ACL 2025 Papers

In generative commonsense reasoning tasks such as CommonGen, generative large language models (LLMs) compose sentences that include all given concepts. However, when focusing on instruction-following capabilities, if a prompt specifies a concept order, LLMs must generate sentences that adhere to the specified order. To address this, we propose Ordered CommonGen, a benchmark designed to evaluate the compositional generalization and instruction-following abilities of LLMs. This benchmark measures ordered coverage to assess whether concepts are generated in the specified order, enabling a simultaneous evaluation of both abilities. We conducted a comprehensive analysis using 36 LLMs and found that, while LLMs generally understand the intent of instructions, biases toward specific concept order patterns often lead to low-diversity outputs or identical results even when the concept order is altered. Moreover, even the most instruction-compliant LLM achieved only about 75% ordered coverage, highlighting the need for improvements in both instruction-following and compositional generalization capabilities.

Arabic has diverse dialects, where one dialect can be substantially different from the others. In the NLP literature, some assumptions about these dialects are widely adopted (e.g., “Arabic dialects can be grouped into distinguishable regional dialects”) and are manifested in different computational tasks such as Arabic Dialect Identification (ADI). However, these assumptions are not quantitatively verified. We identify four of these assumptions and examine them by extending and analyzing a multi-label dataset, where the validity of each sentence in 11 different country-level dialects is manually assessed by speakers of these dialects. Our analysis indicates that the four assumptions oversimplify reality, and some of them are not always accurate. This in turn might be hindering further progress in different Arabic NLP tasks.

The research on classical Chinese event extraction trends to directly graft the complex modeling from English or modern Chinese works, neglecting the utilization of the unique characteristic of this language. We argue that, compared with grafting the sophisticated methods from other languages, focusing on classical Chinese’s inimitable source of __Ancient Literature__ could provide us with extra and comprehensive semantics in event extraction. Motivated by this, we propose a Literary Vision-Language Model (VLM) for classical Chinese event extraction, integrating with literature annotations, historical background and character glyph to capture the inner- and outer-context information from the sequence. Extensive experiments build a new state-of-the-art performance in the GuwenEE, CHED datasets, which underscores the effectiveness of our proposed VLM, and more importantly, these unique features can be obtained precisely at nearly zero cost.

In this work, we identify the “2D-Cheating” problem in 3D LLM evaluation, where these tasks might be easily solved by VLMs with rendered images of point clouds, exposing ineffective evaluation of 3D LLMs’ unique 3D capabilities. We test VLM performance across multiple 3D LLM benchmarks and, using this as a reference, propose principles for better assessing genuine 3D understanding. We also advocate explicitly separating 3D abilities from 1D or 2D aspects when evaluating 3D LLMs.

Large language models (LLMs) have demonstrated significant advancements in code generation, yet they still face challenges when tackling tasks that extend beyond their basic capabilities. Recently, the concept of self-debugging has been proposed as a way to enhance code generation performance by leveraging execution feedback from tests. However, the availability of high-quality tests in real-world scenarios is often limited. In this context, self-debugging with self-generated tests emerges as a promising solution, though its limitations and practical potential have not been fully explored. To address this gap, we investigate the efficacy of self-debugging in code generation tasks. We propose and analyze two distinct paradigms for the self-debugging process: post-execution and in-execution self-debugging. Our findings reveal that post-execution self-debugging struggles with the test bias introduced by self-generated tests, which can lead to misleading feedback. In contrast, in-execution self-debugging enables LLMs to mitigate this bias and leverage intermediate states during program execution. By focusing on runtime information rather than relying solely on potentially flawed self-generated tests, this approach demonstrates significant promise for improving the robustness and accuracy of LLMs in code generation tasks.

The effectiveness of large language models (LLMs) in conversational AI is hindered by their reliance on single-turn supervised fine-tuning (SFT) data, which limits contextual coherence in multi-turn dialogues. Existing methods for generating multi-turn dialogue data struggle to ensure both diversity and quality in instructions. To address this, we propose Review-Instruct, a novel framework that synthesizes multi-turn conversations through an iterative “Ask-Respond-Review” process involving three agent roles: a Candidate, multiple Reviewers, and a Chairman. The framework iteratively refines instructions by incorporating Reviewer feedback, enhancing dialogue diversity and difficulty. We construct a multi-turn dataset using the Alpaca dataset and fine-tune the LLaMA2-13B model. Evaluations on MT-Bench, MMLU-Pro, and Auto-Arena demonstrate significant improvements, achieving absolute gains of 2.9% on MMLU-Pro and 2% on MT-Bench compared to prior state-of-the-art models based on LLaMA2-13B. Ablation studies confirm the critical role of the Review stage and the use of multiple Reviewers in boosting instruction diversity and difficulty. Our work highlights the potential of review-driven, multi-agent frameworks for generating high-quality conversational data at scale.

