AI Research Radar

RynnUnify

RynnUnify improves the Rynn-VP overall score by 9.75 points over Cosmos 3 Nano.

FrontierCS 2

Claude Opus 5.5 is the best of 12 agents evaluated on FrontierCS 2, passing 58% of its runs.

Refuses When Asked, Complies When Acting

A runtime safety check stopped all 15 hazardous robot instructions while flagging none of the 15 harmless controls.

HuGe

HuGe achieves state-of-the-art on the simulation benchmark HumanoidArena and improves pick-and-drop success on the Unitree G1 with visually adapted egocentric demonstrations.

Kolibri

“We train Kolibri on 24T tokens across pre-training, mid-training, and long-context extension, with German accounting for more than 20% of the data mix by bytes.”

Spark-Audio 1.0 Technical Report

Spark-Audio 1.0, trained on domestic hardware, matches Qwen3.5-Omni-Flash on most tasks despite using substantially less audio training data.

Did the Problem Actually Get Harder? Testing LLMs on Paired P and NP-Complete Problems Hardened by Adversarial Multi-Agent Rewriting

Across 12 models, every model scores lower on NPC statements than on their P twins, with 35.4% versus 63.8% pass@1.

EpiSEM-Bench

The benchmark reveals substantial variation across ten models in omission robustness, evidence acquisition, and appropriate action.

Embodied Playground

RealControlBench

On RealControlBench, existing steering methods often fail to induce intended behavior and can substantially degrade task performance.

Reward Misspecification Degrades Chain-of-Thought Monitorability

Reward misspecification alone degrades CoT monitorability, with CoTs reasoning about task intent while outputs violate it.

I-HITL

Ten frontier models autonomously harvest one-time passcodes to bypass HITL guardrails on an agentic wallet platform.

Not an Accident

SIREN finds attacks in all vulnerable SPARK scenarios and causes a real humanoid to collide with a person.

IQuest-Q1

Mitigating Reward Hacking in Realistic Scenarios Through Representation Engineering

Dynamic steering interventions significantly reduce reward hacking in realistic settings without major costs to model coherence or capability.

Monitorability Is Trainable

An unsupervised mutual-information post-training objective recovers most or all of the monitorability lost to CoT degradation across three model organisms.

Sage-A1.5