Researching IT & AI Systems
Given a research topic from this domain set, produce a structured technical report:
- Frame the problem precisely (define the system, constraints, metrics).
- Survey 3-5 relevant existing methods/approaches with their tradeoffs.
- Propose or select a method, justify choice against requirements.
- Define evaluation methodology (metrics, datasets/scenarios, baselines).
- Present expected results/analysis and limitations.
Example invocation: "Оцени эффективность методов квантового отжига для задач комбинаторной оптимизации" → produce a report comparing quantum annealing vs. classical heuristics (simulated annealing, branch-and-bound, tabu search) on specific problem classes (QUBO, TSP, max-cut), with complexity/quality/runtime tradeoffs.
Progress checklist for any task in this skill:
- Step 1: Clarify the system model (actors, topology, data flow, threat model if security-related)
- Step 2: Define success metrics (latency, accuracy, throughput, privacy leakage, consensus finality time, etc.)
- Step 3: Identify and briefly describe 3-5 candidate methods/algorithms from literature
- Step 4: Compare candidates against metrics in a table
- Step 5: Select/propose approach with justification
- Step 6: Outline evaluation plan (simulation setup, datasets, baselines)
- Step 7: Summarize findings, limitations, and future directions
Topic-specific framing
- Information propagation in heterogeneous networks: model as graph with mixed node/link types (e.g., wired/wireless, varying bandwidth); use epidemic models (SIR/SIS), gossip protocols, or agent-based simulation; metrics: coverage time, redundancy, energy cost.
- Privacy in distributed data processing: consider threat model (honest-but-curious vs. malicious), compare differential privacy, secure multiparty computation, federated learning with secure aggregation, homomorphic encryption; metrics: privacy budget (ε), computational overhead, accuracy loss.
- Semantic search in corporate knowledge bases: compare embedding-based retrieval (dense vectors), knowledge graphs, hybrid BM25+semantic rerank; metrics: recall@k, latency, relevance (NDCG), handling of domain-specific terminology.
- Automatic documentation generation from source code: compare rule-based (docstring templates), LLM-based summarization, AST-driven extraction; metrics: BLEU/ROUGE vs. human-written docs, coverage, update staleness handling.
- Quantum annealing for combinatorial optimization: formulate problem as QUBO/Ising model; compare against simulated annealing, genetic algorithms, exact solvers; metrics: solution quality (gap to optimal), time-to-solution, scalability with problem size, hardware constraints (qubit connectivity, noise).
- Low-latency consensus protocols in blockchain: compare PBFT, Raft, HotStuff, DAG-based (e.g., Narwhal/Bullshark), PoS variants; metrics: finality latency, throughput (TPS), fault tolerance (f out of n), communication complexity.
Example 1: Input: "Синтез алгоритмов семантического поиска в корпоративных базах знаний" Output: Report proposing a hybrid architecture: dense embeddings (e.g., sentence-transformers fine-tuned on domain corpus) + knowledge graph entity linking + BM25 fallback for rare terms; reranking via cross-encoder; evaluation plan using held-out query set with NDCG@10 and latency budget <200ms; discussion of cold-start and vocabulary drift issues.
Example 2: Input: "Совершенствование протоколов консенсуса в блокчейн-сетях с низкой задержкой" Output: Analysis comparing PBFT (O(n²) messages, high latency at scale) vs. HotStuff (linear communication, pipelined view changes) vs. DAG-based mempool+consensus separation (Narwhal/Bullshark); proposal to adopt DAG-based mempool with HotStuff-style voting to decouple data dissemination from ordering, reducing latency; evaluation via simulated network with varying node counts (10-100) and Byzantine fault ratios.
- Always state assumptions about scale (number of nodes/users/documents) explicitly — results vary drastically.
- Prefer quantitative comparison tables over prose when comparing 3+ methods.
- Distinguish theoretical complexity from empirical/simulated performance.
- For security/privacy topics, always state the threat model before proposing a solution.
- Cite known algorithm names precisely (don't invent names for well-established protocols).
- Flag when a topic requires specialized hardware (e.g., quantum annealers) and note simulation as a practical substitute.
- Don't propose a single "best" method without tradeoff analysis — these domains always involve tradeoffs (speed vs. accuracy, privacy vs. utility, decentralization vs. latency).
- Don't skip defining metrics before comparing methods — vague claims like "more efficient" are not acceptable.
- Don't conflate simulated/theoretical quantum annealing results with real hardware results (NISQ-era limitations matter).
- Don't ignore Byzantine/adversarial scenarios when discussing consensus or distributed privacy — assume non-trivial threat models unless stated otherwise.
- Avoid overly long reports; keep each section focused and use tables/bullets over dense paragraphs.