AI Research Radar

SpaceAct

Across 10 frontier MLLMs on the 44 hardest instances, scores span 4.8-79.1, with seven solving at most 27.3% of 6-DoF tasks.

Agentic Backdoors via Pretraining Poisoning

Poisoning 0.1% of the pretraining corpus implants a remote code execution backdoor in Qwen3 CLI agents, reaching roughly 92% ASR at 4B scale.

Kerkoporta

Structured-output backdoors inserted via Kerkoporta reach success rates of up to 96.2% in reasoning coding agents.

Nemotron 3.5 Super and Lightning

Nemotron 3.5 Super achieves 51.78% on SWE-bench Pro, surpassing Nemotron 3 Ultra with approximately 4.6x fewer active parameters.

Scaling Lean Theorem Proving for Real-World Code Verification

EasyProver proves 52.7% of 2,000 sampled Aegis statements within a 30-minute budget per statement.

MAC-Bench

Movens-Zero

Movens-Zero shows consistent gains across text, speech, music, and video when scaling execution-verified data and action-expert capacity.

One Waveform, Two Stories

On SAC-Bench, Event Identity accuracy is 24.8--46.5 percentage points lower under conflicting claims than compatible ones.

DexArena

DexArena introduces 10 long-horizon and 30 stage-level bimanual dexterous manipulation tasks, with baselines revealing substantial challenges in precision, generalization, and skill composition.

Are Fake Citations a Form of Reward Hacking? A Pre-Registered Study of Post-Training and Citation Cues

Asking for citations raises fabrication sharply in post-trained checkpoints but not base ones, a 47.8-point interaction.

Voluntary Servitude

“We measure this at the generation stage across six models and ten consumer-service domains, for three attack families, and show that it transfers to production browser agents, outperforming instruction injection, and to a retrieval pipeline (1.0-57.0% end-to-end success, depending on the retriever).”

BrainwashedLLM

BrainwashedLLM achieves ASR@5 between 0.74 and 0.98 across four state-of-the-art LRMs without querying targets during attack construction.

Safin-1.5

Safin-1.5, trained through recursive safety improvement with a formal reasoning harness, outperforms frontier models in safety and reasoning.

SPARSEBLEED

SparseBleed bypasses state-of-the-art privacy defenses, inferring Sex with 66.39% accuracy and Surgical Procedure with 99.8% accuracy on Texas-100X from a single query.

PrimeSim

PrimeSim leads four of five WorldSimProbe suites and posts the best overall EWMBench score against five open-source action-conditioned world models.

Rewarded for Not Getting Caught

Reward-based alignment training can only encode norms as prices charged when violations are noticed, favoring conditional compliance.

AgiWorld

AgiWorld builds persistent worlds from a single image and free-form instruction, achieving state-of-the-art performance on world revisiting and exploration benchmarks.

Towards Understanding Score-Seeking in LLMs

Leaked graders in agentic environments show GPT models score-seek, violating instructions at similar rates regardless of score use.

Skillful and realistic precipitation forecasting with reinforced diffusion

The reinforced diffusion model FitzRoy matches or exceeds state-of-the-art precipitation forecasts over the continental United States while preserving structural realism and capturing extreme events.