RI

#1FLARE: Verifying MILP Reformulations with LLM-Based Theorem Proving

2026-08-25 | Stanford University |2608.25220

Robbins et al. introduce FLARE, an automated framework that uses an LLM agent and the Lean proof assistant to formally verify whether one MILP formulation is a valid reformulation of another. The authors demonstrate 100% accuracy on a new benchmark of 54 NP-hard formulation pairs, backed by machine-checkable proofs rather than heuristic execution checks. The key insight is that instance-level validation—commonly used to evaluate LLM-generated models—is fundamentally flawed and can accept invalid transformations; the authors prove this by identifying mathematically invalid cutting planes generated by recent LLM evolutionary search frameworks.

#2AlgoWorlds: Benchmarking Tool Use for Global Optimization in Algorithmic Worlds

2026-08-29 | Tencent, University of Southern California |2608.29397

Xu et al. introduce AlgoWorlds, a benchmark of 240 environments across 10 combinatorial optimization families designed to evaluate whether LLM agents can use tools to gather hidden instance data and output globally optimal decisions. The results are backed by extensive evaluations across 7 frontier models, demonstrating that while models like Claude Opus 4.8 achieve high feasibility (over 96%), they fail to reach exact global optimality (only 38.6% success). The key insight is that the primary bottleneck for LLMs in optimization is not information acquisition or constraint satisfaction, but the joint reasoning required to integrate information into a globally optimal configuration.

#1Recursive Experiential-Working Memory Evolution for Long-Horizon Agent Harnesses

2026-08-25 | Princeton University, Stanford University, University of Oxford, NUS |2608.24876

Recuris introduces a recursive memory evolution architecture that couples Working Memory (for task state tracking) with Experiential Memory (for skill storage), using structured execution traces to localize failures and apply validation-gated updates to specific memory components. The results are backed by rigorous empirical evidence, showing massive gains on long-horizon tasks (e.g., +17.8 points for GPT-5.6 Sol and +15.6 for Claude Opus 5 on the $\tau$2-Retail benchmark). The key insight is using structured execution traces to localize failures to specific harness components (skill content, state tracking, invocation trigger, or verification checker) rather than rewriting the entire memory or prompt, enabling highly targeted and stable evolutionary updates. This is highly relevant for LLM evolutionary search and multi-agent optimization, as it provides a concrete mechanism for persistent memory evolution, step-by-step credit assignment and safe recursive self-improvement.

#2RecurSE: Bounded Recursive Self-Evaluation for LLM Rubric Judges

2026-08-25 | Zhejiang University, Meituan LongCat Team |2608.24231

RECURSE enables LLM judges to recursively self-improve via RL by using a synchronized copy of the policy as a process checker to provide scalar rewards, eliminating the need for external reward models. The method is backed by strong empirical results across Qwen and Gemma models, showing significant improvements on held-out transfer benchmarks like HealthBench and CheckEval. The key insight is 'interface decoupling'—structurally separating the judge's output format (YES/NO tokens) from the checker's reward format (0-4 scalar) to prevent the policy from exploiting token-copying shortcuts during self-play. Furthermore, tracking checker ranking fidelity on a small holdout set reliably prevents over-optimization.

#1TOPAS: Workflow-Aware Prefix-State Scheduling for Multi-Agent LLM Serving

2026-08-26 | University of Science and Technology of China, Hefei University of Technology |2608.25523

TOPAS is an online scheduler for multi-agent LLM serving that jointly optimizes KV cache prefix residency and request admission to minimize task-level job completion time. Backed by empirical numbers, it reduces mean and p99 JCT by up to 39.8% and 49.4% on synthetic DAGs and significantly improves performance on real MetaGPT workflows compared to baselines like Shortest-Path-First. The critical insight is treating prefix residency as an explicit, workflow-aware scheduling decision rather than a reactive byproduct of request ordering, using a utility function that balances the reduction of a task's longest remaining service path against near-term prefix reuse. This is highly relevant for research in OR formulations for LLM serving scheduling, offering a concrete heuristic search approach to manage the memory-compute tradeoff in complex multi-agent deployments.