Transferring What an AI Learned — by Reading the Training Record Mid-Run: Waypoint-OPD
A model hands off the gains it earned through reinforcement learning to another model, but by consulting the intermediate training record it transfers the final shift weighted by how consistently the two training segments pointed the same way.
Keep Teaching a Small Model While Mitigating the Forgetting — Synthetic Self-Distillation (SynthSelf)
Teach a small model a new skill and it forgets what it knew. Having a slightly larger same-family model rewrite its training examples eases that forgetting — research from LG AI and KAIST.
Blurry but in the Right Place: LEGO, Turning Third-Person Video into First-Person
LEGO converts third-person video into a first-person viewpoint. Following its experiments, we trace why a blurry-but-correctly-placed guide beats a sharp-but-misplaced one.
Where LLM Simulators Get Caught — A Rock-Paper-Scissors Experiment
In a controlled experiment, LLMs reproduced how often each move was played, but struggled to recover and stick to the sequential rules that decide the next move from past ones. They broke down most on rules that require looking two steps back.
CADFather: The Agent That Turns a Mesh Back Into Editable CAD
An autonomous agent that resurrects stale 3D shapes (meshes) back into editable design files. Instead of leaning on a single tool, it conducts several tools and candidates in concert, and it works with no additional training.
Mark It on the Picture and Jot a Note — VibeEdit's Canvas Editing
VibeEdit tackles the hard part of image editing — pointing at "which one" in words — by letting you circle or arrow directly on the picture and add a short note. It outscored a conventional editor that was handed a careful written description on both metrics.
A Test Score Earned Against One Correct Answer, Cut in Half When Re-measured
When tests written by AI are graded against a single reference implementation, their quality gets inflated. Re-measured against multiple valid and invalid implementations, even the best model dropped from 59.67% to 28.00%.
Sharpening Blurry Frames While Keeping Motion Intact — A Close Look at DuoMatching
Fast video generators trade per-frame image quality for speed. DuoMatching adds an image generation model as a second teacher to win that quality back, lifting sharpness while leaving motion largely untouched.
Clean Memories Breed Sycophancy: The Fix MemAdapter Found
It isn't only wrong memories but even accurate, relevant ones that can make LLM agents sycophantic. Instead of deleting memories, MemAdapter retrieves them first and then recalibrates how much to trust each one for the context at hand.
Fixing One Broken Step Mid-Run — No Answer Key Required
In work split across several AIs, InFlowOp picks out the single step that went wrong and repairs it on the fly, without any answer key. But the bigger lever for success turned out to be how the work is composed from the start, more than the repair itself.
Why Agentic SaaS Forces Us to Rethink the Unit We Charge For
A look at where each of the three pricing models — per-seat, usage-based, and outcome-based — slips, and why the market ultimately drifts toward blending all three.
Video AI That Remembers Scenes and Time — How ATAS Keeps Long Videos from Contradicting Themselves
When you build a long video from many shots, earlier scenes and events tend to drift out of sync later on. ATAS tackles this with two kinds of memory — scene memory and event memory — and only commits observations that pass verification.
When the Agent Stumbles: Turning Failure into Recovery Rules with Agent-Reflex
AI agents fall apart when a tool goes dark or a user cuts in mid-task. A team at Renmin University turned those failures into per-environment recovery rules and baked them into a small model beforehand. It lifted ordinary tool work, but handling noise remains unfinished homework.