Esta semana compramos muita pipoca na MultiCortex, pois foi um absoluto cinema de papers no setor de IA. Acreditamos que foi a semana mais insana de 2024 marcada por avanços significativos no campo da inteligência artificial, com a publicação de diversos artigos que abordam desde melhorias em arquiteturas de modelos até questões de segurança e eficiência. A seguir, a lista desses trabalhos:

• Byte Latent Transformer
• Training Large Language Models to Reason in a Continuous Latent Space
• Language Modeling in a Sentence Representation Space
• Phi-4 Technical Report – Best-of-N Jailbreaking
• Forking Paths in Neural Text Generation
• Refusal Tokens – [MASK] is All You Need
• Explore Theory-of-Mind
• Obfuscated Activations Bypass LLM Latent-Space Defenses
• The Pitfalls of Memorization
• How to Merge Your Multimodal Models Over Time?
• Machine Unlearning Doesn’t Do What You Think
• Understanding Gradient Descent through the Training Jacobian
• An Evolved Universal Transformer Memory
• Transformers Struggle to Learn to Search
• Transformers Can Navigate Mazes With Multi-Step Prediction
• Frontier Models are Capable of In-context Scheming
• Mixture of Monosemantic Experts for Transformers
• Learning to Merge LoRAs via Hypernetworks for Subject-Style Conditioned Image Generation
• Scalable Text and Image Conditioned Video Generation
• Hidden in the Noise: Two-Stage Robust Watermarking for Images
• Learned Compression for Compressed Learning
• Learning Flow Fields in Attention for Controllable Person Image Generation
• ProcessBench: Identifying Process Errors in Mathematical Reasoning
• Unraveling the Complexity of Memory in RL Agents
• Expanding Performance Boundaries of Open-Source Multimodal Models with Model, Data, and Test-Time Scaling
• APOLLO: SGD-like Memory, AdamW-level Performance
• Neural LightRig