LLMs & Models
Large language models: GPT, Claude, Gemini, Mistral and open weights.
97 published articles
Video generation
MiniMax scrapped its proven architecture to make H3 do everything
MiniMax says H3 unifies text, image, video and audio generation in one model, prices 2K output below a third of mainstream models, and plans to open the weights within days. The small print is the real story: the company abandoned the architecture that gave it an edge to get there.
2026-08-07
Small-model race
BananaMind 2 Micro: 2.9M parameters, 75B tokens, and an extreme overtraining bet
BananaMind 2 Micro will pack 2.9M parameters and train on 75B tokens, roughly 25,900 tokens per parameter. Training starts August 3, release is estimated August 4 to 6, and a BananaMind 2 Pro public preview arrives the same day. No benchmarks have been shared.
2026-08-06
AI Safety
Shieldstral, the 3B classifier that outguns models seven times its size
A 3B safety classifier matches text models nearly seven times its size and sets a new multimodal moderation state of the art, per a July 2026 arXiv paper. The trick: moderation reframed as binary question answering, trained on roughly 54.1 million samples.
2026-08-05
Artificial Intelligence
LFM2.5-Encoders make the small-model case: 3.7× faster than ModernBERT on CPU
Liquid AI's open-weight LFM2.5-Encoders make the case that production NLP belongs on small models. A 230M encoder beats ModernBERT-base on benchmarks and scans full documents in about 28 seconds on a laptop CPU, roughly 3.7× faster.
2026-08-05
AI Research
Penelope hides its reasoning in a single decoder layer to cut inference costs
Penelope, a latent-reasoning framework for decoder-only Transformers, localizes recurrent computation to a narrow decoder interval, cutting inference latency without a major accuracy hit. The paper describes a curriculum that shifts reasoning from visible tokens to an internal GRU loop.
2026-08-04
AI Research
The monitor that goes silent when AI reasoning fails
Novelis Research shows token log-probability fails as a decoder monitor for quantized reasoning models, being blind to confident loops. They introduce a calibrated e-CUSUM controller that combines uncertainty and repetition signals, achieving selectivity on GSM8K with DeepSeek-R1-Distill-Qwen-1.5B.
2026-08-03
AI Research
The LLM that outsmarted every fixed formula for JD.com warehouse allocation
An LLM trained via solver-guided reinforcement learning picks the best MIP formulation for each inventory allocation instance at JD.com. Hit Ratio@1 jumped from 21% to 50%, and realized allocation accuracy beat every fixed formulation by 12.57 percentage points.
2026-08-03
K12-Bench: New Test of AI Curriculum Understanding
Why your AI tutor can't see how math builds on itself
Peking University's K12-Bench reveals that even the best language models barely understand how school concepts connect. Scores of 57% and 46% show a blind spot in AI's ability to handle prerequisite chains, concept taxonomies, and visual grounding, skills that real tutors use every day.
2026-08-02
Artificial Intelligence
Research idea generation gets 3.89x better with IDEAgent
IDEAgent uses lineages, multi-objective feedback, and sequential memory to balance quality and diversity in LLM-generated research ideas. Tested across 32 topics in 8 CS domains, it achieves 3.89x the Yield (diverse ideas above a quality threshold) of prior methods.
2026-08-02
3D Reasoning
SceneActBench: Why even the best VLMs fail at 3D action
SceneActBench tests eleven VLM configurations on five 3D tasks in a unified agent loop. Overall scores range from 38.6 to 50.2, with no model performing consistently. The benchmark exposes a blind spot in vision-language agents: acting on full scenes, not just describing them.
2026-07-31
Research Digest
The simple topology trick that beats prompt magic in LLM induction tasks
A new methodology called Hourglass reasoning enforces strict context isolation between induction, deduction, and implementation stages, passing only a compressed symbolic rule between them. On ARC-AGI-2, it raises best-of-5 accuracy by up to 14 points over iterative refinement; on ChipBench Verilog synthesis, it nearly doubles accuracy with GPT-5.5. Ablations confirm the topology itself drives the improvement.
2026-07-31
LLM distillation
LLM distillation stalls when context goes static. Flux-OPD makes it evolve
Contexts can carry task preferences into LLM training, but once distilled into a student they add little supervision. Peking University's Flux-OPD keeps the context moving with the student and uses a conflict term to weight teacher corrections. The abstract reports gains over existing OPD paradigms without naming numbers.
2026-07-31