Shikun Zhang

The transition from human-centric assistance to Autonomous Software Engineering (ASE) agents has enabled the resolution of complex real-world SE tasks. However, the trial-and-error nature of these agents generates lengthy interaction trajectories, creating severe bottlenecks in terms of context window limits and cost. While context compression offers a potential remedy, prior approaches suffer from static pruning strategies and granularity mismatches, often failing to preserve the semantic dependencies and syntactic details crucial for SE tasks. To strictly preserve critical task evidence while reducing context length, we introduce AttnCompress, a dynamic attention-guided trajectory compression framework. Unlike existing approaches, AttnCompress bridges the gap between semantic integrity and dynamic adaptability through three key mechanisms: (1) structure-aware segmentation via perplexity (PPL) spikes to preserve the syntactic structure of code and logs; (2) relevance estimation using proxy attention weights to quantify the precise relevance of historical blocks to the agent’s current reasoning; and (3) a dynamic rolling window to re-evaluate and recall historical context as the task evolves. Extensive evaluation on SWE-Bench-Verified and Multi-SWE-Bench demonstrates that AttnCompress achieves a pass rate of 53.17%, outperforming prior state-of-the-art baselines while reducing token consumption by 21.6% and total costs by 33.6%. The framework proves to be model-agnostic and generalizes effectively across diverse programming languages.

Generalist agents are evolving from task-oriented systems into autonomous, self-evolving systems that acquire reusable skills, maintain long-term memory, interact with external environments, and refine themselves through recursive workflows. This shift creates a lifecycle security challenge beyond conventional model-centric or component-wise perspectives: vulnerabilities may arise from acquired external capabilities, be amplified through the orchestration of skills, memory, and tools, materialize as consequential actions, and persist through feedback across future evolution. Security for generalist agents is therefore inherently a lifecycle problem rather than a collection of isolated component- level risks. Motivated by these observations, we present a lifecycle survey organized around three stages: provenance, orchestration, and execution. We trace how risks originate, propagate, accumulate over long-horizon interactions, and persist through feedback. Across this lifecycle, we map four coupled attack surfaces: skill supply chains, user inputs, long-term memory, and external environments, and organize evaluations and defenses by where safety evidence is collected and interventions occur. Evaluations span interactive execution and offline trajectory auditing, while defenses cover input and context filtering, decision and control integrity, and runtime monitoring and enforcement. Finally, motivated by the growing importance of recursive and self-evolving AI, we identify adversarial risk generation, fine-grained safety attribution, and adversarial-feedback-driven continual safety evolution as three connected directions that form a closed loop of risk discovery, diagnosis, and mitigation toward safe recursive self-improvement agents. The latest papers and repositories are maintained at Pandora-Agent.

We present Exp-SWE-Agent, our submission to the FSE-AIWare 2026 MemComp challenge. Exp-SWE-Agent augments the Mini-SWE-Agent framework with a lightweight persistent memory module. Through a continuous retrieve-act-distill cycle, the agent extracts structured factual and experiential knowledge from past issue-resolution trajectories and uses them as strategic hints for subsequent tasks. Evaluated on a sequential 8 × 25 subset of SWE-Bench Pro, Exp-SWE-Agent increases the issue resolution rate (Pass%) from 42.7% to 47.5% while consistently reducing the number of interaction steps. These results demonstrate that structured persistent memory effectively enhances both the accuracy and efficiency of agents in long-horizon, repository-level software engineering workflows.

Automated Code Review (ACR) is crucial for software quality, yet existing benchmarks often fail to reflect real-world complexities, hindering the evaluation of modern Large Language Models (LLMs). Current benchmarks frequently focus on fine-grained code units, lack complete project context, and use inadequate evaluation metrics. To address these limitations, we introduce SWR-Bench, a new benchmark comprising 1000 manually verified Pull Requests (PRs) from GitHub, offering PR-centric review with full project context. SWR-Bench employs an objective LLM-based evaluation method that aligns strongly with human judgment (∼90% agreement) by verifying if issues from a structured ground truth are covered in generated reviews. Our systematic evaluation of mainstream ACR tools and LLMs on SWR-Bench reveals that current systems underperform, and ACR tools are more adept at detecting functional errors. Subsequently, we propose and validate a simple multi-review aggregation strategy that significantly boosts ACR performance, increasing F1 scores by up to 43.67%. Our contributions include the SWR-Bench benchmark, its objective evaluation method, a comprehensive study of current ACR capabilities, and an effective enhancement approach, offering valuable insights for advancing ACR research.

The advent of in-context learning (ICL) allows pretrained large language models (LLMs) to handle unseen inputs by leveraging context, without requiring parameter updates. However, the success of ICL is strongly influenced by factors such as the quality, size, and ordering of demonstrations, often resulting in unstable or less-than-ideal outcomes. This work is the first to address these limitations through the lens of demonstration augmentation. We first propose a simple yet effective ICL method, termed implicit demonstration augmentation-based ICL (IDAICL), that enriches demonstrations by leveraging their deep feature distributions, integrating knowledge from the entire demonstration set to enhance LLM predictions. From a theoretical standpoint, we demonstrate that when the number of augmented samples tends to infinity, our method asymptotically converges to a new form of logit calibration. Building upon this foundation, we further propose a domain-aware IDAICL (D-IDAICL) method, which improves the precision of knowledge integration by identifying and leveraging the most pertinent domain for each test sample during augmentation. Specifically, a hypernetwork is employed to adaptively select the most effective domain based on the deep representation of the test sample. The corresponding domain-specific knowledge is then utilized to augment the demonstrations, resulting in a domain-aware logit calibration function that enhances predictive performance. Comprehensive evaluations across multiple tasks using eight LLMs reveal that both approaches markedly boost overall and worst-case accuracy, leading to improved robustness and predictive capability. In addition, our approaches mitigate performance fluctuations across different demonstrations, orderings, and templates, while also showing effectiveness in handling class imbalance.

The development of LLM-based autonomous agents for end-to-end software development represents a significant paradigm shift in software engineering. However, the scientific evaluation of these systems is hampered by significant challenges, including overly simplistic benchmarks and the difficulty of conducting fair comparisons between different agent architectures due to confounding implementation variables. To address these limitations, we first construct a challenging and dynamically curated E2EDevBench to simulate realistic development scenarios. Second, we propose a hybrid evaluation framework that combines test-case-based functional assessment with fine-grained, LLM-based requirement verification. Using this framework, we conduct a controlled empirical study on three representative agent architectures implemented upon a unified foundation to isolate the impact of workflow design. Our findings reveal that state-of-the-art agents can fulfill approximately 50\% of requirements on \bench{}, but their success is critically dependent on the architectural strategy for task decomposition and collaboration. Furthermore, our analysis indicates that the primary bottleneck is the omission of requirements and inadequate self-verification. This work provides the community with a more realistic benchmark, a comprehensive evaluation framework, and crucial insights into the current capabilities and core challenges of software development agents, guiding future research toward enhancing requirement comprehension and planning.