RI

#1Ask Before You Optimize: Dynamic Pre-Formulation Clarification for Interactive Optimization

2026-09-04 | Shanghai Jiao Tong University, Cardinal Operations |2609.05258

Ge et al. introduce OR-Clarify, a benchmark evaluating whether LLMs can identify missing formulation-critical information and ask clarifying questions before generating an optimization model. The authors also propose InterOPT, a two-stage framework that tracks unresolved formulation gaps to guide questioning, which improves exact requirement recovery by 16.9% over baselines in choice-based settings, though it struggles to dominate free-form interactions. The key insight is the benchmark construction pipeline: systematically withholding specific facts (e.g., objective direction, hard vs. soft constraints) from a complete OR specification to explicitly measure an agent's 'silent assumptions' and premature readiness.

#2Reinforcement Learning Enhanced LLM Agents for Complex Vehicle Routing Problems

2026-09-01 | National University of Singapore, Sun Yat-sen University |2609.00859

This paper proposes RLEA, a multi-agent framework that uses a lightweight neural planner trained via Soft Q-learning to orchestrate LLM actions (refinement, RAG and memory retrieval) for automatically generating solver-ready code for complex VRP variants. The results are backed by strong empirical evidence, achieving a 62.5% success rate on 48 VRP variants using OR-Tools, outperforming the state-of-the-art DRoC by 16.67% while reducing runtime errors. The key insight is that instead of using expensive LLM calls for step-by-step decision making, training a small language model with RL to act as a policy network for selecting agent actions significantly reduces latency and improves exploration. This is highly relevant for LLM evolutionary search and multi-agent optimization, as the architecture of an RL-trained planner combined with an evolving memory pool directly addresses sample efficiency and continuous learning in automated algorithm design.

#1RideSkill: A Hierarchical Algorithm for Generalized Ride Sharing with LLM-Driven Automatic Evolution

2026-09-02 | Huawei, The Hong Kong University of Science and Technology |2609.02250

This paper proposes RideSkill, a hierarchical ride-sharing dispatch algorithm where a skill repository, a skill combiner and a vehicle repositioner are all generated offline via LLM-assisted evolutionary search. The results are backed by strong empirical evidence on large-scale NYC data (up to 1500 vehicles), outperforming MARL and prior LLM-based heuristics by 11-40% in reward while drastically reducing detour times. The key insight is the integration of a self-check audit loop where the LLM evaluates empirical rollout metrics against its intended design, alongside a zero-shot combiner that probes unknown objective functions at runtime to dynamically mix frozen skills. This is highly relevant for LLM evolutionary search and multi-agent optimization, as the self-auditing mechanism and multi-task fitness normalization directly address sample efficiency and search signal quality in automated algorithm design.

#2CoSkill: Joint Reinforcement Learning of Reasoning and Meta-Skill Agents for Hierarchical Skill Evolution

2026-09-04 | ByteDance, Institute of Automation Chinese Academy of Sciences, Renmin University of China |2609.04865

CoSkill introduces a multi-agent reinforcement learning framework where a Reasoning Agent and a Meta-Skill Agent are jointly trained to co-evolve a hierarchical skill library alongside the reasoning policy. The results are rigorously backed by empirical data, achieving state-of-the-art success rates of 98.4% on ALFWorld and 90.6% on WebShop while demonstrating superior sample efficiency. The key insight is treating the skill editor (the Meta-Skill Agent) as a learnable policy rather than a frozen LLM prompt, updating it via RL using post-edit verification success as a delayed reward. This is highly relevant for LLM evolutionary search and multi-agent optimization, as it provides a concrete, RL-backed methodology for 'evolving the evolver' and maintaining a high-utility persistent memory of discovered procedures.

#1Latency-Aware Orchestration for Multi-Agent LLM Workflows on Heterogeneous GPUs

2026-09-03 | Beihang University, University of Leeds |2609.03335

Wang et al. propose a prediction-guided runtime that optimizes the physical execution graph of multi-agent LLM workflows on heterogeneous GPUs by jointly scheduling model lifecycles, placement and execution order. Backed by strong empirical numbers, the system reduces end-to-end makespan by up to 36.8 percent and saves roughly 24 GPU-seconds per session compared to state-of-the-art schedulers like Parrot and Kairos. The key insight is the separation of logical workflow semantics from physical execution, allowing the scheduler to fuse consecutive same-deployment activations and share model-loading actions across isolated workflows to minimize idle residency and repeated admission overhead.

#2hLLM: Single Pass Decoding for Generative Reranking

2026-09-01 | Meta Platforms, Inc. |2609.01807

HLLM replaces the sequential autoregressive decoding of generative rankers with a single forward pass, using a lightweight self-attention head and the Hungarian algorithm to decode the optimal permutation directly from the LLM's prefill hidden states. The results are strongly backed by empirical numbers, demonstrating a 64x speedup (down to 28ms latency) on a proprietary dataset and a 45x speedup on Amazon Beauty, while maintaining lossless ranking quality compared to the autoregressive teacher. The key insight is the architectural pattern of bypassing token-by-token generation for constrained outputs by training with a differentiable relaxation (Sinkhorn) and decoding with an exact combinatorial solver on the hidden states.

#3Atlas: Optimizing Deployment of Compound AI Workflows on Heterogeneous Clusters

2026-09-03 | TU Wien |2609.04513

Atlas optimizes the deployment of compound AI workflows (e.g., RAG, routing, loops) on heterogeneous clusters by using a Mixed-Integer Linear Program (MILP) to select model variants and hardware placements under latency and cost SLOs. The results are backed by strong empirical evidence, showing that their Markovian Accuracy Predictor (MAP) achieves up to 0.947 Spearman correlation with oracle accuracy while reducing profiling costs by 2.6x, allowing the MILP to find plans within 0.03 of oracle accuracy. The key insight is the MAP formulation itself, which models intermediate output quality as a discrete Markov chain between adjacent stages, enabling accurate end-to-end pipeline evaluation without combinatorial exhaustive profiling. This is highly relevant for research in OR formulations for LLM serving scheduling, providing a concrete mathematical framework for optimizing multi-stage LLM systems that could also be adapted for multi-agent coordination pipelines.