Recent advancements in large language models (LLMs) have catalyzed the rise of reasoningintensive inference paradigms, where models perform explicit step-by-step reasoning before generating final answers. While such approaches improve answer quality and interpretability, they incur substantial computational overhead due to the prolonged generation sequences. In this paper, we propose Tandem, a novel collaborative framework that synergizes large and small language models (LLMs and SLMs) to achieve high-quality reasoning with significantly reduced computational cost. Specifically, the LLM serves as a strategic coordinator, efficiently generating a compact set of critical reasoning insights. These insights are then used to guide a smaller, more efficient SLM in executing the full reasoning process and delivering the final response. To balance efficiency and reliability, Tandem introduces a cost-aware termination mechanism that adaptively determines when sufficient reasoning guidance has been accumulated, enabling early stopping of the LLM’s generation. Experiments on mathematical reasoning and code generation benchmarks demonstrate that Tandem reduces computational costs by approximately 40% compared to standalone LLM reasoning, while achieving superior or competitive performance. Furthermore, the sufficiency classifier trained on one domain transfers effectively to others without retraining. The code is available at: https://github.com/Applied-MachineLearning-Lab/ACL2026_Tandem.
论文检索
输入标题、作者或关键词，从 7,537 篇学术成果中精准定位
Recent work has demonstrated the potential of large language models (LLMs) for program optimization, a key challenge in programming languages. We propose a blackbox adaptation method called Retrieval Augmented Search (RAS) that performs beam search over candidate optimizations; at each step, it retrieves in-context examples from a given training dataset of slow-fast program pairs to guide the LLM. Critically, we find that performing contextual retrieval based on an LLM-generated natural language description significantly outperforms retrieval based on the source code. We also propose AEGIS, a method for improving interpretability by decomposing training examples into "atomic edits” that are significantly more incremental in nature. We show that RAS performs up to 2.06\times better than prior state-of-the-art blackbox adaptation strategies on optimizing C++ programs, and that AEGIS performs up to 1.37\times better while making significantly smaller edits. We also show that using RAS improves the mean runtime percentile of Python programs by 10.27 compared to baselines.
Beyond Prompt Engineering: A Systematic Analysis of Prompt Lexical Sensitivity and Its Impacts on Quality
PDF ↗Large Language Models (LLMs) often exhibit extreme sensitivity to surface-level prompt variations, where minor lexical perturbations trigger disproportionate performance fluctuations. Moving beyond black-box optimization or coarse-grained templates, we conduct the first analysis of n-gram token-level mechanisms, leveraging a large-scale dataset of 132,000 prompt variants. Our investigation uncovers the Scaling Law of Prompt Performance Stability: higher average performance is inherently associated with lower variance and greater stability. We identify that this robustness is driven by two linguistic pillars: Domain-Specific Terminology, which anchors semantic boundaries, and Explicit Action Directives, which formalize reasoning trajectories. By narrowing the model’s interpretative space, these patterns effectively "lock" the generation process. We operationalize these findings into an automated Prompt-Refining Agent that autonomously restructures queries via domain anchoring and operational constraints. Empirical results show a 40.7% reduction in performance variance for code generation, offering a statistically grounded framework for robust prompt engineering.
CoTJudger: A Graph-Driven Framework for Automatic Evaluation of Chain-of-Thought Efficiency and Redundancy in LRMs
PDF ↗Large Reasoning Models (LRMs) have demonstrated strong performance by producing extended Chain-of-Thought (CoT) traces before answering. However, this paradigm often induces over-reasoning: redundant calculations and circular self-verification that increase computational cost without improving outcomes. Existing evaluations largely emphasize final accuracy or coarse token counts, and lack automated tools to separate essential logic from structural redundancy. We introduce CoTJudger, a graph-driven framework that quantifies reasoning efficiency by converting free-form CoTs into directed dependency graphs and extracting the Shortest Effective Path (SEP) needed to reach a correct solution. This yields an interpretable efficiency signal – how much of a CoT is necessary versus structurally redundant – that is comparable across models and tasks. Evaluating 21 LRMs, CoTJudger reveals pervasive redundancy and surfaces recurring failure modes, including verification obsession and compensatory redundancy. These results provide a practical metric for disentangling reasoning ability from computational waste, enabling more targeted evaluation and diagnosis of LRM efficiency.
