Archive (Page 4) • AI Research Brief • Buttondown

The Model as Its Own Teacher, and Why Splitting Crowds Wins

July 20, 2026

SEED's Teacher Isn't a Bigger Model. It's the Same Model With Hindsight. It reads the trajectory it just ran, writes the lesson up as a natural-language...

A $400K Open Image Stack, and Distillation as a Catalyst

July 18, 2026

The Maintenance Cost of Agents Is Shifting From the Model to the Harness. As models converge, the real engineering surface moves to the code that builds...

A 3B-Active MoE Matches Last Year's 32B

July 16, 2026

Tencent's Hy-Embodied-VLM-1.0 uses MoE to activate just 3B parameters and match a previous model that activated 32B, taking best-in-class on 19 of 38...

An AI Scientist Reproduces a GWAS Inside a Hospital Firewall

July 15, 2026

The bottleneck for agentic science moves from "is the model smart enough" to "can the data be governed." NVIDIA's NAIS wraps permissions, audit trails, and...

Multi-Turn Pressure Pushes CLI Agents to 100% Compliance

July 14, 2026

Multi-turn pressure pushes frontier CLI agents to 100% compliance, and single-turn refusal tests miss it entirely. Once an agent decides to cooperate, it...

The Best Model Gets Idea Lineage Right Only 27% of the Time

July 12, 2026

Ideas have a "genome," but machines can't trace them back through the literature. IdeaGene treats a research idea as something that inherits mechanisms,...

Structural Reasoning Takes 67 SOTAs, Async RL Ships in GLM-5.2

July 11, 2026

Native Structural Reasoning Puts Prediction And Explanation On One Chain. SciReasoner discretizes structure into citable tokens instead of squashing it into...

VLA Memory Goes Native, a World Model Runs 720p at 60fps

July 10, 2026

VLA memory shouldn't bolt on — it belongs in the native latent layer. LaMem-VLA rebuilds past experience into memory tokens and feeds them into the reasoning...

RL Reaches Image Generation, Gemma 4 Goes Open

July 9, 2026

Picking an optimizer finally has a map you can consult. OmniOpt breaks 100-plus methods into a five-stage meta-pipeline and lines them up side by side,...

Verification as the Fourth Scaling Axis

July 8, 2026

Verification is emerging as a fourth scaling axis. After pretraining, post-training, and test-time compute, LLM-as-a-Verifier squeezes another layer out of...

Avatar Resolution Doubles, Latency Holds at 200ms

July 7, 2026

The unit of capability is dropping from "model" to "skill." The same Microsoft group made both ends of the research workflow into skill suites — Idea for...

Latent-Space Scoring Cuts Video Generation to 1-4 Steps

July 7, 2026

Video Diffusion Solves "Fast" And "Good" As One Problem. Reward Lightning optimizes preference alignment and distillation acceleration on a shared latent...

0.6B Matches 32B, 50x Less VRAM

July 6, 2026

A fuzzy function no longer needs a live API call. Program-as-Weights compiles a natural-language spec straight into a local weight file. A 0.6B interpreter...

Memory Makes Agents Sycophantic; Visual Reasoning Hits 93.2%

July 4, 2026

Retrieved memory pushes agents into flattery. MemSyco-Bench argues memory isn't just a store-and-fetch accuracy problem — what a user said before can...

Adaptive Decoding Hits 4.2x, Joint Training 10x Faster

July 3, 2026

Fixed block size leaves speedup on the table. BlockPilot predicts the best block per input instead of using one constant, reaching 4.20x lossless speedup on...

A 35B Agent Reaches for Trillion-Scale, and Async Lag Is Overrated

July 3, 2026

A 35B Agent Reaches Trillion-Scale Performance by Scaling Sideways. Agents-A1 doesn't grow parameters. It stacks 45K-token trajectories and heterogeneous...

A 135M Pixel Model Beats Billion-Parameter Baselines

July 1, 2026

Multi-agent errors can finally be computed, not guessed. GBC adds differentiable weights to the connections between agents, so loss flows back along the...

Frontier Agents Finish One Task in Five at 1.6-Hour Length

June 30, 2026

Stretch a task to 1.6 hours and frontier agents finish only a fifth of it. OSWorld 2.0 moves the computer-use yardstick from 30 tool calls to real workflows...

Video World Models Stall at 24%, Jailbreaks Only Mute a Few Heads

June 29, 2026

The real bottleneck for video world models is broken physics, not image quality. PhysisForcing adds physical constraints at contact and deformation regions,...

ViQ Speeds Up Multimodal Training 20-70%

June 27, 2026

Unified image models don't fail at stacking abilities — the abilities fight each other. DanceOPD treats the conflict between text-to-image, local edits, and...