Accurate project localization (e.g., files and functions) for issue resolution is a critical first step in software maintenance. However, existing benchmarks for issue localization, such as SWE-Bench and LocBench, are limited. They focus predominantly on pull-request issues and code locations, ignoring other evidence and non-code files such as commits, comments, configurations, and documentation. To address this gap, we introduce MULocBench, a comprehensive dataset of 1,100 issues from 46 popular GitHub Python projects. Comparing with existing benchmarks, MULocBench offers greater diversity in issue types, root causes, location scopes, and file types, providing a more realistic testbed for evaluation. Using this benchmark, we assess the performance of state-of-the-art localization methods and five LLM-based prompting strategies. Our results reveal significant limitations in current techniques: even at the file level, performance metrics (Acc@5, F1) remain below 40%. This underscores the challenge of generalizing to realistic, multi-faceted issue resolution. Modern software projects are inherently complex. They often consist of thousands of files spanning code, configurations, tests, and documentation. The complexity making developers routinely encounter a wide spectrum of issues, ranging from runtime failures and unexpected results to enhancement requests and usage questions. A prerequisite for resolving these issues is to accurately identify the locations, such as the relevant files and functions. Existing benchmarks have advanced research on issue localization. SWE-Bench Jimenez et al. collects 2,294 issues with pull requests from 12 Python projects, primarily targeting bug fixing. To encourage adoption, it releases SWE-bench Lite, a subset of 300 instances.

Recent advancements in training paradigms for Large Language Models (LLMs) have unlocked their remarkable capabilities in natural language processing and cross-domain generalization. While LLMs excel in tasks like programming and mathematical problem-solving, their zero-shot performance in specialized domains requiring expert knowledge, such as cybersecurity, is often suboptimal. This limitation arises because foundational LLMs are designed for general-purpose applications, constraining their ability to encapsulate domain-specific expertise within their parameter space. To address this, we explore fine-tuning strategies to embed cybersecurity knowledge into LLMs, enhancing their performance in cybersecurity question-answering (Q\&A) tasks while prioritizing computational efficiency. Specifically, we investigate Supervised Fine-Tuning (SFT), Low-Rank Adaptation (LoRA), and Quantized Low-Rank Adaptation (QLoRA) using a cybersecurity Q\&A dataset. Our results demonstrate that these fine-tuning approaches significantly outperform the foundational model in cybersecurity Q\&A tasks. Moreover, LoRA and QLoRA achieve comparable performance to SFT with substantially lower computational costs, offering an efficient pathway for adapting LLMs to specialized domains. Our work highlights the potential of low-rank fine-tuning strategies to bridge the gap between general-purpose LLMs and domain-specific applications.

Recent curriculum reinforcement learning for large language models (LLMs) typically rely on difficulty-based annotations for data filtering and ordering. However, such methods suffer from local optimization, where continual training on simple samples in the early steps can cause the policy to lose its exploration. We propose a novel schema, namely Hamiltonian curiosity augmented large language model reinforcement (HAMMER), that transfers diversity metrics, commonly used in dataset evaluation, into the dynamic reinforcement learning procedure, where training samples are ordered via a minimum-semantic Hamiltonian path making the initial training retrain more exploration. From a theoretical perspective of generalization bounds, diversity-driven ordering facilitates stable convergence. Empirical evaluations indicate that HAMMER stimulates model "curiosity" and consistently achieves a 3% to 4% average accuracy gain across diverse inference benchmark.

Tool-augmented language agents frequently fail in real-world deployment due to tool malfunctions--timeouts, API exceptions, or inconsistent outputs--triggering cascading reasoning errors and task abandonment. Existing agent training pipelines optimize only for success trajectories, failing to expose models to the tool failures that dominate real-world usage. We propose \textbf{PALADIN}, a generalizable framework for equipping language agents with robust failure recovery capabilities. PALADIN trains on 50,000+ recovery-annotated trajectories constructed via systematic failure injection and expert demonstrations on an enhanced ToolBench dataset. Training uses LoRA-based fine-tuning to retain base capabilities while injecting recovery competence. At inference, PALADIN detects execution-time errors and retrieves the most similar case from a curated bank of 55+ failure exemplars aligned with ToolScan's taxonomy, then executes the corresponding recovery action. This approach generalizes to novel failures beyond the training distribution, retaining 95.2\% recovery performance on unseen tool APIs. Evaluation across PaladinEval and ToolReflectEval demonstrates consistent improvements in Recovery Rate (RR), Task Success Rate (TSR), Catastrophic Success Rate (CSR), and Efficiency Score (ES). PALADIN improves RR from 32.76% to 89.68% (+57% relative) over ToolBench and outperforms the strongest baseline CRITIC (76.34%) by +13.3%. Against vanilla agents, PALADIN achieves 89.86\% RR (+66% relative improvement from 23.75%). These results establish PALADIN as an effective method for building fault-tolerant agents capable of robust recovery in real-world tool environments.

