Large language models (LLMs) are increasingly deployed with hierarchical instruction schemes, where certain instructions (e.g., system-level directives) are expected to take precedence over others (e.g., user messages). Yet, we lack a systematic understanding of how effectively these hierarchical control mechanisms work. We introduce a systematic evaluation framework based on constraint prioritization to assess how well LLMs enforce instruction hierarchies. Our experiments across six state-of-the-art LLMs reveal that models struggle with consistent instruction prioritization, even for simple formatting conflicts. We find that the widely-adopted system/user prompt separation fails to establish a reliable instruction hierarchy, and models exhibit strong inherent biases toward certain constraint types regardless of their priority designation. Interestingly, we also find that societal hierarchy framings (e.g., authority, expertise, consensus) show stronger influence on model behavior than system/user roles, suggesting that pretraining-derived social structures function as latent behavioral priors with potentially greater impact than post-training guardrails.

Saeed Almheiri, Bilal Elbouardi, Salsabila Zahirah Pranida, Irina Nikishina, Ashwath Rao B, Parameswari Krishnamurthy, Muhammad Cendekia Airlangga, Rifo Ahmad Genadi, Nguyen Phan Gia Bao, Amir Hossein Yari, Hawau Olamide Toyin, Nurdaulet Mukhituly, Mena Attia, Besher Hassan, Ahmad Fathan Hidayatullah, Tatsuki Kuribayashi, Haonan Li, Suma Bhat, Fajri Koto. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Zhuohan Xie, Daniil Orel, Rushil Thareja, Dhruv Sahnan, Hachem Madmoun, Fan Zhang, Debopriyo Banerjee, Georgi Nenkov Georgiev, Xueqing Peng, Lingfei Qian, Jimin Huang, Jinyan Su, Aaryamonvikram Singh, Rui Xing, Rania Elbadry, Chen Xu, Haonan Li, Fajri Koto, Ivan Koychev, Tanmoy Chakraborty, Yuxia Wang, Salem Lahlou, Veselin Stoyanov, Sophia Ananiadou, Preslav Nakov. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Neural Theorem Proving (NTP) employs Large Language Models (LLMs) to automate formal proofs in proof assistants. While LLMs have achieved relatively remarkable success in informal reasoning tasks using natural languages, the transition to mechanized formal theorem proving presents persistent challenges. Mechanized proof languages often contain many syntactic constructs and diverse, specialized proof tactics, which facilitate expert use but have no direct counterpart in informal mathematical proofs. These prover-specific idioms represent an additional burden for LLM-based NTPs that might be otherwise successful in generating informal proofs. Seeking to bridge this gap between formal proof construction and informal reasoning, in order to better facilitate NTP, this work approaches these challenges from a language design perspective. We look at common reasoning patterns in informal proofs and in existing mechanized proofs, and design Minilang (formally named Isabelle/Minilang), a minimalist proof language that captures these reasoning patterns. In contrast to proof languages (informal and formal) that often feature a large collection of operations with unclear semantic boundaries, Minilang is deliberately kept minimalist – its core design comprises only 10 proof operations, each with clear semantic distinctions. We further develop a rule-based translator from Isabelle’s proof language (Isar) to Minilang, translating ∼340,000 existing Isabelle proofs with an ∼85% success rate. Using this translated corpus, we finetune two LLMs to compare machine learning performance on Minilang versus the original Isar language. Experiments show Minilang benefits the two LLMs by improving the pass@1 success rate on the PISA benchmark by up to 20/29 percentage points in comparison to the Isar-based LLMs w/wo Sledgehammer. The pass@1 rate reaches 69.1%, exceeding the prior work Baldur’s pass@64 (65.7%); the pass@8 rate reaches 79.2%, exceeding the state-of-the-art on PISA (71.0%) achieved by Magnushammer.

