Researchers have introduced XVada, a method for black-box context-free grammar inference that outperforms existing approaches like TreeVada in both accuracy and compactness. In empirical evaluations, XVada discovered a CVE in the Python Liquid engine and identified five additional bugs through fuzzing, all of which were subsequently fixed by the developers.
Why it matters: This work demonstrates a significant advance in automated grammar inference with practical impact, as evidenced by the discovery and remediation of real-world software vulnerabilities.
A study of 22 Triton and TileLang GPU kernels reveals that correctness-based evaluation can overlook severe performance issues—one TileLang LayerNorm kernel, for example, passes correctness checks but is over 300× slower than the PyTorch baseline. The underlying causes of inefficiency vary by kernel family, with some due to repairable authoring defects and others stemming from code-generation or autotuning limitations. The authors propose two lightweight screening methods—library-relative efficiency and roofline utilization—that successfully identify all functionally correct but inefficient kernels in their evaluation.
Why it matters: This work exposes a critical flaw in current GPU kernel evaluation practices and offers practical tools to ensure that functionally correct kernels also meet performance expectations.
Researchers have introduced PerfCodeBench, a benchmark designed to evaluate large language models (LLMs) on system-level high-performance code optimization tasks. PerfCodeBench features executable correctness checks, baseline implementations, and reference optimized solutions, enabling assessment of both code correctness and runtime efficiency. Results show a significant gap between model-generated code and expert-optimized solutions, particularly in tasks involving parallelism and GPU operations.
Why it matters: PerfCodeBench exposes the limitations of current LLMs in generating efficient, hardware-aware code for performance-critical systems, emphasizing the need for performance-oriented evaluation in code generation.
RESOURCE2SKILL is a framework that extracts and organizes executable skills for software agents from multimodal resources such as tutorial videos, code repositories, and articles. These skills are structured in a hierarchical Skill Wiki that integrates text, code, visual examples, and metadata, enabling agents to retrieve and compose relevant skills for complex tasks. In evaluations across seven authoring domains, RESOURCE2SKILL improved agent performance by an average of 11.9 percentage points compared to agents without skill libraries, and outperformed strong baselines in most tested scenarios.
Why it matters: This work demonstrates a significant advance in enabling software agents to autonomously acquire and utilize diverse, human-created resources as reusable skills, enhancing their adaptability and effectiveness in complex tasks.
Researchers have identified 'Fisher Rank Inflation,' a spectral signature in neural network gradients that marks the transition from learning clean data to memorizing corrupted labels. The effective rank of per-example gradients expands during memorization and contracts afterward, with the peak rank increasing as label corruption becomes more severe. This phenomenon enables detection of corrupted examples and, in some cases, the onset of memorization can be detected before test performance degrades.
Why it matters: This work provides a principled and interpretable method to detect and quantify memorization of label noise in deep networks, which could improve data auditing and robust training practices.
Researchers introduce ArtifactGuide, a structured rubric for software artifact evaluation, and ArtifactCopilot, an LLM-based agent that automates the evaluation process. In tests on 60 real software artifacts, ArtifactCopilot achieved 70.56% badge-level agreement with human reviewers, outperforming baseline prompt-based approaches. A user study with 8 experienced researchers found that ArtifactCopilot's reports improved reviewer confidence and helped locate supporting evidence.
Why it matters: This work demonstrates a significant advance in automating software artifact evaluation, potentially reducing reviewer workload and improving consistency in the review process.
A new method called ROBIN introduces white-box, head-level fairness debugging for transformer models by ranking attention heads based on their sensitivity to fairness probes and removing a small bias subspace from selected head outputs. In a four-model pilot study, ROBIN reduced the WinoBias gap while preserving language-modeling quality better than whole-head zeroing. The results indicate that both the selection of attention heads and the method of modification are important for effective bias repair.
Why it matters: This work demonstrates a targeted inference-time approach for mitigating bias in transformer models, offering a more precise alternative to retraining or coarse interventions.
A recent preprint demonstrates that providing feedback as line-anchored comments, using the FileMark VSCodium extension, can reduce the number of generated tokens by 22% to 58% compared to holistic prompts in AI code editing tasks. The study found that this approach also improves code correctness, with gains of 2 to 7 points for local models. The benefits are especially pronounced for longer files and for models with more room for improvement.
Why it matters: This work introduces a practical technique that can lower the cost and improve the accuracy of AI-assisted code editing, potentially enhancing developer workflows and model efficiency.
A new preprint introduces 'use-case-oriented regeneration,' a software sourcing paradigm that uses generative AI to synthesize only the specific dependency functionality required by a project, rather than relying on entire external libraries. In an evaluation across 180 repository-dependency pairs, the approach preserved 99.8% of observed behavior and reduced the exported API surface by 93%. This suggests that targeted code regeneration could be a feasible alternative to traditional software supply chains.
Why it matters: If widely adopted, this approach could significantly reduce software supply chain risks by enabling local verification of code rather than relying on external sources.
CT-Repair is an automated program repair framework that leverages Code Property Graphs (CPG) and Temporal Execution Graphs (TEG) to represent static and dynamic evidence. It uses three finite-state-machine-guided agents to analyze bugs from static, dynamic, and hybrid perspectives, and applies a filtering pipeline that reduces runtime evidence by over 94%. On 854 Java bugs from Defects4J v3.0, CT-Repair correctly repairs 489 bugs in a mixed-model configuration and 388 bugs under a controlled GPT-5.4-mini configuration, outperforming ReinFix and RepairAgent.
