IAMJB

IAMJB

@iamjbd.bsky.social

🤗 ML at Hugging Face 🌲 Academic Staff at Stanford University (AIMI Center) 🦴 Radiology AI is my stuff

1.8K followers1.1K following94 posts

IAMJB
@iamjbd.bsky.social

💥 Excited to share our latest work: Structuring Radiology Reports: Challenging LLMs with Lightweight Models In this study, we explore how small, task-specific encoder-decoder models can rival (and sometimes outperform) much larger LLMs; all while being faster, cheaper, and easier to deploy. ons.

View full thread
IAMJB
@iamjbd.bsky.social

Huge thanks to the amazing team at Stanford Center for Artificial Intelligence in Medicine and Imaging (AIMI): Johannes Moll, Louisa Fay, @asfandyar_azhar, @SophieOstmeier, Tim Lueth, Sergios Gatidis, @curtlanglotz

IAMJB
@iamjbd.bsky.social

💥 Excited to share our latest work: Structuring Radiology Reports: Challenging LLMs with Lightweight Models In this study, we explore how small, task-specific encoder-decoder models can rival (and sometimes outperform) much larger LLMs; all while being faster, cheaper, and easier to deploy. ons.

IAMJB
@iamjbd.bsky.social

We benchmark lightweight models (<300M params) against state-of-the-art LLMs (up to 70B params), using human-reviewed test data and clinically grounded evaluation metrics. Our results highlight the strong potential of specialized, efficient models in clinical NLP application.

IAMJB
@iamjbd.bsky.social

💥 We unveil our paper accepted at the #ACL2025 Main Conference: Automated Structured Report Generation Let's revisit automated radiology report generation for CXR. Free-form reports make it hard for AI systems to learn accurate generation, and even harder to evaluate. 🧵👇 @StanfordAIMI @hopprai

View full thread
IAMJB
@iamjbd.bsky.social

4) We conduct a reader study to create a radiologist-validated test set for both the automated structured radiology report task, as well as utterances disease labels from our new ontology. Finally, external evaluation is conducted using out-of-institution data by @hopprai.

IAMJB
@iamjbd.bsky.social

💥 We unveil our paper accepted at the #ACL2025 Main Conference: Automated Structured Report Generation Let's revisit automated radiology report generation for CXR. Free-form reports make it hard for AI systems to learn accurate generation, and even harder to evaluate. 🧵👇 @StanfordAIMI @hopprai

View full thread
IAMJB
@iamjbd.bsky.social

2) Since each reported observation, whether in the findings or impression sections, is expressed as a single utterance (1.5M unique utterances in total), we use a large language model to label each one according to a new ontology comprising 72 critical chest X-ray (CXR) observations.

IAMJB
@iamjbd.bsky.social

3) We fine-tune popular RRG system on this restructured findings and impression, namely: - Chexagent @StanfordAIMI - MAIRA-2 @MSFTResearch - RaDialog @TU_Muenchen - Chexpert-plus @StanfordAIMI As well as a BERT architecture for the disease classification system on our new ontology.

IAMJB
@iamjbd.bsky.social

💥 We unveil our paper accepted at the #ACL2025 Main Conference: Automated Structured Report Generation Let's revisit automated radiology report generation for CXR. Free-form reports make it hard for AI systems to learn accurate generation, and even harder to evaluate. 🧵👇 @StanfordAIMI @hopprai

IAMJB
@iamjbd.bsky.social

1) We leverage LLM to restructure MIMIC-CXR and Chexpert-plus (180K Findings sections and 400K Impression sections) into reports categorized by organ system, under strict rules.

IAMJB
@iamjbd.bsky.social

Just noticed our lightweight RRG model has been downloaded over 92,000 times this months on 🤗HuggingFace. This model was included in the CheXpert-Plus release and contains just 67 million parameters: huggingface.co/IAMJB/chexpe... Its also a top ranking model on RexRank (rexrank.ai)

huggingface.coIAMJB/chexpert-mimic-cxr-impression-baseline · Hugging FaceWe’re on a journey to advance and democratize artificial intelligence through open source and open science.
IAMJB
@iamjbd.bsky.social

🧵 What if AI could learn from millions of unlabeled radiology images and reports—and then flexibly adapt to new clinical tasks? In a new comprehensive review in @radiology_rsna, colleagues at stanford dive into how foundation models (FMs) are set to revolutionize radiology!

IAMJB
@iamjbd.bsky.social

"Second, we develop budget forcing to control test-time compute by forcefully terminating the model's thinking process or lengthening it by appending "Wait" multiple times to the model's generation when it tries to end." What a trick...

IAMJB
@iamjbd.bsky.social

Is this the last benchmark before AGI? Humanity's Last Exam (HLE) 🤯 3,000 expert-level questions across 100+ subjects, created by nearly 1,000 subject matter experts globally.

IAMJB
@iamjbd.bsky.social

𝗔 𝗦𝗶𝗺𝗽𝗹𝗲 𝗚𝘂𝗶𝗱𝗲 𝘁𝗼 𝗔𝗜 𝗔𝗴𝗲𝗻𝘁 𝗠𝗲𝗺𝗼𝗿𝘆 🌟 An agent's memory helps it plan and react by leveraging past interactions or external data via prompt context. Here’s a breakdown: 𝟭. Episodic Memory: Logs past actions/interactions (e.g., stored in a vector database for semantic search).

IAMJB
@iamjbd.bsky.social

𝟮. Semantic Memory: External/grounding knowledge or self-knowledge, similar to RAG context. 𝟯. Procedural Memory: System setup details like prompts, tools, and guardrails (stored in Git/registries). 𝟰. Task Memory: Info retrieved from long-term storage for immediate tasks.

IAMJB
@iamjbd.bsky.social

𝟱. Working Memory: Compiles long-term and task memory to create the final prompt for the LLM. Typically, 1–3 = Long-Term Memory; 5 = Short-Term Memory. Thoughts on agent memory?👇

IAMJB
@iamjbd.bsky.social

🧩 The future of creativity is elemental. ✨ Kling AI just announced Elements 🌎 First, world building: Craft your characters, environments, props. Plan your motion and VFX. 🎛️ Then, remixing: Bring it all together into a cohesive story.

IAMJB
@iamjbd.bsky.social

Amazing. Agent Roles: ⛳ PhD Agent: Conducts literature reviews, interprets results, writes reports. ⛳ Postdoc Agent: Plans research, designs experiments. ⛳ ML Engineer Agent: Prepares data, writes, optimizes code. ⛳ Professor Agent: Oversees, refines reports.

IAMJB
@iamjbd.bsky.social

Neat: Representing Long Volumetric Video with Temporal Gaussian Hierarchy Contrib: Temporal Gaussian Hierarchy representation for long volumetric video.

IAMJB
@iamjbd.bsky.social

Neat. Converts images, PDFs, and Office documents to Markdown or JSON using OCR and LLM models, with features for caching, distributed processing, and PII removal

IAMJB
@iamjbd.bsky.social

🚀 PRIME + Eurus-2 beat Qwen2.5-Math-Instruct with 1/10 the data! ✨ Implicit PRM (no labels) 🔄 Online updates, zero overhead 🎯 Token-level rewards + RLOO Scaling up with 3x more data!

IAMJB
@iamjbd.bsky.social

2 OLMo 2 Furious captures every lesson learned since OLMo 1, featuring in-depth explorations of: • Stable pretraining • LR annealing, data curricula, and soups • Tulu post-training • Compute infrastructure