As large language models (LLMs) advance, their inability to autonomously execute tasks by directly interacting with external tools remains a critical limitation. Traditional methods rely on inputting tool descriptions as context, which is constrained by context length and requires separate, often inefficient, retrieval mechanisms. We introduce ToolGen, a paradigm shift that integrates tool knowledge directly into the LLM’s parameters by representing each tool as a unique token. This enables the LLM to generate tool calls and arguments as part of its next token prediction capabilities, seamlessly blending tool invocation with language generation. Our framework allows the LLM to access and utilize a vast amount of tools with no additional retrieval step, significantly enhancing both performance and scalability. Experimental results with over 47,000 tools show that ToolGen not only achieves superior results in both tool retrieval and autonomous task completion but also sets the stage for a new era of AI agents that can adapt to tools across diverse domains. By fundamentally transforming tool retrieval into a generative process, ToolGen paves the way for more versatile, efficient, and autonomous AI systems. ToolGen enables end-to-end tool learning and opens opportunities for integration with other advanced techniques such as chain-of-thought and reinforcement learning, thereby expanding the practical capabilities of LLMs

Renxi Wang, Xudong Han, Yixuan Zhang, Timothy Baldwin, Haonan Li. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

Reinforcement learning (RL) has shown promise in enhancing large language model (LLM) reasoning, yet progress towards broader capabilities is limited by the availability of high-quality, multi-domain datasets. This work introduces \ours, a 92K RL-for-reasoning dataset designed to address this gap, covering six reasoning domains: Math, Code, Science, Logic, Simulation, and Tabular, each with corresponding verifiers. We build \ours via a careful data-curation pipeline, including sourcing, deduplication, reward design, and domain-specific and difficulty-based filtering, to facilitate the systematic investigation of cross-domain RL generalization. Our study using \ours suggests the efficacy of a simple mixed-domain RL training approach and reveals several key aspects affecting cross-domain transferability. We further train two models {\ours}-7B and {\ours}-32B purely with RL on our curated data and observe largely improved performance over leading open RL reasoning model baselines, with gains of 7.3\% and 7.8\% respectively on an extensive 17-task, six-domain evaluation suite. We are releasing our dataset, code, and evaluation suite to the community, aiming to support further research and development of more general RL-enhanced reasoning models.

Generative models are rapidly gaining popularity and being integrated into everyday applications, raising concerns over their safe use as various vulnerabilities are exposed. In light of this, the field of red teaming is undergoing fast-paced growth, highlighting the need for a comprehensive survey covering the entire pipeline and addressing emerging topics. Our extensive survey, which examines over 120 papers, introduces a taxonomy of fine-grained attack strategies grounded in the inherent capabilities of language models. Additionally, we have developed the “searcher” framework to unify various automatic red teaming approaches. Moreover, our survey covers novel areas including multimodal attacks and defenses, risks around LLM-based agents, overkill of harmless queries, and the balance between harmlessness and helpfulness. Warning: This paper contains examples that may be offensive, harmful, or biased.

Rocktim Das, Simeon Hristov, Haonan Li, Dimitar Dimitrov, Ivan Koychev, Preslav Nakov. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024.

Multimodal large language models (MLLMs) have shown impressive success across modalities such as image, video, and audio in a variety of understanding and generation tasks. However, current MLLMs are surprisingly poor at understanding webpage screenshots and generating their corresponding HTML code. To address this problem, we propose Web2Code, a benchmark consisting of a new large-scale webpage-to-code dataset for instruction tuning and an evaluation framework for the webpage understanding and HTML code translation abilities of MLLMs. For dataset construction, we leverage pretrained LLMs to enhance existing webpage-to-code datasets as well as generate a diverse pool of new webpages rendered into images. Specifically, the inputs are webpage images and instructions, while the responses are the webpage's HTML code. We further include diverse natural language QA pairs about the webpage content in the responses to enable a more comprehensive understanding of the web content. To evaluate model performance in these tasks, we develop an evaluation framework for testing MLLMs' abilities in webpage understanding and web-to-code generation. Extensive experiments show that our proposed dataset is beneficial not only to our proposed tasks but also in the general visual domain. We hope our work will contribute to the development of general MLLMs suitable for web-based content generation and task automation. Our data and code are available at https://github.com/MBZUAI-LLM/web2code.

