These questions test self-awareness, learning ability, and reflective judgment. Use specific examples rather than abstract traits. ## 1. Three words that describe you Choose traits that are true and role-relevant. Example: ### Curious "I tend to dig into why a system or model behaves the way it does. For example, after reading about contrastive learning, I reproduced a simplified experiment on a small dataset to understand why negative sampling affected results." ### Reliable "When I own a project, I try to make progress visible and risks explicit. In one project with a fixed deadline, I set up weekly milestones, tracked blockers, and raised a data dependency risk early enough for the team to adjust scope." ### Data-driven "I try not to rely on intuition alone. In a model-tuning project, I ran ablation tests to compare feature groups and found that a simpler model achieved nearly the same accuracy with lower latency." The exact words should match your real experience. Other good choices include structured, collaborative, resilient, customer-focused, or pragmatic. ## 2. Goal planning Use a framework and cadence. Example answer: "I set goals using a SMART or OKR-style structure. I start by defining the outcome, then break it into milestones and leading indicators. For a model-improvement goal, the objective might be to improve production-quality performance, and the milestones could be data audit, baseline reproduction, feature experiments, error analysis, and deployment review. I track progress weekly, review what is blocked, and adjust scope if the data shows the original plan is not realistic. At the end, I write a short retrospective so I can improve the process next time." Key points: - measurable goal, - milestones, - review cadence, - adjustment, - reflection. ## 3. Recent paper Use a simple structure: 1. Problem. 2. Key idea. 3. Results. 4. Limitations. 5. What you learned. Example: "A recent paper I read focused on improving long-sequence modeling. The problem was that standard self-attention becomes expensive as sequence length grows. The key idea was to reduce attention cost by using more efficient attention patterns or chunked representations. The paper compared against standard Transformer baselines and showed better speed or memory usage while preserving accuracy on long-context tasks. My takeaway was that architecture choices should match the sequence length and task structure; for some problems, preserving all pairwise token interactions is less important than designing a representation that captures the right long-range signals." If using your own paper, replace this with specific details and what you tried afterward. ## 4. Competition reflection The best answer shows humility and concrete learning. Example: "In one competition, my solution performed well but did not win. After reviewing the top solution, I saw three differences. First, the winner had stronger validation: they used a split that matched the test distribution, while my validation was slightly too optimistic. Second, their feature pipeline was more systematic; they ran ablations and knew which features mattered. Third, their code was more modular, so they could run more experiments quickly without breaking the pipeline. The lesson I took was that winning was not just about choosing a stronger model. It was about experiment discipline, leakage prevention, and engineering speed. After that, I changed my workflow: I separated data processing, features, and models; added config-driven experiments; and started tracking ablation results more consistently." Strong points to mention: - better validation, - better features, - stronger model choice for the data, - cleaner code, - faster experimentation, - better error analysis, - no leakage, - better ensembling or calibration. ## Common pitfalls - Choosing generic traits with no examples. - Saying goals are tracked "in my head." - Summarizing a paper without showing what you learned. - Blaming luck or resources for a competition result. Good answers show that you can learn deliberately and improve your process.