AI Research Radar

Wednesday, April 29: 1 frontier paper out of 195 read.

1 frontier22 cleared195 papers read

Frontier

Papers rated as moving the field

Odysseys

“We tested several leading frontier models and find that the strongest models achieve a success rate of 44.5%, which leaves substantial room for future improvements.”

The authors, in the abstract

Benchmarking Web Agents on Realistic Long Horizon Tasks

BenchmarkarXiv PDF

This week in research

Apr 23–29: 73 cleared, 5 frontier

The week’s papers have not been sorted into fields yet.

Models under test

Papers this week that name the model

Claude Sonnet 4.52
GPT-5.42
GPT-52
GPT-4o2
Gemini 3 Pro1
Llama 3.1 8B Instruct1
Mistral Small 3.21
Qwen3 8B1

Labs publishing

Papers this week with an author at the lab

Google3
Microsoft3
NVIDIA2
Alibaba1
AllenAI1
Anthropic1
IBM1
Mistral1

Also cleared

21 papers

MotionBricks: Scalable Real-Time Motions with Modular Latent Generative Model and Smart PrimitivesNVIDIA
Is your AI Model Accurate Enough? The Difficult Choices Behind Rigorous AI Development and the EU AI Act
The LLM Fallacy: Misattribution in AI-Assisted Cognitive Workflows
SketchVLM: Vision language models can annotate images to explain thoughts and guide users
Quantifying and Mitigating Self-Preference Bias of LLM Judges
Agent-Centric Visual Reinforcement Learning under Dynamic Perturbations
Incompressible Knowledge Probes: Estimating Black-Box LLM Parameter Counts via Factual Capacity
asRoBallet: Closing the Sim2Real Gap via Friction-Aware Reinforcement Learning for Underactuated Spherical Dynamics
BenchGuard: Who Guards the Benchmarks? Automated Auditing of LLM Agent Benchmarks
The Dynamics of Delusion: Modeling Bidirectional False Belief Amplification in Human-Chatbot Dialogue
Making AI-Assisted Grant Evaluation Auditable without Exposing the Model
Safe-Support Q-Learning: Learning without Unsafe Exploration

All 21 papers

FED-FSTQ: Fisher-Guided Token Quantization for Communication-Efficient Federated Fine-Tuning of LLMs on Edge Devices
Praxy Voice: Voice-Prompt Recovery + BUPS for Commercial-Class Indic TTS from a Frozen Non-Indic Base at Zero Commercial-Training-Data Cost
Navigating Global AI Regulation: A Multi-Jurisdictional Retrieval-Augmented Generation System
Bye Bye Perspective API: Lessons for Measurement Infrastructure in NLP, CSS and LLM Evaluation
The Surprising Universality of LLM Outputs: A Real-Time Verification Primitive
Cross-Lingual Jailbreak Detection via Semantic Codebooks
Subliminal Steering: Stronger Encoding of Hidden Signals
No Pedestrian Left Behind: Real-Time Detection and Tracking of Vulnerable Road Users for Adaptive Traffic Signal Control
Conditional misalignment: common interventions can hide emergent misalignment behind contextual triggers

SIGNIFICANT ON THE BOARDS

MODEL TRACKER2,648 new benchmark results across 174 models this week DEAL FLOW$158B disclosed AI funding this week across 61 rounds POLICY125 policy events tracked across 7 days