Named entity recognition (NER) is a fundamental task in numerous downstream applications. Recently, researchers have employed pre-trained language models (PLMs) and large language models (LLMs) to address this task. However, fully leveraging the capabilities of PLMs and LLMs with minimal human effort remains challenging. In this paper, we propose GPT4NER, a method that prompts LLMs to resolve the few-shot NER task. GPT4NER constructs effective prompts using three key components: entity definition, few-shot examples, and chain-of-thought. By prompting LLMs with these effective prompts, GPT4NER transforms few-shot NER, which is traditionally considered as a sequence-labeling problem, into a sequence-generation problem. We conduct experiments on two benchmark datasets, CoNLL2003 and OntoNotes5.0, and compare the performance of GPT4NER to representative state-of-the-art models in both few-shot and fully supervised settings. Experimental results demonstrate that GPT4NER achieves the $F_1$ of 83.15\% on CoNLL2003 and 70.37\% on OntoNotes5.0, significantly outperforming few-shot baselines by an average margin of 7 points. Compared to fully-supervised baselines, GPT4NER achieves 87.9\% of their best performance on CoNLL2003 and 76.4\% of their best performance on OntoNotes5.0. We also utilize a relaxed-match metric for evaluation and report performance in the sub-task of named entity extraction (NEE), and experiments demonstrate their usefulness to help better understand model behaviors in the NER task.

Text-to-SQL, the task of translating natural language questions into SQL queries, has significantly advanced with the introduction of Large Language Models (LLMs), broadening database accessibility for a wide range of users. Despite substantial progress in generating valid SQL, current LLMs still struggle with complex queries. To address this limitation, test-time strategies such as Best-of-N (BoN) and Majority Voting (Maj) are often employed, based on the assumption that LLMs can produce correct answers after multiple attempts. However, these methods rely on surface-level heuristics, selecting the syntactically correct query through execution-based BoN (ex-BoN) or the most frequently generated one through Majority Voting. Recently, Outcome Reward Models (ORMs), which assign utility scores to generated outputs based on semantic correctness, have emerged as a promising reinforcement learning approach for improving model alignment. We argue that ORMs could serve as an effective new test-time heuristic, although their application in this context remains largely underexplored. In this work, we propose a unified framework for training ORMs tailored to the Text-to-SQL task and assess their effectiveness as a test-time heuristic within the BoN strategy. We benchmark ORMs against ex-BoN and Maj across the BIRD and Spider datasets, fine-tuning diverse open-source LLMs from the Qwen2, Granite3, and Llama3 families. Results show that ORMs outperform ex-BoN and Maj, achieving execution accuracy gains of +4.33% (BIRD) and +2.10% (Spider) over ex-BoN, and +2.91% (BIRD) and +0.93% (Spider) over Maj. We further demonstrate that finetuning models already aligned with SQL generation, such as OmniSQL, yields superior ORM performance. Additionally, we observe that ORMs achieve competitive results on simple queries and benefit more from an increased number of candidates compared to ex-BoN and Maj.

The rapid advancement of Large Language Models (LLMs) has enabled them to generate complex, multi-step plans and itineraries. However, these generated plans often lack temporal and spatial consistency, particularly in scenarios involving physical travel constraints. This research aims to study the temporal performance of different LLMs and presents a validation framework that evaluates and improves the temporal consistency of LLM-generated travel itineraries. The system employs multiple state-of-the-art LLMs to generate travel plans and validates them against real-world flight duration constraints using the AeroDataBox API. This work contributes to the understanding of LLM capabilities in handling complex temporal reasoning tasks like itinerary generation and provides a framework to rectify any temporal inconsistencies like overlapping journeys or unrealistic transit times in the itineraries generated by LLMs before the itinerary is given to the user. Our experiments reveal that while current LLMs frequently produce temporally inconsistent itineraries, these can be systematically and reliably corrected using our framework, enabling their practical deployment in large-scale travel planning.