Instruction following (IF) is a critical capability for large language models (LLMs). However, handling complex instructions with multiple constraints remains challenging. Previous methods typically select preference pairs based on the number of constraints they satisfy, introducing noise where chosen examples may fail to follow some constraints and rejected examples may excel in certain respects over the chosen ones. To address the challenge of aligning with multiple preferences, we propose a simple yet effective method called Reverse Preference Optimization (RPO). It mitigates noise in preference pairs by dynamically reversing the constraints within the instruction to ensure the chosen response is perfect, alleviating the burden of extensive sampling and filtering to collect perfect responses. Besides, reversal also enlarges the gap between chosen and rejected responses, thereby clarifying the optimization direction and making it more robust to noise. We evaluate RPO on two multi-turn IF benchmarks, Sysbench and Multi-IF, demonstrating average improvements over the DPO baseline of 4.6 and 2.5 points (on Llama-3.1 8B), respectively. Moreover, RPO scales effectively across model sizes (8B to 70B parameters), with the 70B RPO model surpassing GPT-4o.

Large language models (LLMs) have shown remarkable performance in reasoning tasks but face limitations in mathematical and complex logical reasoning. Existing methods to improve LLMs’ logical capabilities either involve traceable or verifiable logical sequences that generate more reliable responses by constructing logical structures yet increase computational costs, or introduces rigid logic template rules, reducing flexibility. In this paper, we propose Reversal of Thought (RoT), a plug-and-play and cost-effective reasoning framework designed to enhance the logical reasoning abilities of LLMs during the warm-up phase prior to batch inference. RoT utilizes a Preference-Guided Reverse Reasoning warm-up strategy, which integrates logical symbols for pseudocode planning through meta-cognitive mechanisms and pairwise preference self-evaluation to generate task-specific prompts solely through demonstrations, aligning with LLMs’ cognitive preferences shaped by RLHF. Through reverse reasoning, we utilize a Cognitive Preference Manager to assess knowledge boundaries and further expand LLMs’ reasoning capabilities by aggregating solution logic for known tasks and stylistic templates for unknown tasks. Experiments across various tasks demonstrate that RoT surpasses existing baselines in both reasoning accuracy and efficiency.

Knowledge editing, which aims to update the knowledge encoded in language models, can be deceptive. Despite the fact that many existing knowledge editing algorithms achieve near-perfect performance on conventional metrics, the models edited by them are still prone to generating original knowledge. This paper introduces the concept of "**superficial editing**” to describe this phenomenon. Our comprehensive evaluation reveals that this issue presents a significant challenge to existing algorithms. Through systematic investigation, we identify and validate two key factors contributing to this issue: (1) the residual stream at the last subject position in earlier layers and (2) specific attention modules in later layers. Notably, certain attention heads in later layers, along with specific left singular vectors in their output matrices, encapsulate the original knowledge and exhibit a causal relationship with superficial editing. Furthermore, we extend our analysis to the task of superficial unlearning, where we observe consistent patterns in the behavior of specific attention heads and their corresponding left singular vectors, thereby demonstrating the robustness and broader applicability of our methodology and conclusions. Our code is available at https://github.com/jiakuan929/superficial-editing.

Large language models (LLMs) have advanced significantly due to the attention mechanism, but their quadratic complexity and linear memory demands limit their performance on long-context tasks. Recently, researchers introduced Mamba, an advanced model built upon State Space Models (SSMs) that offers linear complexity and constant memory. Although Mamba is reported to match or surpass the performance of attention-based models, our analysis reveals a performance gap: Mamba excels in tasks that involve localized key information but faces challenges with tasks that require handling distributed key information. Our controlled experiments suggest that the inconsistency arises from Mamba’s reliance on **local pattern shortcuts** across model scales (10M to 1.4B), which enable Mamba to remember local key information within its limited memory but hinder its ability to retain more dispersed information. Therefore, we introduce a global gate module into the Mamba model to address this issue. Experiments on extensive synthetic tasks, as well as real-world tasks, demonstrate the effectiveness of our method. Notably, with the introduction of only 4M extra parameters, our approach enables the Mamba model (130M) to achieve a significant improvement on tasks with distributed information, increasing its performance from **below 5% to 80%**.

