AI Research Brief

ImageNet-FID Negatively Correlates With Text-to-Image

June 26, 2026

The ImageNet-FID Leaderboard May Point the Wrong Way. DiffusionBench trained 21 diffusion models in one framework and found ImageNet rank and text-to-image...

---

Late Layers Quietly Rewrite Correct Answers for Alignment

June 25, 2026

A model's final layers quietly rewrite correct answers to sound more aligned. Early layers guess, middle layers sharpen the reasoning, but late layers drag...

---

Pose New Objects From a Few Reference Photos

June 24, 2026

Onboarding A New Object Drops From "Prep A CAD Model" To "Snap A Few Photos." PANY swaps single-anchor matching for a multi-view geometric backbone, lifting...

---

An 8B Model Beats a 235B One at Science Reasoning

June 24, 2026

Multi-turn agent training is expensive because of decision density, not horizon length. Mila re-did the math: the real signal-killer is the flood of reward-...

---

Leaderboards Don't Predict Deployment; Robot Arm Self-Trains to 99%

June 20, 2026

Leaderboard Rank Doesn't Predict Deployment. A position paper rebuilds one MCP agent benchmark 14 ways and finds aggregate scores rank unstably out of...

---

A 7B Video Agent Beats a 72B Model by Looking Less

June 19, 2026

Long video understanding doesn't need to watch every frame. OmniAgent models perception as a decision the model makes for itself; its 7B agent hits 50.5% on...

---

Two Loops Take SWE-bench From 43 to 64

June 18, 2026

Looping for deeper compute saturates at two passes. LoopCoder-v2 turns "how many loops" into an engineering knob, and two loops push SWE-bench Verified from...

---

Pruning Context Can Cost More Than It Saves

June 17, 2026

Trimming context to save tokens can backfire when the cache misses. Cutting old text is the obvious way to slim a long-running agent, but TokenPilot shows...

---

A 1.5B Model Beats Sonnet 3.5 at Event Forecasting

June 16, 2026

Agent Memory Forgets the Wrong Things, and the Root Cause Is Architecture, Not Prompts. MIT swept 13 configurations and found that where the LLM sits in the...

---

Best Code Agent Hits 61.1%, VLA Reasoning Runs 6x Faster

June 16, 2026

The real bottleneck for code agents is the hand-off from "find the data" to "write the code." CODA-BENCH drops 1,009 tasks into a Linux sandbox averaging 980...

---

Pruning a Small Model Is a Shortcut Only on a Tight Budget

June 15, 2026

Pruning is not an unconditional shortcut. Princeton compares pruning against training from scratch under matched token budgets. Pruning wins reliably only...

---

Swap the Action Interface, Gain 11 Points on Spatial Reasoning

June 13, 2026

The interface design sets the ceiling on spatial reasoning. SpatialClaw changes nothing about the model and skips fine-tuning. It rewrites the agent's action...

---

Arbor Triples Research Gains; Environments Become the New Scaling Axis

June 12, 2026

Arbor runs the full research loop on its own, accumulating experience across rounds through a Hypothesis Tree. It topped all six real research tasks, with...

---

One Token per Evidence Cuts Generation Cost 3-10x

June 11, 2026

Understanding, generation, and editing in one autoregressive model: ARM bets everything on a discrete tokenizer trained against three goals at once —...

---

DeepSeek V4 Cuts KV to 13.5%, Video Memory Runs 10x Faster

June 10, 2026

DeepSeek V4 bakes "index then attend" into the main architecture. Decoding no longer keeps the full KV cache in VRAM. A Neural Memory Indexer fetches...

---

Video Models Stumble on Composite Edits, MoE Fails at the Router

June 9, 2026

Single edits are good enough; composite instructions fall apart together. CoVEBench breaks multi-point editing into 9,990 fine-grained checklist items, and...

---

Swap the Arm Without Retraining; VLMs See Both the Duck and the Rabbit

June 9, 2026

Swap a robot arm and the whole skill set breaks — the fix is rewiring, not retraining. RECENT writes skills as executable code and locally refactors only the...

---

dots.tts Hits 54ms First Packet, SWE Agent Self-Evolves Past 50%

June 8, 2026

Open-source TTS takes the continuous-latent route, with three design choices all aimed at deployment. dots.tts is a 2B continuous autoregressive speech...

---

Streaming Hand-offs Make Multi-Agent Sharper, ZipSplat Splats With 1/6 the Gaussians

June 6, 2026

Streaming Hand-offs Beat Waiting for the Full Chain. StreamMA pipelines adjacent agents so reliable early signals reach downstream sooner — average +7.3...

---

NVIDIA Packs Five Modalities Into One Set of Weights

June 5, 2026

NVIDIA Crams Language, Image, Video, Audio, and Action Into One Set of Weights. Cosmos 3 bets a single mixture-of-transformers can do every modality, and...

---