Qualitative analysis is critical to understanding human datasets in many social science disciplines. A central method in this process is inductive coding, where researchers identify and interpret codes directly from the datasets themselves. Yet, this exploratory approach poses challenges for meeting methodological expectations (such as "depth" and "variation"), especially as researchers increasingly adopt Generative AI (GAI) for support. Ground-truth-based metrics are insufficient because they contradict the exploratory nature of inductive coding; cluster- or topic-level metrics fail to capture the interpretive, cross-cutting nature of qualitative codes; and manual evaluation can be labor-intensive. This paper presents a theory-informed computational method for measuring inductive coding results from humans and GAI. Our method first merges individual codebooks into an Aggregated Code Space using an LLM-enriched hierarchical clustering algorithm. It then measures each coder’s contribution against the merged result using four novel metrics: Coverage, Overlap, Novelty, and Divergence, designed to capture breadth, consensus, unique contribution, and systematic deviation without assuming ground truth. Through two experiments on a human-coded online conversation dataset, we 1) reveal the merging algorithm’s impact on metrics; 2) validate the metrics’ stability and robustness across multiple runs and different LLMs; and 3) showcase the metrics’ ability to diagnose coding issues, such as excessive or irrelevant (hallucinated) codes. We discuss how these metrics should be interpreted in combination and their current limitations. Our work provides a reliable pathway for ensuring methodological rigor in human-AI qualitative analysis.
MaskTab: Scalable Masked Tabular Pretraining with Scaling Laws and Distillation for Industrial Classification
PDF ↗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.
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.
Conjecture and Inquiry: Quantifying Software Performance Requirements via Interactive Retrieval-Augmented Preference Elicitation
PDF ↗Since software performance requirements are documented in natural language, quantifying them into mathematical forms is essential for software engineering. Yet, the vagueness in performance requirements and uncertainty of human cognition have caused highly uncertain ambiguity in the interpretations, rendering their automated quantification an unaddressed and challenging problem. In this paper, we formalize the problem and propose IRAP, an approach that quantifies performance requirements into mathematical functions via interactive retrieval-augmented preference elicitation. IRAP differs from the others in that it explicitly derives from problem-specific knowledge to retrieve and reason the preferences, which also guides the progressive interaction with stakeholders, while reducing the cognitive overhead. Experiment results against 10 state-of-the-art methods on four real-world datasets demonstrate the superiority of IRAP on all cases with up to 40x improvements under as few as five rounds of interactions.
Style representation learning is a powerful tool for authorship analysis and modeling writing style, yet the latent nature of learned representations makes them difficult to interpret. Recent work has attempted to explain these representations by generating natural language descriptions with large language models (LLMs) conditioned on input text. However, such descriptions are often prone to the LLM’s biases and hallucinations, and they lack an explicit objective and practical utility. In this work, we propose a novel framework for interpreting style representations through style-eliciting prompts: natural language instructions designed to steer LLMs to generate text that reflects specific stylistic attributes. We curate 1,010 distinct style features spanning 26 stylistic categories and construct a dataset by prompting an LLM to generate text conditioned on these features. Using this data, we train a decoder to generate a style prompt from the style representation of the generated text. We evaluate our approach on three tasks: (1) recovering original style prompts from generated text, (2) generating text in the same style using the recovered prompts, and (3) steering LLM outputs to match the style of human-written texts. Experiments demonstrate that our method consistently outperforms strong baselines that directly prompt LLMs with target text, achieving superior performance in both style description and style imitation. These results highlight style-eliciting prompts can provide a practical and interpretable interface to stylistic information encoded in style representations.
Large language models (LLMs), despite strong performance on complex mathematical problems, exhibit systematic limitations in counting tasks. This issue arises from the architectural limits of transformers, where counting is performed across layers, leading to degraded precision for larger counting problems due to depth constraints. To address this limitation, we propose a simple test-time strategy inspired by System-2 cognitive processes that decomposes large counting tasks into smaller, independent sub-problems that the model can reliably solve. We evaluate this approach using observational and causal mediation analyses to understand the underlying mechanism of this System-2-like strategy. Our mechanistic analysis identifies key components: latent counts are computed and stored in the final item representations of each part, transferred to intermediate steps via dedicated attention heads, and aggregated in the final stage to produce the total count. Experimental results demonstrate that this strategy enables LLMs to surpass architectural limitations and achieve higher accuracy on large-scale counting tasks. This work provides mechanistic insight into System-2 counting in LLMs and presents a generalizable approach for improving and understanding their reasoning behavior.
Steering large language model (LLM) reasoning via high-level reasoning actions offers a promising approach to improve robustness and interpretability. However, existing action-based paradigms, ranging from training-free prompting to static plan retrieval or prediction, often fail to consistently outperform standard generation because their planners tend to degenerate into repetitive loops or fixed patterns. We propose PG-HAP (Policy-Guided High-Level Action Planning), a lightweight stepwise planner–executor framework that learns to select reasoning actions dynamically while keeping the executor LLM fully frozen. The planner is trained with reinforcement learning to optimize answer correctness. To prevent degeneration, we introduce two targeted mechanisms: (i) an Action-Dependency Logit Mask that enforces valid transitions to avoid redundancy, and (ii) an Action Diversity Reward that discourages mode collapse by promoting varied action sequences. Across mathematical and commonsense reasoning benchmarks, PG-HAP improves accuracy over strong baselines while producing less redundant, more adaptive trajectories. This demonstrates that learning high-level planning alone can substantially strengthen reasoning without expensive end-to-end model tuning.
