VLSA
“Notably, AEGIS achieves over 50% improvement in obstacle avoidance rate while substantially increasing the task success rate by nearly 10%.”
The authors, in the abstract
Vision-Language-Action Models with Plug-and-Play Safety Constraint Layer
AI Research Radar
November 29 – December 5: 3 frontier papers out of 861 read.
Frontier
Papers rated as moving the field
How we choosePatronus
“Extensive experiments across 15 PLMs and nine tasks demonstrate that Patronus achieves \geq98.3% backdoor detection recall and reduces attack success rates to clean settings, significantly outperforming all state-of-the-art baselines in all settings.”
First, do NOHARM
“Across 28 LLMs, recommendations carried the potential for severe harm in up to 22.6% of cases, with errors of omission accounting for more than 80% of severe errors.”
The week by field
Nov 29 – Dec 5: 31 cleared, 3 frontier
The week’s papers have not been sorted into fields yet.
Models under test
Papers this week that name the model
Labs publishing
Papers this week with an author at the lab
Also cleared
28 papers
- TokSuite: Measuring the Impact of Tokenizer Choice on Language Model BehaviorGoogle, Hugging Face
- AugServe: Adaptive Request Scheduling for Augmented Large Language Model Inference ServingAlibaba
- LLM-Auction: Generative Auction towards LLM-Native AdvertisingAlibaba
- KernelEvolve: Scaling Agentic Kernel Coding for Heterogeneous AI Accelerators at MetaMeta
- CUDA-L2: Surpassing cuBLAS Performance for Matrix Multiplication through Reinforcement LearningCode
- Epistemic Bias Injection: Manipulating LLM Opinion via Selective Context Retrieval
- Reward Auditor: Inference on Reward Modeling Suitability in Real-World Perturbed Scenarios
- COGNITION: From Evaluation to Defense against Multimodal LLM CAPTCHA Solvers
- Locality-Aware Continual Unlearning for Diffusion Models
- jina-vlm: Small Multilingual Vision Language Model
- When AI Takes the Couch: Psychometric Jailbreaks Reveal Internal Conflict in Frontier Models
- 6 Fingers, 1 Kidney: Natural Adversarial Medical Images Reveal Critical Weaknesses of Vision-Language Models
All 28 papers
- Aligned but Stereotypical? How System Prompts Shape Demographic Bias in LLM-Based Text-to-Image Models
- When AI Agents Compete for Jobs: Strategic Capabilities and Economic Dynamics of AI Labour Markets
- BEAVER: An Efficient Deterministic LLM Verifier
- JEEVHITAA -- An HCAI Ecosystem to Support Collective Care
- Semantic Attacks on Tool-Augmented LLMs: Securing the Model Context Protocol Against Descriptor-Level Manipulation
- ReasonBENCH: Benchmarking the (In)Stability of LLM Reasoning
- When Tables Leak: Attacking String Memorization in LLM-Based Tabular Data Generation
- WarmServe: Enabling One-for-Many GPU Prewarming for Multi-LLM Serving
- RobustSora: De-Watermarked Benchmark for Robust AI-Generated Video Detection
- DualGuard: Dual-stream Large Language Model Watermarking Defense against Paraphrase and Spoofing Attack
- AI Security Beyond Core Domains: Resume Screening as a Case Study of Adversarial Vulnerabilities in Specialized LLM Applications
- Learning-Based Automated Adversarial Red-Teaming for Robustness Evaluation of Large Language Models
- Nightjar: Dynamic Adaptive Speculative Decoding for Large Language Models Serving
- VLA-Arena: An Open-Source Framework for Benchmarking Vision-Language-Action Models
- It's a TRAP! Task-Redirecting Agent Persuasion Benchmark for Web Agents
- SB-TRPO: Towards Safe Reinforcement Learning with Hard Constraints