In this paper, we present WEST(WE Speech Toolkit), a speech toolkit based on a large language model (LLM) for speech understanding, generation, and interaction. There are three key features of WEST: 1) Fully LLM-based: Standing on the shoulders of giants by reusing mature architectures, ecosystems (e.g., Hugging Face), and methods (e.g., sequence packing) from large models. 2) Full-stack: Supports tasks such as recognition, synthesis, understanding, dialogue, and multimodal capabilities, with extensibility to incorporate open-source models. 3) Simple and Stupid: A simple and stupid speech toolkit that everyone can Touch. In addition, WEST provides two types of recipes, models, and experimental results. The first is entirely based on open-source models and open-source data, allowing users to fully reproduce the experiments in this paper and serving as a verification system or minimal system baseline. The second is trained on massive data, offering superior performance so the user can directly apply it out of the box. WEST is publicly avilable at https://github.com/wenet-e2e/west/

Abstract--Automating penetration testing is crucial for enhancing cybersecurity, yet current Large Language Models (LLMs) face significant limitations in this domain, including poor error handling, inefficient reasoning, and an inability to perform complex end-to-end tasks autonomously. T o address these challenges, we introduce Pentest-R1, a novel framework designed to optimize LLM reasoning capabilities for this task through a two-stage reinforcement learning pipeline. We first construct a dataset of over 500 real-world, multi-step walkthroughs, which Pentest-R1 leverages for offline reinforcement learning (RL) to instill foundational attack logic. Subsequently, the LLM is fine-tuned via online RL in an interactive Capture The Flag (CTF) environment, where it learns directly from environmental feedback to develop robust error self-correction and adaptive strategies. Our extensive experiments on the Cybench and AutoPenBench benchmarks demonstrate the framework's effectiveness. On AutoPenBench, Pentest-R1 achieves a 24.2% success rate, surpassing most state-of-the-art models and ranking second only to Gemini 2.5 Flash. Ablation studies confirm that the synergy of both training stages is critical to its success.

Recent advancements in language models (LMs) have marked a shift toward the growing importance of post-training. Yet, post-training approaches such as supervised fine-tuning (SFT) do not guarantee the effective use of knowledge acquired during pretraining. We therefore introduce InfoSteer, a lightweight method that encourages parametric information utilization in LMs during post-training. Specifically, InfoSteer treats the feed-forward network (FFN) layer as associate key-value memory and promotes the use of stored memory vectors via forward-pass interventions or regularization during backpropagation. This simple guidance during post-training phase yields consistent performance improvements across diverse model families -- including Qwen, Gemma and Llama -- spanning 15 downstream tasks in both in-distribution (ID) and out-of-distribution (OOD) evaluations. Beyond performance gains, we also find that steered LMs can adaptively allocate information by placing more emphasis on generating semantically meaningful tokens, while using fewer resources on simple transition ones (e.g., `\texttt{,}' or `\texttt{and}'). Our work underscores that vanilla post-training does not fully exploit the potential gained during pre-training, and that steering LMs in latent representation space offers a promising approach to enhance both performance and interpretability. The code is available at: https://github.com/chili-lab/InfoSteer.

Multimodal in-context learning (ICL) remains underexplored despite significant potential for domains such as medicine. Clinicians routinely encounter diverse, specialized tasks requiring adaptation from limited examples, such as drawing insights from a few relevant prior cases or considering a constrained set of differential diagnoses. While multimodal large language models (MLLMs) have shown advances in medical visual question answering (VQA), their ability to learn multimodal tasks from context is largely unknown. We introduce SMMILE, the first expert-driven multimodal ICL benchmark for medical tasks. Eleven medical experts curated problems, each including a multimodal query and multimodal in-context examples as task demonstrations. SMMILE encompasses 111 problems (517 question-image-answer triplets) covering 6 medical specialties and 13 imaging modalities. We further introduce SMMILE++, an augmented variant with 1038 permuted problems. A comprehensive evaluation of 15 MLLMs demonstrates that most models exhibit moderate to poor multimodal ICL ability in medical tasks. In open-ended evaluations, ICL contributes only an 8% average improvement over zero-shot on SMMILE and 9.4% on SMMILE++. We observe a susceptibility for irrelevant in-context examples: even a single noisy or irrelevant example can degrade performance by up to 9.5%. Moreover, we observe that MLLMs are affected by a recency bias, where placing the most relevant example last can lead to substantial performance improvements of up to 71%. Our findings highlight critical limitations and biases in current MLLMs when learning multimodal medical tasks from context. SMMILE is available at https://smmile-benchmark.github.io.