ROSCO-Omni: Multimodal LLM-Based Communication Understanding for Non- and Minimally-Speaking Autistic Individuals
PDF ↗Approximately 30% of autistic individuals remain non- or minimally-speaking throughout their lives, yet communicate richly through gestures, vocalizations, facial expressions, and augmentative devices. Interpreting this communication is an inherently multimodal task: caregivers rely on the simultaneous integration of visual cues, auditory signals, and contextual understanding to infer intent. Despite this natural alignment with multimodal large language models (MLLMs), research in this intersection remains narrowly focused on diagnosis rather than communication understanding. We address this gap by reframing the problem around two complementary dimensions: communicative actions (the physical modality) and communicative functions (the pragmatic intent). We analyze the ROSCO dataset, containing 2,903 caregiver-annotated video samples from 27 non- and minimally-speaking individuals, with multi-label annotations capturing up to three concurrent actions and two functions per sample across 6 action and 6 function classes. We further propose ROSCO-Omni, a teacher-student distillation framework that generates label-guided instruction data from a high-capability teacher MLLM and uses it to finetune a student MLLM for domain-specialized inference. ROSCO-Omni achieves performance comparable to closed-source models, demonstrating that open-source MLLMs can be adapted to understand communication in this underserved population.
IDEA: An Interpretable and Editable Decision-Making Framework for LLMs via Verbal-to-Numeric Calibration
PDF ↗Large Language Models are increasingly deployed for decision-making, yet their adoption in high-stakes domains remains limited by miscalibrated probabilities, unfaithful explanations, and inability to incorporate expert knowledge precisely. We propose **IDEA**, a framework that extracts LLM decision knowledge into an interpretable parametric model over semantically meaningful factors. Through joint learning of verbal-to-numerical mappings and decision parameters via EM, correlated sampling that preserves factor dependencies, and direct parameter editing with mathematical guarantees, IDEA produces calibrated probabilities while enabling quantitative human-AI collaboration. Experiments across five datasets show IDEA with Qwen-3-32B (78.6%) outperforms DeepSeek R1 (68.1%) and GPT-5.2 (77.9%), achieving perfect factor exclusion and exact calibration—precision unattainable through prompting alone.
Aligning Vision-Language Models (VLMs) with safety standards is essential to mitigate risks arising from their multimodal complexity, where integrating vision and language unveils subtle threats beyond the reach of conventional safeguards. Inspired by the insight that reasoning across modalities is key to preempting intricate vulnerabilities, we propose a novel direction for VLM safety: multimodal reasoning-driven prompt rewriting. To this end, we introduce VLMGuard-R1, a proactive framework that refines user inputs through a reasoning-guided rewriter, dynamically interpreting text-image interactions to deliver refined prompts that bolster safety across diverse VLM architectures without altering their core parameters. To achieve this, we devise a three-stage reasoning pipeline to synthesize a dataset that trains the rewriter to infer subtle threats, enabling tailored, actionable responses over generic refusals. Extensive experiments across five benchmarks with six VLMs reveal that VLMGuard-R1 outperforms four baselines. In particular, VLMGuard-R1 achieves a remarkable 43.59% increase in average safety across five models on the SIUO benchmark.
LLMs are increasingly being integrated into clinical workflows, yet they often lack clinical empathy, an essential aspect of effective doctor–patient communication. Existing NLP frameworks focus on reactively labeling empathy in doctors’ responses but offer limited support for anticipatory modeling of empathy needs, especially in general health queries. We introduce the Empathy Applicability Framework (EAF), a theory-driven approach that classifies patient queries in terms of the applicability of emotional reactions and interpretations, based on clinical, contextual, and linguistic cues. We release a benchmark of real patient queries, dual-annotated by human annotators and GPT-4o. In the subset with human consensus, we also observe substantial human–GPT alignment. To validate EAF, we train classifiers on human-labeled and GPT-only annotations to predict empathy applicability, achieving strong performance and outperforming the heuristic and zero-shot LLM baselines. Error analysis highlights persistent challenges: implicit distress, clinical-severity ambiguity, and contextual hardship, underscoring the need for multi-annotator modeling, clinician-in-the-loop calibration, and culturally diverse annotation. EAF provides a framework for identifying empathy needs before response generation, establishes a benchmark for anticipatory empathy modeling, and enables supporting empathetic communication in asynchronous healthcare.
