RI

#1MEGA: Self-Evolving Agent Optimization Infrastructure via Wisdom Graph

2026-08-11 | megacode.ai |2608.10504

MEGA is a three-layer infrastructure for self-evolving agent optimization that distills execution traces into a structured 'Wisdom Graph' and uses a multi-agent loop to iteratively improve agent workflows. The results are backed by strong empirical numbers, achieving a 76.55 aggregate score on four benchmarks (HotpotQA, IFBench, HoVer, PUPA) and outperforming SOTA optimizers like GEPA and TextGrad. The key insight is the 'Seed-Epoch' evaluation regime—which fixes the evaluation seed during an optimization epoch to eliminate data variance and isolate the true impact of strategy changes—alongside the decomposition of memory into atomic Primary-Context-Resultant (PCR) triplets for logical reasoning.

#2Hierarchical Self-Improvement: A Framework for Task-Specific Evolvable Agent Harnesses

2026-08-09 | HKUST |2608.08466

Zhou introduces Hierarchical Self-Improvement (HSI), a framework where a single frozen LLM evolves both its task-specific execution harness and the evolutionary search strategy used to improve that harness. The results are backed by strong empirical evidence on the BALROG benchmark, showing massive gains (e.g., +39.3% on BabyAI, +33.0% on Crafter) using DeepSeek-V4-Flash, with ablation studies proving that evolving the search strategy itself contributes significantly to the performance. The key insight is the three-layer hierarchical architecture (task harness, evolver, meta-evolver) bounded by a frozen outer anchor, which allows the system to safely rewrite its own seed selection and commit policies without unbounded self-reference.

#3PACE: Primitive-Aware Code Evolution for Automated Algorithm Design

2026-08-07 | Southern University of Science and Technology, Shenzhen University |2608.07395

Xie et al. introduce Primitive-Aware Code Evolution (PACE), which improves LLM-based automated algorithm design by extracting and preserving useful local logic as reusable Executable Algorithmic Primitives (EAPs). The method demonstrates strong empirical results, outperforming recent baselines like ReEvo and MCTS-AHD on zero-shot transfer for TSP (e.g., achieving a cost of 28.13 vs 50.29 on 1000-node TSP-ACO) and continuous control tasks. The key insight is using Thompson Sampling to evaluate the parent-relative performance improvement of individual EAPs, effectively solving the credit assignment problem for sub-components without requiring extra evaluation budgets.

#4Relay, Don't Route: Adaptive Population Handoff for Cost-Efficient LLM-Driven Evolution

2026-08-06 | Carnegie Mellon University, The University of Hong Kong, City University of Hong Kong, JIUTIAN Research, China Mobile |2608.05651

RelayEvolve is a training-free framework that reduces the cost of LLM-driven evolutionary search by using a cheap model to explore and curate a quality-diverse population, which is then handed off to a strong model for refinement. The results are backed by solid empirical evidence, showing it outperforms single-model and call-level routing baselines in 11 of 12 budget-benchmark settings across four program evolution tasks. The key insight is that because evolutionary search is stateful, budget allocation should focus on population handoff rather than individual mutation steps; cheap models can efficiently capture front-loaded progress to build a diverse seed bank.

#1ElastiCo: Elastic Configuration and Interference-Aware Orchestration for GPU Clusters

2026-08-08 | Beihang University, University of Leeds, University of Sydney |2608.07971

ElastiCo is a GPU cluster scheduler that co-locates deep learning training and offline LLM inference by treating job configurations as elastic variables and using shadow pricing for multi-resource allocation. The results are backed by strong empirical evidence on a 64-GPU testbed and 512-GPU simulation, demonstrating a 2.94x reduction in average job completion time and an increase in GPU utilization from 25% to 46% compared to static partitioning baselines. The key insight is the combination of 'Resource Shape Transformation'—exposing a family of valid resource-performance profiles for each job—with a Lagrangian decomposition that uses dynamic shadow prices to resolve cluster-wide multi-resource contention.

#2Hybrid-Adaptive Thread Tuning to Mitigate Simulation Execution Bottlenecks in High-Performance Reinforcement Learning Inference

2026-08-06 | National University of Defense Technology |2608.06025

Su et al. propose AutoThread, a hybrid adaptive thread-tuning method that dynamically optimizes worker thread counts for reinforcement learning simulation environments. The authors demonstrate real empirical gains, achieving up to an 83.8% runtime reduction and 1.8x throughput improvement over baseline tuning methods on AMD and Intel platforms. The key insight is the use of a Physics-Informed Neural Operator (PINO) where a finite-source M/M/1 queueing model acts as a structural constraint on the neural network's loss function, ensuring predictions remain physically plausible under dynamic workloads.

#3LazyTrain: Limited-resource Allocation toward Zero-waste Yield Optimization in Large Language Model Training

2026-08-12 | The Hong Kong University of Science and Technology (Guangzhou), IDEA Research, DataArcTech Ltd. |2608.11919

LazyTrain formulates limited-resource LLM training, specifically activation offloading, recomputation and communication overlap, as a mixed-integer linear programming (MILP) scheduling problem. The results are backed by concrete hardware measurements, demonstrating a 1.24x TFLOPS improvement over heuristic baselines on a single H800 for a 27B model. The key insight is that treating heterogeneous memory offloading as a joint MILP path-selection problem, rather than a greedy heuristic, enables the solver to perfectly hide PCIe and NVMe transfer costs within compute windows.

#4Every Cache Entry Earns Its Place: Global Allocation of Resolution and Coverage for KV Cache Compression

2026-08-07 | Tsinghua University |2608.07001

GraceKV formulates KV cache compression for long-context LLMs as a global resource allocation problem, using prototype trees and greedy marginal-utility allocation to balance local resolution and broad coverage. The method is backed by strong empirical results, achieving state-of-the-art performance across 24 of 32 settings on LongBench and RULER while reducing KV cache memory by up to 92% at 30K context lengths. The key insight is treating KV cache compression not as a fixed eviction or merging rule, but as a unified knapsack-like allocation problem across layers, heads and context slots, enabling dynamic, input-conditioned memory optimization.

#5BALANCE: Hybrid Autoregressive-Speculative LLM Inference in Wireless Edge Networks

2026-08-06 | The University of Hong Kong, Imperial College London |2608.05926

Qu et al. propose an edge LLM inference scheduling framework that maximizes task throughput by jointly assigning users to either autoregressive or speculative decoding while partitioning GPU memory and compute. The results are backed by numerical simulations and device measurements, demonstrating ~30-38% throughput improvements over single-mode baselines. The key insight is that the latency-memory trade-off between AD (low memory, high latency) and SD (high memory, low latency) can be effectively optimized by decoupling the modes and applying greedy knapsack heuristics based on user-specific latency contributions.