Weekly Research Intelligence — OR for Generative AI

Adaptive Resource Allocation and Policy Optimization for LLM Inference and Reasoning

This research front explores advanced Operations Research (OR) and AI techniques to optimize various aspects of Large Language Model (LLM) systems, spanning inference, serving, and reasoning. A core theme is the application of mathematical optimization (e.g., Mixed-Integer Programming, Linear Programming, gradient-based scheduling) and reinforcement learning (e.g., GRPO, DynaMO, BandPO) to enhance LLM efficiency, performance, and decision-making. Key frameworks include GRPO for policy optimization, Puzzle (LANA) for NAS, DORA (REBASE) and ETS for test-time search and KV cache management, and Entropic-Time Inference for dynamic compute allocation.

Significant contributions include the Solver-in-the-Loop framework ([1]) using GRPO and Process Reward Models (PRM) for iterative OR model debugging, achieving +11.1% RR@5 over API models. Puzzle ([2]) leverages MIP and blockwise distillation for NAS, yielding 2.17x throughput speedup for Llama-70B with 98.4% accuracy retention. DORA ([3]) optimizes test-time search via embedding-based clustering and resource allocation, outperforming baselines with 3.5x fewer FLOPs on MATH500. Entropic-Time Inference ([4]) dynamically allocates compute based on entropy, showing 30-45% throughput increase over vLLM. BandPO ([12]) improves GRPO by 2-10% on math benchmarks by addressing policy update clipping. Other notable works include VectorLiteRAG ([6]) for RAG serving, GoodSpeed ([7]) for distributed edge inference, and PROBE ([11]) for MoE inference, all demonstrating substantial efficiency gains.

This research front is rapidly emerging and growing, characterized by a strong interdisciplinary fusion of OR and AI. The emphasis on verifiable rewards, solver-in-the-loop evaluation, and mathematically grounded resource allocation indicates a shift towards more robust and efficient LLM systems. The next wave of papers will likely focus on integrating these fine-grained optimization techniques across the entire LLM lifecycle, from training and architecture design to inference and complex reasoning. We can expect more sophisticated RL-infused evolutionary search frameworks that dynamically adapt compute based on uncertainty (e.g., Entropic Time, DynaMO's variance-based allocation) and MILP/ILP-driven control planes for real-time resource management in distributed LLM serving.

AlphaEvolve-Style LLM Evolution for OR Heuristics and Scientific Code

This research front focuses on advancing LLM-driven evolutionary computation and program synthesis for complex Operations Research (OR) and AI for Science (AI4S) problems. The core theme revolves around developing sophisticated frameworks that leverage large language models to automatically generate, optimize, and refine code, heuristics, and scientific models. Key approaches include multi-agent systems, advanced feedback mechanisms, and novel strategies to improve sample efficiency and generalization in LLM-guided search.

Key contributions include ParEVO, which achieves up to 106x speedup in parallel algorithm synthesis for irregular data using MAP-Elites and DPO fine-tuning. AILS-AHD demonstrates 8 new Best-Known Solutions on CVRPLib by evolving ruin heuristics with LLM-driven Chain-of-Thought voting. DHEvo introduces a data-algorithm co-evolution framework for MILP, outperforming FunSearch by ~60% on Setcover. ConstraintLLM fine-tunes LLMs for Constraint Programming modeling, achieving ~51% accuracy on industrial benchmarks via structural retrieval and iterative self-correction. BAMF-AI4S significantly boosts scientific code generation (+54.3% on HumanEval) using Bayesian surrogate models, while StitchCUDA automates GPU program generation with rubric-based reinforcement learning, yielding 1.72x speedup over baselines. MAESTRO employs an 'Exploration Report Agent' for materials discovery, and NetGPT uses SFT-anchored RL to prevent schema collapse in LLM routing.

