Singularity-Bench
Opus 5.5 with Claude Code reaches a Singularity-Index of 1.25, matching or beating human baselines on all six tasks, while remaining agents average 0.58.
Can Agents Build Agents?
AI Research Radar
Friday, November 13: 6 frontier papers out of 595 read.
Frontier
Papers rated as moving the field
How we chooseLazare, veni foras
Editing seven neurons turns Nemotron-Terminal-8B into a sleeper agent that fires the payload on at least 98.8% of triggered test trajectories.
Bait
Bait mitigates harmful agentic reinforcement learning by probing the reward function while still allowing learning from benign rewards.
Benchmarking Multilingual Document Understanding in Vision-Language Models
The best average score on the new multilingual MDU benchmark is 69.0, with only eight of 45 evaluated systems reaching 65.
ManiLadder
ManiLadder benchmark with 112 tasks across 4 difficulty levels calibrates manipulation task difficulty using a unified imitation learning policy.
Reins
Reins lets GPT-6.1 Sol complete 88% of real dual-arm manipulation episodes on five tasks with no task-specific code.
This week in research
Nov 7–13: 281 cleared, 17 frontier
- Security98◆ 4
- Safety37◆ 3
- Agents34◆ 4
- InferenceRising28
- Society21
- Robotics19◆ 4
- 4 moreScience, Reasoning, ML, Training19◆ 1
- Language13
- Vision12◆ 1
- Security98◆ 4
- 8 moreSociety, Robotics, Language, Vision, Science, Reasoning, ML, Training84◆ 6
- Safety37◆ 3
- Agents34◆ 4
- InferenceRising28
Size: cleared papers◆ and tint: frontier papersRising beyond normal weekly swingInference rising: 4.2% → 10.0% of papers
Models under test
Papers this week that name the model
Labs publishing
Papers this week with an author at the lab
Also cleared
68 papers, by field
Security & privacy28
- Intent-Aware Protection for AI AgentsMicrosoft
- guardscope: Measuring Which Actions Runtime Controls Check in LLM Agent SystemsCode
- OverServe: Resource Amplification Attacks on SaaS Agents via Tool--Prompt Co-Design
- When Safety Becomes an Attack Surface: Manipulating Abstention in Tool-Using AI Agents
- Speech Tokenizers are Vulnerable: Transferable Semantic Attack and Robust Tokenizer
All 28 papers in Security & privacy
- Reality Perception Gap: Are AI-Manipulated Video Detectors Ready for the Real World?
- Poison-then-Hide: Finetuning-Activated Backdoor Attack on Pretrained Vision Encoders
- Hijacking a Trusted LLM in Real-World Quantized Deployments Only Requires a Disguised Adapter
- VRMaaS: Verifiable Reward-Model-as-a-Service for Trustworthy LLM Alignment
- Protecting Sensitive Knowledge Under LLM Unlearning: The Proxy Threat and a Superposition-Aware Defense
- One Fabrication Is Enough: A Fake News Article Changes the Forecasts of LLM Forecasting Systems
- SecDVA: A Maliciously Secure Multi-Party Framework for Data Value Assessment in Modern Machine Learning Data Markets
- Your Secrets Are Showing: Deep Image Hiding Models Are Surprisingly Easy to Attack
- Privacy Leakage from Recurrent State in LLM Inference
- Fault Injection Attacks on Q-Learning: Analysis, Vulnerabilities, and Countermeasures
- PolicySniff: Inferring Agent Policy States through Shared-GPU Execution
- Look Here, Then Decide: Small Decision Models in a Harness Find Memory-Safety Vulnerabilities at Scale on Commodity GPUs
- The Unseen Threat: How Complementary Combinations of Acoustic Features Compromise Speech Privacy
- MobileProbe: Adaptive Security Evaluation of Mobile Agents via Trajectory-Aware Injection
- Scale-Induced Semantic Shift in Visual Prompt Injection Against VLMs and Web Agents
