RI

#1Falsification-Based Verification of LLM-Generated Optimization Models: Sound Test Batteries and Their Detection Limits

2026-07-18 | Central University of Finance and Economics |2607.16646

This paper introduces a falsification-based, oracle-free verification framework for LLM-generated optimization models using metamorphic tests derived from optimization theory. The results are rigorously backed by empirical evidence, achieving a 0.0% false-positive rate on faithful models compared to 54.9% for threshold-based baselines, while successfully detecting execution-blind errors across multiple benchmarks. The key insight is that classical optimization theory, such as value function convexity and shadow-price consistency, provides provably valid metamorphic relations that can verify generated models without requiring ground-truth labels.

#2Constrained Path Reasoning: Measuring When Committed Stages Earn Their Cost

2026-07-19 | ShanghaiTech University |2607.17240

This paper introduces Constrained Path Reasoning (CPR), a framework to quantify the utility versus computational cost of intermediate LLM reasoning stages (e.g., formalization, convexification) when solving non-convex optimization problems. Backed by rigorous experiments on over 1,100 QCQPs, the authors demonstrate that while formal transcription with a deterministic solver achieves 90% yield, adding an LLM-proposed convex surrogate drops yield to 20% due to relaxation leakage and over-contraction. The key insight is the use of classical error bounds to map constraint residuals into a continuous feedback signal for triage and local repair, which successfully recovers 63% of feasible yield using only 17% of the computational attempts.

#1Automated Discovery Has No Universally Superior Harness

2026-07-20 | UC Berkeley, MIT, MIT-IBM Watson AI Lab, Weizmann Institute of Science |2607.18235

Gupta et al. systematically decompose and evaluate LLM-guided discovery harnesses across 3.1 million rollouts to show that no single fixed harness universally generalizes across models and problems. The results are highly rigorous, utilizing budget-matched statistical bootstrap tests to prove that complex evolutionary machinery often underperforms simpler tree-search or epsilon-greedy baselines. The key insight is that early search progress strongly predicts final performance, enabling an online adaptive allocation strategy that prunes weak partial runs and reallocates compute to surviving trajectories.

#2How to Guide LLM Generation: Dual-Surrogate Guided Search for Automated Heuristic Design

2026-07-15 | The Hong Kong Polytechnic University, Nankai University, South China Agricultural University |2607.13911

Wang et al. introduce Dual-Surrogate Guided Search (DGS) to improve sample efficiency in LLM-based automated heuristic design by learning which parent heuristics and mutation operators to select before querying the LLM. The results are backed by solid empirical evidence, showing superior performance over strong baselines like FunSearch and EoH across five combinatorial optimization tasks (including TSP and CVRP) within a strict 200-query budget. The key insight is the use of a transition surrogate that predicts the latent distribution of an ungenerated child heuristic based on the parent and operator, coupled with a utility surrogate to estimate its performance.

#3Reward-Free Evolving Agents via Pairwise Validator

2026-07-15 | Accenture |2607.14408

Liu et al. replace the expensive scalar task reward in self-evolving agentic loops with a training-free pairwise LLM validator that compares parent and child candidates. The results are backed by rigorous empirical evaluations across prompt and code evolution engines, including ShinkaEvolve, demonstrating that this reward-free approach matches or exceeds full-reward baselines. The key insight is the 'Soft Elo' mechanism, which uses pairwise LLM verdicts to update candidate ratings for parent selection, entirely eliminating the need for scalar validation rewards.

#4Autonomous Discovery of Wireless Communications Algorithms

2026-07-20 | NVIDIA |2607.17762

Aït Aoudia et al. introduce AITE, an LLM-driven evolutionary search framework that uses a two-tier orchestrator-worker architecture with agentic tool-use to autonomously discover wireless communication algorithms. The framework produces real, empirically validated results, discovering an OTFS equalizer that is 3.6x faster than the state-of-the-art with better error rates and the first explainable pilotless OFDM receiver matching neural network performance. The most transferable methodological insights are the post-processing adherence check (where an LLM verifies if the generated code actually implemented the assigned idea before updating the orchestrator's leaderboard) and the prompt refinement stage (which critiques worker execution journals to improve instructions for the next generation).