This front is rapidly emerging and growing, marked by several

Formal Optimization, Pruning, and Risk Aversion in Multi-Agent LLM Systems

This research front focuses on enhancing the robustness, efficiency, and coordination of multi-agent LLM systems through a combination of formal optimization, test-time pruning, and risk-averse policy learning. Key frameworks include AgentDropoutV2 for error mitigation, BAMAS for budget-constrained structuring, SRPO for partner generalization, and GEPA for system-level prompt optimization.

AgentDropoutV2 demonstrates significant accuracy gains (+7.92% math) by rectifying erroneous outputs using a failure-driven indicator pool. BAMAS achieves substantial cost reductions (~80%) on benchmarks like GSM8K and MBPP by employing Integer Linear Programming for LLM selection and offline reinforcement learning for topology. SRPO eliminates free-riding equilibria in collaborative tasks, improving joint accuracy by up to 19% on GSM8K multi-agent debate. MAMUT GEPA optimizes multi-agent conversational assistants, yielding a +7.6% rubric pass rate by jointly evolving prompt bundles and calibrating the LLM judge.

This front is rapidly growing, demonstrating a clear trajectory towards more principled and robust multi-agent LLM design. Future work will likely focus on integrating these diverse optimization techniques, exploring dynamic adaptation during execution, and extending their application to more complex, real-world agentic AI settings and human-AI collaboration.

ILP and Heuristic Optimization for Disaggregated LLM Serving on Heterogeneous Distributed GPUs

This research front centers on the application of Integer Linear Programming (ILP) and advanced heuristic algorithms to optimize resource allocation for Large Language Model (LLM) serving, particularly within disaggregated prefill/decode architectures and heterogeneous, geographically distributed GPU clusters. Key frameworks like Helix, SageServe, Cascadia, and Dynamo are developed to address challenges in dynamic resource provisioning, multi-tenancy, and carbon-aware caching, moving beyond reactive baselines to proactive, optimized solutions.

Papers in this front present diverse contributions, including SageServe's ILP and ARIMA-based forecasting achieving 25% GPU-hour savings on Microsoft O365 traces, and Helix's max-flow MILP formulation yielding up to 3.3x decode throughput on heterogeneous L4/T4/A100 clusters by enabling per-request pipeline scheduling. Cascadia employs bi-level optimization (MILP + Chebyshev) for LLM cascade serving, showing 2.3x throughput gains over SGLang. Other works like AMPD (Dynamo framework) demonstrate 67-340% SLO attainment improvements for multi-round inference via ILP-based offline planning and adaptive routing, while MIGRator uses ILP for dynamic NVIDIA MIG reconfiguration, improving 'Goodput' by ~20%. GreenCache introduces carbon-aware ILP for KV cache management, reducing carbon emissions by ~15%.

This front is maturing rapidly, driven by the critical need for efficient LLM serving infrastructure. The trajectory indicates a shift from foundational ILP formulations to more sophisticated hybrid approaches, integrating RL (TORTA) or advanced heuristics for dynamic, real-time adaptation. Future work will likely focus on scaling these OR solutions to extremely large clusters, accommodating more complex multi-tenant and multi-modal workloads, and integrating with emerging hardware capabilities like CXL memory, while also refining predictive models for greater robustness against real-world variability.

Reinforcement Learning and Fine-Tuning for Automated OR Modeling

This research front focuses on the intersection of Large Language Models (LLMs) and Operations Research (OR), primarily exploring how LLMs can be fine-tuned and reinforced to automate OR modeling and optimize instruction learning. Key frameworks like ORLM and OR-R1 leverage instruction tuning and test-time reinforcement learning to generate optimization models, while methods like EE-CPO apply linear programming to optimize instruction sets for LLM supervised fine-tuning. Concurrently, a significant portion of the front applies rigorous OR principles to optimize LLM inference scheduling, addressing challenges like KV cache management and throughput.

