Generative AI for Autonomous Driving

Describe how tokenization converts text into sequences of integer token IDs

Explain how embedding layers map token IDs to dense vector representations

Define autoregressive language modeling as the task of predicting the next token given all previous tokens

Compute the softmax function to convert raw logits into a probability distribution over the vocabulary

Compute the cross-entropy loss for a language model's next-token predictions

Compute scaled dot-product attention given query, key, and value matrices

Explain the role of the scaling factor in preventing vanishing gradients through softmax

Describe self-attention and how it allows each token to attend to all other tokens in a sequence

Explain causal (masked) self-attention and why it is required for autoregressive models

Describe multi-head attention and how it enables attending to different representation subspaces simultaneously

Explain the role of learned projection matrices W_Q, W_K, W_V, and W_O in multi-head attention

Describe sinusoidal positional encodings and why position information must be explicitly injected

Describe the structure of a transformer block: self-attention, feed-forward network, residual connections, and layer normalization

Explain LayerNorm and how it stabilizes training in deep networks

Distinguish encoder-decoder transformers from decoder-only transformers and identify their typical use cases

Describe how decoder-only models like GPT perform text generation via iterative next-token sampling

Explain the KV cache mechanism and how it avoids redundant computation during autoregressive inference

Describe the pre-training and fine-tuning paradigm for large language models

Explain RMSNorm and how it simplifies LayerNorm by removing the mean-centering operation

Describe the SwiGLU activation function and how its gating mechanism increases FFN capacity

Explain why RoPE produces attention scores that depend only on relative position, enabling context length extrapolation

Compare pre-norm and post-norm transformer block configurations and their effects on training stability

Quantify the KV cache memory cost for standard multi-head attention as a function of layers, heads, and sequence length

Describe Grouped Query Attention (GQA) and how sharing key-value heads reduces memory usage

Explain Multi-head Latent Attention (MLA) and its use of low-rank compression for the KV cache

Describe the down-projection and on-the-fly up-projection steps in MLA

Explain the decoupled RoPE strategy in MLA that separates content and positional components

Compare MHA, GQA, and MLA in terms of memory efficiency, compute cost, and model quality trade-offs

Explain conditional computation and how sparse activation allows scaling model parameters without proportional compute cost

Describe the architecture of a Mixture-of-Experts layer: router network and parallel expert FFNs

Explain the gating/routing mechanism that assigns tokens to a subset of experts

Describe the load balancing problem in MoE and the auxiliary losses used to mitigate it

Explain DeepSeek-V3's MoE design with shared and routed experts

Distinguish total parameters from active parameters per token in MoE models

Define the components of a Markov Decision Process: states, actions, transition dynamics, and rewards

Define the state-value function V(s) and the action-value function Q(s,a)

State the policy gradient theorem and explain its significance for optimizing parameterized policies

Describe the REINFORCE algorithm and how it uses sampled returns to estimate policy gradients

Explain how subtracting a baseline reduces variance in policy gradient estimates

Describe the importance sampling ratio and its role in off-policy policy gradient methods

Explain supervised fine-tuning (SFT) and its role in teaching a base model to follow instructions

Describe how reward models are trained from pairwise human preference comparisons

Outline the standard RLHF pipeline: SFT, reward model training, and policy optimization

Describe the PPO clipped surrogate objective and why clipping stabilizes policy updates

Explain the KL divergence penalty that prevents the policy from diverging too far from the reference model

Identify the four models required in PPO-based RLHF: actor, critic, reward model, and reference model

Describe Direct Preference Optimization (DPO) and how it eliminates the need for a separate reward model

Explain how GRPO eliminates the critic network by estimating advantages from group-relative reward comparisons

Describe the GRPO algorithm steps: group sampling, reward computation, advantage normalization, and clipped policy update

Write the GRPO objective function and identify the clipped importance ratio and KL divergence terms

Explain why tasks with verifiable correctness (math, code) are especially suited to rule-based rewards in GRPO

Describe the DeepSeek-R1 cold start phase: fine-tuning on chain-of-thought data to initialize structured reasoning

Explain how reasoning behaviors such as self-verification and backtracking emerge during GRPO training on R1-Zero

Describe the rejection sampling stage of the DeepSeek-R1 pipeline and its role in creating a high-quality SFT dataset

Explain the purpose of the final general RL stage in the DeepSeek-R1 pipeline for aligning the model toward general helpfulness and safety

Explain how vision encoders produce visual tokens that can be processed by a language model

Describe unified embedding spaces where text, image, and audio share a common latent representation

Compare GPT-4o's native multimodal approach with pipeline-based systems that chain separate models

Describe how domain-specific pretraining and data filtering create specialist models like Qwen 2.5-Coder and Qwen 2.5-Math

Explain the Code-as-Policy concept where a planner outputs executable code rather than raw trajectory points

Compare the traditional modular AD pipeline (perception, prediction, planning) with unified generative planning

Explain how trajectory planning can be reformulated as an autoregressive sequence modeling task

Describe VQ-VAE: the encoder, learned codebook, nearest-neighbor quantization, and decoder

Explain how VQ-VAE tokenizes Bird's Eye View (BEV) feature maps into discrete scene tokens

Describe DAP's tokenization scheme for ego history (curvature/acceleration), navigation commands, and environment

Explain DAP's autoregressive interleaving: predicting future scene tokens before generating ego action tokens

Describe how predicting the next scene state before acting effectively embeds a world model inside the planner

Explain post-tuning refinement that smooths discrete-token trajectories via jerk minimization

Explain why imitation learning on human driving data can replicate unsafe behaviors like speeding

Describe Plan-R1's two-stage framework: imitation pretraining followed by GRPO fine-tuning

Design a rule-based reward function to penalize collisions and enforce hard safety constraints

Design rule-based reward functions for lane compliance and adherence to speed limits

Design rule-based reward functions to encourage progress towards the goal and ensure passenger comfort via jerk minimization

Explain Variance-Decoupled GRPO (VD-GRPO) and why replacing normalization with centering prevents reinforcing unsafe actions

Compare GRPO and PPO for trajectory planning, explaining how verifiable rewards enable reference-free safety alignment

Distinguish open-loop from closed-loop evaluation metrics for autonomous driving planners (NAVSIM, nuPlan)