Heterogeneous Large Language Model (LLM) fusion integrates the strengths of multiple source LLMs with different architectures into a target LLM with low computational overhead. While promising, existing methods suffer from two major limitations: 1) reliance on real data from limited domain for knowledge fusion, preventing the target LLM from fully acquiring knowledge across diverse domains, and 2) fixed data allocation proportions across domains, failing to dynamically adjust according to the target LLM's varying capabilities across domains, leading to a capability imbalance. To overcome these limitations, we propose Bohdi, a synthetic-data-only heterogeneous LLM fusion framework. Through the organization of knowledge domains into a hierarchical tree structure, Bohdi enables automatic domain exploration and multi-domain data generation through multi-model collaboration, thereby comprehensively extracting knowledge from source LLMs. By formalizing domain expansion and data sampling proportion allocation on the knowledge tree as a Hierarchical Multi-Armed Bandit problem, Bohdi leverages the designed DynaBranches mechanism to adaptively adjust sampling proportions based on the target LLM's performance feedback across domains. Integrated with our proposed Introspection-Rebirth (IR) mechanism, DynaBranches dynamically tracks capability shifts during target LLM's updates via Sliding Window Binomial Likelihood Ratio Testing (SWBLRT), further enhancing its online adaptation capability. Comparative experimental results on a comprehensive suite of benchmarks demonstrate that Bohdi significantly outperforms existing baselines on multiple target LLMs, exhibits higher data efficiency, and virtually eliminates the imbalance in the target LLM's capabilities. Our code is available at https://github.com/gjq100/Bohdi.git.

Computer use agents are LLM-based agents that can directly interact with a graphical user interface, by processing screenshots or accessibility trees. While these systems are gaining popularity, their safety has been largely overlooked, despite the fact that evaluating and understanding their potential for harmful behavior is essential for widespread adoption. To address this gap, we introduce OS-Harm, a new benchmark for measuring safety of computer use agents. OS-Harm is built on top of the OSWorld environment and aims to test models across three categories of harm: deliberate user misuse, prompt injection attacks, and model misbehavior. To cover these cases, we create 150 tasks that span several types of safety violations (harassment, copyright infringement, disinformation, data exfiltration, etc.) and require the agent to interact with a variety of OS applications (email client, code editor, browser, etc.). Moreover, we propose an automated judge to evaluate both accuracy and safety of agents that achieves high agreement with human annotations (0.76 and 0.79 F1 score). We evaluate computer use agents based on a range of frontier models - such as o4-mini, Claude 3.7 Sonnet, Gemini 2.5 Pro - and provide insights into their safety. In particular, all models tend to directly comply with many deliberate misuse queries, are relatively vulnerable to static prompt injections, and occasionally perform unsafe actions. The OS-Harm benchmark is available at https://github.com/tml-epfl/os-harm.

Jailbreak transferability is the surprising phenomenon when an adversarial attack compromising one model also elicits harmful responses from other models. Despite widespread demonstrations, there is little consensus on why transfer is possible: is it a quirk of safety training, an artifact of model families, or a more fundamental property of representation learning? We present evidence that transferability emerges from shared representations rather than incidental flaws. Across 20 open-weight models and 33 jailbreak attacks, we find two factors that systematically shape transfer: (1) representational similarity under benign prompts, and (2) the strength of the jailbreak on the source model. To move beyond correlation, we show that deliberately increasing similarity through benign only distillation causally increases transfer. Our qualitative analyses reveal systematic transferability patterns across different types of jailbreaks. For example, persona-style jailbreaks transfer far more often than cipher-based prompts, consistent with the idea that natural-language attacks exploit models' shared representation space, whereas cipher-based attacks rely on idiosyncratic quirks that do not generalize. Together, these results reframe jailbreak transfer as a consequence of representation alignment rather than a fragile byproduct of safety training.