LLMs & Models
Large language models: GPT, Claude, Gemini, Mistral and open weights.
97 published articles
AI Research
The AI that tracks what students skip, not just what they click
MR-ConceptGCN combines personal knowledge graphs with multi-relational GCNs and SBERT embeddings to preserve the difference between a student's correct answers and skipped content. In a 31-person study, it improved recommendation accuracy, usefulness, diversity, and satisfaction over baseline approaches.
2026-07-26
Cost vs. capability
The $1.40 model that crushed its $2.82 sibling on physics
Four AI models were asked to build interactive destruction scenes with real physics. Opus 5 passed all three at the lowest cost among top performers, while Fable 5 failed each one at twice the price. A reminder that price tags don't predict physics.
2026-07-25
Agentic RL research
RL's sparse-reward blind spot meets SEED: agents write their own lessons
SEED (Self-Evolving On-Policy Distillation) lets an LLM analyze its own past trajectories, extract reusable natural-language skills from them in hindsight, and distill those lessons back into its policy during RL. The result: denser, on-policy supervision that improves success rates by up to 22% on long-horizon tasks.
2026-07-25
AI Models
Claude Opus 5 nearly matches Fable 5 across benchmarks, with cybersecurity as a deliberate blind spot
Claude Opus 5 launches with near-Fable-level scores on coding and knowledge benchmarks at half the price. But its deliberately weakened cybersecurity capabilities create a clear trade-off for developers.
2026-07-24
Document AI
The 0.8B model that beat every pipeline document parser at its own game
Alibaba's OvisOCR2, a compact 0.8B end-to-end model, achieves state-of-the-art scores on OmniDocBench (96.58) and PureDocBench (75.06), overtaking larger pipeline-based parsers. Its success comes from a data engine mixing filtered real documents with synthetic pages, reinforcement learning on a 4B teacher, and on-policy distillation into the small model.
2026-07-24
AI Research
Four minds, one answer: why AI that thinks differently beat the biggest models at humanity's hardest test
PoTRE breaks inference into four agents working in parallel, then reconciles their answers dynamically. It hit 49.92% on Humanity's Last Exam, beating the previous official best. The approach works with fewer tokens than scaled-up baselines.
2026-07-24
Alibaba Cloud's persistent memory for AI coding
The AI coding partner that finally remembers what you told it last week
Alibaba Cloud's new framework gives AI coding assistants a long-term memory that remembers preferences, history, and project knowledge across sessions. No more starting from zero every time.
2026-07-24
Multimodal AI
The three-stage rhythm that stops AI from seeing things that aren't there
New research reveals a stable three-stage redistribution of multimodal attention in VLMs, operationalized as the Visual Relay Window (VRW). The TRACE framework uses lightweight trained modules to schedule this window per task, improving grounding-sensitive benchmarks by 4.33 points on average and up to 6.6 points.
2026-07-23
LLMs & Inference
Speculative decoding is the nearest free lunch in LLM inference
Speculative decoding accelerates autoregressive generation by proposing candidate tokens with a cheap drafting mechanism and verifying them in a single target-model forward pass, without changing the output. This analysis breaks down the math, the main methods (draft models, EAGLE-3, DFLASH, Multi-Token Prediction, n-gram), and what acceptance rates actually mean for real-world latency.
2026-07-21
Formal agent orchestration with stack-based execution
The 96% context cut that finally makes AI agent tool hierarchies work
A hierarchical, skill-based architecture for AI agents solves the bottleneck of flat tool registries. By organizing capabilities as a tree and using a stack-based execution loop, it scales sub-linearly with tool count while providing execution isolation and auditability for regulated domains like digital payments.
2026-07-21
AI Research
Qwen just taught diffusion models a trick from the LLM playbook: RL beats supervised fine-tuning
The Qwen-Image-2.0-RL report details a post-training pipeline that combines RLHF, on-policy distillation, and composite reward models to improve text-to-image and image editing quality. The approach yields measurable gains across aesthetic quality, instruction following, and face identity preservation, borrowing techniques from LLM alignment research.
2026-07-20
Alibaba Qwen-Music Model and Melody Chain-of-Thought
Alibaba's Qwen-Music composes melodies first, then fills in the band, and out-sings Suno
Qwen-Music breaks new ground by explicitly planning melodies before generating full songs, a technique the team calls Melody-CoT. The model achieves state-of-the-art results on 13 of 16 objective metrics and, in blind tests, wins against Suno V5 and MiniMax Music 2.6.
2026-07-20