The Robot Already Knew How. We Just Didn't Know How to Tell It.
The robot already knows how—what was missing was a way to tell it. Freeze the body and evolve only the instruction sheet (a "reward program"), with a language model proposing and a tamper-proof verifier judging, and the success rate climbs 2.5x.
GraphForge: Training Practical AI Agents on Real Files
An evidence-graph approach that pulls both tasks and their grading rubrics straight out of real work files. Using 2,169 training trajectories, it lifted a Qwen model's scores on three work-focused benchmarks.
Teaching GUI Agents Without Installing the Software — A Look Inside AutoGUIWorld
A study that synthesized GUI-agent training data with an image generator instead of actually running any software, then used that fake data to measurably improve real-world performance on desktop and scientific tasks.
Robot Agents That Run on Code: 65% Fewer Tokens, 14% Higher Success Rate
By having robot agents bundle actions into Python code instead of calling tools one at a time, success rate climbed from 63.1% to 71.7% under the same call budget, while token usage dropped 65%.
MILO: Letting an AI Agent's Harness Evolve Itself
모델을 감싸는 실행 하네스를 사람 손이 아니라 탐색으로 자동 발굴하고, 그 탐색 전략까지 함께 진화시키는 프레임워크. 세 벤치마크에서 기존 최고 설계를 앞섰다.
Like a Coach Who Keeps Revising the Training Plan, Let the Agent's Curriculum Evolve Too
When you optimize an agent, locking in what it learns from as a fixed upfront order leaves gains on the table. Evolve the training plan along with the fix method, and the same budget buys a better result.
An AI Agent's Greatest Virtue Isn't Acting — It's Holding Back
Even the best of six voice AIs stayed silent on only 14% of switched-speaker commands; holding back is a separate skill.
The AI PM That Knows Your Project Looks Things Up Like a New Consultant
Follow a single payment-timeout question from the chat box to a sourced answer, and see — through the lens of a new consultant — exactly how an AI PM manages to reply as if it already knew your project inside out.
The Trap in Merging Experts: It Wasn't About "Who Teaches," but "How Hard" (DN-MOPD)
An analysis of a Qwen3.5 experiment showing that in multi-teacher distillation — fusing math, coding, and instruction-following experts into a single model — normalizing the strength of each domain's feedback is what actually moved the needle.
A VLM That Learns Space Through Interaction: A Deep Dive into Spatial-Interactor
A walkthrough of Spatial-Interactor, a paper that teaches vision-language models to reason about space not from still images but from "observe–act–observe" interaction trajectories, organized into a three-stage curriculum.
Realtime Full-Duplex AI, Realtime-Venus: A Conversational System That Listens, Watches, and Delegates Work
An abstract-level deep dive into Realtime-Venus, a system that handles real-time full-duplex conversation and background task delegation using one of two 9B model variants per session — the speech-and-video Omni and the audio-only Audio.
Full Participation, Light Compute: How IntBMoE Resolves the Three-Way Dilemma of Mixture-of-Experts
Mixture-of-Experts could never tune participation, execution, and materialization independently. IntBMoE decouples all three with block-level conditioning, and lifted UVCTR by a relative 2.4% in AMap's recommendation service.