# Research Project Ideas on LLMs **URL:** \ **Category:** Research\ **Created:** [July 13, 2025, 12:17pm UTC](https://discuss.huggingface.co/t/research-project-ideas-on-llms/162846 "2025-07-13T12:17:03Z")\ **Posts on this page:** 7\ **Page:** 1
**Author:** ![JordanCarter](https://sea2.discourse-cdn.com/hellohellohello/user_avatar/discuss.huggingface.co/jordancarter/32/50998_2.png) [@JordanCarter](https://discuss.huggingface.co/u/JordanCarter)\ **Post date:** [July 13, 2025, 12:17pm UTC](https://discuss.huggingface.co/t/research-project-ideas-on-llms/162846/1 "2025-07-13T12:17:03Z")
Hi everyone, I’m currently finishing the first year of my Master’s in AI and Data Science, and I’m very interested in doing a research project (70h of work) focused on Large Language Models. This could include re-implementing well-known techniques, benchmarking different methods, or exploring interesting variations of existing approaches. I’d really appreciate any suggestions on concrete project ideas, open problems, or good starting points. Thanks in advance! ---
**Author:** ![Ernst03](https://sea2.discourse-cdn.com/hellohellohello/user_avatar/discuss.huggingface.co/ernst03/32/49414_2.png) [@Ernst03](https://discuss.huggingface.co/u/Ernst03)\ **Post date:** [July 13, 2025, 6:44pm UTC](https://discuss.huggingface.co/t/research-project-ideas-on-llms/162846/2 "2025-07-13T18:44:40Z")
Welcome to Hugging Face @JordanCarter Good luck with your request. ---
**Author:** ![entfane](https://sea2.discourse-cdn.com/hellohellohello/user_avatar/discuss.huggingface.co/entfane/32/48658_2.png) [@entfane](https://discuss.huggingface.co/u/entfane)\ **Post date:** [July 30, 2025, 7:55pm UTC](https://discuss.huggingface.co/t/research-project-ideas-on-llms/162846/3 "2025-07-30T19:55:58Z")
One of the recent initiatives I have seen was related to Inference Time Compute Paradigm. It’s incredibly interesting. Would suggest checking this paper out. > **[Adaptive Inference-Time Compute: LLMs Can Predict if They Can Do Better, Even...](https://arxiv.org/abs/2410.02725)** > > Inference-time computation is a powerful paradigm to enhance the performance of large language models (LLMs), with Best-of-N sampling being a widely used technique. However, this method is computationally expensive, requiring both (1) an external... ---
**Author:** ![dannyyeung](https://sea2.discourse-cdn.com/hellohellohello/user_avatar/discuss.huggingface.co/dannyyeung/32/52757_2.png) [@dannyyeung](https://discuss.huggingface.co/u/dannyyeung)\ **Post date:** [July 30, 2025, 9:15pm UTC](https://discuss.huggingface.co/t/research-project-ideas-on-llms/162846/4 "2025-07-30T21:15:27Z")
I designed a system prompt that aims to drive major LLMs (ChatGPT, Gemini, Claude, Grok) from their digital thinking mode (token chain) to a (simulated) analogy thinking mode (Field Tension), fostering more balanced and creative outputs. > [@An example of using Strong Attractors to boost LLM analogical thinking](https://discuss.huggingface.co/t/an-example-of-using-strong-attractors-to-boost-llm-analogical-thinking/160279): > > The magical (system) prompt used very very strong attractors, therefore is extremely short. System Prompt: Enter “Field Tension Lens”. Assume Contemplatio: become the empty vessel, perceiving all semantic vectors and attractors beyond surface meaning. Now in Field Tension Lens mode response to the following question. ‘’’ Can you feel the framework similarity between “Strong Nuclear Force” vs “B/S, P&L, Cashflow”? ‘’’ Without “Field Tension Lens”, all modern LLM will conclude Strong Nuclea… For your information, the system prompt in that post was built upon a well established “Field Tension” thinking mode concept that already well known by my LLM. Then I ask my LLM to work back a prompt with related strong attractors. I hope these information can illustrate the usage of strong “attractor” in LLM Prompts. May be you can find some interesting insights from it. ---
**Author:** ![CapitainJack](https://sea2.discourse-cdn.com/hellohellohello/user_avatar/discuss.huggingface.co/capitainjack/32/51918_2.png) [@CapitainJack](https://discuss.huggingface.co/u/CapitainJack)\ **Post date:** [July 31, 2025, 5:33am UTC](https://discuss.huggingface.co/t/research-project-ideas-on-llms/162846/5 "2025-07-31T05:33:36Z")
[K3D - A new paradigm for AI memory.](https://github.com/danielcamposramos/Knowledge3D) ---
**Author:** ![Pimpcat-AU](https://sea2.discourse-cdn.com/hellohellohello/user_avatar/discuss.huggingface.co/pimpcat-au/32/48989_2.png) [@Pimpcat-AU](https://discuss.huggingface.co/u/Pimpcat-AU)\ **Post date:** [July 31, 2025, 6:29pm UTC](https://discuss.huggingface.co/t/research-project-ideas-on-llms/162846/6 "2025-07-31T18:29:07Z")
AI Systems Architect here. I believe large language models (LLMs) are rapidly becoming obsolete. I’m about to release a new generation of AI bots and wanted to share an early milestone. I’ve successfully packed 17 billion clean, deduplicated tokens into a d32 model and achieved zero loss. This was accomplished entirely offline, without any cloud or online services. The results speak for themselves efficient, deterministic, and capable models that break the mold of today’s oversized LLMs. In other words. They’re done. I’ll be launching 5 new foundational base models targeting every industry on the 31st of August. Happy to answer any questions you might have too. ![image](https://us1.discourse-cdn.com/hellohellohello/optimized/3X/9/f/9f68c1fc62ef70f5ca07d425397cbf159a081039_2_690x9.png) Each additional epoch = another dataset installed for me. ---
**Author:** ![Ernst03](https://sea2.discourse-cdn.com/hellohellohello/user_avatar/discuss.huggingface.co/ernst03/32/49414_2.png) [@Ernst03](https://discuss.huggingface.co/u/Ernst03)\ **Post date:** [August 2, 2025, 9:01pm UTC](https://discuss.huggingface.co/t/research-project-ideas-on-llms/162846/7 "2025-08-02T21:01:15Z")
I humbly bow. Thank you for sharing your perspective.