ACL 2026 Papers

Quality management when creating large-scale speech datasets is essential for building reliable downstream models, yet verification pipelines are often brittle, domain-specific, and expertise-intensive. We introduce SpeechQM-Agent, a natural language-driven agentic framework that compiles user requirements into dependency-aware DAG workflows over modular tools for audio, transcript, and metadata verification. A central planner LLM enforces prerequisites and supports execution-time replanning (e.g., re-running failed steps or swapping tools), reducing manual pipeline engineering and improving robustness across heterogeneous vendor formats and multilingual settings. We also release SpeechQM-Dataset, a multilingual benchmark with controlled, vendor-inspired quality artifacts spanning 24 verification tasks. Across experiments, SpeechQM-Agent attains 80-90% agreement with expert verification while requiring <20% of the cost and time of manual QC, and we further validate transfer to real vendor-supplied corpora. Planner LLM comparisons highlight fidelity-efficiency trade-offs.

Recent advancements in large language models (LLMs) have demonstrated emergent capabilities in complex reasoning, largely spurred by rule-based Reinforcement Learning (RL) techniques applied during post-training. This has raised the question of whether similar methods can instill more nuanced, human-like social intelligence, such as a Theory of Mind (ToM), in LLMs. This paper investigates whether small- scale LLMs can acquire a robust and generalizable ToM capability through RL with verifiable rewards (RLVR). We conduct a systematic evaluation by training models on various combinations of prominent ToM benchmarks (HiToM, ExploreToM, FANToM) and testing for generalization on held-out benchmarks (e.g., Open- ToM). Our findings indicate that small LLMs struggle to develop a generic ToM capability. While performance on in-distribution tasks improves, this capability fails to transfer to unseen ToM tasks with different characteristics. Even observed out-of-distribution (OOD) performance improvements occur unpredictably across the training run, and don’t generalize across other OOD benchmarks. Furthermore, we conduct analysis to show that the learned behavior is likely a form of narrow overfitting rather than the acquisition of a true, abstract ToM capability.

Building datasets for dialogue tasks is expensive and time-consuming, requiring recruitment, training, and data collection from study participants. In response, much recent work has sought to use large language models (LLMs) to simulate both human-human and human-LLM interactions, as they have been shown to generate convincingly human-like text in many settings. However, how well do LLM-based simulations reflect real human dialogue? In this work, we answer this question by generating a large-scale dataset of 100,000 paired LLM-LLM and human-LLM dialogues from the WildChat dataset and quantifying how well the LLM simulations align with their human counterparts. Overall, we find relatively low alignment between simulations and human interactions, with systematic differences in multiple textual properties, including style and conversational dynamics. Further, we find that models perform similarly in simulating English, Chinese, and Russian dialogues. Our results also suggest that LLMs only simulate a narrow range of the overall distribution of human dialogue, as they perform better on the subset of humans who write similarly to the LLM’s own style.

Large language models (LLMs) have demonstrated strong performance in translation tasks. Subtitle translation presents unique challenges, such as preserving the original work’s worldview and the distinctive speaking styles of its characters. Achieving high-quality translations that reflect these stylistic nuances typically requires bilingual data for a specific movie, which is often scarce or unavailable. Thus, we propose a data augmentation method that uses LLMs to improve translation performance for specific movies, even when only a few hundred bilingual sentence pairs are available. The method expands source-side data by rewriting original subtitles using information that can be extracted from the context, such as character profiles and scene descriptions, to maintain the tone and thematic consistency of the movie. For translation, the augmented sentences are aligned with manually translated originals using structural similarity, which enables style-preserving bilingual data generation via one-shot learning. Experimental results show that data augmented using the proposed method effectively improves BLEU scores for film subtitle translation, and achieves superior stylistic quality in human evaluation.

Large Language Model(LLM)-based agents have shown strong capabilities in web information seeking, with reinforcement learning (RL) becoming a key optimization paradigm. However, planning remains a bottleneck, as existing methods struggle with long-horizon strategies. Our analysis reveals a critical phenomenon—plan anchor—where the first reasoning step disproportionately impacts downstream behavior in long-horizon web reasoning tasks. Current RL algorithms, fail to account for this by uniformly distributing rewards across the trajectory.To address this, we propose Anchor-GRPO, a two-stage RL framework that decouples planning and execution. In Stage 1, the agent optimizes its first-step planning using fine-grained rubrics derived from self-play experiences and human calibration. In Stage 2, execution is aligned with the initial plan through sparse rewards, ensuring stable and efficient tool usage. We evaluate Anchor-GRPO on four benchmarks: BrowseComp, BrowseComp-Zh, GAIA, and XBench-DeepSearch. Across models from 3B to 30B, Anchor-GRPO outperforms baseline GRPO and First-step GRPO, improving task success and tool efficiency. Notably, WebAnchor-30B achieves 46.0% pass@1 on BrowseComp and 76.4% on GAIA. Anchor-GRPO also demonstrates strong scalability, getting higher accuracy as model size and context length increase.

