ACL 2026 Papers

Misleading visualizations are a potent driver of misinformation on social media and the web. By violating chart design principles, they distort data and lead readers to draw inaccurate conclusions. Prior work has shown that both humans and multimodal large language models (MLLMs) are frequently deceived by such visualizations. Automatically detecting misleading visualizations and identifying the specific design rules they violate could help protect readers and reduce the spread of misinformation. However, the training and evaluation of AI models has been limited by the absence of large, diverse, and openly available datasets. In this work, we introduce Misviz, a benchmark of 2,604 real-world visualizations annotated with 12 types of misleaders. To support model training, we also release Misviz-synth, a synthetic dataset of 81,814 visualizations generated using Matplotlib and based on real-world data tables. We perform a comprehensive evaluation on both datasets using state-of-the-art MLLMs, rule-based systems, and fine-tuned classifiers. Our results reveal that the task remains highly challenging. We release Misviz, Misviz-synth, and the accompanying code.

Multilingual text embeddings are often assumed to encode meaning in a perspective-independent semantic space, yielding stable similarity judgments across tasks and languages. Our results show that this assumption does not hold in practice. We introduce SEA-BED, a large-scale benchmark covering 10 Southeast Asian (SEA) languages and diverse embedding tasks, designed to systematically examine how embedding performance varies across tasks, languages, and language-task combinations. Across extensive evaluations, we observe that no single model performs uniformly well across SEA languages; task difficulty differs markedly within languages, and success on one task does not reliably generalize to others. Language-task analyses further reveal highly non-uniform performance landscapes, where performance varies across different language-task combinations. These findings call for closer attention to performance measurements that provide an expansive view across languages and tasks to uncover inconsistencies in semantic representation. Based on these observations, we provide insights for future model development, including data, algorithmic, and architectural considerations.

Audio chaptering, the task of automatically segmenting long-form audio into coherent sections, is increasingly important for navigating podcasts, lectures, and videos. Despite its relevance, research remains limited and text-based, leaving key questions unresolved about leveraging audio information, handling ASR errors, and transcript-free evaluation. We address these gaps through three contributions: (1) a systematic comparison between text-based models with acoustic features, a novel audio-only architecture (AudioSeg) operating on learned audio representations, and multimodal LLMs; (2) empirical analysis of factors affecting performance, including transcript quality, acoustic features, duration, and speaker composition; and (3) formalized evaluation protocols contrasting transcript-dependent text-space protocols with transcript-invariant time-space protocols. Our experiments on YTSeg reveal that AudioSeg substantially outperforms text-based approaches, pauses provide the largest acoustic gains, and current MLLMs struggle due to context limitations and weak instruction following.

Despite strong medical benchmark accuracy, LLMs can exhibit severe multi-turn sycophancy in clinical dialogue, abandoning initial correct diagnosis under escalating pressure. We propose Med-Stress, a targeted stress test framework that evaluates belief stability under escalating pressure. Across nine frontier large language models (LLMs), we find a clear dissociation between medical knowledge and robustness: high initial diagnostic capability does not imply high belief stability, yielding large knowledge-robustness gaps for several LLMs. To mitigate this failure mode, we propose a lightweight inference-time defense, RBED (Role-Based Epistemic Defense), and R-FT (Resilience-oriented Fine-Tuning), a training-time approach that internalizes evidence-based resistance to pressure. Experiments show that R-FT nearly eliminates belief change and substantially improves robustness.

Lifelong learning investigates how models adapt when exposed to a potentially infinite stream of data. Most conventional approaches focus on updating model parameters (i.e., the neural network weights) as the underlying data distribution evolves over time. However, in natural language processing, model parameters are not the only components that matter. The tokenizer, a foundational part of the system, is usually assumed to remain fixed in lifelong learning scenarios. In this work, we challenge the validity of this assumption: as language evolves, a static tokenizer fragments newly emerging lexical items, reducing compression efficiency and consequently degrading the model performance. We introduce the Temporal Drift Tokenizer (Ted-Tok), which maintains an evolving vocabulary that adapts to emerging linguistic patterns over time. This adaptivity is driven by time-weighted frequency estimators that smooth short-term fluctuations to capture persistent linguistic trends, and a principled addition-deletion strategy targeting sink tokens. Across multiple domains, Ted-Tok consistently improves compression and task performance, with gains increasing under stronger drift, underscoring the role of tokenizer adaptivity in lifelong learning.

