The Alignment Auditor
A Bayesian Framework for Verifying and Refining LLM Objectives
AI Research Radar
Wednesday, October 1: 7 frontier papers out of 1,042 read.
Frontier
Papers rated as moving the field
How we chooseScaling Latent Reasoning via Looped Language Models
“We hope our results show the potential of LoopLM as a novel scaling direction in the reasoning era.”
How to Train Your Advisor
“Advisor Models improve GPT-5.2's performance on RuleArena (Taxes) by 27.4%, reduce Gemini 3 Pro's steps taken in SWE agent tasks by 24.6%, and outperform static prompt optimizers in personalizing GPT-5 to user preferences (85-100% vs. 40-60%).”
MetaBreak
“However, when there is content moderation, MetaBreak outperforms SOTA solutions PAP and GPTFuzzer by 11.6% and 34.8%, respectively.”
UltraCUA
“Cross-domain validation on WindowsAgentArena demonstrates robust generalization with 21.7% success rate, surpassing Windows-trained baselines.”
StatEval
“Experiments show that while LLMs perform reasonably on foundational tasks, they struggle with rigorous research-level reasoning.”
Bypassing Prompt Guards in Production with Controlled-Release Prompting
“We find our attack to be successful on four major chat platforms (Google Gemini, DeepSeek Chat, xAI Grok, and Mistral Le Chat) where baseline methods fail.”
This week in research
Sep 25 – Oct 1: 41 cleared, 7 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
34 papers
- Breaking the Code: Security Assessment of AI Code Agents Through Systematic Jailbreaking AttacksAmazon, Meta
- InvThink: Premortem Reasoning for Safer Language ModelsGoogle
- A global log for medical AIGoogle
- Teaching LLM to be Persuasive: Reward-Enhanced Policy Optimization for Alignment from Heterogeneous RewardsAlibaba
- When Personalization Tricks Detectors: The Feature-Inversion Trap in Machine-Generated Text DetectionByteDance
- Beyond Reasoning Gains: Mitigating General-Capability Forgetting in Large Reasoning ModelsMeta
- SoK: Systematizing LLM Prompt Security: Taxonomies, Datasets, and Unified Evaluation of Attacks and DefensesCode
- VERA-V: Variational Inference Framework for Jailbreaking Vision-Language ModelsCode
- TetraJet-v2: Accurate NVFP4 Training for Large Language Models with Oscillation Suppression and Outlier ControlCode
- DecepChain: Inducing Deceptive Reasoning in Large Language Models
- Make a Video Call with LLM: A Measurement Campaign over Six Mainstream Apps
- Eyes-on-Me: Scalable RAG Poisoning through Transferable Attention-Steering Attractors
All 34 papers
- Where Do Backdoors Live? A Component-Level Analysis of Backdoor Propagation in Speech Language Models
- LapSurgie: Humanoid Robots Performing Surgery via Teleoperated Handheld Laparoscopy
- Speculative Actions: A Lossless Framework for Faster Agentic Systems
- Membership Inference Attacks on Tokenizers of Large Language Models
- Humanoid Everyday: A Comprehensive Robotic Dataset for Open-World Humanoid Manipulation
- Breaking the Self-Confirming Loop: Diagnosing and Mitigating Systemic Reward Bias in Self-Rewarding RL
- GREAT: Generalizable Backdoor Attacks in RLHF via Emotion-Aware Trigger Synthesis
- Safety Game: Inference-Time Alignment of Black-Box LLMs via Constrained Optimization
- Auditing medical multi-agent AI reveals risks of false consensus
- RAG-Pull: Turning Retrieval into a Code-Injection Channel via Invisible Unicode Perturbations
- Characterizing Web Search in The Age of Generative AI
- Generative AI and Sales Productivity: Field Experiments in Online Retail
- Scheming Ability in LLM-to-LLM Strategic Interactions
- Attention Misses Visual Risk: Risk-Adaptive Steering for Multimodal Safety Alignment
- Echoes of Human Malice in Agents: Benchmarking LLMs for Multi-Turn Online Harassment Attacks
- A First Look at the Security Issues in the Model Context Protocol Ecosystem
- UNDREAM: Bridging Differentiable Rendering and Photorealistic Simulation for End-to-end Adversarial Attacks
- BadScientist: Can a Research Agent Write Convincing but Unsound Papers that Fool LLM Reviewers?
- Generating the Modal Worker: A Cross-Model Audit of Race and Gender in LLM-Generated Personas Across 41 Occupations
- Sentra-Guard: A Real-Time Multilingual Defense Against Adversarial LLM Prompts
- When Compression Becomes an Attack Surface: Black-Box Attacks on Prompt-Compressed LLM Agents
- Model Inversion meets Cryptographic Fuzzy Extractors