Although large language models (LLMs) are often pre-trained on large-scale multilingual texts, their reasoning abilities and real-world knowledge are mainly evaluated based on English datasets.Assessing LLM capabilities beyond English is increasingly vital but hindered due to the lack of suitable datasets.In this work, we introduce IndoMMLU, the first multi-task language understanding benchmark for Indonesian culture and languages, which consists of questions from primary school to university entrance exams in Indonesia.By employing professional teachers, we obtain 14,981 questions across 64 tasks and education levels, with 46% of the questions focusing on assessing proficiency in the Indonesian language and knowledge of nine local languages and cultures in Indonesia.Our empirical evaluations show that GPT-3.5 only manages to pass the Indonesian primary school level, with limited knowledge of local Indonesian languages and culture.Other smaller models such as BLOOMZ and Falcon perform at even lower levels.1

Most existing pre-trained language representation models (PLMs) are sub-optimal in sentiment analysis tasks, as they capture the sentiment information from word-level while underconsidering sentence-level information.In this paper, we propose SentiWSP, a novel Sentiment-aware pre-trained language model with combined Word-level and Sentence-level Pre-training tasks.The word level pre-training task detects replaced sentiment words, via a generator-discriminator framework, to enhance the PLM's knowledge about sentiment words.The sentence level pre-training task further strengthens the discriminator via a contrastive learning framework, with similar sentences as negative samples, to encode sentiments in a sentence.Extensive experimental results show that SentiWSP achieves new state-of-the-art performance on various sentence-level and aspectlevel sentiment classification benchmarks.We have made our code and model publicly available at https://github.com/XMUDM/SentiWSP.

Haonan Li, Martin Tomko, Maria Vasardani, Timothy Baldwin. Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2022.

In this work, we explore character-level neural syntactic parsing for Chinese with two typical syntactic formalisms: the constituent formalism and a dependency formalism based on a newly released character-level dependency treebank. Prior works in Chinese parsing have struggled with whether to de ne words when modeling character interactions. We choose to integrate full character-level syntactic dependency relationships using neural representations from character embeddings and richer linguistic syntactic information from human-annotated character-level Parts-Of-Speech and dependency labels. This has the potential to better understand the deeper structure of Chinese sentences and provides a better structural formalism for avoiding unnecessary structural ambiguities. Specifically, we first compare two different character-level syntax annotation styles: constituency and dependency. Then, we discuss two key problems for character-level parsing: (1) how to combine constituent and dependency syntactic structure in full character-level trees and (2) how to convert from character-level to word-level for both constituent and dependency trees. In addition, we also explore several other key parsing aspects, including di erent character-level dependency annotations and joint learning of Parts-Of-Speech and syntactic parsing. Finally, we evaluate our models on the Chinese Penn Treebank (CTB) and our published Shanghai Jiao Tong University Chinese Character Dependency Treebank (SCDT). The results show the e effectiveness of our model on both constituent and dependency parsing. We further provide empirical analysis and suggest several directions for future study.

Existing metonymy resolution approaches rely on features extracted from external resources like dictionaries and hand-crafted lexical resources.In this paper, we propose an end-to-end word-level classification approach based only on BERT, without dependencies on taggers, parsers, curated dictionaries of place names, or other external resources.We show that our approach achieves the state-of-the-art on 5 datasets, surpassing conventional BERT models and benchmarks by a large margin.We also show that our approach generalises well to unseen data.

This paper presents a truly full character-level neural dependency parser together with a newly released character-level dependency treebank for Chinese, which has suffered a lot from the dilemma of defining word or not to model character interactions. Integrating full character-level dependencies with character embedding and human annotated character-level part-of-speech and dependency labels for the first time, we show an extra performance enhancement from the evaluation on Chinese Penn Treebank and SJTU (Shanghai Jiao Tong University) Chinese Character Dependency Treebank and the potential of better understanding deeper structure of Chinese sentences.