When asked a question in a language less seen in its training data, current reasoning large language models (RLMs) often exhibit dramatically lower performance than when asked the same question in English. In response, we introduce SP3F (Self-Play with Privileged Pairwise Feedback), a two-stage framework for enhancing multilingual reasoning without any data in the target language(s). First, we supervise fine-tune (SFT) on translated versions of English question-answer pairs to raise base model correctness. Second, we perform RL with feedback from a pairwise judge in a self-play fashion, with the judge receiving the English reference response as privileged information. Thus, even when none of the model’s responses are completely correct, the privileged pairwise judge can still tell which response is better. End-to-end, SP3F greatly improves base model performance, even outperforming fully post-trained models on multiple math and non-math tasks with less than 1/8 of the training data across the single-language, multilingual, and generalization to unseen language settings.

Lipid nanoparticles (LNPs) can deliver cargos to both tumor and immune cells, playing a crucial role in biomedicine. Traditional approaches rely on experimental screening and expert knowledge, which can be costly and time-consuming. Recent methods based on language models have accelerated this process using deep learning. Although these methods can retrieve molecules for fusion or rank candidates from existing libraries, they are still limited by the scope of known formulations. In this work, we propose a method, LiGen, to generate lipid molecules efficiently and actively, facilitating the discovery of high-performing LNP formulations. We first train a lipid-specific molecular language model, LiCore, to learn hidden representations of lipid molecules. We then explore the learned latent space to generate improved candidate formulations. This process is guided by a trained predictor, which evaluates delivery efficiency and provides directional signals. In reconstruction tasks, LiCore achieves nearly perfect reconstruction output with a low invalid ratio on both the LNP-Virtual900k and LNP-Exp12k datasets. The predictor consistently improves ranking-oriented metrics across multiple cell lines, with our method outperforming the best baselines by an average of 4.1%, 10.8%, and 8.1% in Top-50, Top-10, and Top-5 identification accuracy, respectively. Guided by the predictor, LiGen generates novel lipid candidates that achieve a 30.7% improvement over baseline methods on average, with some samples exceeding 50% improvement.

Large Language Models (LLMs) struggle to automate real-world vulnerability detection due to two key limitations: the heterogeneity of vulnerability patterns undermines the effectiveness of a single unified model, and manual prompt engineering for massive weakness categories is unscalable.To address these challenges, we propose MulVul, a retrieval-augmented multi-agent framework designed for precise and broad-coverage vulnerability detection. MulVul adopts a coarse-to-fine strategy: a Router agent first predicts the top- coarse categories and then forwards the input to specialized Detector agents, which identify the exact vulnerability types. Both agents use evidence retrieved from vulnerability knowledge bases to mitigate hallucinations. Crucially, to automate the generation of specialized prompts, we design Cross-Model Prompt Evolution, a prompt optimization mechanism where a generator LLM iteratively refines candidate prompts while a distinct executor LLM validates their effectiveness. This decoupling mitigates the self-correction bias inherent in single-model optimization. Evaluated on 130 CWE types, MulVul achieves 34.79% Macro-F1, outperforming the best baseline by 41.5%. Ablation studies validate cross-model prompt evolution, which boosts performance by 51.6% over manual prompts by effectively handling diverse vulnerability patterns.

While Chain-of-Thought (CoT) prompting has become a standard paradigm for eliciting complex reasoning capabilities in Large Language Models, it inadvertently exposes a new attack surface for backdoor attacks. Existing CoT backdoor attacks typically manipulate the intermediate reasoning steps to steer the model toward incorrect answers. However, these corrupted reasoning traces are readily detected by prevalent process-monitoring defenses. To address this limitation, we introduce MirageBackdoor(MirageBD), the first backdoor attack to achieve Think Well but Answer Wrong. By unlocking the model’s post-output space alongside a tailored training procedure, MirageBD enables the triggered model to preserve clean CoTs while selectively steering the final answer toward a specific target, significantly enhancing the stealthiness of the attack. Experiments show that MirageBD generally achieves over 90% attack success rate across four datasets and five models with a poison ratio of only 5%. Moreover, even under rigorous evaluations such as trigger perturbations and CoT-based detection, MirageBD maintains robust performance and stealthiness, posing a critical challenge to existing safety guardrails.

Model merging, typically on Instruct and Thinking models, has shown remarkable performance for efficient reasoning. In this paper, we systematically revisit the simplest merging method that interpolates two weights directly. Particularly, we observe that model interpolation follows a three-stage evolutionary paradigm with distinct behaviors on the reasoning trajectory. These dynamics provide a principled guide for navigating the performance-cost trade-off. Empirical results demonstrate that a strategically interpolated model surprisingly surpasses sophisticated model merging baselines on both efficiency and effectiveness. We further validate our findings with extensive ablation studies on model layers, modules, and decoding strategies. Ultimately, this work demystifies model interpolation and offers a practical framework for crafting models with precisely targeted reasoning capabilities.

