← Lila Sciences Interview Insights

Lila Sciences·Software Engineer·Technical Phone Screen·Senior

Senior
May 2026

Summary

Research engineer interview at Lila Sciences focused heavily on LLM-driven scientific automation, specifically around atomic manipulation and crystal structure tasks. The technical depth surprised me a bit, less coding and more conceptual design around AI systems for physical science.

Questions Asked (3)

Q1

How would you use LLMs to drive automation in experimental or atomic manipulation workflows, and what are the key limitations you'd need to address?

System DesignTechnical Trade-offsAdaptability & Ambiguity
Author's notes

This is where I stumbled a bit.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by framing LLMs as a high-level reasoning layer that orchestrates experimental workflows, not as direct controllers of atomic manipulation. Then walk through a concrete architecture (e.g., LLM parses protocols, generates control code, and adapts to feedback) while explicitly calling out limitations like hallucination, latency, and safety. Close by proposing mitigation strategies such as simulation, human-in-the-loop, and formal verification.

Pro tip: Emphasize that LLMs should be used for planning and code generation, but the actual atomic operations must be validated by physics-based simulations and real-time control systems. Show you understand the boundary between AI and deterministic control.

1. Clarify the workflow and LLM's role

Define the experimental or atomic manipulation workflow (e.g., scanning probe microscopy, optical tweezers) and identify where LLMs add value: protocol translation, parameter optimization, or anomaly detection. Avoid claiming LLMs directly control hardware.

2. Propose an architecture

Describe a layered system: LLM as a planner that generates high-level steps or code, a middleware that translates to device-specific commands, and a low-level controller for real-time execution. Include feedback loops where experimental results are fed back to the LLM for iterative refinement.

3. Address key limitations

Discuss limitations such as hallucination (generating unsafe or invalid commands), latency (LLM inference too slow for real-time control), lack of physical grounding, and data scarcity for fine-tuning. Also mention safety and reproducibility concerns.

4. Propose mitigations and safeguards

Suggest solutions: use LLMs only for offline planning, validate outputs with physics simulators, implement human-in-the-loop approval, and constrain generation with formal grammars or domain-specific languages. Highlight the need for extensive testing and fail-safes.

5. Conclude with trade-offs and future work

Summarize the trade-off between automation and reliability, and suggest incremental adoption starting with low-risk tasks. Mention potential improvements like fine-tuning on experimental data or integrating with reinforcement learning.

Key Points to Mention

  • LLM as a high-level orchestrator, not a real-time controller
  • Use of code generation (e.g., Python for instrument control) with validation
  • Feedback loops from experimental results to refine LLM prompts or fine-tune models
  • Safety and reliability: simulation, formal verification, and human oversight
  • Latency and determinism challenges in atomic-scale manipulation
  • Data scarcity and the need for domain-specific fine-tuning or retrieval-augmented generation

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.

Q2

How would you introduce causal reasoning and physical constraints into an LLM-based agent that operates in a scientific domain?

System DesignTechnical Trade-offsAlgorithms & Data Structures
Author's notes

Trickier than it sounds.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying the scientific domain and the agent's tasks, then propose a hybrid architecture that combines LLM flexibility with symbolic causal models and physics-based simulators. Emphasize iterative refinement: use the LLM to generate hypotheses, validate them with causal inference and physical constraints, and feed results back to improve the agent.

Pro tip: Ground your answer in a concrete example from a scientific field (e.g., drug discovery or materials science) to show practical understanding, and discuss how you'd measure success (e.g., reduced invalid hypotheses, faster convergence).

1. Clarify domain and requirements

Ask questions to understand the specific scientific domain, the types of causal relationships, and the physical constraints involved. Identify the agent's goals and success metrics.

2. Design hybrid architecture

Propose integrating the LLM with a causal reasoning module (e.g., causal graphs, do-calculus) and a physics simulator or constraint solver. Define interfaces for the LLM to query these modules.

3. Implement constraint checking and feedback

Use the physics simulator to validate LLM-generated hypotheses or plans, and employ causal inference to test relationships. Provide feedback to the LLM to iteratively refine outputs.

4. Address trade-offs and scalability

Discuss trade-offs between accuracy and computational cost, and how to scale the approach. Consider caching, approximate simulators, or learned surrogates.

5. Evaluate and iterate

Define evaluation metrics (e.g., validity of hypotheses, discovery rate) and a plan for continuous improvement, including human-in-the-loop validation.

Key Points to Mention

  • Causal inference techniques (e.g., causal graphs, do-calculus, counterfactuals)
  • Physical constraints representation (e.g., differential equations, conservation laws, simulation)
  • Hybrid AI architecture combining LLMs with symbolic reasoning and simulators
  • Iterative refinement and feedback loops between LLM and constraint checkers
  • Trade-offs between computational cost and accuracy, and scalability considerations
  • Evaluation metrics and validation against ground truth or expert review

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.

Q3

How would you design an evaluation benchmark to assess whether an AI agent can reason logically about crystal structure operations?

A/B Testing & ExperimentationSystem DesignProduct Analytics & Metrics
Author's notes

Probably my best answer of the session.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying the goal: to measure an AI agent's logical reasoning about crystal structure operations, not just its knowledge. Then propose a benchmark that combines synthetic tasks with known ground truth, real-world crystal data, and a scoring rubric that isolates reasoning from memorization. Emphasize iterative design, validation, and metrics that reflect practical utility for materials science.

Pro tip: Frame the benchmark around failure modes you expect (e.g., confusing symmetry operations or misapplying translations) and design tasks that explicitly test those, because a benchmark that only confirms success is less informative than one that diagnoses errors.

1. Define the reasoning scope and operations

Specify which crystal structure operations (e.g., symmetry transformations, unit cell manipulations, coordination analysis) and logical reasoning types (deduction, induction, analogy) the benchmark will assess. Ground this in real tasks from materials science to ensure relevance.

2. Design task types with controllable complexity

Create a hierarchy of tasks: from simple (e.g., identify if a transformation is valid) to complex (e.g., predict the outcome of a sequence of operations). Use synthetic crystals with known symmetries to generate unlimited examples with exact ground truth.

3. Establish ground truth and scoring

For each task, define the correct answer and a scoring method that rewards correct reasoning steps, not just final answers. Include partial credit for intermediate steps to diagnose where reasoning fails.

4. Validate with human experts and baseline models

Have materials scientists review a sample of tasks for clarity and correctness. Run baseline AI models (e.g., LLMs, specialized models) to ensure the benchmark is challenging but not impossible, and to set performance expectations.

5. Iterate and expand with real-world data

After initial validation, incorporate real crystal structures from databases (e.g., ICSD) to test generalization. Continuously update the benchmark as models improve and new reasoning challenges emerge.

Key Points to Mention

  • Use of synthetic data with known ground truth to avoid ambiguity and enable large-scale generation.
  • Inclusion of both forward and inverse problems (e.g., given a transformation, infer the operation; given an operation, predict the result).
  • Metrics that go beyond accuracy: reasoning step correctness, robustness to noise, and generalization to unseen crystal systems.
  • Avoiding data leakage by ensuring test tasks are not solvable by memorization of common crystal structures.
  • Incorporation of multi-step reasoning tasks that require chaining operations, reflecting real scientific workflows.
  • Collaboration with domain experts to define valid operations and edge cases.

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.