Rising now
The live ranking for 16 Nov 2026. It changes every day; each Monday the top ten are frozen from it into the ledger.
25 ideas · new = entered since the day before
weight-space programmingin the ten
Using an LLM to directly edit neural network weights to diagnose and fix model behaviour without gradient training.
zero-knowledge proofs of machine learning inferencein the ten
Cryptographic proofs that a model inference was computed correctly without revealing the model or inputs.
recursive self-improvementin the ten
Agent autonomously improving its own capabilities through iterative learning cycles that revise its own knowledge or models.
NVFP4 quantizationin the ten
A 4-bit floating-point number format with block scaling for quantizing neural network weights and activations on NVIDIA hardware.
Agent Client Protocolin the ten
A protocol standardizing communication between code editors and AI coding agents.
pixel-space diffusionin the ten
Diffusion image generation operating directly on pixels rather than in a latent autoencoder space.
diffusion language modelin the ten
Language model that generates text by iteratively denoising masked or noisy token sequences.
driving world model
A generative model that predicts future autonomous driving sensor observations given control signals.
EEG visual decodingin the ten
Reconstructing or classifying visual content from electroencephalography brain signals.
LLM agent for roboticsnew
Placing a large language model planner in the robot control loop to compose perception and motion tools.
agent harness evolution
Iteratively optimizing the surrounding scaffolding of an LLM agent using execution feedback.
on-policy distillationin the ten
Distilling a student model from teacher signals computed on the student's own generated trajectories.
spec-driven development
Development practice where machine-readable specifications and acceptance criteria drive implementation and verification.
multi-agent coding
Running several AI coding agents in parallel to carry out software development tasks.
expert pruningnew
Removing less useful experts from a mixture-of-experts model after training to reduce its size and cost.
world action model
Models that jointly forecast future scenes and ego trajectories to support planning in autonomous driving.
specialized decision modelin the ten
A model that selects among given options instead of generating free-form text, trained for decision making.
self-evolving agents
LLM-based agents that autonomously modify their own runtime instructions or tools to adapt to changing conditions.
feed-forward 3D reconstruction
Predicting 3D scene geometry such as point maps directly in one forward pass without optimization.
looped transformer
A transformer that reuses shared blocks over repeated recurrent loops to increase depth without adding parameters.
GPU kernel optimization
Automatically writing and tuning GPU kernels for speed and correctness.
differential privacynew
Formal privacy guarantee limiting how much any single record can influence a released output.
conditional value-at-risknew
A risk measure equal to the expected loss in the worst tail of a distribution, used to penalise rare severe outcomes.
parameter-efficient fine-tuningnew
Adapting large pretrained models by training only a small number of added or modified parameters.
vLLMnew
High-throughput serving engine for large language models using paged attention and continuous batching.