Specific contributions include ORLM-LLaMA-3, which, trained with the OR-Instruct framework, significantly outperforms GPT-4 on benchmarks like NL4OPT and IndustryOR for automated optimization modeling. OR-R1 further advances this by integrating Supervised Fine-tuning with Test-Time Group Relative Policy Optimization (TGRPO), achieving +7.6% average accuracy over LLMOPT with 1/10th of the training data and improving Pass@1 consistency. In instruction optimization, EE-CPO, utilizing causal intervention and linear programming, consistently outperforms DEITA by up to +1.73 on AlpacaEval2.0. For LLM inference, papers introduce algorithms like Memory Constrained Shortest First (MC-SF) for near-optimal latency, Nested WAIT for multi-stage scheduling with high-probability memory bounds, Staggered Batch Scheduling (SBS) reducing Time-to-First-Token by 30-40%, and Justitia, a fair scheduler reducing average job completion time by ~60% over VTC.

This front is stable, exhibiting a dual trajectory: the application of LLMs for automated OR modeling is rapidly emerging with novel RL-based approaches, while the use of OR for LLM inference scheduling is maturing, establishing rigorous theoretical foundations and practical, high-performance algorithms. Future work will likely see a convergence, with LLMs potentially generating or optimizing OR-based scheduling policies. The next generation of papers will focus on integrating advanced reinforcement learning techniques (e.g., RLHF, DPO) for solution ranking in automated OR modeling, developing multi-agent LLM systems for complex problem-solving, and extending OR scheduling frameworks to dynamic, multi-GPU, and multi-tenant LLM inference environments.

Optimal OR for LLM Inference Scheduling with Heterogeneous Workloads

This research front unifies advanced Operations Research methodologies, including Mixed-Integer Programming, stochastic control, and fine-grained heuristic scheduling, to address the complex challenges of Large Language Model (LLM) inference and training optimization. The core theme revolves around efficiently managing computational resources, particularly GPUs, under diverse conditions such as prefill-decode contention, heterogeneous hardware, and specialized model architectures like Mixture-of-Experts (MoE) and Early-Exit LLMs.

Key contributions include Pang et al.'s hybrid offline-online MIP and Lagrangian heuristic, boosting vLLM utilization by 8.86% and reducing inference time by 10.42s. Lin et al. introduce asymptotically optimal 'Gate-and-Route' policies via stochastic control for prefill-decode contention, while Bari et al.'s RAD/SLAI schedulers achieve a 53% reduction in median TTFT and 26% capacity increase over Sarathi-Serve. She et al. apply MIP to operator-level parallelism, reducing pipeline bubbles by 50% for DeepSeek V3. FinDEP optimizes MoE inference with 1.02x-1.61x throughput gains, and HetRL leverages multi-level search for 3-9x throughput increases in heterogeneous RLHF scheduling. DREX improves Early-Exit LLM throughput by 2-12% through dynamic rebatching, and Trident integrates MILP and Bayesian optimization for up to 2.01x throughput speedup in multimodal pipelines.

This front is rapidly maturing, characterized by a strong emphasis on rigorous mathematical modeling and empirical validation. The trajectory indicates a shift towards integrating diverse OR techniques to tackle increasingly complex, real-world LLM serving scenarios. Future work will likely focus on developing unified frameworks that combine stochastic programming, adaptive control, and advanced heuristics to manage dynamic, bursty workloads across highly heterogeneous and distributed GPU infrastructures, potentially incorporating LLM-guided meta-heuristics for online adaptation.

OR-Driven Distributed LLM System Optimization for Inference and Training Efficiency

This front unifies research on applying Operations Research (OR) techniques, specifically Integer Linear Programming (ILP), Mixed-Integer Linear Programming (MILP), and Mixed Integer Programming (MIP), to optimize distributed Large Language Model (LLM) serving and training infrastructure. Key systems like DualPath, MoETuner, FlexSP, and MEMO address critical bottlenecks in agentic inference, Mixture-of-Expert (MoE) serving, flexible sequence parallelism, and ultra-long context training by formulating resource allocation and scheduling as optimization problems.

