RI

#1Improving Natural-Language Combinatorial-Optimization Accuracy in Resource-Constrained Language Models via Formal Abstractions

2026-08-19 | UC Berkeley, Independent Researcher |2608.18409

Sharma et al. introduce SDDL, a 7-primitive domain-specific language that allows resource-constrained LLMs to translate natural-language scheduling problems into formal representations, which are then deterministically compiled to CP-SAT. The results are backed by rigorous evaluation, showing that Qwen3.5-27B improves from 21.7% feasibility when generating raw solver code to 55.3% when using SDDL on a 300-instance benchmark. The key insight is that bounding the LLM's generation space to a closed vocabulary of domain constructs removes error-prone variable declarations and API syntax, shifting the burden of correct solver encoding to a deterministic compiler.

#1FormuEvo: LLM-Guided Evolution for Discovering Solver-Efficient Mixed-Integer Programming Formulations

2026-08-24 | Tsinghua University, Nanyang Technological University, Singapore Management University |2608.23353

FormuEvo is an LLM-guided evolutionary framework that discovers solver-efficient mixed-integer programming (MIP) formulations by searching the symbolic space of executable modeling programs. Backed by strong empirical results, it accelerates solvers by up to 5.5x compared to expert-designed formulations and existing LLM-based cutting plane generation baselines across diverse OR benchmarks. THE KEY INSIGHT is the 'solver-informed diagnosis' mechanism, which translates fine-grained solver statistics (e.g., root gap, branch-and-bound node count, presolve reductions) into interpretable verbal gradients to guide LLM mutation, coupled with a structured memory of Condition-Strategy-Effect triplets.

#2Eureka: Task-Conditioned Meta-Agent Orchestration for Scientific Discovery

2026-08-19 | Duke University, Shanghai Jiao Tong University, Hokkaido University, South China Normal University, Guangdong University of Technology, ManXis |2608.19047

ManXis introduces Eureka, a task-conditioned Meta-Agent architecture that dynamically compiles long-horizon tasks into obligation graphs and forms specialized Macro-Agents with tailored state, memory and verifiers during execution. The results are backed by extensive empirical validation, including a 57.8% reduction in median context usage across 12,000 dependency-update tasks and successful application to open mathematical conjectures. The key insight is governed evolution: rather than continuously mutating the agent, the system uses a mathematically grounded cost-benefit threshold to trigger architecture evolution only when the expected amortized savings exceed the fixed cost of diagnosis and migration. This is highly relevant to LLM evolutionary search and multi-agent optimization, as it provides a rigorous, actionable framework for improving sample efficiency, managing persistent memory and dynamically scaling agent architectures.

#3AsmEvo: Agentic Assembly-Level Optimization of AMD GPU Kernels with Functional Equivalence Verification

2026-08-21 | Advanced Micro Devices, Inc. (AMD), Southern University of Science and Technology |2608.20711

AsmEvo uses a long-horizon LLM agent to optimize compiled AMD GPU kernels at the assembly level, relying on the original binary as a differential oracle to verify functional equivalence before measuring performance. The paper presents strong empirical results on real hardware (MI308X/MI300X), achieving up to 3.88x speedups on KernelBench and 1.18x geometric-mean speedups on production vLLM/SGLang Triton kernels. The key insight is the strict separation between the LLM proposer and a deterministic, correctness-gated verification harness (using real-dispatch capture for a non-hackable fitness signal), which prevents the search from exploiting fast but incorrect programs.

#1PowerSlider: Exploiting Phase Asymmetry for LLM Serving under Demand Response

2026-08-22 | Microsoft Azure Research, Cornell University, Cornell Tech |2608.21719

POWERSLIDER optimizes LLM serving under dynamic power caps by disaggregating the pipeline into Prefill, Think and Answer stages and using a KKT online solver to dynamically adjust GPU allocation and frequencies based on phase asymmetry. The results are rigorously backed by empirical data, demonstrating 1.64x higher goodput than the best baseline at a 30% power cap reduction while solving the allocation problem in just 7.7ms. The key insight is that reasoning workloads break standard prefill-decode disaggregation due to massive KV-cache accumulation during the 'thinking' phase; isolating this into a three-stage pipeline allows an online convex solver to aggressively scale down frequencies for memory-bound decode stages without starving compute-bound prefill.