#1Robust KV Cache Management for LLM Serving under Output Token Length Uncertainty

2026-07-18 | Arizona State University |2607.16892

Cheng et al. formulate KV cache reservation in LLM serving as a Wasserstein distributionally robust optimization (DRO) problem to jointly optimize GPU parallelism, routing and prefix caching under output length uncertainty. The results are rigorously backed by trace-driven simulations on production workloads (BurstGPT, Azure), demonstrating up to 56% lower costs compared to fixed-quantile baselines while maintaining competitive P99 latency. The key insight is the derivation of a critical fractile structure that automatically adapts the reservation quantile based on the ratio of preemption to memory-waste costs, eliminating the need for manual tuning.

#2ExpertPlex: A High-Goodput Disaggregated Serving System for MoE LLMs with Adaptive Persistent Kernels

2026-07-20 | Peking University, Independent Researcher |2607.18002

ExpertPlex introduces a hybrid LLM serving architecture for Mixture-of-Experts (MoE) models that shares massive MoE weights across prefill and decode phases while disaggregating attention, supported by tile-level GPU scheduling and a cross-stack placement optimizer. The system demonstrates strong empirical results, achieving up to 2.01x higher goodput over standard prefill-decode disaggregation on real-world MoE models like MiniMax-M2.7 and GLM-5.1-FP8. The key insight is the tile-aware latency model for MoE computation: because sparse routing changes latency based on per-expert packing rather than just token volume, resource allocation models must account for active expert counts rather than raw token counts.

#3WAR: Workload-Aware Rollouts for Synchronous Agentic Reinforcement Learning

2026-07-19 |2607.17299

WAR optimizes synchronous agentic RL rollouts by dynamically applying model-free speculative decoding (SuffixDecoding) under low load and cache-aware request scheduling under high load. The results are backed by strong empirical evidence, demonstrating 1.4x to 1.6x throughput improvements on 16 H100 GPUs using Qwen3-32B for software engineering agent tasks. The key insight is that the effectiveness of rollout optimizations is highly workload-dependent; speculative decoding wastes compute under high load where standard batched decoding already saturates GPUs, making cache-aware scheduling (prioritizing KV-cache locality and shorter trajectories) the better lever.

#4LMEdge: QoS-Aware LLM Inference Orchestration on Edge Clusters

2026-07-19 | Vienna University of Technology, University of Klagenfurt, University of Messina |2607.17175

LMEdge orchestrates LLM inference across heterogeneous edge devices by using ML predictors and a lightweight heuristic to jointly optimize model selection, quantization and placement under QoS constraints. The results are backed by empirical measurements from a 57-device Kubernetes testbed, demonstrating improved latency and resource utilization over standard load-aware baselines. The key insight is the practical integration of query-specific XGBoost predictors for accuracy and latency directly into the scheduling constraints, alongside a released dataset of 59,000 inference measurements that could be highly valuable for offline scheduling simulations. This work is directly relevant to operations research formulations for LLM serving scheduling, though its methods are relatively standard and primarily target edge environments rather than large-scale cloud GPU clusters.

#5JoyNexus: Service-Oriented Multi-Tenant Post-Training for VLA Models

2026-07-17 | Tsinghua University, Peking University, Beihang University, Beijing Institute of Technology, JDT AI Infra |2607.16074

JoyNexus proposes a multi-tenant service architecture for Vision-Language-Action model post-training that decouples training, inference and environment interaction into separate, independently scheduled services. The results are backed by empirical simulations, demonstrating a 28.3% reduction in aggregate GPU time and up to 3.24x shared-forward throughput speedup. The key insight is the use of group batching for heterogeneous requests, which canonicalizes different tenant data schemas to share a single forward pass through a frozen base model before splitting features for tenant-specific updates. This systems engineering pattern is highly relevant for scaling LLM evolutionary search infrastructure, as decoupling generation, evaluation and update phases allows for asynchronous scaling and better GPU utilization during bursty search workloads.