Adam: understand the update, not just the defaults
Why two moving averages can change the geometry of a training problem. Work through the first step and inspect the effect of bias correction.
Kingma & Ba · arXiv:1412.6980
Biblioteca de artículos
17 papers. Seven topics. Each analysis connects the original claim to a worked example, an interactive explanation and a falsifiable experiment.
Detailed technical articles are available in English and Chinese. This locale shows the English article with localized navigation.
Optimization & training
How updates move parameters, and why regularization changes the result.
AdamW: where you apply weight decay matters
Separate a loss penalty from parameter shrinkage, then test the difference with an intentionally unequal preconditioner.
Loshchilov & Hutter · arXiv:1711.05101
Attention & efficient inference
From token interactions to GPU memory traffic.
Transformer: turn token compatibility into information flow
Track Q, K and V, interpret a causal attention row, and distinguish parallel training from autoregressive generation.
Vaswani et al. · arXiv:1706.03762
FlashAttention: optimize data movement before counting FLOPs
See why not storing an attention matrix can matter as much as arithmetic, and reconstruct softmax one tile at a time.
Dao et al. · arXiv:2205.14135
Efficient adaptation & alignment
Low-rank updates, quantized backbones and preference objectives.
LoRA: constrain the update, not the whole model
Factor a weight update, count trainable parameters and explain the asymmetric first gradient.
Hu et al. · arXiv:2106.09685
QLoRA: a memory budget is more than weight precision
Separate stored weights, computation precision, adapters and training state. Build an honest memory ledger.
Dettmers et al. · arXiv:2305.14314
DPO: preference learning through relative log probabilities
Derive the pairwise margin and see what β changes in a single preference example.
Rafailov et al. · arXiv:2305.18290
Representations & contrastive learning
What makes a positive pair, and what a representation should preserve.
CPC: learn by identifying the right future
Treat contrastive prediction as a classification problem and inspect what negative sampling assumes.
van den Oord, Li & Vinyals · arXiv:1807.03748
SimCLR: augmentation defines what similarity means
Follow two views through a shared encoder and projection head, then audit whether the invariance is valid.
Chen et al. · arXiv:2002.05709
Evaluation & reproducible research
Calibration, judges, uncertainty and fair compute budgets.
Calibration: being right is different from knowing how often
Adjust temperature without changing class ranking, and learn why one confidence score cannot establish calibration.
Guo et al. · arXiv:1706.04599
Deep RL that Matters: a score is a distribution
Separate seed variation, implementation choices and model selection before declaring an improvement.
Henderson et al. · arXiv:1709.06560
Show Your Work: compare search budgets, not just winners
A better winning score can come from more attempts. Plot the selection process as part of the result.
Dodge et al. · arXiv:1909.03004
LLM-as-a-judge: evaluate the evaluator
Separate answer preference from correctness and reveal how order effects can alter a reported win rate.
Zheng et al. · arXiv:2306.05685
Agents & tool use
Turn observations into actions, then test the entire loop.
ReAct: close the loop between actions and observations
Follow an external-evidence loop, then add operational limits that a research demonstration does not automatically provide.
Yao et al. · arXiv:2210.03629
Reinforcement learning & reasoning
From policy updates to feedback-driven language-model improvement.
PPO: inspect the clipped objective one sample at a time
Understand the sign of advantage, the rollout-policy ratio and the limits of clipping.
Schulman et al. · arXiv:1707.06347
InstructGPT: change the training signal to change behavior
Connect demonstrations, preference labels, reward learning and policy optimization.
Ouyang et al. · arXiv:2203.02155
DeepSeek-R1: rewards, reasoning and what the evidence establishes
Study R1-Zero and R1 separately, then build a controlled test of reasoning improvement. This analysis is pinned to arXiv v1.
DeepSeek-AI · arXiv:2501.12948