Skip to content
VECTOR WIREAI INTELLIGENCE
UTC

AI Research Radar

November 13–19: 19 frontier papers out of 2,597 read.

19 frontier330 cleared2,597 papers read

Frontier

Papers rated as moving the field

How we choose

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?

AgentsNVIDIA, Amazon, MetaBenchmarkarXivPDF

CompoundingKernel

CompoundingKernel ranks #1 on SOL-ExecBench L1, L2 and Quantization boards, 4.40x faster than the reference baseline.

AgentsAmazon

Lazare, 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.

SecurityMicrosoft

Bait

Bait mitigates harmful agentic reinforcement learning by probing the reward function while still allowing learning from benign rewards.

SecurityGoogleCode

TabBioMed

On 94 biomedical tabular datasets, foundation models perform best overall, though tree models and tuned multilayer perceptrons dominate under computational constraints.

ScienceBenchmarkCode

WebArena-Pro

On WebArena-Pro, Codex with GPT 6.1 Sol and Claude Code with Claude Opus 5.5 both reach 62.3%, the best among evaluated agents.

AgentsBenchmark

ManiLadder

ManiLadder benchmark with 112 tasks across 4 difficulty levels calibrates manipulation task difficulty using a unified imitation learning policy.

RoboticsBenchmark

K2-Horizon

The 36B-A4B model nearly matches a 32B dense model trained on the same data with one-seventh the active parameters.

Training

Optimistic Verifiers Get Hacked

Optimistic verifiers are hacked by padding: a sufficiency proxy reaches 0.940 while true faithfulness falls from 0.731 to 0.667.

Safety

ScaleVLA

ScaleVLA with a 552B backbone outperforms a 2B backbone in sample efficiency, generalization, and precise manipulation.

Robotics

Soteria

Soteria achieves 1.7 times the average task success rate of the best baselines across tool-use benchmarks, with no violations of policy specifications.

Safety

Reins

Reins lets GPT-6.1 Sol complete 88% of real dual-arm manipulation episodes on five tasks with no task-specific code.

Robotics

The week by field

Nov 13–19: 330 cleared, 19 frontier

Also cleared

311 papers, by field

Security & privacy101
All 101 papers in Security & privacy
Agents47
All 47 papers in Agents
Safety & alignment43
All 43 papers in Safety & alignment
AI & society26
All 26 papers in AI & society
Robotics & embodied AI22
All 22 papers in Robotics & embodied AI
Inference & efficiency20
All 20 papers in Inference & efficiency
Vision & multimodal18
All 18 papers in Vision & multimodal
Training & data14
All 14 papers in Training & data
Language & speech8
All 8 papers in Language & speech
Science & health5
Machine learning5
Reasoning & math2