RI

#1ModelEquivBench: Certifying Multi-Relational Evaluation of LLM-Generated Optimization Models

2026-07-31 | University of the Chinese Academy of Sciences |2607.29431

This paper introduces ModelEquivBench, a certifying evaluation system that assesses LLM-generated optimization models across a seven-dimensional semantic profile (E0-E6) using exact-rational mathematical certificates rather than simple execution or structural matching. The results are rigorously backed by numbers on 173 Bench4Opt problems, demonstrating that standard execution-success metrics overestimate correctness by up to 49 cases per model, while structural baselines falsely reject mathematically equivalent models. The key insight is that the evaluation of generated OR models cannot be reduced to a single scalar or structural graph match; instead, using exact-rational certificates (e.g., Farkas lemma for feasible set containment) provides a mathematically sound, multi-relational profile of model correctness.

#2OptGraph: Large Language Models Enhanced Evolutionary Optimization Via Graph Retrieval-Augmented Generation

2026-07-30 | Shanghai University, Sun Yat-sen University, Guangxi University |2607.27918

OptGraph introduces a GraphRAG-based multi-agent workflow for automated operations research modeling, translating natural language problem descriptions into executable solver code. Backed by strong empirical results, it achieves an average exact accuracy of 75.4% across six benchmarks (including NL4Opt and OptMATH), outperforming recent baselines like OptiTree and Lean-LLM-OPT by approximately 9%. The key insight is the use of a dynamic, heterogeneous graph memory that links problem variants, math models, code snippets and typical errors, which is adaptively updated with execution traces and validation feedback without requiring LLM fine-tuning.

#3Large Language Model for Operations Research Formulation Selection in Multi-Warehouse Inventory Allocation

2026-07-28 | JD.com |2607.25956

Xu et al. propose an LLM-based selector trained via SFT, IPO and GRPO to route multi-warehouse inventory allocation instances to the most suitable mixed-integer programming (MIP) formulation. Backed by real-world data from JD.com, the GRPO-trained selector improves top-1 expert selection accuracy by 29 percentage points over an SFT+IPO baseline, translating to a 12.5% allocation accuracy gain over the incumbent system. The key insight is the offline GRPO metadata construction: pre-computing and caching solver evaluations (scores, rankings, best expert) for historical instances to serve as fast, reliable reward signals during iterative RL post-training, which avoids expensive solver calls in the RL loop.

#1Frontis-MA1: Training an AI4AI Model towards Recursive Self-Improvement in Machine Learning Engineering

2026-07-30 | Horizon Research, Frontis.AI, Tsinghua University, Zhejiang University, Shanghai Jiao Tong University, Georgia Institute of Technology |2607.28568

Yang et al. introduce OpenMLE, a full-stack framework that post-trains an LLM via execution-grounded RL to act as the variation engine for long-horizon evolutionary search in machine learning engineering. The results are highly rigorous and backed by extensive numbers, showing their 35B model improving from a 39.4% to 71.2% medal rate on MLE-Bench Lite, surpassing GPT-5.5. The key insight is the meta-evolutionary loop: instead of using a frozen LLM, they explicitly train atomic operators (Draft, Improve, Debug, Crossover) using RL with adaptive bounds and entropic advantages and deploy them in a search harness that uses multi-factor parent selection (quality, progress, novelty) and lazy, operator-conditioned memory synthesis. This is a for LLM evolutionary search, as the structured experience cards and RL-infused operator training directly address the community's core challenges in sample efficiency, memory and scalability.

#2MOT-SR: Multi-Objective Tool-Augmented Scientific Equation Discovery with Large Language Models

2026-07-31 | Institute of Automation, Chinese Academy of Sciences, University of the Chinese Academy of Sciences, Goethe University Frankfurt |2607.29561

Wang et al. propose MOT-SR, an LLM-based evolutionary search framework for symbolic regression that uses a meta-strategy LLM to analyze data residuals with external tools and maintains a multi-objective Pareto front to guide equation generation. The results are rigorously backed by numbers, achieving state-of-the-art NMSE on 40 benchmark tasks and a 26.8x error reduction on a real-world astrophysics dataset compared to prior LLM-SR methods. The key insight is the 'Meta Strategy Generator' which autonomously invokes external analysis tools on the residuals of the current Pareto-optimal candidates to synthesize a natural language search strategy for the generator LLM, effectively acting as an automated planner for the evolutionary process.

#3Budget-Aware LLM Discovery via Cost-Calibrated Frontier Utility

2026-07-29 | Mohamed bin Zayed University of Artificial Intelligence (MBZUAI), City University of Hong Kong, International Digital Economy Academy (IDEA) |2607.26828

This paper introduces CostAda, an adaptive controller for LLM evolutionary search that allocates compute across search frontiers based on cost-calibrated utility (progress divided by realized token cost) and remaining budget. The results are rigorously backed by empirical evaluations across 8 benchmarks using GLM-5 and GPT-5.4, demonstrating that CostAda reaches the full-budget quality of state-of-the-art baselines (AdaEvolve, EvoX) using only half the token budget. The key insight is that progress-only credit is fundamentally flawed when search actions (e.g., short refinements vs. long-context guidance) have vastly different token costs; dividing fitness gains by the log of realized cost and shifting weight from local to global progress as the budget drains drastically improves search efficiency.

