AI Research Brief

AI Tutoring Matches Expert GRE Learning Gains

September 27, 2026

One AI Tutor Matched Expert GRE Learning Gains at Far Lower Cost. It spent about $0.0052 per percentage point gained, versus $4.81 for human tutoring. The...

---

RRSI Gains Up to 4.7 Points Across Five OOD Benchmarks

September 25, 2026

Constrained Harness Editing Delivers Limited but Transferable Gains. RRSI gained up to 14.1 points on its evolution split and 4.7 points across five out-of-...

---

When FP8 Training Stops Punishing Bad Answers

September 23, 2026

FP8 RL Can Lose Its Negative Feedback Signal. Quantization noise can push negative-advantage tokens beyond clipping boundaries and erase their gradients....

---

Semantic Rewrites Shift 23 of 29 Review Metrics

September 23, 2026

Meaning-Preserving Rewrites Shift 23 of 29 Review Metrics. Researchers turned 674 ICLR and NeurIPS reviews into 4,044 versions, with broadly consistent...

---

Agents Score 58.1%, but Only 2.8% Pass Every Test

September 21, 2026

Hybrid Agents Can Rebuild Interfaces but Still Cannot Deliver Reliably: RecreationWorld joins exploration, coding, execution, and visual review in one loop....

---

7B Searches the Web, BVB Recovers 53.7% of Facts

September 21, 2026

A 7B Model Uses Search to Offset Limited Capacity. ZGCM-1 combines internal reasoning, external tools, and a 256K context. Its 16K pretraining reaches the...

---

XConf Cuts Costs Tenfold; Agora Logs 165 Replications

September 19, 2026

Confidence Needs a Track Record: XConf matched or beat 10-sample self-consistency in 23 of 24 comparisons at one-tenth the generation cost. Scientific Code...

---

SAS Learns What Sparse Attention Should Keep

September 17, 2026

SAS Trains Sparse Selectors on the Final Task Loss: Continuous scores enter attention logits, bypassing hard Top-K's gradient block. Its advantage grows as...

---

Equivalent Tasks Break Reasoning, Agents Reallocate Training

September 16, 2026

Structural Equivalence Does Not Guarantee Equal Performance. A model can solve the original yet depend on its wording. Grouped isomorphic variants test...

---

The First Query Sets the Search Trajectory

September 15, 2026

Panoramic State And Local Rendering Can Be Optimized Separately. AlayaVista preserves off-screen content in a 360-degree latent state, then generates each...

---

NCP Nearly Halves Training Tokens as T1 Reaches 64%

September 14, 2026

Dynamic Gating Separates Reasoning Flaws From Ordinary Language. NSD identifies reasoning-critical tokens automatically, then limits negative-training...

---

VLMs Control Robots, World State Accuracy Hits 98%

September 11, 2026

Semantic Action Interfaces Unlock Existing VLMs for Robot Control. Closed models work zero-shot, while small open models need only hours of GPU fine-tuning....

---

A 4B Model Gains 5.93 Points as AuK Runs 4.5× Faster

September 10, 2026

NeoHorse-1 Turns Routing Logs Into a Training Flywheel. Its 4B model’s macro average across 11 benchmarks rose from 58.94 to 64.87, offering a measurable...

---

Iris Turns Links Into Questions, Layer Dropout Saves 25%

September 9, 2026

Iris Turns Web Links Into Training Questions. Multi-hop entity chains and difficulty filters produce fresh supervision for iterative “SFT-RL climbing.”...

---

Active Learning Cuts Retraining Time by 49%

September 9, 2026

Active Learning Needs Stage-Aware Training: Retrain from scratch early, then fine-tune from checkpoints after stabilization. HybridAL cuts retraining time by...

---

Neural Functions Hit 83.6%, Random KV Eviction Boosts Throughput

September 7, 2026

Natural-Language Specifications Can Compile Into Local Neural Functions. Compile by Training spends about one minute training reusable capabilities and...

---

Repositories Become Skills as Models Manage Their Own Context

September 6, 2026

Repository Skills Boost Research Agents. DisCo extracted over 5,000 skills from 1,000 machine learning repositories. It raised MLE-bench scores by 134.3% and...

---

Production Traffic, Video Pretraining, and Smarter Distillation

September 4, 2026

Production Traffic Should Set Post-Training Priorities: One self-hosted model now handles 116 million monthly requests after training against demand from...

---

Stable Agent Protocols and 18% Cheaper Loops

September 4, 2026

Separating Control Flow From Prompts Delivers 100% Protocol Validity. Multi-agent systems can improve content without corrupting routing, formatting, or...

---

OPSA Lifts AIME24 by 35.41 Points

September 2, 2026

Bigger Teachers Do Not Guarantee Better Distillation Feedback. Teacher scores in OPD can grow noisier with scale, while teacher-free OPSA lifts AIME24 by...

---