Multimodal Large Language Models (MLLMs) are increasingly deployed as social agents, yet their ability to integrate conflicting identity cues remains underexplored. We audit gender bias in ten recent MLLMs using a counterfactual cooperative gaming task that pairs synthetic voices with avatars of varying gender presentation and visual fidelity. Our analysis reveals distinct bias patterns that can occur independently: closed-source models (e.g., Gemini 2.5/3) exhibit a near-deterministic “voice-matching” bias that enforces binary alignment between voice and appearance, whereas open-weight models (e.g., Qwen-2.5-Omni-7B) show limited responsiveness to vocal cues and instead exhibit context-driven stereotypes, such as preferring male avatars in combat scenarios. We further find that reducing visual realism attenuates matching tendencies in some models. These findings demonstrate that multimodal fairness is not monolithic; models may appear unbiased on one dimension while enforcing strict identity congruence or role-based stereotypes on another. Code and data are available at https://github.com/halfhoon/whose-voice-whose-avatar.

Recent Large Audio-Language Models (LALMs) integrate acoustic capabilities into reasoning, yet whether they reliably ground clinical judgments in audible evidence remains unproven. We introduce CliniCAST (Clinical Controlled Acoustic Synthetic Triage), a controlled benchmark that disentangles clinically meaningful acoustic cues from lexical content and speaker demographics. CliniCAST comprises 5,856 synthetic samples across 12 disease conditions: 4,800 audio samples forming 2,400 tagged–untagged pairs for five-level emergency triage, and 1,056 audio–text inconsistent samples in which reassuring speech is paired with high-risk acoustic cues. Evaluating a diverse suite of audio-capable foundation models, we find that LALMs exhibit fragile acoustic grounding and a pronounced “text dominance” failure mode: reassuring lexical content suppresses response to audible distress signals even under safety-critical conditions. Age and gender interactions are weak across conditions, indicating that the primary failure mode is insufficient cross-modal integration rather than demographic bias. These results suggest current LALMs are not yet robust enough for high-stakes medical triage, and motivate training objectives that explicitly enforce reliance on clinically grounded audible evidence.

Sentiment classification is a crucial task in natural language processing (NLP). To mitigate the spurious correlation, the causal word identification method estimates the impact of treatment words on sentence sentiment and removes those with low treatment effects. However, previous works regard the presence or absence of a specific word in a sentence as a binary treatment. This approach limits the generalizability to novel words and the robustness of low-frequency words. To bridge this gap, we propose a meta-causal approach that achieves causal word identification for arbitrary words with a single training task. Specifically, we begin by clustering contexts based on their embeddings obtained from a pre-trained language model. Subsequently, for each cluster, a representation and multi-head prediction networks are trained to estimate the treatment effect of each word to distinguish causally related words from spuriously correlated ones. The trained word classifier is then used to give weights for different words to train a more robust and generalizable sentiment classification model. Extensive experiments on public datasets demonstrate the effectiveness of our method in identifying causal words and improving the performance of sentiment classification.

Large Language Models (LLMs) have demonstrated effectiveness as zero-shot time series (TS) forecasters. While existing work often relies on fine-tuning specialized modules to bridge this gap, a distinct, yet challenging, paradigm aims to leverage truly off-the-shelf LLMs without any fine-tuning whatsoever, relying solely on strategic tokenization of numerical sequences. However, the parameters of these fully frozen models cannot adapt to distribution shifts. Thus, we introduce a novel yet highly effective strategy to overcome this brittleness: injecting noise into the raw TS before tokenization. This non-invasive intervention acts as a form of inference-time augmentation, compelling the frozen LLM to extrapolate based on robust underlying temporal patterns rather than superficial numerical artifacts. We theoretically analyze this phenomenon and empirically validate its effectiveness across diverse benchmarks. Notably, to fully eliminate potential biases from data contamination during LLM pre-training, we introduce multiple novel real-world TS datasets that fall outside all utilized LLMs’ pre-training scopes, and consistently observe improved performance. This study provides a further step in directly leveraging off-the-shelf LLMs for TS forecasting[https://github.com/jkumh/NLTS].

