Learn the papers¶
简体中文 · English
Work from basic understanding → core mechanism → method comparison → implementation details. Try answering first, then reveal explanations and teaching examples. You can filter by level.
Study by theme¶
Use a subtopic bank for paper-by-paper study, or a deep dive for comparisons, trade-offs, and cross-paper Q&A.
6 themes · 31 subtopics · 4253 questions across 1061 papers
Search, ads, and recommendation
Follow the path from finding candidates to ranking and serving: semantic representations, user knowledge, long histories, generative recommendation, feedback, and deployment.
Open subtopic question banks
- Candidate generation and search matching148 questionsDeep dive
- Semantic representations, user knowledge, and cold start240 questionsDeep dive
- User histories and ranking backbones299 questionsDeep dive
- Generative retrieval and reasoning-based recommendation248 questionsDeep dive
- Learning signals, multitask learning, and feedback bias144 questionsDeep dive
- Slates and business control: reranking, bidding, and long-term value148 questionsDeep dive
- Automated research and efficient serving76 questionsDeep dive
Foundation models
Understand context access, allocation of compute and training budgets, and trade-offs in compression, generation, and evaluation.
Open subtopic question banks
- Long context: attention, recurrent state, and memory136 questionsDeep dive
- Conditional computation: experts, elastic depth and width, and lookup memory104 questionsDeep dive
- Training budgets: data mixtures, objectives, and optimizers120 questionsDeep dive
- Inference compression: weights, KV cache, and serving bandwidth128 questionsDeep dive
- Generation paths: draft verification, parallel denoising, and few-step refinement60 questionsDeep dive
- Capability evidence: sampling, readout, and reliability56 questionsDeep dive
Structured Decisions
Use language models for decisions over finite candidates: start with distribution readout, then test invariance, reasoning budgets, and ongoing feedback.
Open subtopic question banks
Multimodal models
Study three questions: how to align modalities, how to budget visual evidence, and how to predict states and the consequences of actions.
Open subtopic question banks
LLM post-training
Start by identifying the feedback source, then study policy updates, credit assignment, teacher supervision, and data and compute allocation.
Open subtopic question banks
- Preference and alignment: identify the feedback first100 questionsDeep dive
- Policy optimization: which policy samples and which policy learns?220 questionsDeep dive
- Reward and credit: which steps deserve a good final score?148 questionsDeep dive
- Trajectory distillation: students, proposals, and prefix distributions168 questionsDeep dive
- Teachers and correction: when extra context can help188 questionsDeep dive
- Data and budgets: where to spend the next unit of compute152 questionsDeep dive
Agents
Follow the path from the next action to a complete execution workflow: planning, memory, tool collaboration, self-improvement, training, and independent evaluation.
Open subtopic question banks
- Planning and search: what should happen next?64 questionsDeep dive
- Memory and skills: retrieval, reuse, and invalidation228 questionsDeep dive
- Tools and collaboration: turning plans into inspectable actions197 questionsDeep dive
- Harnesses and self-improvement: which layer changes?188 questionsDeep dive
- Interactive policy learning: converting trajectory feedback into updates212 questionsDeep dive
- Evaluation and reliability: how do we know it worked?176 questionsDeep dive
Recently added¶
- When the draft uses top-k support, how does renormalizing the teacher there change the comparison?2026-11-04 · specopd
- Why does reordering identical token-overlap values change the SpecOPD loss?2026-11-04 · specopd
- Why is SpecOPD on-policy if its reference remains fixed?2026-11-04 · specopd
- Why check prediction rows separately when the proposal tree already has a causal mask?2026-11-04 · specopd
- Does RLiD's variance theorem prove a newly trained LDM must beat a generative LLM?2026-11-04 · rlid
- Do executable menus and clipped PPO complete RLiD agent training?2026-11-04 · rlid
- Is KL from uniform to the menu marginal equivalent to making every decision more random?2026-11-04 · rlid
- How does a RLiD action differ fundamentally from an LLM writing an action command?2026-11-04 · rlid