category":13},"3f886925-6381-4770-980a-1001203cf245","claude-4-5-ai-progress-still-accelerating-en","Claude 4.5 proves AI progress is still accelerating","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1786257168594-qg8m.png","2026-08-09T06:32:27.560501+00:00",{"id":59,"slug":60,"title":61,"cover_image":62,"image_url":62,"created_at":63,"category":13},"8675701f-d283-4d69-8818-bdf59e5ae09e","mage-vl-cuts-visual-tokens-by-reading-codecs-en","Mage-VL Cuts Visual Tokens by Reading Codecs","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1786170812146-kqtv.png","2026-08-08T06:33:06.377112+00:00",{"id":65,"slug":66,"title":67,"cover_image":68,"image_url":68,"created_at":69,"category":13},"7368755d-86ca-461e-9d95-d7e74e95b561","astra-turns-long-math-tasks-into-multi-agent-work-en","Astra turns long math tasks into multi-agent work","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1786088006851-29p0.png","2026-08-07T07:32:52.200646+00:00",{"id":71,"slug":72,"title":73,"cover_image":74,"image_url":74,"created_at":75,"category":13},"e69199db-e1f8-4e12-aaf2-ea92eeb2e0cc","evidence-linked-feature-engineering-heart-failure-en","Evidence-linked feature engineering for heart failure","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1786086190016-bykl.png","2026-08-07T07:02:31.382531+00:00",[77,82,87,92,97,102,107,112,117,122],{"id":78,"slug":79,"title":80,"created_at":81},"a2715e72-1fe8-41b3-abb1-d0cf1f710189","ai-predictions-2026-big-changes-en","AI Predictions for 2026: Brace for Big Changes","2026-03-26T01:25:07.788356+00:00",{"id":83,"slug":84,"title":85,"created_at":86},"8404bd7b-4c2f-4109-9ec4-baf29d88af2b","ml-papers-of-the-week-github-research-desk-en","ML Papers of the Week Turns GitHub Into a Research Desk","2026-03-27T01:11:39.480259+00:00",{"id":88,"slug":89,"title":90,"created_at":91},"87897a94-8065-4464-a016-1f23e89e17cc","ai-ml-conferences-to-watch-in-2026-en","AI\u002FML Conferences to Watch in 2026","2026-03-27T01:51:54.184108+00:00",{"id":93,"slug":94,"title":95,"created_at":96},"6f1987cf-25f3-47a4-b3e6-db0997695be8","openclaw-agents-manipulated-self-sabotage-en","OpenClaw Agents Can Be Manipulated Into Failure","2026-03-28T03:03:18.899465+00:00",{"id":98,"slug":99,"title":100,"created_at":101},"a53571ad-735a-4178-9f93-cb09b699d99c","vega-driving-language-instructions-en","Vega: Driving with Natural Language Instructions","2026-03-28T14:54:04.698882+00:00",{"id":103,"slug":104,"title":105,"created_at":106},"a34581d6-f36e-46da-88bb-582fb3e7425c","personalizing-autonomous-driving-styles-en","Drive My Way: Personalizing Autonomous Driving Styles","2026-03-28T14:54:26.148181+00:00",{"id":108,"slug":109,"title":110,"created_at":111},"2bc1ad7f-26ce-4f02-9885-803b35fd229d","training-knowledge-bases-writeback-rag-en","Training Knowledge Bases with WriteBack-RAG","2026-03-28T14:54:45.643433+00:00",{"id":113,"slug":114,"title":115,"created_at":116},"71adc507-3c54-4605-bbe2-c966acd6187e","packforcing-long-video-generation-en","PackForcing: Efficient Long-Video Generation Method","2026-03-28T14:55:02.646943+00:00",{"id":118,"slug":119,"title":120,"created_at":121},"675942ef-b9ec-4c5f-a997-381250b6eacb","pixelsmile-facial-expression-editing-en","PixelSmile Framework Enhances Facial Expression Editing","2026-03-28T14:55:20.633463+00:00",{"id":123,"slug":124,"title":125,"created_at":126},"6954fa2b-8b66-4839-884b-e46f89fa1bc3","adaptive-block-scaled-data-types-en","IF4: Smarter 4-Bit Quantization That Adapts to Your Data","2026-03-31T06:00:36.65963+00:00"]