We introduce Blueprint-Bench, a benchmark designed to evaluate spatial reasoning capabilities in AI models through the task of converting apartment photographs into accurate 2D floor plans. While the input modality (photographs) is well within the training distribution of modern multimodal models, the task of spatial reconstruction requires genuine spatial intelligence: inferring room layouts, understanding connectivity, and maintaining consistent scale. We evaluate leading language models (GPT-5, Claude 4 Opus, Gemini 2.5 Pro, Grok-4), image generation models (GPT-Image, NanoBanana), and agent systems (Codex CLI, Claude Code) on a dataset of 50 apartments with approximately 20 interior images each. Our scoring algorithm measures similarity between generated and ground-truth floor plans based on room connectivity graphs and size rankings. Results reveal a significant blind spot in current AI capabilities: most models perform at or below a random baseline, while human performance remains substantially superior. Image generation models particularly struggle with instruction following, while agent-based approaches with iterative refinement capabilities show no meaningful improvement over single-pass generation. Blueprint-Bench provides the first numerical framework for comparing spatial intelligence across different model architectures. We will continue evaluating new models as they are released and welcome community submissions, monitoring for the emergence of spatial intelligence in generalist AI systems.

Safety and controllability are critical for large language models. A central question is whether undesirable behaviors like deception are localized functions that can be removed, or if they are deeply intertwined with a model's core cognitive abilities. We introduce "cyclic ablation," an iterative method to test this. By combining sparse autoencoders, targeted ablation, and adversarial training on DistilGPT-2, we attempted to eliminate the concept of deception. We found that, contrary to the localization hypothesis, deception was highly resilient. The model consistently recovered its deceptive behavior after each ablation cycle via adversarial training, a process we term functional regeneration. Crucially, every attempt at this "neurosurgery" caused a gradual but measurable decay in general linguistic performance, reflected by a consistent rise in perplexity. These findings are consistent with the view that complex concepts are distributed and entangled, underscoring the limitations of direct model editing through mechanistic interpretability.

Recent advancements in large language models (LLMs) have shown promise in feature engineering for tabular data, but concerns about their reliability persist, especially due to variability in generated outputs. We introduce a multi-level diagnosis and evaluation framework to assess the robustness of LLMs in feature engineering across diverse domains, focusing on the three main factors: key variables, relationships, and decision boundary values for predicting target classes. We demonstrate that the robustness of LLMs varies significantly over different datasets, and that high-quality LLM-generated features can improve few-shot prediction performance by up to 10.52%. This work opens a new direction for assessing and enhancing the reliability of LLM-driven feature engineering in various domains.

Abstract-- Incorporating empathetic behavior into robots can improve their social effectiveness and interaction quality. In this paper, we present whEE (when and how to express empathy), a framework that enables social robots to detect when empathy is needed and generate appropriate responses. Using large language models, whEE identifies key behavioral empathy cues in human interactions. We evaluate it in human-robot interaction scenarios with our social robot, Haru. Results show that whEE effectively identifies and responds to empathy cues, providing valuable insights for designing social robots capable of adaptively modulating their empathy levels across various interaction contexts. In most scenarios, Large Language Models (LLMs) represent the state-of-the-art approach for classifying empathy [1], [2] and generating empathetic responses [3], [4]. However, the development of robots capable of dynamically adjusting their level of empathy based on the context remains an underexplored area [5]. To this end, we introduce whEE (when and how to express empathy), an empathy framework that provides guidelines on when robots should respond empathetically and how to achieve it. Using our framework, we analyze the utterances of speakers and listeners in dyadic and group conversations with varying levels of empathy. Our analysis identifies key empathy cues that indicate when a speaker seeks an empathetic response and the cues exhibited by listeners displaying high levels of empathy. We approach empathy by focusing on observable behaviors that individuals exhibit when demonstrating an understanding of others' emotions and engaging deeply with their experiences--referred to as behavioral empathy [6].

Recent advancements in large language models (LLMs) suggest great promises in code and proof generations. However, scaling automated formal verification to real-world projects requires resolving cross-module dependencies and global contexts, which are crucial challenges overlooked by existing LLM-based methods with a special focus on targeting isolated, function-level verification tasks. To systematically explore and address the significant challenges of verifying entire software repositories, we introduce RVBench, the first verification benchmark explicitly designed for repository-level evaluation, constructed from four diverse and complex open-source Verus projects. We further introduce RagVerus, an extensible framework that synergizes retrieval-augmented generation with context-aware prompting to automate proof synthesis for multi-module repositories. RagVerus triples proof pass rates on existing benchmarks under constrained model inference budgets, and achieves a 27% relative improvement on the more challenging RVBench benchmark, demonstrating a scalable and sample-efficient verification solution.

APIs are central to modern software development, yet composing new APIs from large libraries is difficult due to the exponential search space; traditional component-based synthesis relies on costly exploration and hand-crafted specifications. While large language models (LLMs) can generate implementations from natural language, hallucinations and limited access to up-to-date contextual information often yield incorrect code. In this paper, we present APRIL, an approach that combines LLM-based synthesis with Automatic Prompt Optimization (APO) and Reinforcement Learning from Verifiable Rewards (RL VR): APO iteratively refines prompts for a frozen model, while RLVR fine-tunes the policy toward functional correctness, producing an efficient synthesis pipeline. Evaluated on 81 real-world APIs from widely used scientific Python libraries and benchmarked against instruction-tuned but unfine-tuned LLMs guided by expert prompts, APRIL achieves substantial improvements. These results indicate that integrating APO and RLVR provides a robust, scalable path for component-based API synthesis in large libraries.