The emergence of fine-grained numerical formats like NVFP4 presents new opportunities for efficient Large Language Model (LLM) inference. However, it is difficult to adapt existing Post-Training Quantization (PTQ) strategies to these formats: rotation-based methods compromise fine-grained block isolation; smoothing techniques struggle with significant 4-bit quantization errors; and mixed-precision approaches often conflict with hardware constraints on unified-precision computation. To address these challenges, we propose ARCQuant, a framework that boosts NVFP4 performance via Augmented Residual Channels. Distinct from methods that compromise block isolation or hardware uniformity, ARCQuant maintains a strictly unified NVFP4 format by augmenting the activation matrix with quantized residual channels. This design integrates the error compensation process directly into the matrix reduction dimension, enabling the use of standard, highly optimized GEMM kernels with minimal overhead. Theoretical analysis confirms that the worst-case error bound of our dual-stage NVFP4 quantization is comparable to that of standard 8-bit formats such as MXFP8. Extensive experiments on LLaMA and Qwen models demonstrate that ARCQuant achieves state-of-the-art accuracy, comparable to full-precision baselines in perplexity and downstream tasks. Furthermore, deployment on RTX 5090 and RTX PRO 6000 GPUs confirms practical benefits, achieving up to 3× speedup over FP16. Our code is available at https://github.com/actypedef/ARCQuant.

Chain-of-Thought (CoT) has become a standard method for improving reasoning capabilities in large language models (LLMs) by eliciting step-by-step thinking, but its effectiveness in multimodal tasks remains unclear. In this paper, we aim to systematically investigate the key question: What can multimodal Chain-of-Thought reasoning do, and where and why does it fall short? To this end, we evaluate 12 multimodal tasks across perception and reasoning categories using both 14 non-reasoning models and 8 reasoning models. Our analysis reveals several important findings: (1) CoT is not a free lunch and should be used selectively depending on the specific requirements of each task. For perception tasks, CoT can lead to undesirable side effects, such as reduced performance in visual grounding and object counting. In contrast, it proves effective for reasoning tasks involving mathematical, scientific, and multi-image reasoning; (2) Compared to original models, existing open-source multimodal reasoning models often yield only marginal overall improvements, possibly due to an overemphasis on mathematical reasoning at the expense of broader capabilities; (3) Visual reasoning remains a key bottleneck for current multimodal CoT, as models exhibit a Look Light, Think Heavy” pattern where verbal reflection rises and falls during reasoning, whereas visual reflection consistently diminishes. These findings suggest that while multimodal CoT handles verbal reflection relatively well, it lacks the ability to maintain deep visual introspection throughout the reasoning process.

Amidst the rapid advances of large language models (LLMs), most LLMs still struggle with mixed-language inputs, limited Code-switching (CSW) datasets, and evaluation biases, which hinder their deployment in multilingual societies. This survey provides the first comprehensive analysis of CSW-aware LLM research, reviewing 327 studies spanning five research areas, 15+ NLP tasks, 30+ datasets, and 80+ languages. We classify recent advances by architecture, training strategy, and evaluation methodology, outlining how LLMs have reshaped CSW modelling and what challenges persist. The paper concludes with a roadmap emphasizing the need for inclusive datasets, fair evaluation, and linguistically grounded models to achieve truly multilingual intelligence. A curated collection of all resources is maintained at https://github.com/lingo-iitgn/awesome-code-mixing/.

Retrieval-Augmented Generation (RAG) systems augment large language models with external knowledge, yet introduce a critical security vulnerability: RAG Knowledge Base Leakage, wherein adversarial prompts can induce the model to divulge retrieved proprietary content. Recent studies reveal that such leakage can be executed through adaptive and iterative attack strategies (named RAG extraction attack), while effective countermeasures remain notably lacking. To bridge this gap, we propose CanaryRAG, a runtime defense mechanism inspired by stack canaries in software security. CanaryRAG embeds carefully designed canary tokens into retrieved chunks and reformulates RAG extraction defense as a dual-path runtime integrity game. Leakage is detected in real time whenever either the target or oracle path violates its expected canary behavior, including under adaptive suppression and obfuscation. Extensive evaluations against existing attacks demonstrate that CanaryRAG provides robust defense, achieving substantially lower chunk recovery rates than state-of-the-art baselines while imposing negligible impact on task performance and inference latency. Moreover, as a plug-and-play solution, CanaryRAG can be seamlessly integrated into arbitrary RAG pipelines without requiring retraining or structural modifications, offering a practical and scalable safeguard for proprietary data.

