The AI That Thinks Three Times: SanSi, a Decision Model That Rewrites Its Judgment Instead of Its Words
SanSi is a decision model that rewrites its judgment by looping the same layers several times — without writing a single token. It hit 72.0% accuracy, 13.5 points higher than a structure-matched non-looping model.
Opera: A Coding Critic That Doesn't Buy "Fixed It"
Opera is a critic that, on long coding tasks, separates "fixed it" from "actually gone" and stays on a problem until it's closed. Across three benchmarks it beat the no-critic baseline by up to 15 percentage points.
One Score Can't Capture a Good Image: The Reward Recipe That Lifts Image AI
Chase a single score and real quality actually buckles. We take apart the recipe that researchers at Arena and UCLA used to stably combine rewards of different temperaments and push an open-source image model to the top.
Grading the Process Made the Answers Better — OneSearch-VL
It doesn't just look at whether the answer is right. OneSearch-VL also grades where the model looked and what evidence it reasoned from — and adding that process grading to the reward pushed the benchmark average from 57.3 to 61.1.
Bringing a Frozen 3D World to Endless Motion — A Close Look at OuroWorld
OuroWorld turns a static 3D scene into an endlessly looping video with no masks required. The loop is guaranteed mathematically, and it even absorbs the mismatches between videos generated from different viewpoints.
SparseEngine: An Inference Engine Rebuilt Around "Forgetting" for Long Agent Conversations
How SparseEngine, which packs 15 different sparse methods into a single engine, chips away at the slowdown of long agent conversations, and why deciding which memories are worth keeping is the part that stays on our plate.
The Model Stays, Only the Procedure Changes: Gan Jiang, an X-ray Analysis Agent
An X-ray diffraction agent that never retrains its language model—instead it turns each analysis failure into a reusable "procedure" to run on the next sample. Even without composition information, it beat the previous state-of-the-art method by a wide margin.
Even With Help, It's Still a Long Way: Flagging the Decisions an Autonomous Agent Should Run Past You
A study on surfacing the critical decisions a human should review in long-running autonomous work, each backed by its own evidence. Weaving that evidence into a graph helps models pinpoint such decisions better, yet even the best models still have a long way to go.
It Renders Beautifully — Just Not What You Asked: The Editing-AI Bottleneck RISEBench++ Exposed
RISEBench++ pitted 58 image-editing AIs against each other, and even the top model aced only 56.6% of its tasks across every evaluation axis. Its sharpest relative weakness wasn't rendering quality — it was reasoning about the instruction.
The Robot World Model That Only Imagines Success — and How DreamTrue Curbed Its Optimism
A robot world model trained only on success keeps predicting success even when the gripper misses. DreamTrue post-trains it to imagine failures and grades them, cutting its own early version's interaction-defect rate from the 48% range to the 6% range.
The Robot That Left Its Model Alone and Fixed Everything Around It — PhysEvo
Do you rewire a genius's brain, or hand the genius better tools? A robotics project froze the model that drives the robot and rebuilt only the surrounding system from its own failure logs — and the exact same model's success rate climbed from 26% to 68%.
Should German RAG Think in German? The Payoff of Reasoning-Language Alignment
In a RAG setup where the question, the documents, and the answer are all German, aligning the model's thinking language to German too beat another forced language (French) on accuracy. But it never got past the model's native English reasoning (measured on a lightweight model).