体系的な学習ルート
読む → 実装する → 評価する
最先端の研究
研究のアイデアから実装・検証へ。厳選した8記事。
注目記事Robot recovery: when retrying helps, and when it hides failure
Posted
Read Backtrack, DexPort and ViT-H through a practical lens: frozen policies, action labels, executable motion and bounded recovery. Includes a local state-machine experiment.
研究ノートSelf-improving agents: reward gains are not capability gains
Posted
A research reading guide and local experiment on adaptive harnesses, reward collapse, frozen-policy comparisons and held-out evaluation.
研究ノートFrom LLM Calls to Agent Harnesses
Posted
An English reading brief and hands-on tutorial on agent runtimes: context, tool permissions, checkpoints, bounded execution and reproducible failure tests.
研究ノートTest-time compute: search, verification and budgets
Posted
Prerequisites: probability, evaluation and inference cost. Learn when more attempts help and when a verifier becomes the bottleneck.
研究ノートSelective state spaces: Mamba and sequence memory
Posted
Prerequisites: linear algebra, recurrence and attention. Compare compressed state with explicit token retrieval.
研究ノートDiffusion language models: masking and iterative decoding
Posted
Prerequisites: cross-entropy, Transformers and generative modeling. Understand LLaDA-style masked generation without equating parallelism with speed.
論文解説DeepSeek-R1: rewards, reasoning and what the evidence establishes
Posted
Study R1-Zero and R1 separately, then build a controlled test of reasoning improvement. This analysis is pinned to arXiv v1.
論文解説LLM-as-a-judge: evaluate the evaluator
Posted
Separate answer preference from correctness and reveal how order effects can alter a reported win rate.
学習ライブラリ
学習ルートを選びましょう。
最新のチュートリアル
ライブラリの新着記事
Write Multi-Head Attention from scratch
Trace Q, K and V, split and merge heads, then prove your causal mask works.
記事を読む ↗Softmax that stays stable
Derive the gradient, work through a small example and avoid overflow.
記事を読む ↗K-Means beyond the happy path
Implement assignment and updates; handle empty clusters and convergence.
記事を読む ↗Adam, one update at a time
Understand persistent moments, bias correction and the first parameter update.
記事を読む ↗Beam Search without hidden mistakes
Keep global top-k candidates, track EOS and compare against greedy decoding.
記事を読む ↗ローカルで実験
検証したいことからツールを選びます。
入力はブラウザ内で処理します。APIキー不要、入力のアップロードなし。
面接とプロジェクト
暗記ではなく、実装と根拠で理解を確かめる。
最近の演習
面接対策と実践プロジェクト
Handwrite Beam Search, then defend every choice
Practice global top-k selection, EOS handling and length bias with a worked example and reference code.
演習を見る ↗INTERVIEW / OPTIMIZATIONDerive Adam and trace the first update
A paper-and-pencil drill in moving averages and bias correction, followed by a tested implementation.
演習を見る ↗PROJECT / TRANSFORMERSProve your KV cache is correct
Build causal attention and compare full-sequence outputs with incremental decoding using numerical invariants.
演習を見る ↗PROJECT / RAGBuild retrieval with a measured baseline
Measure retrieval accuracy, answerable recall and correct abstention separately; inspect the failure cases.
演習を見る ↗