AI Research Brief

Context Shifts Concentrate Occupational Stereotypes

September 2, 2026

Context Shifts Fail to Weaken Occupational Stereotypes. ContextBias tests 92 occupations across 66,240 images. Irrelevant settings increase attribute...

---

Models Detect Impossible Questions but Answer Anyway

September 1, 2026

Legal RAG Should Prioritize Retrieval: An Uzbek-language study used 178 retrieval queries and 504 question-answer pairs. Targeted retriever tuning could...

---

How Agents Learn to Manage Context and Credit

August 31, 2026

ContextPilot Lets Agents Restructure Context Proactively: It turns planning, long-term memory, and soft offloading into learnable operations, then trains...

---

Eleven World Models Miss the Odds; ES Covers More Paths

August 30, 2026

World Models Must Match Future Frequencies: PAWBench tested 11 systems across 50 scenarios. None captured multiple valid outcomes at the reference...

---

300 Tasks Test Visual Reasoning, Retrieval Runs 12.4× Faster

August 28, 2026

Visual Reasoning Becomes Trainable and Verifiable. VBVR-Pro compares image, video, and interleaved reasoning across 300 procedural tasks with task-specific...

---

On-Policy Feedback Makes Training Signals Useful

August 27, 2026

DiffusionOPSD Leads 19 of 20 Settings. It turns final image rewards into stepwise denoising targets on current-policy trajectories. It beats the strongest...

---

Voice RAG’s F1 Drop Widens by 67%

August 26, 2026

Agent Capability Is Becoming Sustained Delivery: Apodex 1.1 builds state retention, failure recovery, multi-agent coordination, and result verification into...

---

Learning Workflows Lift Design Agents Across Seven Setups

August 16, 2026

Recursive Feedback Improves Paper-to-Poster Workflows. After training DesignHarness, all seven code-agent and model setups improved. The average score rose...

---

Thirteen Skills Push Hallucination Detection to 92%

August 14, 2026

Thirteen Skills Make Research Agents Auditable. Spark-to-Paper raises hallucination detection from 14% to 92%, with 99.5% citation validity. Weak Models...

---

Life Agents, 4D Video, and Sparse Domain Updates

August 14, 2026

Life Agents Must Keep Working After Users Go Silent: VibeLifeBench tests proactive checks, timing, and hidden constraints across 200 multiweek tasks. All...

---

SimWAM Hits 91.5 as LoRA Cuts Memory 43.9%

August 13, 2026

SimWAM Removes the Video Expert From Deployment: It learns temporal patterns through video generation during training, keeps only a lightweight action branch...

---

Runtime Control Moves Into the Open

August 10, 2026

WorldTrace Makes Long Video Memory Addressable: It assigns in-distribution virtual positions to compressed memory slots. Two variants improve LoopBench...

---

Skill Entropy Lifts Qwen3-4B to 68.4%

August 9, 2026

Skill Entropy Lifts Qwen3-4B From 34.4% to 68.4%: Long-horizon tasks depend on individual skills and reliable handoffs between them. Earlier Multimodal...

---

Knowledge Refresh, Agent Memory, and Unified 3D

August 7, 2026

KGD Separates Knowledge Refresh From Task Adaptation. It beats strong baselines by 4%–12% across eight public benchmarks and sustains gains over a 90-day...

---

DAPD Gains Two Points as Diffusion Hits 1,500 Tokens/s

August 6, 2026

Privileged Information Can Make Self-Distillation Worse: DAPD constrains irreproducible teacher behavior with dual-path and dual-source anchoring, averaging...

---

User Code Edits Cut Agent Success by 7.7 Points

August 5, 2026

User Edits Cut Agent Solve Rates by 7.7 Percentage Points: SWE-Touch finds that nine coding models often preserve conflicts or overwrite user work....

---

Compress Tokens Before They Become Operational Debt

August 5, 2026

3DZip Preserves 94.7% Performance With 128 Tokens. It reduces 3D scene redundancy through feature diversity and spatial constraints, while raising inference...

---

A 4B Agent Ties the Giants by Verifying, Not Searching Harder

August 2, 2026

The real lever for a deep-research agent is the gap between verifying and finding, not a stronger base model. AREX keeps partly-verified intermediate states...

---

Mage-Flow Renders in 0.59s on One A100

July 23, 2026

Efficiency isn't squeezed out afterward by distillation. Mage-Flow co-designs the tokenizer and backbone as a pair that accommodate each other. At 4B, on a...

---

Treat Denoising as Filtering, Rebuild Viewer and View

July 23, 2026

Treat diffusion denoising as state estimation, not ODE solving. DiFA borrows the Kalman filter idea, fusing past predictions along the reverse trajectory to...

---