DualPath ([1]) introduced a dual-path KV-Cache loading architecture with CNIC-centric traffic management, achieving up to 1.96x average online serving throughput for agentic LLM inference on datasets up to 64K MaxLen. MoETuner ([2]) leveraged ILP for balanced expert placement and token routing in MoE models, demonstrating a 17.5% end-to-end speedup on multi-node H200 clusters. FlexSP ([3]) utilized MILP and dynamic programming for adaptive sequence parallelism, accelerating LLM training by up to 1.98x on A100 clusters. MEMO ([4]) applied bi-level MIP for fine-grained tensor management, enabling 7B LLMs with 1M context on 8 GPUs and achieving up to 1.97x MFU improvement.

This front is maturing, demonstrating the significant impact of formal OR methods on complex LLM system design. The consistent use of ILP/MILP/MIP for problems like KV-cache management, expert placement, sequence parallelism, and tensor management indicates a robust and validated approach. Future work will likely focus on integrating these static optimization techniques with dynamic, online adaptation strategies, exploring approximate solvers for larger scales, and extending holistic optimization across different parallelism types. The next papers will likely combine these fine-grained OR solutions into more comprehensive, adaptive LLM serving and training frameworks.

Robust Multi-Objective LLM Alignment and Efficient Fine-Tuning

This research front explores the application of Operations Research techniques, including game theory, convex optimization, linear programming, and dynamic programming, to enhance Large Language Models. The core themes revolve around achieving robust multi-objective LLM alignment, ensuring safety and helpfulness, and optimizing the efficiency of multi-task LLM fine-tuning in datacenter environments.

Key contributions include the Robust Multi-Objective Decoding (RMOD) framework, which uses a maximin two-player game and convex optimization to improve worst-case reward by +1.2% on alignment benchmarks. The PAMA algorithm offers a computationally efficient convex optimization approach for multi-objective RLHF, demonstrating stable convergence where baselines like MGDA-UB oscillate. The Safety Game formulates LLM response selection as a zero-sum LP, achieving up to two-fold accuracy improvements on SafetyBench by dynamically balancing helpfulness and safety. Furthermore, MuxTune introduces a hierarchical scheduler leveraging dynamic programming for multi-tenant PEFT, yielding up to 5x throughput and 5.29x memory reduction compared to NeMo and S-LoRA.

This front is actively maturing, demonstrating a clear trajectory towards more robust and efficient LLM systems through principled OR methodologies. Future work will likely focus on extending these game-theoretic and optimization frameworks to handle more complex, dynamic interactions in LLM agents, such as sequential dialogues and multi-player scenarios. Additionally, there will be continued efforts to integrate these techniques with advanced cluster scheduling policies and adaptive resource management to guarantee performance SLOs and energy efficiency in large-scale LLM deployments.

MC/MC# Framework for MoE LLM Mixed-Precision Quantization and Dynamic Pruning

This research front centers on the MC and MC# frameworks, which leverage Integer Programming (IP) and Linear Programming (LP) to achieve extreme compression of Mixture-of-Experts (MoE) Large Language Models (LLMs) and Vision-Language Models (VLMs). The core unifying theme is the hybrid approach of combining static mixed-precision quantization with dynamic expert pruning, formulated as an optimization problem to minimize quantization error under memory constraints.

Key contributions include the development of Pre-Loading Mixed-Precision Quantization (PMQ) via LP and Online Top-any Pruning (OTP) via Gumbel-Softmax sampling. Paper [1] demonstrated MC# achieving 6.2x weight reduction on DeepSeek-VL2 with less than 2% accuracy loss, and PMQ showing 18.4% better performance than BSP 2.54-bit on Mixtral 8x7b LM-Eval. Paper [2] compressed Mixtral 8x7b to approximately 16GB (fitting on a single RTX 3090) with only a ~4% drop in zero-shot accuracy, significantly outperforming uniform GPTQ. Benchmarks used include C4, M4, WikiText2, and EleutherAI LM Harness.

This front is emerging, with both papers published in 2025, indicating active development in applying OR techniques to MoE compression. The trajectory suggests a focus on refining these hybrid compression strategies. Future work will likely address performance drops on challenging tasks like GSM8K and HumanEval, as well as adapting the frameworks for specific multimodal applications and optimizing them for diverse hardware platforms.