Same Output, Up to 4.76x Faster Generation for Coding Agents — KAIST's AgSpec
KAIST's AgSpec speeds up only the stretch where a coding agent writes out its answer. Under greedy decoding it leaves the output untouched, and its generation throughput ran 18.0% higher on average than the previous best method.
On-Policy Wasn't the Magic Bullet — What Decided Distillation Was KL Direction and Learning Rate
A Cambridge team put the supposed superiority of on-policy distillation to a controlled test. What actually decided final performance was the KL direction, not the rollout policy; and forgetting and sparsity were governed by the learning rate.
How Far Should You Trust an AI PM's "Done"?
Don't trust the word "done" — trust the evidence sitting next to it and the list of what's still outstanding. Following a single deadline-change request from start to finish, here's how to run an AI PM by tracing its reports back to their proof.
OneStreamer: A Streaming Video AI That Remembers the Past as Notes, Without Clouding the Present
OneStreamer, a streaming video AI from Nanjing University and others, writes captions of what it sees so it can carry the past as notes instead of video — and at just 4B parameters, it tops all eight benchmarks.
Just Labeling the Models Makes AIs Pick Sides — The Discovery of Factionalism
When agents were told each other's model families, a "factionalism" emerged where those sharing a label clustered together. It was measured across two cooperation games and a reasoning benchmark, and simply hiding the labels made it vanish.
A deep research agent that doesn't plan everything upfront proved more accurate and cheaper under identical conditions
A new deep research AI agent skips committing to a full plan upfront, instead growing the plan piece by piece as it reads the results of finished tasks. In head-to-head tests under identical conditions, it raised accuracy while cutting tokens, tool calls, and steps all at once.
The Same AI, Run Three Times, Swung More Than Different AIs Did — BIABench
BIABench turns published biology studies back into tasks. Agents cleared the 2D ones, but on some tasks that added depth or time the best score topped out at 0.19 — and a single agent's run-to-run spread was wider than the spread between agents.
Handing Context Between Different LLM Families Without Re-Reading It — HeteroFold
HeteroFold lets two LLMs from different families share a long context without re-reading it, passing along only precomputed internal state. At 32K it ran 10.7x faster than re-prefilling, and it matched text-based communication in multi-agent settings.
Don't "Write" Actions, "Choose" Them — How JevSpawn Made Agents Fast
An analysis of JevSpawn, which stops spelling out each action token by token and instead selects from a finite set of candidates — while letting the model build and revise that candidate list itself from the task.
Why AI Collapses on the Last Line — Dissecting KaliBench
A benchmark that splits the skill of turning natural-language security requests into Kali commands into "tool selection" and "argument construction" found that models pick the right tool but fall apart on the arguments.
Can a Coding AI Become a Robotics Engineer? Inside RLE-Bench's Qualifying Exam
Harvard·Georgia Tech researchers tested, entirely in simulation, how well coding AIs handle part of a robot-learning engineer's job across 51 tasks—and on design stability margin every model scored zero.
The Footage Looks Convincing, But the Text Is Wrong — VTR-Bench Measures How Well Video Models Write Words
A benchmark that uses 300 prompts to measure how accurately video-generation AI renders on-screen text. Even the best model scored a normalized edit-error of 0.250, and a video looking convincing turned out to be a separate matter from whether its text was correct.