Interesting Talks I Heard at ASE 2026 that Inspired Me

Created on October 15, 2026

2026   ·   ase,   conference,   research,   inspiration

It was great to have the chance to join the ASE 2026 conference taking place in Munich. The following talks/papers were really inspiring.

Talk Summary
Trust the Delta: Rethinking Software Validation in the AI Era Michael Pradel (CISPA Helmholtz Center for Information Security, University of Stuttgart) argues that as AI makes code generation cheap, the bottleneck shifts to validation, and we should reframe software correctness from absolute validation to validating behavioral deltas between program versions.
ActionNex: A Production-Grade System for Next-Best-Action Recommendation in Cloud Outage Management Zhenfeng Lin et al. (Microsoft) present ActionNex, a production-grade memory-augmented system for next-best-action recommendation in cloud outage management that grounds recommendations in structured operational context and three complementary memories.
Stop When It Matters: Detectability-Guided Microbenchmarking for Performance Regression Testing Zongxiong Chen et al. (Fraunhofer FOKUS, TU Munich, Rochester Institute of Technology, Wuhan University, Ontario Tech University, University of Waterloo) present detectability-guided microbenchmarking, a framework that reformulates benchmark execution as a Minimum-Detectable-Effect problem while maintaining statistical guarantees.
Li-WaC: A Lightweight WebAssembly Container for Independently Upgradable IoT Applications in Xiaomi Vela Bin Liao et al. (Sun Yat-sen University, Beijing Xiaomi Mobile Software) present Li-WaC, a lightweight WebAssembly container for Xiaomi’s Vela IoT OS that offloads performance-intensive tasks from JavaScript to Wasm while enabling modular OTA updates.
WiseSword: An Automated Computation Graph Partitioning Tool for Industrial DLRM Deployment on Ascend NPUs Ke Cheng et al. (Nanjing University, Huawei Technologies) present WiseSword, an automated computation graph partitioning tool for DLRM deployment on Ascend NPUs that uses progressive subgraph search to offload VC-intensive operators to CPUs while satisfying CPU utilization and tail-latency constraints, and is deployed in Huawei’s production environment.
When Compression Becomes an Attack Surface: Black-Box Attacks on Prompt-Compressed LLM Agents Zesen Liu et al. (Hong Kong University of Science and Technology) present COMA, a transfer-based black-box attack on prompt-compressed LLM agents that exploits adversarial information loss during compression, demonstrating that compression creates a new attack surface distinct from prompt injection or jailbreaks.
To Think or Not to Think: Evaluating LLM Reasoning and Agents in Vulnerability Detection Hua Beng Tan et al. (Singapore Management University, SINTEF, Ozyegin University) present a phased empirical study showing that LLM reasoning improves vulnerability detection mainly through recall, dual agents achieve high detection rates but may flag many functions as vulnerable, and multi-agents perform better on pairwise discrimination while using more energy.
IAM Policy Autopilot: Static Analysis for Policy Generation from Application Code Antonio Filieri et al. (AWS, Imperial College London) present IAM Policy Autopilot, a deterministic static-analysis tool that generates IAM policies from application code across Python, Java, Go, and TypeScript/JavaScript by parsing SDK calls, mapping to permissions, and synthesizing policies with provenance.
LLMs as Hackers: Autonomous Linux Privilege Escalation Attacks Andreas Happe et al. (TU Wien) present hackingBuddyGPT, a fully automated LLM-driven prototype for Linux privilege escalation that demonstrates effectiveness with high-level guidance and state management, while cost analysis indicates competitive performance.
Customizing an LLM for Enterprise Software Engineering (presented by Saksham Thakur) Aditya Kini et al. (Google, Meta, Google DeepMind) present Customizing an LLM for Enterprise Software Engineering, introducing Gemini for Google, an adaptation of Gemini specialized for internal software engineering ecosystems, detailing end-to-end development from curating proprietary data to implementing training strategies, and providing a comprehensive blueprint for enterprise model adaptation covering data extraction, preparation, full-stack tuning, and application deployment.
Cochise: A Reference Harness for Autonomous Penetration Testing Andreas Happe presented Cochise, a Python reference harness for autonomous penetration testing that implements a Planner-Executor architecture for SSH-based attacks, evaluated on Game of Active Directory (GOAD), and includes replay/analysis tools and a corpus of trajectory logs to enable reproducible research.

Last updated: October 15, 2026