Week 14: Autonomous Science, Enterprise LLMs, and Medical Simulations

OpenAI introduces PaperBench to evaluate AI agents' ability to replicate cutting-edge machine learning research papers from scratch.

• Tests paper reproduction across 20 ICML 2024 papers with ~8,316 fine-grained tasks
• Uses LLM-based judges with high human agreement (F1 = 0.83) for scalable evaluation
• Best model (Claude 3.5 Sonnet) achieved only 21.0% score vs. human 41.4%

Cohere's 111B parameter open-weights model features modular expert merging and hybrid architecture for enterprise applications.

• Interleaves sliding window and full attention for efficient 256k context support
• Outperforms GPT-4o and Claude 3.5 on RAG, tool use, and enterprise tasks
• Excels in 23 languages with 94.2% pass rate on real-world generative tasks

CodeScientist automatically generates and tests scientific hypotheses via code-based experimentation with minimal human input.

• Five-step pipeline: Ideation → Planning → Code Execution → Reporting → Meta-Analysis
• Produced 6 scientifically sound and novel findings from 50 AI research papers
• Discovered that LLM self-confidence often mismatches actual accuracy

RARE shifts LLM training from memorizing knowledge to applying and evaluating it, separating domain knowledge from domain thinking.

• Uses retrieved knowledge during training to teach reasoning patterns
• Small RARE models outperform GPT-4+RAG on medical benchmarks
• Achieves up to +20% accuracy boosts on complex medical QA tasks

This paper explains why LLMs focus attention on the first token, showing it prevents representational collapse in deep Transformers.

• Attention sinks act as no-ops that preserve representation diversity
• Larger models (LLaMA 3.1 405B) show stronger sink behavior
• Sinks emerge naturally due to position, not the token itself

MedAgentSim creates a fully automated, open-source hospital simulation where LLM-powered agents simulate doctor-patient interactions.

• Features multi-turn consultations with lab requests in a 2D game environment
• Uses memory buffers and kNN retrieval for self-improvement
• Improves performance by 6-37% across medical benchmarks

ODS provides an open-source search AI framework that rivals proprietary systems like GPT-4o Search Preview and Perplexity Sonar.

• Combines search tools with reasoning agents in two variants: ReAct and CodeAct
• Outperforms GPT-4o Search Preview by +9.7% on FRAMES benchmark
• Adapts search queries dynamically for optimal balance of cost and accuracy

Z1 improves LLM compute efficiency by training models with varying reasoning trajectories and dynamically adjusting depth during inference.

• 107K-sample dataset with short and long reasoning paths for coding problems
• Matches larger models while using ~30% of the reasoning tokens
• Code reasoning generalizes well to science and math tasks

This survey analyzes how to balance deep reasoning performance with computational cost in large language models.

• Reviews inefficiencies and behavioral patterns in LLM reasoning
• Explores solutions at both post-training and inference stages
• Provides framework for measuring reasoning economy

This study introduces a framework to measure hidden knowledge in LLMs, revealing significant gaps between internal encoding and expressed outputs.

• Models encode up to 40% more factual information than they express
• Some answers, though known internally, are never generated
• Highlights limitations in test-time sampling for QA tasks