Why it matters: This work demonstrates that structured runtime evidence and multi-perspective reasoning can significantly improve automated program repair effectiveness without relying solely on larger patch-generation budgets.
A new method called E3 (Estimate, Execute, Expand) enables large language model (LLM) agents to estimate task difficulty before execution, reducing unnecessary context processing. On the MSE-Bench benchmark, E3 matches the strongest baseline's 100% task success rate while reducing costs by 85%, tokens by 91%, and inspected files by 92%. Live tests with gpt-4o on a real open-source library confirm that E3 remains leaner and faster than alternatives at comparable task success.
Why it matters: This work demonstrates a significant advance in reducing computational redundancy and operational costs for LLM agents, improving their scalability and efficiency.
A new method called EVITA uses multi-objective optimization to automatically generate driving scenarios that test interactions among multiple autonomous vehicles (AVs). Unlike traditional approaches that focus on single-AV testing, EVITA is designed to uncover safety-critical behaviors that emerge only when multiple AVs interact. Experimental results indicate that EVITA produces a greater diversity of AV interactions compared to existing scenario generation methods.
Why it matters: This approach addresses a key gap in AV safety testing by enabling the discovery of critical multi-vehicle interaction scenarios that single-vehicle tests may miss.
Researchers introduce Code-MUE, a black-box framework for measuring uncertainty in code large language models (LLMs) by analyzing execution-based semantic interaction graphs. The method quantifies semantic diversity using Von Neumann entropy and demonstrates a strong negative correlation with functional correctness (Spearman's up to -0.98) across eight models. Code-MUE outperforms lexical and embedding-based baselines for risk detection and selective prediction, addressing limitations of existing black-box uncertainty metrics for code.
Why it matters: This work provides a practical and effective approach to uncertainty estimation for closed-source code LLMs, supporting safer and more reliable automation in software engineering.
Researchers have introduced LapSurgie, described as the first humanoid-robot-based laparoscopic teleoperation framework. The system uses an inverse-mapping strategy to control standard surgical tools without requiring additional setup, and a user study provides initial evidence of its effectiveness.
Why it matters: This work suggests a potential path to expanding access to minimally invasive surgery in underserved regions by enabling humanoid robots to operate in existing operating rooms without infrastructure changes.
Researchers introduce PREC, a framework that clusters users by preference to learn representative reward models from sparse and noisy feedback. In simulated locomotion tasks, PREC groups users into preference-coherent clusters more accurately than baseline methods and improves social welfare metrics compared to both single shared-policy and per-user alignment approaches.
Why it matters: This work proposes a practical solution for aligning robot policies with diverse human preferences, addressing challenges of sparse and noisy feedback and reducing deployment validation burden.
Researchers present MAMMOTH, a unified end-to-end navigation policy that integrates RGB, thermal, 3D point cloud, and ego velocity data for autonomous off-road navigation. The system employs a modality dropout training scheme to ensure robustness to missing sensor inputs and uses a diffusion policy for safer, terrain-aware trajectory planning. Real-world experiments demonstrate improved collision avoidance, terrain-aware planning, and generalization to missing modalities, including during night-time operation.
Why it matters: This work advances autonomous off-road navigation by enabling robust performance even when some sensors fail or degrade, addressing a key challenge in the field.
A new framework called Instance-Enriched Semantic Maps (IESM) has been proposed for Visual Language Navigation (VLN). IESM uses 2.5D maps with instance-level object details and LLM-based query processing, enabling more robust navigation in complex indoor environments. The method achieves approximately 96% storage reduction compared to 3D scene graphs, and demonstrates over 17% improvement in object retrieval and over 23% in navigation success rates compared to baseline methods.
Why it matters: This work represents a significant advance in efficient and robust robot navigation using natural language instructions in complex environments.
Researchers introduce TrustVLA, an inference-time defense designed to detect and mitigate backdoor attacks in Vision-Language-Action (VLA) models. TrustVLA identifies abnormal evidence evolution and localizes compact causal footprints associated with visual triggers, enabling recovery of clean behavior without retraining. The method operates using only a small clean calibration set and demonstrates reduced attack success while maintaining clean-task performance.
Why it matters: This work provides a practical, retraining-free defense against backdoor attacks in VLA models, addressing a significant security risk in robotics applications.
A new approach called Model-Based Diffusion Optimal Control (MDOC) is introduced for multi-robot motion planning. MDOC generates dynamically feasible, collision-free trajectories without relying on demonstration data by integrating known dynamics models with Control Barrier Function-constrained projections. The method scales to multi-robot scenarios using Conflict-Based Search and, in simulation experiments, outperforms baseline planners in sample efficiency, smoothness, and success rate.
Why it matters: This work demonstrates a significant advance in multi-robot motion planning by removing the need for demonstration data while rigorously enforcing dynamics and safety constraints, potentially enabling more scalable and reliable robot coordination.
Researchers present ExToken, a framework that conditions vision-language-action (VLA) reinforcement learning policies on discrete behavioral priors derived from offline demonstrations. By encouraging exploration of diverse trajectory modes, ExToken addresses exploration stagnation and improves sample efficiency, leading to faster convergence and better task performance in both simulated and real-world robotic manipulation tasks, especially under limited interaction budgets.
Why it matters: Improving sample efficiency in VLA reinforcement learning could reduce the cost and time required to train robotic systems, facilitating broader real-world deployment.