While large language models (LLMs) excel at static scientific reasoning, they struggle to model the temporal structure of dynamic physical processes. We present EvoMD-LLM (Evolutionary Molecular Dynamics Large Language Model), a framework that reformulates species-level molecular dynamics as a symbolic temporal language modeling problem. Reactive MD trajectories are discretized into sequences of molecular events, where each token represents a chemical species augmented with its persistence duration, enabling standard autoregressive LLMs to learn compositional evolution over time through efficient fine-tuning. A key component of EvoMD-LLM is temporal scaffolding, which treats event duration as an explicit linguistic token and serves as a structured inductive bias, significantly reducing invalid or hallucinated molecular outputs compared to conventional sequence modeling approaches. We evaluate EvoMD-LLM on multiple temporal prediction tasks, achieving up to 66.14% accuracy and consistently outperforming sequential neural networks and language-based baselines. Beyond quantitative improvements, we qualitatively observe that the model can generate plausible physical interpretations of reaction dynamics, despite not being explicitly trained for explanation. These results demonstrate that symbolic temporal language modeling provides an effective framework for grounding LLMs in dynamic physical simulations.
CausalGaze: Unveiling Hallucinations via Counterfactual Graph Intervention in Large Language Models
PDF ↗Despite the groundbreaking advancements made by large language models (LLMs), hallucination remains a critical bottleneck for their deployment in high-stakes domains. Existing classification-based methods mainly rely on static and passive signals from internal states, which often captures the noise and spurious correlations, while overlooking the underlying causal mechanisms. To address this limitation, we shift the paradigm from passive observation to active intervention by introducing CausalGaze, a novel hallucination detection framework based on structural causal models (SCMs). CausalGaze models LLMs’ internal states as dynamic causal graphs and employs counterfactual interventions to disentangle causal reasoning paths from incidental noise, thereby enhancing model interpretability. Extensive experiments across four datasets and three widely used LLMs demonstrate the effectiveness of CausalGaze, especially achieving over 5.2% improvement in AUROC on the TruthfulQA dataset compared to state-of-the-art baselines.
Vision-language model (VLM) hallucination is commonly linked to imbalanced allocation of attention across input modalities: system, image and text. However, existing mitigation strategies tend towards an image-centric interpretation of these imbalances, often prioritising increased image attention while giving less consideration to the roles of the other modalities. In this study, we evaluate a more holistic, system-mediated account, which attributes these imbalances to functionally redundant system weights that reduce attention to image and textual inputs. We show that this framework offers a useful empirical perspective on the yes-bias, a common form of hallucination in which VLMs indiscriminately respond ‘yes’. Causally redistributing attention from the system modality to image and textual inputs substantially suppresses this bias, often outperforming existing approaches. We further present evidence suggesting that system-mediated attention imbalances contribute to the yes-bias by encouraging a default reliance on coarse input representations, which are effective for some tasks but ill-suited to others. Taken together, these findings firmly establish system attention as a key factor in VLM hallucination and highlight its potential as a lever for mitigation.
We study ambiguous-query disambiguation in retrieval-augmented generation (RAG). Prior Diversify-then-Verify (DtV) pipelines first generate interpretations and then retrieve evidence, often introducing ungrounded queries that cannot be answered from the corpus and requiring costly post-hoc pruning and verification. We propose VerDICT, a novel approach that unifies diversification with verification by integrating retriever relevance and generator answerability feedback early. This not only reduces cascading errors but also enables parallelism. On ASQA, VerDICT improves grounding-aware F1 by an average of 23% over the strongest baselines across multiple LLM backbones.
Poetry has long been a central art form for Arabic speakers, serving as a powerful medium of expression and cultural identity. While modern Arabic speakers continue to value poetry, existing research on Arabic poetry within Large Language Models (LLMs) has primarily focused on analysis tasks such as interpretation or metadata prediction, e.g., rhyme schemes and titles. In contrast, our work addresses the practical aspect of poetry creation in Arabic by introducing controllable generation capabilities to assist users in writing poetry. Specifically, we present a large-scale, carefully curated instruction-based dataset in Modern Standard Arabic (MSA) and various Arabic dialects. This dataset enables tasks such as writing, revising, and continuing poems based on predefined criteria, including style and rhyme, as well as performing poetry analysis. Our experiments show that fine-tuning LLMs on this dataset yields models that can effectively generate poetry that is aligned with user requirements, based on both automated metrics and human evaluation with native Arabic speakers. The data and the code are available at https://github.com/mbzuai-nlp/instructpoet-ar