Tabular data forms the backbone of high-stakes decision systems in finance, healthcare, and beyond. Yet industrial tabular datasets are inherently difficult: high-dimensional, riddled with missing entries, and rarely labeled at scale. While foundation models have revolutionized vision and language, tabular learning still leans on handcrafted features and lacks a general self-supervised framework. We present MaskTab, a unified pre-training framework designed specifically for industrial-scale tabular data. MaskTab encodes missing values via dedicated learnable tokens, enabling the model to distinguish structural absence from random dropout. It jointly optimizes a hybrid supervised pre-training scheme—utilizing a twin-path architecture to reconcile masked reconstruction with task-specific supervision—and an MoE-augmented loss that adaptively routes features through specialized subnetworks. On industrial-scale benchmarks, it achieves +5.04% AUC and +8.28% KS over prior art under rigorous scaling. Moreover, its representations distill effectively into lightweight models, yielding +2.55% AUC and +4.85% KS under strict latency and interpretability constraints, while improving robustness to distribution shifts. Our work demonstrates that tabular data admits a foundation-model treatment—when its structural idiosyncrasies are respected.

Mitigating social bias in Large Language Models (LLMs) presents a distinct alignment challenge: unlike verifiable tasks, social bias lacks a single ground truth, creating a high-variance, subjective reward landscape. Previous preference-based fine-tuning methods have major trade-offs: Direct Preference Optimization (DPO) is limited by the lack of exploration inherent in offline training, while Proximal Policy Optimization (PPO) can lead to training instability due to potentially unreliable critic estimates. In this paper, we propose BiasGRPO, an adaptation of Group Relative Policy Optimization (GRPO) that stabilizes alignment by normalizing rewards across a group of sampled completions. By substituting the value function with a group-relative baseline, our approach reduces instability while maintaining the exploration benefits of online reinforcement learning. To adapt GRPO, we curate and synthetically extend a dataset spanning multiple domains and contexts, and create a custom, bias-specific reward model for effectively guiding generation while avoiding knowledge degradation. We find that BiasGRPO outperforms DPO and PPO across multiple benchmarks, indicating its effectiveness as an alignment technique that can overcome the limitations of previous preference-based methods.

Despite the growing adoption of large language models (LLMs) in scientific research workflows, automated support for academic rebuttal, a crucial step in academic communication and peer review, remains largely underexplored. Existing approaches typically rely on off-the-shelf LLMs or simple pipelines, which struggle with long-context understanding and often fail to produce targeted and persuasive responses. In this paper, we propose DRPG, an agentic framework for automatic academic rebuttal generation that operates through four steps: Decompose reviews into atomic concerns, Retrieve relevant evidence from the paper, Plan rebuttal strategies, and Generate responses accordingly. Notably, the Planner in DRPG reaches over 98% accuracy in identifying the most feasible rebuttal direction. Experiments on data from top-tier conferences demonstrate that DRPG significantly outperforms existing rebuttal pipelines and achieves performance beyond the average human level using only an 8B model. Our analysis further demonstrates the effectiveness of the planner design and its value in providing multi-perspective and explainable suggestions. We also showed that DRPG works well in a more complex multi-round setting. These results highlight the effectiveness of DRPG and its potential to provide high-quality rebuttal content and support the scaling of academic discussions.

Zero-shot Named Entity Recognition (ZS-NER) remains brittle under domain and schema shifts, where unseen label definitions often misalign with a large language model’s (LLM’s) intrinsic semantic organization. As a result, directly mapping entity mentions to fine-grained target labels can induce systematic semantic drift, especially when target schemas are novel or semantically overlapping. We propose SAM-NER, a three-stage framework based on Semantic Archetype Mediation that stabilizes cross-domain transfer through an intermediate, domain-invariant archetype space. SAM-NER: (i) performs Entity Discovery via cooperative extraction and consensus-based denoising to obtain high-coverage, high-fidelity entity spans; (ii) conducts Abstract Mediation by projecting entities into a compact set of universal semantic archetypes distilled from high-level ontological abstractions; and (iii) applies Semantic Calibration to resolve archetype-level predictions into target-domain types through constrained, definition-aligned inference with a frozen LLM. Experiments on the CrossNER benchmark show that SAM-NER consistently outperforms strong prior ZS-NER baselines in cross-domain settings.

Graph problems are fundamentally challenging for large language models (LLMs). While LLMs excel at processing unstructured text, graph tasks require reasoning over explicit structure, permutation invariance, and computationally complex relationships, creating a mismatch with the representations of text-based models. Our work investigates how LLMs can be effectively applied to graph problems despite these barriers. We introduce a human-interpretable structural encoding strategy for graph-to-text translation that injects graph structure directly into natural language prompts. Our method involves computing a variant of Weisfeiler–Lehman (WL) similarity classes and maps them to human-like color tokens rather than numeric labels. The key insight is that semantically meaningful and human-interpretable cues may be more effectively processed by LLMs than opaque symbolic encoding. Experimental results across multiple graph algorithms and predictive tasks show significant improvements from our method on both synthetic and real-world datasets. By capturing both local and global-range dependencies, our method enhances LLM performance, especially on graph tasks that require reasoning over global graph structure.