The ability of Natural Language Processing (NLP) methods to categorize text into multiple classes has motivated their use in online content moderation tasks, such as hate speech and fake news detection. However, there is limited understanding of how or why these methods make such decisions, or why certain content is moderated in the first place. To investigate the hidden mechanisms behind content moderation, we explore multiple directions: 1) training classifiers to reverse-engineer content moderation decisions across countries; 2) explaining content moderation decisions by analyzing Shapley values and LLM-guided explanations. Our primary focus is on content moderation decisions made across countries, using pre-existing corpora sampled from the Twitter Stream Grab. Our experiments reveal interesting patterns in censored posts, both across countries and over time. Through human evaluations of LLM-generated explanations across three LLMs, we assess the effectiveness of using LLMs in content moderation. Finally, we discuss potential future directions, as well as the limitations and ethical considerations of this work.

Multi-turn interactions between large language models (LLMs) and users naturally include implicit feedback signals. If an LLM responds in an unexpected way to an instruction, the user is likely to signal it by rephrasing the request, expressing frustration, or pivoting to an alternative task. Such signals are task-independent and occupy a relatively constrained subspace of language, allowing the LLM to identify them even if it fails on the actual task. We introduce ReSpect, a method to learn from such signals in past interactions via retrospection without additional annotations. We deploy ReSpect in a new multimodal interaction scenario, where humans instruct a multimodal LLM to solve an abstract reasoning task with a combinatorial solution space. Through thousands of interactions with humans, we show how ReSpect gradually improves task completion rate from 31% to 82%, all without any external annotation.

Current language models (LMs) use a fixed, static subword tokenizer. This default choice typically results in degraded efficiency and language capabilities, especially in languages other than English. To address this issue, we challenge the static design and propose retrofitting LMs with dynamic tokenization: a way to dynamically decide on token boundaries based on the input text via a subword-merging algorithm inspired by byte-pair encoding. We merge frequent subword sequences in a batch, then apply a pre-trained embedding-prediction hypernetwork to compute the token embeddings on-the-fly. For encoder-style models (e.g., XLM-R), this on average reduces token sequence lengths by >20% across 14 languages while degrading performance by less than 2%. The same method applied to pre-filling and scoring in decoder-style models (e.g., Mistral-7B) results in minimal performance degradation at up to 17% reduction in sequence length. Overall, we find that dynamic tokenization can mitigate the limitations of static tokenization by substantially improving inference speed and promoting fairness across languages, enabling more equitable and adaptable LMs.

Large language models (LLMs) exhibit remarkable generative capabilities but often suffer from hallucinations. Retrieval-augmented generation (RAG) offers an effective solution by incorporating external knowledge, but existing methods still face several limitations: additional deployment costs of separate retrievers, redundant input tokens from retrieved text chunks, and the lack of joint optimization of retrieval and generation. To address these issues, we propose RetroLLM, a unified framework that integrates retrieval and generation into a single, auto-regressive process, enabling LLMs to directly generate fine-grained evidence from the corpus with constrained decoding. Moreover, to mitigate false pruning in the process of constrained evidence generation, we introduce (1) hierarchical FM-Index constraints, which generate corpus-constrained clues to identify a subset of relevant documents before evidence generation, reducing irrelevant decoding space; and (2) a forward-looking constrained decoding strategy, which considers the relevance of future sequences to improve evidence accuracy. Extensive experiments on five open-domain QA datasets demonstrate RetroLLM’s superior performance across both in-domain and out-of-domain tasks. The code is available at https://anonymous.4open.science/r/RetroLLM-D95A.

