Translating textual problem descriptions into mathematical programming formulations (LP/MILP/CP), including variables, domains, constraints, global constraints, and MiniZinc/CP-SAT/Pyomo-style models

Planner-verifier-debugger agents for formulation decomposition, constraint validation, solver routing, and iterative repair

RL, DPO, GRPO, SFT, and execution-reward training for generating correct and solvable optimization models

Datasets and evaluation suites for optimization modeling, solver execution, formulation correctness, and LLM's OR reasoning quality

Understanding user needs from problem descriptions, extracting entities, objectives, constraints, assumptions, and explaining modeling choices

Supporting what-if analysis, why-not questions, infeasibility diagnosis, sensitivity analysis, and iterative updates to existing optimization models

Evolutionary search where the search process itself evolves, such as evolving mutation, selection, memory, prompts, operators, or evaluation strategies

Evolving algorithms and heuristics tested on classical optimization problems, especially combinatorial optimization, routing, scheduling, packing, and search

Evolving low-level kernels, tensor programs, compiler schedules, and hardware-aware code for accelerators

Reinforcement learning methods that tune evolutionary search policies, mutation choices, population control, exploration, or selection during algorithm design

Methods that adapt, improve, or specialize algorithm-design models during inference or deployment using problem-specific feedback

Multi-agent LLM workflows that propose, critique, test, debug, and refine algorithms through structured collaboration

Evolving algorithms, hypotheses, symbolic rules, or computational procedures for mathematics, health, materials, chemistry, biology, and other scientific domains

Evolving algorithms for cloud infrastructure, scheduling, resource allocation, compilers, databases, networking, storage, and system-level optimization

Execution feedback, process reward models, self-reflection, verbal gradients, unit tests, and trajectory scoring for improving generated algorithms

Methods that reduce overfitting through diverse benchmarks, adversarial instances, validation protocols, co-evolution, and out-of-distribution testing

Batching, routing, admission control, prefill/decode scheduling, SLO-aware serving, and throughput optimization

Queueing-theoretic models, stability guarantees, request admission, latency control, and traffic shaping for LLM workloads

KV eviction, prefix caching, memory allocation, disaggregated serving, and long-context resource control

Heterogeneous GPU allocation, training/serving placement, workload packing, parallelism, and dynamic replanning

OR methods for routing, budgeting, scheduling, and topology design in multi-agent and test-time scaling systems

Optimizing reasoning budgets, rollout allocation, best-of-N, tree search, speculative decoding, and adaptive compute

Cost-quality optimization for model selection, cascades, offloading, mixture-of-experts, and multi-model serving