AI Research Brief

A 20B Search Agent Ties the Frontier by Offloading Its Bookkeeping

June 4, 2026

Deleting stale observations to save context follows an inverted-U, not a straight line. Sweep across 4B to 284B models and three retrievers: strong...

---

A 4B Web Agent Catches Up to Closed CUAs on a Few Thousand Trajectories

June 3, 2026

PEFT isn't just cheap fine-tuning — it's per-user persistent state. A framing paper recasts small adapters as local state attached to a shared trillion-...

---

Move to See: Top Model Reaches the Target Just 12% of the Time

June 2, 2026

Spatial intelligence shifts from passive understanding to active perception. TVR asks an agent to turn and step through a 3D room until its view matches a...

---

MoE Safety Lives in a Few Experts, Exclusive Batching Adds 42%

June 2, 2026

Lab VLM scores don't survive robot deployment. RoboStressBench breaks physical rendering into material, lighting, viewpoint, and geometry stress, showing...

---

LoRA as a Ruler for Memory, Reversed Video as a Free Counterfactual

June 1, 2026

Flip LoRA into a measuring stick and the real capacity of parametric memory falls out: it follows a power law you can estimate in advance, and a token-...

---

Reasoning Search Adds a Second Direction, World Models Add Agents

May 31, 2026

BES Splits Inference-Time Search From One-Way Expansion. Forward search gains evolutionary operators to escape the model prior's "entropy shell," and...

---

Agents Start Improving Themselves, and Reaching for Fewer Tools

May 30, 2026

A Chinese MoE puts "self-evolution" on the roadmap. MiniMax-M2 runs 230B params with only 9.8B active, built end-to-end for agent work, and its latest...

---

Agent Trajectories Let a 30B Match a 235B

May 25, 2026

ACC Repackages Agent Tool-Use Trajectories as Long-Context QA Pairs. Qwen3-30B trained on them lifts MRCR from 50.2 to 68.3, matching Qwen3-235B-A22B at...

---

Gated DeltaNet-2 Splits the Gate, Maestro Outscores GPT-5

May 23, 2026

Linear Attention's Real Bottleneck Is State-Edit Granularity, Not Speed. Gated DeltaNet-2 splits the scalar gate into channel-wise erase and write gates. It...

---

Optimizer Choice Stretches Capacity Scaling 2.3x

May 22, 2026

Three Classes of Physical 3D Assets Merge Into One Pipeline. PhysX-Omni puts rigid, deformable, and articulated objects into a single framework. Output...

---

$15 Per Paper, Healthcare Agents Cap at 28%

May 22, 2026

Auto-Research Cost Curve Has Crossed. $15 produces a full paper, but frontier LLMs still fabricate results and miss errors. End-to-end autonomy still falls...

---

8% of Tokens Decide the Reasoning Gap

May 19, 2026

"Unlearnable" Samples in RLVR. A set of hard examples never gets learned across training, even though rollouts produced correct answers. The reward curve...

---

Real-Time Video's Bottleneck Moved Past Step Count

May 17, 2026

Real-Time AR Video's Bottleneck Is Shifting. Causal Forcing++ pushes frame-wise distillation to 1-2 steps. RAVEN attacks long-rollout history distribution...

---

Olympiad Gold Becomes a Two-Step Recipe

May 16, 2026

Olympiad Gold Becomes a Portable Two-Step Recipe. SU-01 combines reverse-perplexity curriculum SFT with two-stage RL. A 30B-A3B backbone clears IMO and IPhO...

---

Readable Rules Don't Belong in LLM Weights

May 15, 2026

Readable Dynamics Don't Belong in Weights. Enterprise World Models use CascadeBench to show that cross-tenant business rules get more brittle the better...

---

δ-mem Trades Long Context for an 8×8 State Matrix

May 14, 2026

δ-mem Bolts an 8×8 State Matrix Onto a Frozen Backbone. A delta-rule online update lifts memory-heavy tasks 10–15% over baseline. Reframes long context from...

---

Flow-OPD Lifts GenEval From 63 to 92

May 13, 2026

Image Generation Alignment and LLM Post-Training Now Share One Toolbox. Flow-OPD ports On-Policy Distillation to flow matching. SD 3.5 Medium hits GenEval 92...

---

Geometry Conflict Predicts Continual Fine-Tuning Forgetting

May 12, 2026

Geometry Conflict Predicts Continual Fine-Tuning Forgetting. Treating each task's parameter-update covariance as a measurable signal, GCWM beats data-free...

---

Lorem Ipsum Rescues GRPO's Wasted Hard Samples

May 10, 2026

Skill1 Unifies Skill Retrieval, Use, and Distillation in One Policy. A single task reward co-trains all three, avoiding interference between competing reward...

---

10.6k SFT Trajectories Match Full RL Pipeline; Mamba Beats LZMA

May 8, 2026

10.6k Curated Trajectories Match a Four-Stage RL Pipeline. OpenSeeker-v2 expands knowledge graph and tool set, applies strict low-step filtering. Pure SFT on...

---