- Look It Up: Multi-bit LLM Watermarking via Codebook Search
- Batch, Pack, and Prove: More Efficient Verifiable Computation for CKKS
- S1-Evo: Experience-Guided Evidence Competition Attacks on Typed System-1 Decision Models
- Prevalence Bounds for Edited Watermarked Text: Power Envelopes and Calibration Sensitivity
- Infrastructure for Verifiable AI: Where Proof Costs Go and How Hardware Can Help
- CapLite: Prompt-Injection Defense for Small Language Model Agents without Model-Written Security Artifacts
- Preserve the Vulnerability, Break the Alert: Testing the Robustness of CodeQL Taint Tracking with Natural Rewrites
- Synthesis Smuggling: Maliciously Hiding Inferred Personal Attributes in LLM Outputs
Agents8
- DISCERN: Preserve at Write, Resolve at Read for Persistent LLM-Agent MemoryAmazon
- Discovering Effective Agent Memory Designs through Program EvolutionCode
- Probe What-If Worlds Before You Ship: Persona Simulation for Real-World Decision Making
- FOE-Dreamer: Deployment-Efficient Learning of Cyber Defense Policies in Operational Networks
- ToolTurns: Benchmarking LLM Agents in Multi-Turn Dialogue
All 8 papers in Agents
- Incentive-Responsive Agents: A Principal-Agent Framework, Metric Suite, and Runtime Architecture for Long-Lived Personal AI Agents
- Contain, Conform, Compare (Extended Version): A Red-Teamed Evaluation Kit for Small Coding Agents and the Full Record of Its Experiments
- Auditing AI peer reviewers: a benchmark on real scientific papers
Safety & alignment8
- Loyalty Capture: Reporting Relationships and Structural Sycophancy in Frontier AI ModelsGoogle
- Data-Efficient LLM Safety Alignment Guided by Interpretable FeaturesTencent
- ReA-Bench: Benchmarking Responsible Agent AuditingAlibaba
- DETER: Suppression During Pretraining for Tamper ResistanceCode
- Leaderboard Hacking: Preference-Based Model Evaluations are Vulnerable to Manipulation
AI & society6
- Terms of Care: Designing for Participation in Data Governance by Disability CommunitiesMicrosoft
- Scaling public input on AI policy with pluralistic agentsGoogle
- Measuring Preferred Representation in AI-generated Images with Community ScorersMicrosoft
- Do People Trust AI Assistants Even When They Are Consistently Wrong?
- Do Large Language Models Show Racial and Sex Bias in Pain Management? A Counterfactual Audit of GPT, Claude, Gemini, Llama, DeepSeek and Qwen
- A Code-Level Audit of X
Inference & efficiency5
- Scaling VLA Serving Under Robotic Execution ConstraintsCode
- TraceKV: Policy-Preserving Low-Rank KV Cache Compression for Reasoning Vision-Language ModelsCode
- metalSPARSE: A 2:4 Semi-Structured Sparse GEMV Kernel for Apple Silicon GPUs
- Together Compiler: Continual Compilation for AI Inference
- Smart Model Routing for Order-of-Magnitude LLM Cost Reduction
Robotics & embodied AI4
- Can MLLMs Reason over Surround-View Driving Videos? A Multi-View Autonomous Driving QA Benchmark
- Runtime Safety and Emergency Stops on a Full-Size Humanoid: Deployment Lessons and a Case for Graduated Force Authority
- Alloy: A Multi-Material Physics Engine for Robot Learning
- ACE-CaP-0: Real-World Code-as-Policies in the Era of Experience
Vision & multimodal4
- Beyond the Explicit: Benchmarking and Detecting Implicit Harm in Short Videos
- Verifier Choice is a Benchmark Design Variable: Auditing Structural Counting Evaluation in Text-to-Image Models
- DEDCA: Test-Time Adaptation for Generalized AI-Generated Image Detection
- AURA 2.0: From User-Driven Responses to Model-Initiated Conversations in Streaming Video