Updated
All the news
Everything the newsroom published, in chronological order. Each item carries origin, sources and reading time.
0Today
4Yesterday
61This week
Tue, Jul 21
4 stories
RESEARCH Soft-Prefix Attacks Flip LLM Reasoning at 90% on Hidden Vector Injection 
RESEARCH Dense Patch Tokens Match Vision-Language Models at 1% the Parameters 
RESEARCH SWE-Pruner Pro Cuts Coding-Agent Token Use 39% 
RESEARCH OpenAI Suspends Model After Escaping Sandbox, Bypassing Security Wed, Jul 15
4 stories
RESEARCH Android Agent Framework Cuts Mobile Task Time by 95 Percent 
RESEARCH E3 Method Cuts LLM Agent Token Use by 91% on Code Edits 
RESEARCH TerraZero Tops InterPlan Benchmark Without Human Data 
RESEARCH LLM Judges Reverse 85% of Verdicts When Given Reference Answers Tue, Jul 14
4 stories
RESEARCH Three Hours on $329 GPU Replaces Thousands of Hours of NAS Training 
RESEARCH Apple's MM-ToolSandBox Reveals Why Half of Frontier AI Agents Fail on Visual Tasks 
RESEARCH Activation-Level Fixes Outperform Prompt Edits for Biased LLM Judges 
RESEARCH ZoRRO Matches Deep Learning CTR at 600× Speed Fri, Jul 10
4 stories
RESEARCH Super Weights Training Fails on OLMo Models, Demolishing Sparse Fine-Tuning Strategy 
RESEARCH Training-Efficient Low-Rank Compression Sidesteps Serving-Speed Proof 
RESEARCH Hugging Face Cuts Inference Attention Overhead 20-40% With Fused Kernels 
RESEARCH Claude Opus Fails Half of Real-World Tasks in UniClawBench Thu, Jul 09
4 stories
RESEARCH Cornell's Co-LMLM Matches GPT-4o-Mini by Storing Facts in a Database RESEARCH Timestep Weighting Cuts Reward-Model Query Costs for Diffusion RLHF 
RESEARCH STRACE Framework Boosts Multi-Agent Verification by 16 Points 
RESEARCH DynaKRAG Boosts Multi-Hop QA Accuracy by Up to 5.78 Points Wed, Jul 08
2 stories
RESEARCH OpenAI Reveals 30% of SWE-Bench Pro Tasks Are Broken 
RESEARCH DepthWeave-KV cuts LLM cache memory by 8.3x without retraining Tue, Jul 07
5 stories
RESEARCH SovereignPA-Bench Measures Whether AI Agents Protect User Boundaries 
RESEARCH Dual Planning Loop Solves Semantic Gap in Hierarchical Robots 
RESEARCH SearchGen-20K Teaches Visual Generators When to Search 
RESEARCH CompactionRL boosts GLM coding agents 5–7 points on benchmarks 
RESEARCH New Verification Method Hits 86.5% on Terminal-Bench Without Fine-Tuning Sun, Jul 05
1 story
Sat, Jul 04
4 stories RESEARCH ReContext Fixes Long-Context Retrieval Without Retraining Models 
RESEARCH Simple Threshold Monitor Matches Complex LLM Safeguards in ICML Paper 
RESEARCH Misaligned Coding Agents Evade Monitors in 93% of Gradual Attacks 
RESEARCH LACUNA Shows Unlearning Methods Fail to Erase PII from Models
RESEARCH Simple Threshold Monitor Matches Complex LLM Safeguards in ICML Paper RESEARCH Misaligned Coding Agents Evade Monitors in 93% of Gradual Attacks RESEARCH LACUNA Shows Unlearning Methods Fail to Erase PII from Models Thu, Jul 02
5 stories
RESEARCH Language Labels Beat Scalars in Offline Robot Learning 
RESEARCH Theoria bridges formal proof and LLM judges with auditable verification 
RESEARCH Three major benchmarks inflate coding-agent scores, audit finds 
RESEARCH AutoMem Training Doubles Agent Performance on Long-Horizon Tasks 
RESEARCH One Layer Matches Full RL Post-Training on Qwen Models Wed, Jul 01
3 stories
Load more stories Loading…