Hallucinations—generating responses inconsistent with the visual input—remain a critical limitation of large vision-language models (LVLMs), especially in open-ended tasks such as image captioning and visual reasoning. In this work, we probe the layer-wise generation dynamics that drive hallucinations and propose a training-free mitigation strategy. Employing the Logit Lens, we examine how LVLMs construct next-token distributions across decoder layers, uncovering a pronounced commitment-depth gap: truthful tokens accumulate probability mass on their final candidates earlier than hallucinatory ones. Drawing on this discovery, we introduce Context Embedding injection (CEI), a lightweight method that harnesses the hidden state of the last input token—the context embedding—as a grounding signal to maintain visual fidelity throughout decoding and curb hallucinations. Evaluated on the CHAIR, AMBER, and MMHal-Bench benchmarks (with a maximum token length of 512), CEI outperforms state-of-the-art baselines across three LVLMs, with its dynamic variant yielding the lowest overall hallucination rates. By integrating novel mechanistic insights with a scalable intervention, this work advances the mitigation of hallucinations in LVLMs. Data and code are available at https://github.com/mehrdadfazli/CEI.

Multimodal tables i.e. tabular layouts interleaved with charts, maps, icons, and color encodings are ubiquitous in real applications yet remain difficult for Multimodal Large Language Models (MLLMs). Despite advances in text and image understanding, systematic evaluation of table-centric multimodal reasoning is limited. We introduce MMTabReal, a MultiModal Table Benchmark, human-curated suite of 500 real-world tables paired with 4021 question–answer pairs. MMtabReal spans four question types, five reasoning categories, and eight structural archetypes. Evaluations of state-of-the-art models reveal substantial gaps, especially in visual grounding, spatial alignment, and multi-step inference, with 20–40% performance drops relative to existing benchmarks. These results highlight the need for architectures that more tightly fuse vision with tabular structure and support explicit numeric/logical operations. MMtabReal is released for evaluation only, providing a rigorous, reproducible testbed that reflects the linguistic, structural, and reasoning complexity of real-world multimodal tables. Code and data are available at: https://coral-lab-asu.github.io/mmtabreal/

Despite its simplicity and efficacy, the high token expenditure of self-consistency can limit its practical utility. We investigate whether early hypothesis pruning can improve the token efficiency of self-consistency for long chain-of-thought reasoning tasks, while preserving its parallelism. Concretely, we generate all solutions in parallel but periodically prune intermediate hypotheses based on two lightweight indicators: (a) the model’s confidence in each hypothesis, and (b) the lexical coverage of all current hypotheses by candidate subsets. We design a fast weighted set cover algorithm that utilizes the two indicators; evaluation of five LLMs on three math benchmarks shows that our method improves token efficiency in most cases, with reductions of 10-35% in many.

Retrieval-augmented generation (RAG) can enhance large language models (LLMs) by providing external knowledge and helping reduce hallucinations. In multimodal RAG, however, retrieval remains challenging because a single retriever may fail to capture fine-grained multimodal semantics, and visually or semantically similar entities may still contain misleading information for answer generation. We propose a progressive multimodal re-ranking framework with curriculum learning to improve CLIP-based visual coarse-grained retrieval. Our framework progressively refines retrieval results through two stages: fine-grained section-level re-ranking and multimodal section reassessment. To better align re-ranking with multimodal queries, we introduce a curriculum-learning strategy that trains the model with hard negatives that are visually or semantically similar but contain misleading information. Experiments on InfoSeek and Enc-VQA show that our method achieves state-of-the-art answer accuracy and competitive retrieval performance.

Automatic evaluation of book-length stories remains underexplored, particularly for non-English literature. We introduce ChangJuan, the first benchmark for book-length Chinese story evaluation, comprising 300 novels with metadata, human ratings, and large-scale user reviews. To mitigate the subjectivity of raw reviews, we propose a distillation method to aggregate them into generally agreed viewpoints (pros and cons) across key evaluation aspects such as plot and character. We conduct systematic experiments to benchmark current LLMs, analyze aspect importance, and examine genre differences. For book-length story evaluation, we propose an enhanced summary-based method that leverages length-detail balanced summaries and representative excerpts, generates aspect-specific reviews, and considers genre-aware aspect weighting to assign a final score. Using this framework and our distilled viewpoints, we fine-tune an 8B model, CLEM, which outperforms open-source baselines and raises Qwen3’s Kendall’s tau correlation with human judgments from 24.8 to 34.1. Our datasets and codes are available at https://github.com/DingyiYang/ChangJuan.