#4EvoPINN: Agentic Discovery of Executable Algorithms for Physics-Informed Neural Networks

2026-07-29 | Institute of Automation, Chinese Academy of Sciences, University of Chinese Academy of Sciences |2607.26490

EvoPINN is an LLM-driven evolutionary framework that autonomously discovers novel neural architectures and training programs for Physics-Informed Neural Networks (PINNs). Backed by rigorous empirical results, the framework evaluated only 56 candidates to discover algorithms that outperformed expert-designed baselines by up to 35% on complex PDE benchmarks. The key insight is its decoupled, module-wise evolution strategy: it uses a UCB scheduler to dynamically allocate LLM mutation budget between different algorithm components, combined with AST-based structural verification to filter out cosmetic LLM edits before triggering expensive evaluations. This is highly relevant for LLM evolutionary search research, as the telemetry-guided generation and modular credit assignment directly address the sample efficiency and sparse reward challenges inherent in automated heuristic discovery.

#5FunL2O: LLM-Guided Feature Function Design for Learning to Optimize

2026-07-29 | IBM Research, University of Southern California, University of Wisconsin–Madison, Michigan State University |2607.27389

FunL2O uses an LLM in an evolutionary loop to automatically generate executable feature extraction code for Learning-to-Optimize (L2O) pipelines, replacing hand-crafted representations. Results are rigorously backed by numbers, showing consistent improvements across eight L2O baselines (e.g., reducing MILP solver wall time by 11-55% and LP objective gaps by 16-49%) while controlling for feature width and LLM sampling budgets. The key insight is the use of 'semantic feature contracts' to statically and dynamically validate LLM-generated tensor operations before triggering expensive model retraining, effectively constraining the search space.

#6A Control System, a Dataset and a Recipe for Making Frozen LLM Agents Learn a Domain

2026-07-28 |2607.25415

This paper formulates LLM agent harness optimization (prompts, tools, memory, planning) as a discrete control problem and evaluates online RL (bandits, REINFORCE) against a static DSPy-optimized baseline. Backed by rigorous empirical results across multiple domains and models, the authors find that online RL consistently underperforms the static DSPy baseline within practical episode budgets (e.g., 62% vs 96% success on tool-use). The key insight is that starting online controllers from a uniform prior over a discrete harness space is severely sample-inefficient; dynamic harness optimization must be bootstrapped from a metric-guided static optimizer to be viable.

#1Energy-Efficient LLM Serving via Disaggregated Attention--FFN and Flexible Frequency Scaling

2026-08-03 | University of Science and Technology of China, China Telecom Cloud Computing Research Institute, Xidian University, SKLP, ICT, CAS |2608.01891

AFlex optimizes energy consumption in LLM serving by disaggregating Attention and FFN operators and applying an Integer Linear Program (ILP) alongside a local controller to dynamically scale GPU frequencies. The results are empirically validated on NVIDIA A800 GPUs, demonstrating up to a 49% reduction in energy per token compared to state-of-the-art disaggregated serving systems while maintaining strict latency SLOs. The key insight is that Attention and FFN operators exhibit distinct frequency sensitivities; formulating an ILP to independently provision and scale frequencies for these operators yields massive energy savings over coarse-grained phase-level controls.

#2HorizonServe: Coordinating Request Scheduling with GPU Sharing for Omni-Model Serving

2026-08-03 | The University of Sydney |2608.01785

HorizonServe introduces a joint temporal-spatial scheduler for single-GPU omni-model serving that coordinates request admission and streaming multiprocessor (SM) allocation to meet heterogeneous service-level objectives (SLOs). Backed by strong empirical results, it improves SLO attainment by up to 7.0x and reduces p95 first-response latency by up to 63.7% compared to vLLM-Omni and EDF baselines on RTX 6000 GPUs. The key insight is that the critical bottleneck in omni-model serving is cross-stage memory bandwidth contention between the shared multimodal backbone and downstream generators; bounding the shared-stage SM allocation during co-running prevents goodput collapse for tight-SLO text requests.

#3BANDMAS: Causality-Inspired Semantic Packet Scheduling for Bandwidth-Efficient Multi-Agent Collaboration

2026-08-01 | The Hong Kong Polytechnic University |2608.00458

BANDMAS optimizes multi-agent LLM communication by breaking messages into semantic packets and scheduling their transmission based on a causality-inspired value predictor and resource constraints. Backed by strong empirical results, it reduces transmitted bytes by 53-77% on QA benchmarks while maintaining or improving task accuracy compared to pruning baselines. The key insight is the 'causality-inspired replay valuation'—using offline counterfactual removal (testing sufficiency and necessity of individual message chunks) to train a lightweight predictor of a message's contribution to the final outcome.