Language models hold incredible promise for enabling scientific discovery by synthesizing massive research corpora. Many complex scientific research questions have multiple plausible answers, each supported by evidence of varying strength. However, existing language models lack the capability to quantitatively and faithfully compare answer plausibility in terms of supporting evidence. To address this, we introduce Retrieve to Explain (R2E), a retrieval-based model that scores and ranks all possible answers to a research question based on evidence retrieved from a document corpus. The architecture represents each answer only in terms of its supporting evidence, with the answer itself masked. This allows us to extend feature attribution methods such as Shapley values, to transparently attribute answer scores to supporting evidence at inference time. The architecture also allows incorporation of new evidence without retraining, including non-textual data modalities templated into natural language. We developed R2E for the challenging scientific discovery task of drug target identification, a human-in-the-loop process where failures are extremely costly and explainability paramount. When predicting whether drug targets will subsequently be confirmed as efficacious in clinical trials, R2E not only matches non-explainable literature-based models but also surpasses a genetics-based target identification approach used throughout the pharmaceutical industry.

While large language models (LLMs) have significantly advanced mathematical reasoning, Process Reward Models (PRMs) have been developed to evaluate the logical validity of reasoning steps. However, PRMs still struggle with out-of-distribution (OOD) challenges. This paper identifies the OOD issues including step OOD, arising from differences in reasoning patterns across model types and sizes, and question OOD, due to dataset shifts between training and real-world problems. To address these issues, we introduce Retrieval-Augmented Process Reward Model (RetrievalPRM), a novel framework designed to tackle these OOD issues. By utilizing a two-stage retrieval-enhanced mechanism, RetrievalPRM retrieves semantically similar questions and steps for PRM as a warmup to stimulate its potential to judge target steps, improving generalization and reasoning consistency across different models and problem types. Our extensive experiments demonstrate that RetrievalPRM outperforms existing baselines across multiple real-world datasets. Our open-source contributions include a retrieval-enhanced dataset, a tuning framework for PRM training, and the RetreivalPRM model, establishing a new standard for PRM performance.

Visual programming languages (VPLs) allow users to create programs through graphical interfaces, which results in easier accessibility and their widespread usage in various domains. To further enhance this accessibility, recent research has focused on generating VPL code from user instructions using large language models (LLMs). Specifically, by employing prompting-based methods, these studies have shown promising results. Nevertheless, such approaches can be less effective for industrial VPLs such as Ladder Diagram (LD). LD is a pivotal language used in industrial automation processes and involves extensive domain-specific configurations, which are difficult to capture in a single prompt. In this work, we demonstrate that training-based methods outperform prompting-based methods for LD generation accuracy, even with smaller backbone models. Building on these findings, we propose a two-stage training strategy to further enhance VPL generation. First, we employ retrieval-augmented fine-tuning to leverage the repetitive use of subroutines commonly seen in industrial VPLs. Second, we apply direct preference optimization (DPO) to further guide the model toward accurate outputs, using systematically generated preference pairs through graph editing operations. Extensive experiments on real-world LD data demonstrate that our approach improves program-level accuracy by over 10% compared to supervised fine-tuning, which highlights its potential to advance industrial automation.

Despite significant advancements in Large Vision-Language Models, Object Hallucination (OH) remains a persistent challenge. Building upon prior studies on contrastive decoding that address this issue without requiring additional model training, we introduce RVCD (Retrieval Visual Contrastive Decoding), an advanced method to suppress OH. RVCD leverages both negative and positive images at the logit level, explicitly referencing AI-generated images designed to represent a single concept. Our approach demonstrates substantial improvements over existing decoding-based methods.

Tool learning aims to augment large language models (LLMs) with diverse tools, enabling them to act as agents for solving practical tasks. Due to the limited context length of tool-using LLMs, adopting information retrieval (IR) models to select useful tools from large toolsets is a critical initial step. However, the performance of IR models in tool retrieval tasks remains underexplored and unclear. Most tool-use benchmarks simplify this step by manually pre-annotating a small set of relevant tools for each task, which is far from the real-world scenarios. In this paper, we propose ToolRet, a heterogeneous tool retrieval benchmark comprising 7.6k diverse retrieval tasks, and a corpus of 43k tools, collected from existing datasets. We benchmark six types of models on ToolRet. Surprisingly, even the models with strong performance in conventional IR benchmarks, exhibit poor performance on ToolRet. This low retrieval quality degrades the task pass rate of tool-use LLMs. As a further step, we contribute a large-scale training dataset with over 200k instances, which substantially optimizes the tool retrieval ability of IR models.