惯性聚合 高效追踪和阅读你感兴趣的博客、新闻、科技资讯
阅读原文 在惯性聚合中打开

推荐订阅源

IT之家
IT之家
Microsoft Azure Blog
Microsoft Azure Blog
人人都是产品经理
人人都是产品经理
博客园 - 聂微东
博客园_首页
阮一峰的网络日志
阮一峰的网络日志
V
V2EX
小众软件
小众软件
F
Fortinet All Blogs
Microsoft Security Blog
Microsoft Security Blog
奇客Solidot–传递最新科技情报
奇客Solidot–传递最新科技情报
H
Hackread – Cybersecurity News, Data Breaches, AI and More
量子位
Google DeepMind News
Google DeepMind News
Jina AI
Jina AI
让小产品的独立变现更简单 - ezindie.com
让小产品的独立变现更简单 - ezindie.com
aimingoo的专栏
aimingoo的专栏
B
Blog RSS Feed
Cyber Security Advisories - MS-ISAC
Cyber Security Advisories - MS-ISAC
宝玉的分享
宝玉的分享
有赞技术团队
有赞技术团队
J
Java Code Geeks
WordPress大学
WordPress大学
The Cloudflare Blog

cs.NE updates on arXiv.org

MPCS: Neuroplastic Continual Learning via Multi-Component Plasticity and Topology-Aware EWC Combining Trained Models in Reinforcement Learning Training Non-Differentiable Networks via Optimal Transport ShiftLIF: Efficient Multi-Level Spiking Neurons with Power-of-Two Quantization Probe-Geometry Alignment: Erasing the Cross-Sequence Memorization Signature Below Chance Benchmarking local Hebbian learning rules for memory storage and prototype extraction Robust volatility updates for Hierarchical Gaussian Filtering Spiking Sequence Machines and Transformers Affinity Is Not Enough: Recovering the Free Energy Principle in Mixture-of-Experts Scalable Learning in Structured Recurrent Spiking Neural Networks without Backpropagation Geometric and dynamical analysis of attractor boundaries and storage limits in kernel Hopfield networks Attractor FCM Physical Foundation Models: Fixed hardware implementations of large-scale neural networks When Does Structure Matter in Continual Learning? Dimensionality Controls When Modularity Shapes Representational Geometry Learning to Forget: Continual Learning with Adaptive Weight Decay Causal Learning with Neural Assemblies NORACL: Neurogenesis for Oracle-free Resource-Adaptive Continual Learning Text-Utilization for Encoder-dominated Speech Recognition Models EdgeSpike: Spiking Neural Networks for Low-Power Autonomous Sensing in Edge IoT Architectures EvoTSC: Evolving Feature Learning Models for Time Series Classification via Genetic Programming Analysis and Explainability of LLMs Via Evolutionary Methods Deployment-Aligned Low-Precision Neural Architecture Search for Spaceborne Edge AI SeaEvo: Advancing Algorithm Discovery with Strategy Space Evolution Primitive Recursion without Composition: Dynamical Characterizations, from Neural Networks to Polynomial ODEs MAEO: Multiobjective Animorphic Ensemble Optimization for Scalable Large-scale Engineering Applications Necessary and sufficient conditions for universality of Kolmogorov-Arnold networks Learn&Drop: Fast Learning of CNNs based on Layer Dropping Architecture-Induced Recoverability Bias in Differentiable Symbolic Regression Collocation-based Robust Physics Informed Neural Networks for time-dependent simulations of pollution propagation under thermal inversion conditions on Spitsbergen Structure-Guided Diffusion Model for EEG-Based Visual Cognition Reconstruction
PACEvolve: Enabling Progress-Aware Consistent Evolution
[Submitted on 15 Jan 2026 (v1), last revised 10 Sep 2026 (this v · 2026-01-16 · via cs.NE updates on arXiv.org

View PDF HTML (experimental)

Abstract:Self-evolving agents powered by Large Language Models (LLMs) have emerged as a promising direction across diverse domains, including code optimization and scientific discovery, yet their core failure modes remain underexplored. Through a comprehensive empirical study, we identify that the model's reasoning becomes anchored to the local context of current hypotheses, overemphasizing low-level details while neglecting the broader search landscape. As a result, such agents become prone to context pollution and mode collapse, repeatedly revisiting flawed hypotheses and converging on suboptimal solutions. To address this challenge, we propose Progress-Aware Consistent Evolution (PACEvolve), a systematic framework for governing agent memory and search dynamics. PACEvolve overcomes these limitations through three key techniques: (1) Hierarchical Context Management (HCM), which structures historical trajectories while dynamically pruning branches to preserve a high-signal memory state; (2) Momentum-Based Backtracking (MBB), which monitors optimization progress to escape local minima; and (3) a self-adaptive Collaborative Evolution policy (CE) that balances intra-trajectory refinement with inter-trajectory knowledge transfer. By decoupling high-level idea generation from low-level code evaluation, PACEvolve maintains a global view of search momentum and achieves state-of-the-art results across complex evolutionary benchmarks.

Submission history

From: Minghao Yan [view email]
[v1] Thu, 15 Jan 2026 18:25:23 UTC (758 KB)
[v2] Fri, 16 Jan 2026 22:31:40 UTC (758 KB)
[v3] Thu, 10 Sep 2026 20:40:32 UTC (560 KB)