Build a reasoning model
Spend compute at inference, score answers reliably, then change the weights with RL or distillation.
See what changes. Understand why.
Explore the geometry, memory, and learning signals behind modern models.
The Transformer architecture
The complete encoder–decoder, grounded in the original paper.
Explore the architecturePreparing the 3D view…
LoRA update space
Change the adapter rank and watch its outputs open from a line into a plane and a volume.
Explore in 3DPreparing the 3D view…
Rotary position encoding
Turn three coordinate pairs at different frequencies and inspect how position changes their phase.
Explore in 3DPreparing the 3D view…
KV-cache memory
Step through decoding and see keys and values accumulate in separate banks for each layer.
Explore in 3DPreparing the 3D view…
PCA projection
Move a three-dimensional point cloud onto its first principal component and inspect the lost directions.
Explore in 3DSupervised fine-tuning
Choose the demonstrated response. Watch cross-entropy move probability toward it, one gradient update at a time.
Explore the mechanismDirect preference optimization
Pick a preferred and rejected response. Compare their log-probability ratios against a frozen reference.
Explore the mechanismKnowledge distillation
Reveal a teacher’s softened distribution and watch a student learn its relative preferences.
Explore the mechanismPPO clipping
Change the clip range and inspect where the policy surrogate stops rewarding movement in the same direction.
Explore the mechanismGroup-relative optimization
Score six candidate responses, normalize their rewards, and trace the clipped policy update.
Explore the mechanismPolicy-gradient bandit
A seeded agent samples actions, receives their rewards, and updates its categorical policy with REINFORCE.
Explore the mechanismTeacher-to-reward guidance
Explore ATOD’s annealing and turn-level gates using editable synthetic signals.
Explore the mechanismFollow an idea into the details.
Worked examples that connect the picture, the calculation, and the implementation.
How current models make attention cheaper
Separate efficient kernels, recurrent state, sparse selection and hybrid architectures.
From tokens to a prediction
Follow the computation, then inspect the operations that make it work.
Change a model. Understand the constraint.
Connect gradients, low-rank updates, and the geometry of projection.
Where the ideas are going.
Further reading discovered through alphaXiv.
ATOD: teacher guidance that changes with training
The Studio preset exposes the schedule and turn-weighting mechanism on synthetic inputs. It does not reproduce the training runs.
Explore the interactive mechanismRead on alphaXivDistilled Reinforcement Learning for LLM Post-training
This July 2026 paper, revised in September, investigates selective teacher signals inside reward-based learning. It is further reading, rather than an implemented training algorithm in Studio.
Discover on alphaXivRead the primary paper