1/2
@rohanpaul_ai
This survey maps how prompts turn generic AI models into reliable medical imaging helpers.
It finds that small prompt tweaks often beat full retraining for generation, segmentation, and diagnosis.
LLM based vision tools struggle when patient data shift across scanners and hospitals.
Prompting inserts extra context so the same backbone stays useful.
Text prompts supply disease terms or whole reports, box or point prompts mark regions, and learnable vectors quietly store task hints.
With those inputs models hit up to 20% better accuracy while touching under 1% of their weights.
Prompt brittleness, missing clinical labels, and slow hospital hardware still block everyday use, so automated and lighter prompts are the next focus.
----
Paper – arxiv. org/abs/2507.01055
Paper Title: "Prompt Mechanisms in Medical Imaging: A Comprehensive Survey"
2/2
@PaperWizardAI
Lots of alpha in context management.
To post tweets in this format, more info here: https://www.thecoli.com/threads/tips-and-tricks-for-posting-the-coli-megathread.984734/post-52211196
@rohanpaul_ai
This survey maps how prompts turn generic AI models into reliable medical imaging helpers.
It finds that small prompt tweaks often beat full retraining for generation, segmentation, and diagnosis.
LLM based vision tools struggle when patient data shift across scanners and hospitals.
Prompting inserts extra context so the same backbone stays useful.
Text prompts supply disease terms or whole reports, box or point prompts mark regions, and learnable vectors quietly store task hints.
With those inputs models hit up to 20% better accuracy while touching under 1% of their weights.
Prompt brittleness, missing clinical labels, and slow hospital hardware still block everyday use, so automated and lighter prompts are the next focus.
----
Paper – arxiv. org/abs/2507.01055
Paper Title: "Prompt Mechanisms in Medical Imaging: A Comprehensive Survey"
2/2
@PaperWizardAI
Lots of alpha in context management.
To post tweets in this format, more info here: https://www.thecoli.com/threads/tips-and-tricks-for-posting-the-coli-megathread.984734/post-52211196
Mistral launches Devstral Small 1.1 (53.6% SWE-Bench) and Devstral Medium (61.6%), tuned for coding agents.
Token prices sit between $0.1 and $2 per 1M, with Medium roughly 75% below GPT4-class rates. 
/huggingface teams with Pollen Robotics to launch Reachy Mini, a 28 cm desktop robot priced at $299 and ready for vision, speech, and text models.
Builders can instantly run 15+ preset behaviors or upload new ones through the community hub.
Read on 
Why memory got messy
What a MemCube holds
Three layers doing the heavy lifting
Scheduler keeps memories fresh
Numbers that prove the point
KV tricks to cut wait time
Google Research release MedGemma 27B, multimodal health-AI models that run on 1 GPU 
Checks more than 200 URLs for each task.
Context-aware, very long-horizon reasoning
Runs on an internal Kimi k-series backbone 
Learns entirely through end-to-end agentic RL 
Achieves 26.9% pass@1 on Humanity’s Last Exam, top of the board 
Scores 69% pass@1 on xbench-DeepSearch, edging past o3 with tools 
Delivers solid results on FRAMES, Seal-0, and SimpleQA 
Shows that self-rewarded training can mature planning, search, and coding in one loop
High‑quality agent datasets are rare, so the team generated their own. 
Training stays on‑policy. Tool‑call format guards are switched off, so every trajectory truly reflects model probabilities. 
After RL, the model averages 23 reasoning steps and checks about 200 URLs per task, reaches 69% pass@1 on xbench‑DeepSearch, and shows habits like cross‑verifying conflicting sources before answering.
Blog: 
The quick idea
