AI Breakthrough Digest

New Benchmarks

• 🔥 UniClawBench: UniClawBench introduces the first capability-driven benchmark for proactive agents with 400 bilingual real-world...

UniClawBench tackles limitations of prior agent benchmarks by introducing the first capability-driven evaluation for proactive agents in live, dynamic...

Two distinct paths emerge for long-context LLMs:

• Linear attention (DeltaNet variants) replaces quadratic self-attention with recurrent memory,...

A new policy roadmap calls for a learning sciences benchmark to evaluate generative AI on human flourishing, not just task efficiency.

Key elements...

Three approaches are pushing generative models toward more controllable, stable, and interactive outputs.

• ARDY combines autoregressive diffusion...

Flash-BoN accelerates inference-time scaling for text-to-image diffusion models by generating cheap draft candidates via timestep truncation, layer...

New Benchmarks for Evaluation

• 🔥 UniClawBench: Introduces a capability-driven benchmark with 400 bilingual real-world tasks for proactive...

Agent Memory Advances

• 🔥 NapMem: Paper reframes memory as a structured action space with multi-granularity pyramid and memory-tool RL, remaining...

Three projects push large-scale interactive simulation for AI:

• MIRA trains a 5B model on 10k hours of Rocket League data for fully generated 20fps...

• SciReasoner delivers transparent structure-property reasoning across proteins, molecules and crystals, lifting Gene Ontology Fmax from 0.42 to 0.55...

Researchers show rapid AI adoption at 58% but only 22% call it trustworthy, citing needs for source citations, peer-reviewed training data, and...

RoboDojo delivers a unified benchmark spanning 42 simulation tasks and 18 real-world tasks to rigorously test generalist robot manipulation policies...

Two recent works show memory evolving from passive retrieval to a learned, integrated component of agent reasoning.

• LaMem-VLA embeds historical...

Pharo-specialized models created via targeted data curation, continued pre-training, and fine-tuning beat both their base checkpoints and...

A comprehensive thesis shows medical image domain adaptation fails under limited supervision due to intertwined limits in model capacity, semantic...

LingBot-Video overcomes domain mismatch in video generative models—where visual fidelity trumps physical realism—via a Mixture-of-Experts DiT...

Flow Sampling flips the conditioning: instead of using target data samples for supervising velocity, it derives the drift from the source. This...

AI Agent Safety and Vulnerability Research

• 🔥 Safety Testing LLM Agents at Scale: Presents methods for risk discovery to evidence-grounded...

• Vera automates safety testing via literature-driven risk taxonomies, combinatorial case generation, and evidence-based verification in sandboxes
  -...

Three papers show how generative techniques address offline learning, delays, and evaluation in RL.

• Offline adaptation: Goal-conditioned test-time...