When an AI PM Stops Working — Knowing When to Push and When to Ask
A good AI project manager isn't defined by how much it gets done, but by where it stops. An operational look at how to draw the line between autonomy and asking for confirmation.
Tailoring a Quantum Circuit for Every Token: How HyperQ Widens the Qubits to Boost Performance
Bolting per-token quantum circuits onto a frozen 1.1-billion-parameter model and widening the qubits from 16 to 64 lifted the benchmark average from 47.65 to 54.30 (+4.71 over the backbone alone). The computation is almost entirely classical, and no quantum advantage is claimed.
Does a World Model Hold Up Once You Interact With It? Meet HappyWorld-Bench
A unified benchmark that brings video, spatial, and embodied world models under one yardstick — measuring not surface-level visual quality, but how reliably a world holds together when you step inside, move around, and poke at it.
Reining In Self-Improving AI Agents — Google's RRSI
스스로 하네스를 고치는 AI 에이전트는 연습 과제에 과적합된다. 구글 RRSI는 자기개선에 '정규화'를 걸어 미지 과제 성능을 지키고 비용도 줄인다.
EvoOntology: A Data Map That Agents Grow on Their Own
Instead of a hand-written data dictionary, this is an ontology layer that an agent builds from real data and repairs from its own execution logs. It clearly beats an ontology-free baseline, and in some conditions it draws even with a statically injected agent-built semantic layer.
The AI That Reads Tables, LimiX-2: Beyond Prediction, Mapping How Variables Connect
LimiX-2, an AI built for tabular data, doesn't just get the prediction right—it also sketches a map of how the variables are entangled. What it won't tell you, though, is which way causation flows.
PhysBrain 1.5: Understanding Proven, Real Robot Execution Not Yet — A Physical Foundation Model
An 8B open-source model that unifies understanding, action, and prediction. It topped open-source rankings with a 72.5 average across 28 understanding benchmarks, but its action capability rests only on offline prediction metrics — real robot execution remains unverified.
How to Get Great Work From an AI PM — Good Requests vs. Bad Requests
A request that carries four things — purpose, scope, format, and examples — changes the quality of what an AI PM produces. We walk through one ONTLE case from request to first draft to revision.
Giving Your Agent Every Modality Without Touching the Model — Inside Omni-IO Skills
A study that bolts a single harness layer onto a general-purpose agent that only knew text well, letting it handle images, audio, video, and even 3D. Both host agents saw their input support rate climb to 100%.
RULER: How to Reward Open-Ended SVG Generation With a Scoring Rubric
For SVG generation tasks that have no single ground-truth image, RULER automatically builds a six-criterion scoring rubric for every prompt and uses it as the reward signal for reinforcement learning. Here's how it works.
Knowing the Answer and Teaching It Are Different Skills — Meet OmniEdu, a Model Built for Education
Solving a problem and tutoring a student are not the same ability. OmniEdu is an open education model that reorganizes training data around "competencies" rather than "sources," so it learns to do both well.
AIs Keep a Mistake Log Too: How XConf Grades Confidence by Looking at Past Performance
XConf stops judging an answer by its reasoning alone and instead pulls up the model's track record to score how likely that answer is to be right. It calibrates confidence better than existing methods while costing a tenth of what self-consistency does.