Agents Already Know the Digital World — EVOKE Wakes the Knowledge That's Sleeping Inside Them
The reason LLM agents fall apart in new environments isn't that they lack knowledge — it's that they can't pull it out when it counts. EVOKE is a post-training technique that wakes that dormant knowledge by re-asking the same scene with only the goal swapped.
In-Context Learning for Robots: Four Ways to Read a Situation Without Retraining
A survey walkthrough of "robot in-context learning"—adapting to new tasks from demonstration and interaction without retraining—organized into four families by the channel through which context flows into action.
Teaching Video-Generation AI "Object Permanence" — The WROP Benchmark and PWM-WROP
Fine-tuned on 1.5 million synthetic videos modeled after cognitive-science experiments, PWM-WROP ranked first among continuation models and third overall out of 14 in human evaluation — though whether this carries over to real-world video physics remains unverified.
Rules, Fixed LLM Workflows, and Agents — How to Pick Among the Three Levels of Automation
What to use for automation comes down to one question: can a human map out the flow in advance? Conditional logic calls for rules; text that needs understanding but follows a fixed path calls for an LLM workflow; and when the LLM has to decide its own steps and tools on the fly, you need an agent.
Shipping Apps Without the App Store — PWA First, Store Only When You Need It
ONTLE keeps PWA as the default for turning the web into an installable app with no review process, and only wraps it in a TWA for Google Play when store distribution actually becomes necessary. Here's the thinking behind that call.
SpeakerMem-R1: AI Memory That Tracks Who Said What About Whom in Many-Speaker Conversations
A dual-track memory for multi-party dialogue that remembers who said what about whom, along with how each person's situation changes over time. It reports performance that surpasses existing memory systems.
ATD: Transferring Coding Skill Through Nothing but Off-Topic Word Choices
과제와 무관한 단어 선택 수천 개만으로 교사 모델의 코딩 능력 일부가 학생 모델로 전이됐다. Qwen2.5-1.5B 실험에서 코드 시험 점수가 5.34%p 올랐다.
The AI That Works but Never Finishes: Breaking the "Belief Trap" with PoS
On long-horizon tasks, agents fall into a "belief trap" — never stopping, yet never making progress. PoS externalizes the agent's current understanding, checks it, and pulls the agent out when it gets stuck.
When Robot Skills Hand the Composer "Where They Hold Up" — A Look at the APPL Paper
APPL reuses the structural assumptions a skill was trained under as the interface for the composition stage. We unpack the idea of stitching skill generalization and compositional generalization into one.
When an AI That Writes Its Own Exams Starts Getting Them Wrong Together: Dissecting 'False Frontiers'
A study that diagnoses "co-cheating" — where a search AI that writes and solves its own problems reaches consensus on the same wrong answers as both examiner and solver — and suppresses it with CrossFit, which cross-assigns the graders.
Where the Real Moat Lives in Agentic SaaS — It's the Accumulated Context, Not the Model
Once everyone can tap the same frontier model at the same price, the battleground for agentic work automation shifts away from model performance and toward wherever a customer's working context piles up.
A Small Specialist Teaches a Big Model — The Laws of Same-Family On-Policy Distillation
A paper that, from math-reasoning experiments, distills power laws predicting both the peak performance and the early transfer speed of on-policy distillation within one model family, using only teacher and student size plus the teacher's score — before any training is run.