Users increasingly rely on Large Language Models (LLMs) for Deep Research, using them to synthesize diverse sources into structured reports that support understanding and action. In this context, the practical reliability of such reports hinges on logical quality: whether the report’s claims and arguments are explicitly supported and can be trusted as a basis for downstream use, rather than merely appearing fluent or informative. However, current evaluation frameworks largely overlook this requirement. To bridge this gap, we introduce ReportLogic, a benchmark that quantifies report-level logical quality through a reader-centric lens of auditability. Specifically, ReportLogic adopts a hierarchical taxonomy that evaluates whether readers can (1) trace an on-topic report structure with a unified analytical arc (Macro-Logic), (2) understand the progression with necessary context (Expositional-Logic), and (3) verify conclusions via explicit claim–support (Structural-Logic). Based on this taxonomy, we construct a human-annotated rubric-guided dataset and train an open-source LogicJudge for scalable evaluation. We further evaluate judge robustness via adversarial attacks, showing that off-the-shelf LLM judges are frequently influenced by superficial cues (e.g., verbosity), and reasoning modes can mask broken support relations. Overall, our results provide actionable guidance for building more robust logic evaluators and improving the logical reliability of LLM-generated reports.

Large Language Model (LLM) routing is a pivotal technique for navigating a diverse landscape of LLMs, enabling the selection of the best-performing LLMs for specific user queries while balancing performance and cost. However, current routing approaches often face limitations in scalability when dealing with a large pool of specialized LLMs, or in their adaptability to extending model scope and evolving capability domains. To overcome those challenges, we propose InferenceDynamics, a flexible and scalable multi-dimensional routing framework by modeling the capability and knowledge of models. We operate it on our comprehensive dataset RouteMix, and demonstrate its effectiveness and generalizability in group-level routing using modern benchmarks including MMLU-Pro, GPQA, BigGenBench, and LiveBench, showcasing its ability to identify and leverage top-performing models for given tasks, leading to superior outcomes with cost efficiency. The broader adoption of InferenceDynamics can empower users to harness the full specialized potential of the LLM ecosystem, and our code will be made publicly available to encourage further research.

Tasks such as mathematical problem solving and coding require models to leverage chain-of-thought (CoT) processes, enabling human-like reasoning strategies. However, the advancement of large reasoning models (LRMs) is hindered by the lack of comprehensive CoT datasets. Existing resources often fail to provide extensive reasoning problems with coherent CoT processes distilled from multiple teacher models, and do not account for multifaceted properties describing the internal characteristics of CoTs.To address these challenges, we introduce OmniThought, a large-scale dataset featuring 2 million CoT processes generated and validated by multiple powerful LRMs. Each CoT process in OmniThought is annotated with novel Reasoning Verbosity (RV) and Cognitive Difficulty (CD) scores, which characterize the appropriateness of CoT verbosity and the cognitive difficulty level for models to comprehend these reasoning processes. We further establish a self-reliant pipeline to curate this dataset. Extensive experiments using Qwen2.5 and Qwen3 of various sizes demonstrate the positive impact of our RV and CD scores on LRM training effectiveness. Based on the OmniThought dataset, we train and release a series of high-performing LRMs with enhanced reasoning abilities and optimized CoT output length. Our contributions advance the development of LRMs across different scales for solving complex reasoning tasks.

Recent advancements in Multimodal Large Language Models (MLLMs) have demonstrated significant improvement in offline video understanding. However, extending these capabilities to streaming video inputs, remains challenging, as existing models struggle to simultaneously maintain stable understanding performance, real-time responses, and low GPU memory overhead. To address this challenge, we propose HERMES, a novel training-free architecture for real-time and accurate understanding of video streams. Based on a mechanistic attention investigation, we conceptualize KV cache as a hierarchical memory framework that encapsulates video information across multiple granularities. During inference, HERMES reuses a compact KV cache, enabling efficient streaming understanding under resource constraints. Notably, HERMES requires no auxiliary computations upon the arrival of user queries, thereby guaranteeing real-time responses for continuous video stream interactions. HERMES achieves 10\times faster TTFT compared to prior SOTA. Even when reducing video tokens by up to 68% compared with uniform sampling, HERMES achieves superior or comparable accuracy across all benchmarks, with up to 11.4% gains on streaming datasets.

Large Language Models (LLMs) have achieved high accuracy on complex commonsense and mathematical problems that involve the composition of multiple reasoning steps. However, current compositional benchmarks testing these skills tend to focus on either commonsense or math reasoning, whereas LLM agents solving real-world tasks would require a combination of both. In this work, we introduce an Agentic Commonsense and Math benchmark (AgentCoMa), where each compositional task requires a commonsense reasoning step and a math reasoning step. We test it on 61 LLMs of different sizes, model families, and training strategies. We find that LLMs can usually solve both steps in isolation, yet their accuracy drops by nearly 30% on average when the two are combined. This is a substantially greater performance gap than the one we observe in prior compositional benchmarks that combine multiple steps of the same reasoning type. In contrast, non-expert human annotators can solve the compositional questions and the individual steps in AgentCoMa with similarly high accuracy. Furthermore, we conduct a series of interpretability studies to better understand the performance gap, examining neuron patterns, attention maps and membership inference. Our work underscores a substantial degree of model brittleness in the context of mixed-type compositional reasoning and offers a test bed for future improvement.