AI Research Daily

Emerging papers highlight memory-enhanced, controllable video generation:

• CT-1 uses vision-language-camera models to transfer spatial reasoning for...

Trend alert: Engineering tweaks unlock massive efficiency in generative AI.

• TRL's on-policy distillation now handles 100B+ teachers, 40x faster than...

FORGE is a fine-grained multimodal evaluation benchmark designed specifically for manufacturing scenarios. It highlights key gaps in current multimodal AI for industrial use.

Rethinking neural efficiency:

• TD-MCL mimics brain synaptic pruning: grows connections for new tasks, then prunes weak ones to fight catastrophic...

Phantom-0 benchmark exposes multimodal AIs hallucinating detailed descriptions—like license plates or medical conditions—without any images...

Generative Model Advances

• 🔥 Flow Map Language Models Update: Big update to the flow map language models paper introduces a new class of...

DRAM PIM hits power delivery walls, threatening efficiency gains from reduced data movement:

• Taxonomy: PIM currents classified by temporal (burst...

Flow map language models receive a 🤯 major update, introducing a new class of continuous flow-based models positioned as the future of non-autoregressive text generation. Dive into the blog and paper for details.

• LLMs replicate human decisions with 83.9% agreement
• Human errors high: 15.2% wrong removals, 13% wrong approvals (mostly 84.3% rule mis-cites)
  -...

Core problem: AI coding assistants reset to zero each session, forgetting prefs like Streamlit use or port 8505. Users repeat...

Meta AI and KAUST propose Neural Computers (NCs) as learned runtimes integrating computation and memory. What if the model became the computer itself, not just used it?

No-escape theorem confirms the intuition: filesystems outperform semantic retrieval like RAG, knowledge graphs, embeddings, and parametric memory for agent long-term memory.

DISCO, a steerable, multimodal protein diffusion model, generates enzymes for new-to-nature chemistry—pushing AI-driven protein design into groundbreaking applications.

Explainability powers adversarial defense: New framework extracts DeepSHAP decision logic to summarize common critical neurons, distinguishing normal...

PIArena fills a critical void in prompt injection evaluation, enabling reliable comparison of defenses across diverse attacks and benchmarks.

• Past...

Agent Benchmarks

• ClawBench: ClawBench evaluates whether AI agents can complete everyday online tasks.
• 🔥 CompBioBench: CompBioBench is a...

Handy hack for daily arXiv updates: Swap arxiv.org → arxivtldr.org or paste any paper URL/ID.

• Today's spotlight (Computer Vision): ETCH-X robustly...

HY-Embodied-0.5 debuts as a new embodied foundation model tailored for real-world agents, pushing advancements in deployable AI systems.

Surprising replication result: UK AISI Model Transparency team replicated Anthropic's steering approach to suppress evaluation awareness in models....

AI evals surge with specialized tests exposing limits:

• KellyBench pushes long-horizon agent skills beyond SWE tasks, adding baselines and stats for...